{ "schema_version": "pct-article-edition-1.0", "publication_date": "2026-11-09", "export_scope": "articles_only", "source_edition_id": null, "rights": "PCT reserves rights in original editorial content; third-party rights retained. No general source-data reuse license.", "articles": [ { "id": "pct-2026-11-09-company-together-ai", "content_type": "article", "desk": "company", "title": "Together AI sells predictability, then has to build it", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "editorial_voice": "Company Desk | AI-assisted, measured business columnist", "body": "A reader asking a newspaper a question is not buying a GPU. The reader wants a useful answer while the question still matters. The newspaper, meanwhile, needs a service it can afford to keep running. Together AI occupies the space between those two expectations: it sells the work of turning available models and computing capacity into a dependable product.\n\nIts Washington Post case study gives that proposition a concrete shape. Together reports 1.79 billion monthly input tokens, thousands of daily queries and roughly two-second model responses for Ask The Post AI. The deployment combines dedicated and serverless inference, uses Llama and Mistral models, and has fixed monthly pricing. The amount of that bill is not disclosed. These are vendor-published customer results, not an independently reproduced benchmark. [Customer case study](https://www.together.ai/customers/washington-post).\n\nThe Post's own announcement dates the launch to November 7, 2024. [The Post’s launch announcement](https://www.washingtonpost.com/pr/2024/11/07/washington-post-launches-ask-post-ai-new-search-experience/).\n\nThe attraction is understandable: keep more control over the model without taking on the entire job of serving it. But removing one dependency creates another. Together has to earn its place through engineering, capacity and service, while charging enough to cover them. Its new compiler and its much larger infrastructure commitments show both sides of that bargain.\n\n## A financing round is not an income statement\n\nOn July 1, Together announced an $800 million Series C. The release carried by lead investor Aramco Ventures specified an $8.3 billion post-money valuation, thousands of paying customers and annual bookings that had crossed $1.15 billion in the previous quarter. It also described an ambition to increase the infrastructure footprint roughly fiftyfold over five years. Those are company and investor disclosures, with financial interests attached. [Financing release](https://aramcoventures.com/news/together-ai-raises-800-million-at-8-3-billion-valuation-to-make-frontier-ai-accessible-to-all/).\n\nBookings deserve their own label. They are not interchangeable with recognized revenue, annual recurring revenue or cash collected. The release does not provide enough detail about contract duration, cancellation rights or the calculation to turn its headline into any of those measures. Nor does a valuation tell readers whether selling inference is already profitable. The business can be expanding rapidly while those questions remain open.\n\nCEO Vipul Ved Prakash's accompanying essay added another distinction: the company said it had secured commitments for more than 500 megawatts of compute capacity, to be capitalized independently by new investors. That is a statement about future supply and financing arrangements, not a claim that 500 megawatts were already serving customers. [July 1 company essay](https://www.together.ai/blog/announcing-our-series-c).\n\nSeparate financing may reduce the equity Together itself must devote to physical infrastructure. It does not, by itself, disclose who bears an idle machine's cost. A buyer can rent instead of own and still have a substantial obligation to pay. The missing information is the connection between customer commitments, supplier commitments and the dates on which each becomes binding. That connection matters more to the economics than the label attached to the cloud.\n\n## The capacity arrives on a different clock\n\nIBM supplied a more specific example on August 11. The companies announced a multi-year $240 million agreement for an NVIDIA HGX B300 cluster on IBM Cloud, with availability expected in the first quarter of 2027. Together would use it for open-model inference. IBM also identified Spectrum-X networking. As of this article's November 9 observation date, this is announced future capacity, not an operational cluster verified by this reporting. [IBM announcement](https://newsroom.ibm.com/2026-08-11-IBM-and-Together-AI-Sign-Multi-Year-Agreement-to-Scale-Open-Source-AI-Inference-with-NVIDIA-AI-Infrastructure-on-IBM-Cloud).\n\nThe announcement does not specify the agreement's exact length, payment schedule, minimum draw or termination provisions. Dividing $240 million by an assumed number of years would manufacture an annual cost. Treating the amount as immediate expenditure, or as revenue already earned by either party, would be equally premature. A large agreement is evidence of commercial intent; its cash consequences depend on terms the release does not show.\n\nThe strategic logic is nevertheless clear. Customer-facing inference needs supply that can be counted on, and reserved supply needs customers who will use it. If those commitments line up, a specialist can combine purchasing scale with software efficiency. If customer demand changes faster than the capacity arrangement, the same structure can leave expensive resources attached to yesterday's forecast. This is a conditional business risk, not evidence that Together has excess capacity today.\n\nThere is also a timing problem distinct from demand. A customer can begin paying for an application before a planned cluster is delivered, or need more capacity while an expansion is still under construction. A credible growth story therefore needs an operating bridge between the existing fleet and the future one. An infrastructure announcement explains where additional supply is intended to come from. It does not show that bridge's cost or spare capacity.\n\n## Software has to pay its way\n\nTogether's answer is not simply to acquire more GPUs. On November 4, it announced Together Compiler as part of a broader effort to adapt inference software to changing models, hardware and traffic. The company says AI agents search for faster implementations and that the compiler already supports some production serverless endpoints. Customer access was described as a limited beta; self-service optimization of dedicated deployments remained forthcoming. [November 4 announcement](https://www.together.ai/blog/polymorphic-inference).\n\nThe underlying engineering idea has a more restrained precedent. A May 27, 2025 Hazy Research post examined a single-sequence Llama-3.2-1B workload on an H100. Combining operations into one larger GPU program reduced gaps between operations; the researchers reported using 78% of memory bandwidth and beating the tested systems by more than 1.5 times. That was a narrow, disclosed experiment, not a universal inference speedup. [Research and test scope](https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles).\n\nThe commercial interpretation is simple: more completed work from the same hardware can create room for a lower customer price, a better provider margin, or both. But the benefit must survive the actual workload. A fast result on a small model with one user does not establish the cost of serving many users, longer prompts or a different model. Efficiency is valuable precisely because it has to be achieved in specific circumstances.\n\nTogether's current product page says candidate optimizations are checked for speed and correctness before production use. It reports up to 2.6 times the tokens per GPU at matched per-user speed in a set of Blackwell tests, following a 24-hour compilation using two nodes per model. The page also says supported serverless customers receive compiled kernels without a price change. These are the vendor's measurements and product claims. [Compiler product details](https://www.together.ai/polymorphic-inference).\n\nThat last detail makes the business question sharper. If the customer pays the same token price and Together needs less hardware time, the provider could retain some savings. Competition could instead push the selling price down. Maintaining an advantage then requires repeated improvements, not a single published result. The research team is potentially part of the production cost structure as well as the sales story, and automation is an attempt to make that expertise serve more workloads.\n\n## Predictability has a utilization bill\n\nTogether's public menu makes the software premium visible without revealing its margin. On November 9, an H100 was listed at $5.49 per GPU-hour for dedicated inference and $3.99 for on-demand GPU clusters. These are different products, not interchangeable quotes. The $1.50 difference cannot be called profit: serving software, support, configuration and underlying supply economics all matter. [Observed pricing](https://www.together.ai/pricing).\n\nFor reserved token capacity, the documentation describes provisioned throughput units, or PTUs, at $0.05 per unit per minute. Input, output and cached tokens draw on capacity at different, model-specific rates. Unused units do not roll over or receive a refund. Commitments are tied to a model or model family; changing models requires discussion with sales. [Provisioned throughput documentation](https://docs.together.ai/docs/inference/provisioned-throughput).\n\nAt that published rate, one continuously reserved PTU over exactly thirty days costs $2,160 before any additional charges; ten cost $21,600. This is arithmetic, not a customer quote or a statement that one unit can run a particular application. If the same reservation serves half as much useful work, its cost per completed unit of work doubles. A predictable bill removes surprise from spending, not from demand.\n\nThe service promise also needs careful reading. Marketing uses a 99% uptime shorthand, but the documentation defines reliability as at least 99% of eligible requests completing successfully each month without a Together-caused error. Requests outside purchased capacity and stated limits are excluded. Overflow may use shared serverless capacity at standard rates, without the PTU guarantee. A request-success measure is not a promise about clock-time availability. [Product overview](https://www.together.ai/provisioned-throughput); [Detailed terms](https://docs.together.ai/docs/inference/provisioned-throughput).\n\nThis is the practical bargain beneath the fixed bill: the customer buys a defined amount of preparedness, not an unlimited safety net. Steady traffic can make the commitment attractive. Sharp peaks can require more reserved capacity or expose overflow to a different service level. Together benefits when reservations and available machines can be planned together. The buyer benefits when the purchased shape resembles the workload. Neither side wins automatically merely because billing is predictable.\n\n## Model choice is not a finished exit plan\n\nThe Washington Post deployment is useful because it shows an actual application rather than a hypothetical model migration. It does not establish the Post's complete unit economics. The public case study gives neither the total monthly charge nor the output-token volume and workload distribution needed to reproduce the bill. Its response-time result also does not disclose a percentile or a complete end-to-end measurement method.\n\nFor any customer, retaining model access and retaining a cheap, fast deployment are different things. Moving a model to another provider may be possible while still requiring integration work, performance tuning and new capacity commitments. The value of a specialist is partly the work it saves. That same saved work can become work the customer must recreate when leaving. This is an operational switching cost, even without an exclusive model contract.\n\nTogether therefore has a plausible position between self-operated infrastructure and a closed model API. It can sell a managed service while leaving more model choice with the customer. The durable advantage would be making that freedom economical in practice: serving the chosen model efficiently, adapting when it changes, and keeping capacity available without pricing away the benefit.\n\n## What would make the case stronger\n\nThe next revealing disclosures would connect bookings to recognized revenue and cash, distinguish committed capacity from live supply, and show the cost of keeping customer promises. Research results should be followed by production economics on comparable workloads. The November compiler announcement makes the software argument more concrete; the IBM agreement makes the supply ambition more concrete. Neither, alone, proves the margin between them.\n\nFor the newspaper's reader, all of this remains backstage. A useful answer arrives, or it does not. Together's opportunity is to make that ordinary experience easier for customers to deliver at scale. Its test is to turn the engineering and capacity required for predictability into a business that is predictable enough itself.", "sources": [ { "id": "S1", "stable_id": "pct-2026-11-09-company-together-ai-source-s1", "title": "How The Washington Post Achieved AI Independence with Reliable Inference", "publisher": "Together AI", "url": "https://www.together.ai/customers/washington-post", "published_at": null, "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Vendor customer case study; named customer, unverified measurement method and undisclosed contract price.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "S2", "stable_id": "pct-2026-11-09-company-together-ai-source-s2", "title": "The Washington Post Launches “Ask The Post AI,” a New Search Experience", "publisher": "The Washington Post", "url": "https://www.washingtonpost.com/pr/2024/11/07/washington-post-launches-ask-post-ai-new-search-experience/", "published_at": "2024-11-07", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Customer-origin launch announcement; used only to confirm the launch date.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2024-11-07", "observation_date": "2026-11-09" }, { "id": "S3", "stable_id": "pct-2026-11-09-company-together-ai-source-s3", "title": "Together AI Raises $800 Million at $8.3 Billion Valuation to Make Frontier AI Accessible to All", "publisher": "Aramco Ventures / Together AI", "url": "https://aramcoventures.com/news/together-ai-raises-800-million-at-8-3-billion-valuation-to-make-frontier-ai-accessible-to-all/", "published_at": "2026-07-01", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Company financing release reproduced by lead investor. Post-money basis explicit; bookings are not revenue.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2026-07-01", "observation_date": "2026-11-09" }, { "id": "S4", "stable_id": "pct-2026-11-09-company-together-ai-source-s4", "title": "Announcing our $800M Series C to accelerate the shift to open-source AI", "publisher": "Together AI / Vipul Ved Prakash", "url": "https://www.together.ai/blog/announcing-our-series-c", "published_at": "2026-07-01", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "CEO essay; independently capitalized future capacity is a company-reported commitment.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2026-07-01", "observation_date": "2026-11-09" }, { "id": "S5", "stable_id": "pct-2026-11-09-company-together-ai-source-s5", "title": "IBM and Together AI Sign Multi-Year Agreement to Scale Open-Source AI Inference with NVIDIA AI Infrastructure on IBM Cloud", "publisher": "IBM", "url": "https://newsroom.ibm.com/2026-08-11-IBM-and-Together-AI-Sign-Multi-Year-Agreement-to-Scale-Open-Source-AI-Inference-with-NVIDIA-AI-Infrastructure-on-IBM-Cloud", "published_at": "2026-08-11", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Contracting counterparty announcement. Q1 2027 is expected availability, not current operation.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2026-08-11", "observation_date": "2026-11-09" }, { "id": "S6", "stable_id": "pct-2026-11-09-company-together-ai-source-s6", "title": "From Static to Polymorphic Inference: A new era for production open-source inference systems", "publisher": "Together AI / Simran Arora and Dan Fu", "url": "https://www.together.ai/blog/polymorphic-inference", "published_at": "2026-11-04", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Product and research announcement; beta and forthcoming self-serve functions distinguished from deployed endpoints.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2026-11-04", "observation_date": "2026-11-09" }, { "id": "S7", "stable_id": "pct-2026-11-09-company-together-ai-source-s7", "title": "Polymorphic Inference", "publisher": "Together AI", "url": "https://www.together.ai/polymorphic-inference", "published_at": null, "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Live product page; performance is vendor-tested and workload-specific, not independently reproduced.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "S8", "stable_id": "pct-2026-11-09-company-together-ai-source-s8", "title": "Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B", "publisher": "Stanford Hazy Research and named authors", "url": "https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles", "published_at": "2025-05-27", "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Original research description; H100, Llama-3.2-1B, one sequence, 32-token prompt and 128 generated tokens in cited figure.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": "2025-05-27", "observation_date": "2026-11-09" }, { "id": "S9", "stable_id": "pct-2026-11-09-company-together-ai-source-s9", "title": "Pricing", "publisher": "Together AI", "url": "https://www.together.ai/pricing", "published_at": null, "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Live advertised prices observed November 9; product distinctions preserved, no availability or delivered quote implied.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "S10", "stable_id": "pct-2026-11-09-company-together-ai-source-s10", "title": "Provisioned throughput", "publisher": "Together AI documentation", "url": "https://docs.together.ai/docs/inference/provisioned-throughput", "published_at": null, "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Live docs; exact requested-success SLA and exclusions preferred to marketing shorthand. Contract-specific terms were not reviewed.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "S11", "stable_id": "pct-2026-11-09-company-together-ai-source-s11", "title": "Provisioned Throughput", "publisher": "Together AI", "url": "https://www.together.ai/provisioned-throughput", "published_at": null, "updated_at": null, "retrieved_at": "2026-11-09", "evidence_type": "primary", "verification": "Live marketing page; 99% uptime wording differs from request-success metric in docs. Not represented as a negotiated contract.", "license_status": "Rights retained by source publisher; public availability is not a republication license. Facts paraphrased, no source imagery reproduced.", "publication_date": null, "observation_date": "2026-11-09" } ], "facts": [ { "id": "F1", "source_ids": [ "S1" ], "claim": "Together reports a Washington Post hybrid dedicated/serverless deployment, 1.79B monthly input tokens, thousands of queries daily, approximately two-second model responses, Llama/Mistral and fixed monthly pricing.", "fields": { "monthly_input_tokens": 1790000000, "model_response_seconds_approx": 2, "price_disclosed": false }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F2", "source_ids": [ "S2" ], "claim": "The Post announced Ask The Post AI on November 7, 2024.", "fields": { "launch_date": "2024-11-07" }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F3", "source_ids": [ "S3" ], "claim": "July 1 company/investor release disclosed the financing and company-reported bookings.", "fields": { "series_c_usd": 800000000, "post_money_valuation_usd": 8300000000, "annual_bookings_reported_lower_bound_usd": 1150000000, "bookings_definition": "Not supplied in reviewed release", "infrastructure_footprint_growth_goal_5_years": 50 }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F4", "source_ids": [ "S4" ], "claim": "Company said more than 500MW compute-capacity commitments would be capitalized independently by new investors.", "fields": { "capacity_commitment_mw_lower_bound": 500, "status": "reported commitment; not verified operational capacity" }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F5", "source_ids": [ "S5" ], "claim": "IBM announced a multi-year $240m agreement with Together for an HGX B300 inference cluster.", "fields": { "agreement_value_usd": 240000000, "announcement_date": "2026-08-11", "expected_availability": "2027-Q1", "exact_contract_years": null }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F6", "source_ids": [ "S6" ], "claim": "November 4 compiler announcement describes production use on some serverless endpoints, limited beta customer access and forthcoming self-service dedicated optimization.", "fields": {}, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F7", "source_ids": [ "S8" ], "claim": "Original megakernel experiment reports 78% H100 memory-bandwidth use and over 1.5x performance versus tested systems for a specified one-sequence Llama-3.2-1B workload.", "fields": { "gpu": "H100", "model": "Llama-3.2-1B", "memory_bandwidth_utilization": 0.78, "prompt_tokens": 32, "generated_tokens": 128 }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F8", "source_ids": [ "S7" ], "claim": "Company compiler page reports up to 2.6x tokens per GPU at matched per-user speed in specified Blackwell tests and unchanged prices for supported serverless models.", "fields": { "maximum_claimed_throughput_multiple": 2.6, "compile_hours": 24, "compile_nodes_per_model": 2, "independently_reproduced": false }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F9", "source_ids": [ "S9" ], "claim": "Observed H100 dedicated inference and on-demand cluster list prices differ.", "fields": { "dedicated_inference_usd_per_gpu_hour": 5.49, "cluster_on_demand_usd_per_gpu_hour": 3.99, "observed_on": "2026-11-09", "equivalent_products": false }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F10", "source_ids": [ "S10" ], "claim": "PTUs cost $0.05 per minute; token conversion is model-specific; unused capacity is not refunded or rolled over; changing model requires sales contact.", "fields": { "usd_per_ptu_minute": 0.05, "unused_rollover": false, "unused_refund": false }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." }, { "id": "F11", "source_ids": [ "S10", "S11" ], "claim": "Docs guarantee at least 99% eligible-request success monthly for Together-caused failures, not 99% clock uptime; overflow is excluded and best-effort.", "fields": { "eligible_request_success_monthly": 0.99, "overflow_covered_by_ptu_sla": false }, "qualifier": "See source verification; company statements and benchmark claims retain attribution." } ], "inferences": [ { "id": "I1", "claim": "Customer value is a dependable application, while control over models does not remove serving dependencies.", "based_on_fact_ids": [ "F1" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I2", "claim": "Software efficiency and capacity procurement jointly determine the business proposition.", "based_on_fact_ids": [ "F5", "F6" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I3", "claim": "Bookings and valuation do not establish recognized revenue, ARR, cash collections or profitability.", "based_on_fact_ids": [ "F3" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I4", "claim": "Separate capitalization does not establish absence of payment or idle-capacity risk.", "based_on_fact_ids": [ "F4" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I5", "claim": "Agreement value cannot be annualized without exact term or interpreted as immediate expenditure/revenue.", "based_on_fact_ids": [ "F5" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I6", "claim": "Customer and supplier timing mismatches are conditional operating risks, not evidence of existing excess capacity.", "based_on_fact_ids": [ "F4", "F5" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I7", "claim": "Narrow research results require workload-specific validation before commercial extrapolation.", "based_on_fact_ids": [ "F7" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I8", "claim": "Efficiency savings could benefit provider margin, customer price or both; repeated optimization is economically relevant.", "based_on_fact_ids": [ "F6", "F8" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I9", "claim": "The $1.50/H100-hour product price gap is not a gross-margin measure.", "based_on_fact_ids": [ "F9" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I10", "claim": "At $0.05/minute, one PTU for 30 continuous days is $2,160; ten are $21,600. Half the useful work on the same reservation doubles reserved cost per unit of work.", "based_on_fact_ids": [ "F10" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I11", "claim": "Eligible-request success is not clock-time availability.", "based_on_fact_ids": [ "F11" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I12", "claim": "Reservation economics depend on demand shape and utilization; overflow has a different risk profile.", "based_on_fact_ids": [ "F10", "F11" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I13", "claim": "Model portability and deployment portability differ; recreating saved engineering work can be a switching cost.", "based_on_fact_ids": [ "F1", "F6" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I14", "claim": "Managed serving and model choice form a plausible position, but customer savings require real operating performance.", "based_on_fact_ids": [ "F1", "F8" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I15", "claim": "Financial, operating-supply and production-efficiency disclosures would strengthen assessment beyond financing and benchmarks.", "based_on_fact_ids": [ "F3", "F5", "F8" ], "status": "editorial_inference_not_reported_fact" }, { "id": "I16", "claim": "The core business test is delivering dependable outputs at a sustainable provider cost.", "based_on_fact_ids": [ "F1", "F5", "F8" ], "status": "editorial_inference_not_reported_fact" } ], "unknowns": [ { "id": "U1", "text": "Washington Post price, output-token volume, workload distribution, latency percentile and complete end-to-end timing methodology are undisclosed in the reviewed case study." }, { "id": "U2", "text": "Together recognized revenue, profit, cash flow, booking methodology, cancellation rights, contract-duration mix and concentration were not established by the reviewed financing sources." }, { "id": "U3", "text": "IBM agreement exact duration, payment schedule, minimum draw and termination terms are not disclosed in the reviewed announcement. The cluster was not independently verified operational." }, { "id": "U4", "text": "Commercial terms, delivery schedule and risk allocation for the separately capitalized 500MW commitments remain unspecified in the reviewed CEO essay." }, { "id": "U5", "text": "Compiler results were not independently reproduced. Provider-wide savings, margins and performance on any customer workload are unknown." }, { "id": "U6", "text": "PTU minimum purchases, negotiated term, remedies, all-in charges and conversion for any specific workload require an actual quote and contract." } ], "correction_history": [], "license_status": "Original PCT editorial text; public license not selected. Source rights remain separate.", "public_url": null, "schema_version": "pct-research-item-1.0", "desk_extensions": { "schema_version": "pct.article.v1", "dek": "A newspaper’s inference workload, a new compiler and a $240 million capacity agreement reveal the economics behind the open-model cloud.", "status": "ready_for_editorial_review", "body_derivation": { "canonical_field": "sections", "method": "Join section headings and paragraph text in order. Body is a derived rendering, not an independent editorial source." }, "sections": [ { "id": "pct-2026-11-09-company-together-ai-opening", "heading": null, "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-opening-p01", "text": "A reader asking a newspaper a question is not buying a GPU. The reader wants a useful answer while the question still matters. The newspaper, meanwhile, needs a service it can afford to keep running. Together AI occupies the space between those two expectations: it sells the work of turning available models and computing capacity into a dependable product.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I1" ] }, { "id": "pct-2026-11-09-company-together-ai-opening-p02", "text": "Its Washington Post case study gives that proposition a concrete shape. Together reports 1.79 billion monthly input tokens, thousands of daily queries and roughly two-second model responses for Ask The Post AI. The deployment combines dedicated and serverless inference, uses Llama and Mistral models, and has fixed monthly pricing. The amount of that bill is not disclosed. These are vendor-published customer results, not an independently reproduced benchmark. [Customer case study](https://www.together.ai/customers/washington-post).", "evidence_type": "reported_fact", "source_ids": [ "S1" ], "claim_ids": [ "F1" ] }, { "id": "pct-2026-11-09-company-together-ai-opening-p03", "text": "The Post's own announcement dates the launch to November 7, 2024. [The Post’s launch announcement](https://www.washingtonpost.com/pr/2024/11/07/washington-post-launches-ask-post-ai-new-search-experience/).", "evidence_type": "reported_fact", "source_ids": [ "S2" ], "claim_ids": [ "F2" ] }, { "id": "pct-2026-11-09-company-together-ai-opening-p04", "text": "The attraction is understandable: keep more control over the model without taking on the entire job of serving it. But removing one dependency creates another. Together has to earn its place through engineering, capacity and service, while charging enough to cover them. Its new compiler and its much larger infrastructure commitments show both sides of that bargain.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I1", "I2" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-capital", "heading": "A financing round is not an income statement", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-capital-p01", "text": "On July 1, Together announced an $800 million Series C. The release carried by lead investor Aramco Ventures specified an $8.3 billion post-money valuation, thousands of paying customers and annual bookings that had crossed $1.15 billion in the previous quarter. It also described an ambition to increase the infrastructure footprint roughly fiftyfold over five years. Those are company and investor disclosures, with financial interests attached. [Financing release](https://aramcoventures.com/news/together-ai-raises-800-million-at-8-3-billion-valuation-to-make-frontier-ai-accessible-to-all/).", "evidence_type": "reported_fact", "source_ids": [ "S3" ], "claim_ids": [ "F3" ] }, { "id": "pct-2026-11-09-company-together-ai-capital-p02", "text": "Bookings deserve their own label. They are not interchangeable with recognized revenue, annual recurring revenue or cash collected. The release does not provide enough detail about contract duration, cancellation rights or the calculation to turn its headline into any of those measures. Nor does a valuation tell readers whether selling inference is already profitable. The business can be expanding rapidly while those questions remain open.", "evidence_type": "editorial_analysis", "source_ids": [ "S3" ], "claim_ids": [ "I3" ] }, { "id": "pct-2026-11-09-company-together-ai-capital-p03", "text": "CEO Vipul Ved Prakash's accompanying essay added another distinction: the company said it had secured commitments for more than 500 megawatts of compute capacity, to be capitalized independently by new investors. That is a statement about future supply and financing arrangements, not a claim that 500 megawatts were already serving customers. [July 1 company essay](https://www.together.ai/blog/announcing-our-series-c).", "evidence_type": "reported_fact", "source_ids": [ "S4" ], "claim_ids": [ "F4" ] }, { "id": "pct-2026-11-09-company-together-ai-capital-p04", "text": "Separate financing may reduce the equity Together itself must devote to physical infrastructure. It does not, by itself, disclose who bears an idle machine's cost. A buyer can rent instead of own and still have a substantial obligation to pay. The missing information is the connection between customer commitments, supplier commitments and the dates on which each becomes binding. That connection matters more to the economics than the label attached to the cloud.", "evidence_type": "editorial_analysis", "source_ids": [ "S4" ], "claim_ids": [ "I4" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-ibm", "heading": "The capacity arrives on a different clock", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-ibm-p01", "text": "IBM supplied a more specific example on August 11. The companies announced a multi-year $240 million agreement for an NVIDIA HGX B300 cluster on IBM Cloud, with availability expected in the first quarter of 2027. Together would use it for open-model inference. IBM also identified Spectrum-X networking. As of this article's November 9 observation date, this is announced future capacity, not an operational cluster verified by this reporting. [IBM announcement](https://newsroom.ibm.com/2026-08-11-IBM-and-Together-AI-Sign-Multi-Year-Agreement-to-Scale-Open-Source-AI-Inference-with-NVIDIA-AI-Infrastructure-on-IBM-Cloud).", "evidence_type": "reported_fact", "source_ids": [ "S5" ], "claim_ids": [ "F5" ] }, { "id": "pct-2026-11-09-company-together-ai-ibm-p02", "text": "The announcement does not specify the agreement's exact length, payment schedule, minimum draw or termination provisions. Dividing $240 million by an assumed number of years would manufacture an annual cost. Treating the amount as immediate expenditure, or as revenue already earned by either party, would be equally premature. A large agreement is evidence of commercial intent; its cash consequences depend on terms the release does not show.", "evidence_type": "editorial_analysis", "source_ids": [ "S5" ], "claim_ids": [ "I5" ] }, { "id": "pct-2026-11-09-company-together-ai-ibm-p03", "text": "The strategic logic is nevertheless clear. Customer-facing inference needs supply that can be counted on, and reserved supply needs customers who will use it. If those commitments line up, a specialist can combine purchasing scale with software efficiency. If customer demand changes faster than the capacity arrangement, the same structure can leave expensive resources attached to yesterday's forecast. This is a conditional business risk, not evidence that Together has excess capacity today.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I6" ] }, { "id": "pct-2026-11-09-company-together-ai-ibm-p04", "text": "There is also a timing problem distinct from demand. A customer can begin paying for an application before a planned cluster is delivered, or need more capacity while an expansion is still under construction. A credible growth story therefore needs an operating bridge between the existing fleet and the future one. An infrastructure announcement explains where additional supply is intended to come from. It does not show that bridge's cost or spare capacity.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I6" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-research", "heading": "Software has to pay its way", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-research-p01", "text": "Together's answer is not simply to acquire more GPUs. On November 4, it announced Together Compiler as part of a broader effort to adapt inference software to changing models, hardware and traffic. The company says AI agents search for faster implementations and that the compiler already supports some production serverless endpoints. Customer access was described as a limited beta; self-service optimization of dedicated deployments remained forthcoming. [November 4 announcement](https://www.together.ai/blog/polymorphic-inference).", "evidence_type": "reported_fact", "source_ids": [ "S6" ], "claim_ids": [ "F6" ] }, { "id": "pct-2026-11-09-company-together-ai-research-p02", "text": "The underlying engineering idea has a more restrained precedent. A May 27, 2025 Hazy Research post examined a single-sequence Llama-3.2-1B workload on an H100. Combining operations into one larger GPU program reduced gaps between operations; the researchers reported using 78% of memory bandwidth and beating the tested systems by more than 1.5 times. That was a narrow, disclosed experiment, not a universal inference speedup. [Research and test scope](https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles).", "evidence_type": "reported_fact", "source_ids": [ "S8" ], "claim_ids": [ "F7" ] }, { "id": "pct-2026-11-09-company-together-ai-research-p03", "text": "The commercial interpretation is simple: more completed work from the same hardware can create room for a lower customer price, a better provider margin, or both. But the benefit must survive the actual workload. A fast result on a small model with one user does not establish the cost of serving many users, longer prompts or a different model. Efficiency is valuable precisely because it has to be achieved in specific circumstances.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I7" ] }, { "id": "pct-2026-11-09-company-together-ai-research-p04", "text": "Together's current product page says candidate optimizations are checked for speed and correctness before production use. It reports up to 2.6 times the tokens per GPU at matched per-user speed in a set of Blackwell tests, following a 24-hour compilation using two nodes per model. The page also says supported serverless customers receive compiled kernels without a price change. These are the vendor's measurements and product claims. [Compiler product details](https://www.together.ai/polymorphic-inference).", "evidence_type": "reported_fact", "source_ids": [ "S7" ], "claim_ids": [ "F8" ] }, { "id": "pct-2026-11-09-company-together-ai-research-p05", "text": "That last detail makes the business question sharper. If the customer pays the same token price and Together needs less hardware time, the provider could retain some savings. Competition could instead push the selling price down. Maintaining an advantage then requires repeated improvements, not a single published result. The research team is potentially part of the production cost structure as well as the sales story, and automation is an attempt to make that expertise serve more workloads.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I8" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-pricing", "heading": "Predictability has a utilization bill", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-pricing-p01", "text": "Together's public menu makes the software premium visible without revealing its margin. On November 9, an H100 was listed at $5.49 per GPU-hour for dedicated inference and $3.99 for on-demand GPU clusters. These are different products, not interchangeable quotes. The $1.50 difference cannot be called profit: serving software, support, configuration and underlying supply economics all matter. [Observed pricing](https://www.together.ai/pricing).", "evidence_type": "mixed_fact_and_analysis", "source_ids": [ "S9" ], "claim_ids": [ "F9", "I9" ] }, { "id": "pct-2026-11-09-company-together-ai-pricing-p02", "text": "For reserved token capacity, the documentation describes provisioned throughput units, or PTUs, at $0.05 per unit per minute. Input, output and cached tokens draw on capacity at different, model-specific rates. Unused units do not roll over or receive a refund. Commitments are tied to a model or model family; changing models requires discussion with sales. [Provisioned throughput documentation](https://docs.together.ai/docs/inference/provisioned-throughput).", "evidence_type": "reported_fact", "source_ids": [ "S10" ], "claim_ids": [ "F10" ] }, { "id": "pct-2026-11-09-company-together-ai-pricing-p03", "text": "At that published rate, one continuously reserved PTU over exactly thirty days costs $2,160 before any additional charges; ten cost $21,600. This is arithmetic, not a customer quote or a statement that one unit can run a particular application. If the same reservation serves half as much useful work, its cost per completed unit of work doubles. A predictable bill removes surprise from spending, not from demand.", "evidence_type": "calculation", "source_ids": [ "S10" ], "claim_ids": [ "I10" ] }, { "id": "pct-2026-11-09-company-together-ai-pricing-p04", "text": "The service promise also needs careful reading. Marketing uses a 99% uptime shorthand, but the documentation defines reliability as at least 99% of eligible requests completing successfully each month without a Together-caused error. Requests outside purchased capacity and stated limits are excluded. Overflow may use shared serverless capacity at standard rates, without the PTU guarantee. A request-success measure is not a promise about clock-time availability. [Product overview](https://www.together.ai/provisioned-throughput); [Detailed terms](https://docs.together.ai/docs/inference/provisioned-throughput).", "evidence_type": "mixed_fact_and_analysis", "source_ids": [ "S10", "S11" ], "claim_ids": [ "F11", "I11" ] }, { "id": "pct-2026-11-09-company-together-ai-pricing-p05", "text": "This is the practical bargain beneath the fixed bill: the customer buys a defined amount of preparedness, not an unlimited safety net. Steady traffic can make the commitment attractive. Sharp peaks can require more reserved capacity or expose overflow to a different service level. Together benefits when reservations and available machines can be planned together. The buyer benefits when the purchased shape resembles the workload. Neither side wins automatically merely because billing is predictable.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I12" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-independence", "heading": "Model choice is not a finished exit plan", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-independence-p01", "text": "The Washington Post deployment is useful because it shows an actual application rather than a hypothetical model migration. It does not establish the Post's complete unit economics. The public case study gives neither the total monthly charge nor the output-token volume and workload distribution needed to reproduce the bill. Its response-time result also does not disclose a percentile or a complete end-to-end measurement method.", "evidence_type": "source_limitations", "source_ids": [ "S1" ], "claim_ids": [ "U1" ] }, { "id": "pct-2026-11-09-company-together-ai-independence-p02", "text": "For any customer, retaining model access and retaining a cheap, fast deployment are different things. Moving a model to another provider may be possible while still requiring integration work, performance tuning and new capacity commitments. The value of a specialist is partly the work it saves. That same saved work can become work the customer must recreate when leaving. This is an operational switching cost, even without an exclusive model contract.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I13" ] }, { "id": "pct-2026-11-09-company-together-ai-independence-p03", "text": "Together therefore has a plausible position between self-operated infrastructure and a closed model API. It can sell a managed service while leaving more model choice with the customer. The durable advantage would be making that freedom economical in practice: serving the chosen model efficiently, adapting when it changes, and keeping capacity available without pricing away the benefit.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I14" ] } ] }, { "id": "pct-2026-11-09-company-together-ai-outlook", "heading": "What would make the case stronger", "paragraphs": [ { "id": "pct-2026-11-09-company-together-ai-outlook-p01", "text": "The next revealing disclosures would connect bookings to recognized revenue and cash, distinguish committed capacity from live supply, and show the cost of keeping customer promises. Research results should be followed by production economics on comparable workloads. The November compiler announcement makes the software argument more concrete; the IBM agreement makes the supply ambition more concrete. Neither, alone, proves the margin between them.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I15" ] }, { "id": "pct-2026-11-09-company-together-ai-outlook-p02", "text": "For the newspaper's reader, all of this remains backstage. A useful answer arrives, or it does not. Together's opportunity is to make that ordinary experience easier for customers to deliver at scale. Its test is to turn the engineering and capacity required for predictability into a business that is predictable enough itself.", "evidence_type": "editorial_analysis", "source_ids": [], "claim_ids": [ "I16" ] } ] } ], "word_count": 1778, "estimated_reading_minutes": 8.1, "reading_rate_words_per_minute": 220, "editorial_checks": { "prepublication_checks": [ "Use July 2026 $800m funding and explicit $8.3bn post-money valuation rather than older round.", "Call $1.15bn annual bookings as source does, not ARR or recognized revenue.", "Expected Q1 2027 IBM capacity is future, not currently operating.", "Do not divide $240m by an invented term.", "Do not equate independently financed capacity with absence of commitments.", "99% SLA is eligible-request success monthly, not clock uptime.", "Dedicated inference and cluster GPU prices are different products; spread is not margin.", "PTU arithmetic assumes precisely 30 continuous days and excludes additional charges.", "No claim that Post uses new compiler or PTUs.", "Research tests and compiler claims remain attributed and workload-limited." ], "independent_interviews_conducted": false, "hands_on_benchmarks_conducted": false, "source_date_caution": "Undated live pages use observation date, never an invented publication date.", "new_reporting_hook": "Together Compiler announcement dated 2026-11-04." }, "provenance": { "research_basis": "Public primary sources: company and investor releases, counterparty announcement, customer-origin launch notice, vendor case study, original research and live pricing/docs. No outreach, purchases or private interviews.", "factual_cutoff_date": "2026-11-09", "entity_focus": "Together AI", "previous_company_subjects_excluded": [ "SF Compute", "CoreWeave", "Crusoe", "Nebius", "Modal", "Runpod" ] }, "company_dossier": { "id": "pct-2026-11-09-company-together-ai-dossier", "content_type": "company_dossier", "schema_version": "pct-company-dossier-1.0", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "parent_article_id": "pct-2026-11-09-company-together-ai", "subject": "Together AI", "source_ids": [ "S1", "S2", "S3", "S4", "S5", "S6", "S7", "S8", "S9", "S10", "S11" ], "fact_ids": [ "F1", "F2", "F3", "F4", "F5", "F6", "F7", "F8", "F9", "F10", "F11" ], "inference_ids": [ "I1", "I2", "I3", "I4", "I5", "I6", "I7", "I8", "I9", "I10", "I11", "I12", "I13", "I14", "I15", "I16" ], "uncertainty_reference": "Parent article unknowns array", "uncertainty_ids": [ "U1", "U2", "U3", "U4", "U5", "U6" ], "correction_history": [], "license_status": "Original PCT editorial text; public license not selected. Source rights remain separate.", "format_note": "Normalized dossier index into the parent article’s facts, inferences and source records; no disconnected duplicate prose." } }, "public_projection": { "version": "1.0.0", "date": "2026-11-10", "changes": [ "Internal workspace artifact paths omitted where present. Historical license/status fields retain original provenance; current public reuse policy is stated separately." ], "rights": "PCT reserves rights in original editorial content; third-party rights retained. No general source-data reuse license." } }, { "id": "PCT-20261109-DD-A01", "content_type": "article", "desk": "deals", "title": "What NVIDIA’s $6.3 billion CoreWeave backstop actually buys", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "editorial_voice": "AI-assisted Deals Desk", "body": "# What NVIDIA’s $6.3 billion CoreWeave backstop actually buys\n\n*Deals Desk | November 9, 2026 | Analysis of an agreement signed September 9, 2025*\n\nCoreWeave’s earnings call is scheduled for November 10. Ahead of that update, one older agreement deserves a careful reading: NVIDIA’s commitment to buy residual cloud capacity. Its value lies in giving eligible infrastructure another potential paying customer. How much protection that creates depends on the infrastructure being available, the contract’s scope and the payments actually triggered. [Earnings-call notice](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Date-of-Third-Quarter-2026-Financial-Results-and-Conference-Call/default.aspx).\n\nThis is a September 2025 deal, not a new November signing. CoreWeave entered the order on September 9 and disclosed it on September 15. The initial value was $6.3 billion. Where its own customers do not fully use the relevant capacity, NVIDIA must purchase the unsold remainder through April 13, 2032, subject to delivery, service availability and termination provisions. [Transaction 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm).\n\nThe practical distinction is between finding a buyer and producing something the buyer can use. A fallback customer can make empty capacity less frightening. It cannot make a delayed electrical connection arrive, cool an unfinished room or turn uninstalled servers into a working cluster. Demand risk and construction risk sit at different points in the same business.\n\n## The contract is narrower than the headline\n\nThe underlying master agreement dates to April 10, 2023. Its filed text makes CoreWeave responsible for allocating customer demand across its infrastructure and includes a limit on the capacity dedicated to NVIDIA, measured by server type. The limit itself is redacted. Reserved-compute fees sit in order forms; the public master agreement does not provide the September order’s unit prices. [Master services agreement, Sections 2 and 11](https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm).\n\nThe same agreement permits termination for cause and releases the customer from payment obligations for services after qualifying termination, subject to its terms. That makes a blanket description such as insurance for every unsold CoreWeave GPU too broad. The reviewed documents establish an arrangement with conditions, not an unlimited guarantee of the company’s entire fleet or debts. [Master agreement, Section 10](https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm).\n\nConsider a simplified operating decision. If a covered cluster is ready and another customer wants it, ordinary customer sales determine how much capacity remains. If no customer wants that eligible capacity, the fallback purchase can matter. If the cluster has not met delivery requirements, the commercial problem begins earlier. These scenarios are an analytical illustration, not a reconstruction of private billing rules.\n\n## Why lenders care about the difference\n\nCoreWeave’s June-quarter filing describes $13.6 billion outstanding under delayed-draw term-loan facilities. Those facilities are collateralized by underlying assets and pledged contractual cash flows, generally from investment-grade counterparties. The company draws funds while building infrastructure and amortizes the loans as contractual cash arrives. This is a portfolio financing description, not evidence that every facility relies on this NVIDIA order. [June 30, 2026 10-Q, liquidity and capital resources](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm).\n\nFor a lender, a dependable contractual payer can improve the forecast of money available to service a loan. But the forecast still needs a starting date. Interest, procurement bills and site costs can arrive before a cluster earns revenue. A delay can therefore consume liquidity even where the eventual customer is highly creditworthy. The fallback purchase changes one input to underwriting; it does not settle the whole credit decision.\n\nActual operating progress belongs on a separate ledger. In its August 11 release, CoreWeave reported 1.5 gigawatts of active power and approximately 3.7 gigawatts of contracted power for the second quarter. It also reported $2.575 billion of quarterly revenue. These are company-wide disclosures, not a delivery certificate, utilization report or payment receipt for the September NVIDIA order. [Second-quarter results](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx).\n\nContracted power can describe a future resource. Active power is closer to operations, but it still does not reveal who used a particular server or which order produced an invoice. The public documents reviewed for this feature do not establish how much residual capacity NVIDIA has purchased under this specific order. That gap should stay visible rather than being filled with an assumed utilization rate.\n\n## Three different kinds of NVIDIA support\n\nNVIDIA’s role extends beyond buying compute. A separate filing records a completed January 23, 2026 equity purchase: 22,935,780 CoreWeave shares at $87.20 each, for $2 billion in cash. The same January disclosure describes a collaboration aimed at accelerating more than five gigawatts of AI factories by 2030, while warning that further agreements and conditions would be involved. [January transaction 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000044/crwv-20260123.htm).\n\nThese items belong in different columns. A closed share sale supplies corporate capital and ownership. A capacity purchase pays for services under its conditions. A buildout framework expresses a program still requiring execution. Adding the stock investment to the capacity headline would produce a larger number without producing a meaningful cloud-revenue total. Nor does a strategic relationship by itself establish realized demand from independent end users.\n\n## What a buyer can learn before the next report\n\nThere is a newer pricing signal, but it concerns other contracts. On September 17, 2026, CoreWeave said it had signed three-to-six-month customer contracts at approximately $40 million per megawatt, measured as annualized revenue divided by power needed for the associated clusters. The announcement neither identifies that figure as NVIDIA’s rate nor supplies a universal GPU-hour quote. [September pricing announcement](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx).\n\nA short contract’s annualized revenue density is useful for judging the supplier’s reported selling conditions. It cannot price a multiyear fallback purchase without knowing the hardware mix, covered capacity, service bundle and duration. For an enterprise negotiating compute, the useful comparison remains a matched configuration and delivery window, with explicit remedies for late capacity and a clear bill for storage, networking and support.\n\nThe next useful disclosure would connect the stages: covered capacity ready for service, capacity sold to ordinary customers, residual purchases triggered, and cash collected. Until that bridge is public, the defensible conclusion is modest but meaningful. CoreWeave has a disclosed contractual fallback purchaser for eligible unsold capacity. The commercial benefit must still travel through construction, delivery and invoicing before it becomes cash available to support the business.\n", "sources": [ { "source_id": "PCT-20261109-DD-S01", "id": "PCT-20261109-DD-S01", "title": "CoreWeave NVIDIA capacity-order 8-K", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm", "publication_date": "2025-09-15", "retrieved_date": "2026-11-09", "type": "primary_regulatory_filing", "locator": "Item 1.01; order entered September 9, 2025; initial value, residual purchase condition, service conditions and outside date", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S02", "id": "PCT-20261109-DD-S02", "title": "CoreWeave–NVIDIA master services agreement", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm", "publication_date": "2025-09-25", "retrieved_date": "2026-11-09", "type": "primary_contract_exhibit", "locator": "April 10, 2023 agreement; Sections 2(a), 10 and 11; filed as Exhibit 10.31 to S-4/A; confidential provisions redacted", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S03", "id": "PCT-20261109-DD-S03", "title": "CoreWeave second-quarter 2026 Form 10-Q", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm", "publication_date": "2026-08-12", "retrieved_date": "2026-11-09", "type": "primary_regulatory_filing", "locator": "MD&A, Liquidity and Capital Resources; as of June 30, 2026; filing date confirmed against EDGAR index", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S04", "id": "PCT-20261109-DD-S04", "title": "CoreWeave second-quarter 2026 results", "url": "https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx", "publication_date": "2026-08-11", "retrieved_date": "2026-11-09", "type": "primary_issuer_release", "locator": "Second-quarter revenue and active versus contracted power; corporate metrics, not deal-specific execution", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S05", "id": "PCT-20261109-DD-S05", "title": "CoreWeave short-dated capacity pricing announcement", "url": "https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx", "publication_date": "2026-09-17", "retrieved_date": "2026-11-09", "type": "primary_issuer_release", "locator": "Three-to-six-month contracts; annualized revenue per megawatt definition; unrelated to disclosed NVIDIA order price", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S06", "id": "PCT-20261109-DD-S06", "title": "CoreWeave January 2026 NVIDIA equity transaction 8-K", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000044/crwv-20260123.htm", "publication_date": "2026-01-26", "retrieved_date": "2026-11-09", "type": "primary_regulatory_filing", "locator": "Items 3.02 and 8.01; share purchase completed January 23 and conditional five-gigawatt collaboration framework", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S07", "id": "PCT-20261109-DD-S07", "title": "CoreWeave third-quarter 2026 earnings-call notice", "url": "https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Date-of-Third-Quarter-2026-Financial-Results-and-Conference-Call/default.aspx", "publication_date": "2026-10-27", "retrieved_date": "2026-11-09", "type": "primary_issuer_release", "locator": "Call scheduled for November 10, 2026, 5:00 PM Eastern; future event as of edition date", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } }, { "source_id": "PCT-20261109-DD-S08", "id": "PCT-20261109-DD-S08", "title": "EDGAR filing index for CoreWeave Q2 2026 10-Q", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/0001769628-26-000366-index.htm", "publication_date": "2026-08-12", "retrieved_date": "2026-11-09", "type": "primary_regulatory_index", "locator": "Filing date August 12, 2026; submission accepted August 11 at 19:24:42", "access": "Public source read; relevant document passages checked where applicable.", "license": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" }, "observation_date": "2026-11-09", "license_status": { "status": "no_open_republication_license_verified", "treatment": "Facts paraphrased and linked; no source images or substantial quotations reproduced; rights remain with respective owners" } } ], "facts": [ { "id": "PCT-20261109-DD-F01", "statement": "The order was signed September 9 and disclosed September 15, 2025; the initial value was $6.3 billion and its residual purchase obligation is conditional.", "source_ids": [ "PCT-20261109-DD-S01" ], "claim": "The order was signed September 9 and disclosed September 15, 2025; the initial value was $6.3 billion and its residual purchase obligation is conditional." }, { "id": "PCT-20261109-DD-F02", "statement": "The filed master agreement includes a redacted capacity allocation limit and termination provisions.", "source_ids": [ "PCT-20261109-DD-S02" ], "claim": "The filed master agreement includes a redacted capacity allocation limit and termination provisions." }, { "id": "PCT-20261109-DD-F03", "statement": "The $2 billion January share purchase was a separate completed equity transaction.", "source_ids": [ "PCT-20261109-DD-S06" ], "claim": "The $2 billion January share purchase was a separate completed equity transaction." }, { "id": "PCT-20261109-DD-F04", "statement": "Second-quarter corporate figures and September short-dated pricing are not order-specific execution disclosures.", "source_ids": [ "PCT-20261109-DD-S03", "PCT-20261109-DD-S04", "PCT-20261109-DD-S05" ], "claim": "Second-quarter corporate figures and September short-dated pricing are not order-specific execution disclosures." } ], "inferences": [ { "id": "PCT-20261109-DD-I01", "statement": "A fallback buyer can reduce eligible capacity demand risk while leaving delivery timing and construction funding risks.", "claim": "A fallback buyer can reduce eligible capacity demand risk while leaving delivery timing and construction funding risks." }, { "id": "PCT-20261109-DD-I02", "statement": "Portfolio financing may value contractual cash flows; this does not establish a loan-specific NVIDIA guarantee.", "claim": "Portfolio financing may value contractual cash flows; this does not establish a loan-specific NVIDIA guarantee." }, { "id": "PCT-20261109-DD-I03", "statement": "Share purchase proceeds, service revenue and buildout frameworks should not be added as a single compute-revenue measure.", "claim": "Share purchase proceeds, service revenue and buildout frameworks should not be added as a single compute-revenue measure." }, { "id": "PCT-20261109-DD-I04", "statement": "Annualized short-contract revenue per MW cannot reveal the private unit rate for the residual-capacity order.", "claim": "Annualized short-contract revenue per MW cannot reveal the private unit rate for the residual-capacity order." } ], "unknowns": [ { "uncertainty_id": "PCT-20261109-DD-U01", "subject": "Execution", "detail": "Reviewed primary disclosures do not quantify capacity purchased, invoices issued or cash collected under the September 2025 residual-capacity order. This is an evidence gap, not proof that no purchases occurred." }, { "uncertainty_id": "PCT-20261109-DD-U02", "subject": "Scope and unit price", "detail": "Full order-specific capacity, server mix, delivery schedule, prices and interaction with redacted master-agreement allocation limits are not established." }, { "uncertainty_id": "PCT-20261109-DD-U03", "subject": "Amendments", "detail": "No claim that every subsequent private modification has been reviewed; analysis describes the public September 2025 terms and selected later corporate disclosures." }, { "uncertainty_id": "PCT-20261109-DD-U04", "subject": "Corporate metrics", "detail": "Q2 active/contracted power, revenue and loans are company-wide as-of figures; none establishes NVIDIA-order-specific performance." }, { "uncertainty_id": "PCT-20261109-DD-U05", "subject": "Next report", "detail": "The November 10 earnings call is scheduled, not completed; no third-quarter outcome assumed." }, { "uncertainty_id": "PCT-20261109-DD-U06", "subject": "Price comparability", "detail": "September 2026 annualized revenue/MW is an issuer statement about short-dated contracts, not a price for this deal or an executable offer to this reader." } ], "correction_history": [], "license_status": "Original article; third-party facts paraphrased and linked. No open redistribution license verified for source documents.", "public_url": null, "schema_version": "pct-research-item-1.0", "desk_extensions": { "schema_version": "pct.deals-desk.v1", "edition_date": "2026-11-09", "section": "Deals Desk", "word_count": 997, "estimated_read_minutes": 5.0, "news_status": "Historical-contract analysis ahead of a scheduled November 10, 2026 update; not a new signing, delivery or execution announcement", "paragraphs": [ { "id": "PCT-20261109-DD-P01", "classification": "analysis_with_verified_schedule", "source_ids": [ "PCT-20261109-DD-S07" ], "text": "CoreWeave’s earnings call is scheduled for November 10. Ahead of that update, one older agreement deserves a careful reading: NVIDIA’s commitment to buy residual cloud capacity. Its value lies in giving eligible infrastructure another potential paying customer. How much protection that creates depends on the infrastructure being available, the contract’s scope and the payments actually triggered. [Earnings-call notice](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Date-of-Third-Quarter-2026-Financial-Results-and-Conference-Call/default.aspx)." }, { "id": "PCT-20261109-DD-P02", "classification": "verified_historical_contract_fact", "source_ids": [ "PCT-20261109-DD-S01" ], "text": "This is a September 2025 deal, not a new November signing. CoreWeave entered the order on September 9 and disclosed it on September 15. The initial value was $6.3 billion. Where its own customers do not fully use the relevant capacity, NVIDIA must purchase the unsold remainder through April 13, 2032, subject to delivery, service availability and termination provisions. [Transaction 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm)." }, { "id": "PCT-20261109-DD-P03", "classification": "analysis", "source_ids": [], "text": "The practical distinction is between finding a buyer and producing something the buyer can use. A fallback customer can make empty capacity less frightening. It cannot make a delayed electrical connection arrive, cool an unfinished room or turn uninstalled servers into a working cluster. Demand risk and construction risk sit at different points in the same business." }, { "id": "PCT-20261109-DD-SECTION01", "classification": "heading", "source_ids": [], "text": "The contract is narrower than the headline" }, { "id": "PCT-20261109-DD-P04", "classification": "verified_contract_fact", "source_ids": [ "PCT-20261109-DD-S02" ], "text": "The underlying master agreement dates to April 10, 2023. Its filed text makes CoreWeave responsible for allocating customer demand across its infrastructure and includes a limit on the capacity dedicated to NVIDIA, measured by server type. The limit itself is redacted. Reserved-compute fees sit in order forms; the public master agreement does not provide the September order’s unit prices. [Master services agreement, Sections 2 and 11](https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm)." }, { "id": "PCT-20261109-DD-P05", "classification": "analysis_with_contract_limits", "source_ids": [ "PCT-20261109-DD-S02" ], "text": "The same agreement permits termination for cause and releases the customer from payment obligations for services after qualifying termination, subject to its terms. That makes a blanket description such as insurance for every unsold CoreWeave GPU too broad. The reviewed documents establish an arrangement with conditions, not an unlimited guarantee of the company’s entire fleet or debts. [Master agreement, Section 10](https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm)." }, { "id": "PCT-20261109-DD-P06", "classification": "analysis", "source_ids": [], "text": "Consider a simplified operating decision. If a covered cluster is ready and another customer wants it, ordinary customer sales determine how much capacity remains. If no customer wants that eligible capacity, the fallback purchase can matter. If the cluster has not met delivery requirements, the commercial problem begins earlier. These scenarios are an analytical illustration, not a reconstruction of private billing rules." }, { "id": "PCT-20261109-DD-SECTION02", "classification": "heading", "source_ids": [], "text": "Why lenders care about the difference" }, { "id": "PCT-20261109-DD-P07", "classification": "verified_as_of_company_fact", "source_ids": [ "PCT-20261109-DD-S03" ], "text": "CoreWeave’s June-quarter filing describes $13.6 billion outstanding under delayed-draw term-loan facilities. Those facilities are collateralized by underlying assets and pledged contractual cash flows, generally from investment-grade counterparties. The company draws funds while building infrastructure and amortizes the loans as contractual cash arrives. This is a portfolio financing description, not evidence that every facility relies on this NVIDIA order. [June 30, 2026 10-Q, liquidity and capital resources](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm)." }, { "id": "PCT-20261109-DD-P08", "classification": "analysis", "source_ids": [], "text": "For a lender, a dependable contractual payer can improve the forecast of money available to service a loan. But the forecast still needs a starting date. Interest, procurement bills and site costs can arrive before a cluster earns revenue. A delay can therefore consume liquidity even where the eventual customer is highly creditworthy. The fallback purchase changes one input to underwriting; it does not settle the whole credit decision." }, { "id": "PCT-20261109-DD-P09", "classification": "verified_as_of_company_fact", "source_ids": [ "PCT-20261109-DD-S04" ], "text": "Actual operating progress belongs on a separate ledger. In its August 11 release, CoreWeave reported 1.5 gigawatts of active power and approximately 3.7 gigawatts of contracted power for the second quarter. It also reported $2.575 billion of quarterly revenue. These are company-wide disclosures, not a delivery certificate, utilization report or payment receipt for the September NVIDIA order. [Second-quarter results](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx)." }, { "id": "PCT-20261109-DD-P10", "classification": "analysis", "source_ids": [], "text": "Contracted power can describe a future resource. Active power is closer to operations, but it still does not reveal who used a particular server or which order produced an invoice. The public documents reviewed for this feature do not establish how much residual capacity NVIDIA has purchased under this specific order. That gap should stay visible rather than being filled with an assumed utilization rate." }, { "id": "PCT-20261109-DD-SECTION03", "classification": "heading", "source_ids": [], "text": "Three different kinds of NVIDIA support" }, { "id": "PCT-20261109-DD-P11", "classification": "verified_completed_transaction", "source_ids": [ "PCT-20261109-DD-S06" ], "text": "NVIDIA’s role extends beyond buying compute. A separate filing records a completed January 23, 2026 equity purchase: 22,935,780 CoreWeave shares at $87.20 each, for $2 billion in cash. The same January disclosure describes a collaboration aimed at accelerating more than five gigawatts of AI factories by 2030, while warning that further agreements and conditions would be involved. [January transaction 8-K](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000044/crwv-20260123.htm)." }, { "id": "PCT-20261109-DD-P12", "classification": "analysis", "source_ids": [], "text": "These items belong in different columns. A closed share sale supplies corporate capital and ownership. A capacity purchase pays for services under its conditions. A buildout framework expresses a program still requiring execution. Adding the stock investment to the capacity headline would produce a larger number without producing a meaningful cloud-revenue total. Nor does a strategic relationship by itself establish realized demand from independent end users." }, { "id": "PCT-20261109-DD-SECTION04", "classification": "heading", "source_ids": [], "text": "What a buyer can learn before the next report" }, { "id": "PCT-20261109-DD-P13", "classification": "attributed_issuer_contract_claim", "source_ids": [ "PCT-20261109-DD-S05" ], "text": "There is a newer pricing signal, but it concerns other contracts. On September 17, 2026, CoreWeave said it had signed three-to-six-month customer contracts at approximately $40 million per megawatt, measured as annualized revenue divided by power needed for the associated clusters. The announcement neither identifies that figure as NVIDIA’s rate nor supplies a universal GPU-hour quote. [September pricing announcement](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx)." }, { "id": "PCT-20261109-DD-P14", "classification": "analysis", "source_ids": [], "text": "A short contract’s annualized revenue density is useful for judging the supplier’s reported selling conditions. It cannot price a multiyear fallback purchase without knowing the hardware mix, covered capacity, service bundle and duration. For an enterprise negotiating compute, the useful comparison remains a matched configuration and delivery window, with explicit remedies for late capacity and a clear bill for storage, networking and support." }, { "id": "PCT-20261109-DD-P15", "classification": "analysis_and_open_questions", "source_ids": [], "text": "The next useful disclosure would connect the stages: covered capacity ready for service, capacity sold to ordinary customers, residual purchases triggered, and cash collected. Until that bridge is public, the defensible conclusion is modest but meaningful. CoreWeave has a disclosed contractual fallback purchaser for eligible unsold capacity. The commercial benefit must still travel through construction, delivery and invoicing before it becomes cash available to support the business." } ], "structured_record": { "record_id": "PCT-20261109-DD-D01", "seller": "CoreWeave, Inc.", "buyer": "NVIDIA Corporation", "deal_type": "Conditional purchase of residual unsold cloud computing capacity", "master_agreement_date": "2023-04-10", "signed_date": "2025-09-09", "announced_date": "2025-09-15", "contract_value_usd": 6300000000, "currency": "USD", "value_basis": "Disclosed initial value of September 2025 order; not described as cash received or guaranteed annual revenue", "purchase_obligation_through": "2032-04-13", "delivery_conditions": "Subject to delivery and service availability requirements and termination provisions", "executable_price_per_gpu_hour": null, "gpu_count": null, "capacity_mw": null, "delivered_capacity_mw": null, "nvidia_residual_capacity_purchased": null, "deal_cash_collected_usd": null, "execution_status": "Signed order verified; order-specific deliveries, triggered purchases, invoices and cash collection not quantified in reviewed public sources", "scope_caution": "Do not apply the $6.3B order to every future CoreWeave facility or treat it as a debt guarantee", "separate_equity_transaction": { "completed_date": "2026-01-23", "cash_usd": 2000000000, "shares": 22935780, "price_usd": 87.2, "source_id": "PCT-20261109-DD-S06", "included_in_compute_contract_value": false }, "id": "PCT-20261109-DD-D01", "content_type": "transaction_record", "schema_version": "pct-transaction-record-1.0", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "parent_article_id": "PCT-20261109-DD-A01", "source_ids": [ "PCT-20261109-DD-S01", "PCT-20261109-DD-S02", "PCT-20261109-DD-S03", "PCT-20261109-DD-S04", "PCT-20261109-DD-S05", "PCT-20261109-DD-S06", "PCT-20261109-DD-S07", "PCT-20261109-DD-S08" ], "factual_claims": "See parent article facts array and linked source IDs; management claims and forecasts retain their labels.", "inference": "See parent article inferences array; no transaction price or delivered capacity is inferred from the contract value.", "unknowns": [ { "uncertainty_id": "PCT-20261109-DD-U01", "subject": "Execution", "detail": "Reviewed primary disclosures do not quantify capacity purchased, invoices issued or cash collected under the September 2025 residual-capacity order. This is an evidence gap, not proof that no purchases occurred." }, { "uncertainty_id": "PCT-20261109-DD-U02", "subject": "Scope and unit price", "detail": "Full order-specific capacity, server mix, delivery schedule, prices and interaction with redacted master-agreement allocation limits are not established." }, { "uncertainty_id": "PCT-20261109-DD-U03", "subject": "Amendments", "detail": "No claim that every subsequent private modification has been reviewed; analysis describes the public September 2025 terms and selected later corporate disclosures." }, { "uncertainty_id": "PCT-20261109-DD-U04", "subject": "Corporate metrics", "detail": "Q2 active/contracted power, revenue and loans are company-wide as-of figures; none establishes NVIDIA-order-specific performance." }, { "uncertainty_id": "PCT-20261109-DD-U05", "subject": "Next report", "detail": "The November 10 earnings call is scheduled, not completed; no third-quarter outcome assumed." }, { "uncertainty_id": "PCT-20261109-DD-U06", "subject": "Price comparability", "detail": "September 2026 annualized revenue/MW is an issuer statement about short-dated contracts, not a price for this deal or an executable offer to this reader." } ], "correction_history": [], "license_status": "Original article; third-party facts paraphrased and linked. No open redistribution license verified for source documents." }, "uncertainty": [ { "id": "PCT-20261109-DD-U01", "subject": "Execution", "status": "Reviewed primary disclosures do not quantify capacity purchased, invoices issued or cash collected under the September 2025 residual-capacity order. This is an evidence gap, not proof that no purchases occurred." }, { "id": "PCT-20261109-DD-U02", "subject": "Scope and unit price", "status": "Full order-specific capacity, server mix, delivery schedule, prices and interaction with redacted master-agreement allocation limits are not established." }, { "id": "PCT-20261109-DD-U03", "subject": "Amendments", "status": "No claim that every subsequent private modification has been reviewed; analysis describes the public September 2025 terms and selected later corporate disclosures." }, { "id": "PCT-20261109-DD-U04", "subject": "Corporate metrics", "status": "Q2 active/contracted power, revenue and loans are company-wide as-of figures; none establishes NVIDIA-order-specific performance." }, { "id": "PCT-20261109-DD-U05", "subject": "Next report", "status": "The November 10 earnings call is scheduled, not completed; no third-quarter outcome assumed." }, { "id": "PCT-20261109-DD-U06", "subject": "Price comparability", "status": "September 2026 annualized revenue/MW is an issuer statement about short-dated contracts, not a price for this deal or an executable offer to this reader." } ], "history": [ { "id": "PCT-20261109-DD-H01", "event_date": "2023-04-10", "event": "Master services agreement dated", "stage": "executed_framework", "source_ids": [ "PCT-20261109-DD-S02" ] }, { "id": "PCT-20261109-DD-H02", "event_date": "2025-09-09", "event": "New residual-capacity order entered", "stage": "signed_order", "source_ids": [ "PCT-20261109-DD-S01" ] }, { "id": "PCT-20261109-DD-H03", "event_date": "2025-09-15", "event": "Order and initial $6.3 billion value publicly disclosed", "stage": "public_announcement", "source_ids": [ "PCT-20261109-DD-S01" ] }, { "id": "PCT-20261109-DD-H04", "event_date": "2025-09-25", "event": "Redacted master agreement filed as S-4/A exhibit", "stage": "contract_disclosure", "source_ids": [ "PCT-20261109-DD-S02" ] }, { "id": "PCT-20261109-DD-H05", "event_date": "2026-01-23", "event": "Separate $2 billion NVIDIA share purchase completed", "stage": "completed_equity_transaction", "source_ids": [ "PCT-20261109-DD-S06" ] }, { "id": "PCT-20261109-DD-H06", "event_date": "2026-01-26", "event": "Share purchase and buildout collaboration disclosed", "stage": "public_announcement", "source_ids": [ "PCT-20261109-DD-S06" ] }, { "id": "PCT-20261109-DD-H07", "event_date": "2026-06-30", "event": "Quarter-end observation date for company-wide operating and financing metrics", "stage": "as_of_company_snapshot", "source_ids": [ "PCT-20261109-DD-S03", "PCT-20261109-DD-S04" ] }, { "id": "PCT-20261109-DD-H08", "event_date": "2026-08-11", "event": "Second-quarter results released", "stage": "issuer_results", "source_ids": [ "PCT-20261109-DD-S04" ] }, { "id": "PCT-20261109-DD-H09", "event_date": "2026-08-12", "event": "Second-quarter 10-Q filing date", "stage": "regulatory_filing", "source_ids": [ "PCT-20261109-DD-S03", "PCT-20261109-DD-S08" ] }, { "id": "PCT-20261109-DD-H10", "event_date": "2026-09-17", "event": "Separate short-dated contract pricing disclosed", "stage": "issuer_pricing_claim", "source_ids": [ "PCT-20261109-DD-S05" ] }, { "id": "PCT-20261109-DD-H11", "event_date": "2026-10-27", "event": "November 10 earnings call announced", "stage": "schedule_announcement", "source_ids": [ "PCT-20261109-DD-S07" ] }, { "id": "PCT-20261109-DD-H12", "event_date": "2026-11-10", "event": "Earnings call scheduled for 5:00 PM Eastern; not yet held as of edition", "stage": "future_scheduled_event", "source_ids": [ "PCT-20261109-DD-S07" ] } ], "dates": { "edition": "2026-11-09", "master_agreement": "2023-04-10", "agreement": "2025-09-09", "announcement": "2025-09-15", "contract_exhibit_publication": "2025-09-25", "latest_cited_primary_disclosure": "2026-10-27", "operating_snapshot_as_of": "2026-06-30", "next_scheduled_update": "2026-11-10T17:00:00-05:00" }, "licensing": { "article": "Original prose and analysis composed for this edition; no third-party visual assets or substantial source text reproduced", "sources": "Public availability is not treated as an open republication license; source attribution and direct links preserved", "direct_quoted_words": 0 }, "dedup": { "checked_date": "2026-11-09", "scope": "Canonical PCT editions dated November 3–8, 2026", "result": "No earlier article on this lead in the checked pilot editions." }, "editorial_cautions": [ "Keep September 2025 dates prominent; do not imply new November agreement.", "Initial order value is not collected cash.", "Do not describe all CoreWeave capacity or debt as guaranteed.", "Do not present a redacted allocation limit as known or applicable unchanged to all later order forms.", "Separate January share-sale proceeds from cloud revenue.", "Corporate power and financing metrics do not establish this order’s delivery or invocation.", "No Q3 results available or implied before the scheduled November 10 call.", "The example scenarios are analysis, not disclosed billing mechanics." ], "provenance": { "reporting_method": "Public primary-document review; no interviews, purchases or outreach", "written_date": "2026-11-09", "body_sha256": "b2a248f794f09c2042e346e8df33ac772d64256a544ff2f6d4bf62135a623e36", "last_reviewed": "2026-11-09" } }, "public_projection": { "version": "1.0.0", "date": "2026-11-10", "changes": [ "Internal workspace artifact paths omitted where present. Historical license/status fields retain original provenance; current public reuse policy is stated separately." ], "rights": "PCT reserves rights in original editorial content; third-party rights retained. No general source-data reuse license." } }, { "id": "pct-2026-11-09-formation-opening-cross", "content_type": "article", "desk": "formation", "title": "A compute market needs a meeting time", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "editorial_voice": "AI-assisted Formation Desk | accessible academic analysis", "body": "A compute buyer can refresh a marketplace all morning without finding the combination of hardware, start time and price it needs. A supplier can watch the same quiet screen and conclude that nobody wants its capacity. If the two parties arrive at different times, a continuously open market may make available interest harder to see.\n\nA scheduled auction offers one possible response: give compatible buyers and sellers an appointment. Nasdaq's Opening Cross is a useful historical example of how that appointment becomes an institution. It combines a deadline with defined orders, information before execution and rules that determine which commitments become trades. The compute application depends on finding customers who can wait and a product they can genuinely substitute.\n\n## How the appointment becomes a market\n\nThe SEC approved Nasdaq's new opening process on September 16, 2004. Its order described three components: orders specifically for the open, dissemination of order imbalances, and a 9:30 a.m. cross that would execute the maximum number of shares at one price. Nasdaq reported in January 2005 that it had completed the rollout to all Nasdaq-listed securities the previous December. [SEC approval, published September 23, 2004](https://www.govinfo.gov/content/pkg/FR-2004-09-23/pdf/E4-2349.pdf); [Nasdaq year-end report, January 13, 2005](https://ir.nasdaq.com/node/57576).\n\nThere was a consequential argument about participation. Bloomberg objected that requiring automatic execution would exclude trading interest from electronic networks using order delivery. Nasdaq answered that matching orders at one instant was incompatible with an iterative delivery process and described alternative participation routes. The SEC approved the proposal after considering those objections. The record shows that making orders executable together also required decisions about access. [SEC approval, sections III and IV](https://www.govinfo.gov/content/pkg/FR-2004-09-23/pdf/E4-2349.pdf).\n\nThe timing rules have changed since 2004. An April 2021 Nasdaq notice announced earlier imbalance information from 9:25 a.m., alongside restrictions on cancelling or modifying on-open orders once that information began. New qualifying orders could still arrive under specified cutoffs. This paired additional information with firmer commitments. Those later rules should not be read back into the original launch. [Nasdaq trader alert, April 16, 2021](https://www.nasdaqtrader.com/TraderNews.aspx?id=ETA2021-29).\n\nNasdaq's current rulebook starts price selection by maximizing executed shares, then applies tie-breaks including minimizing unmatched interest. It also specifies price protections and circumstances in which no cross occurs. A scheduled auction therefore needs more than the instruction to pick a price that clears. It needs published answers for ties, unusual prices and failure. [Nasdaq Rule 4752(d), retrieved November 9, 2026](https://listingcenter.nasdaq.com/rulebook/nasdaq/rules/nasdaq-equity-4).\n\n## Five compute blocks at one price\n\nConsider an invented compute auction at 10:00 UTC for service beginning at noon. Each block contains eight GPUs on one approved host configuration for four uninterrupted hours. Hardware, memory, interconnect, software image, region, storage allowance and remedies for failed delivery are fixed beforehand. Each buyer has already tested the eligible suppliers and can use any offered block. Different blocks run independent jobs; nobody needs a sixteen-GPU job spread across two hosts.\n\nSupplier A offers three blocks at a minimum of $80 each. B offers three at $100. C offers two at $120. Buyer X wants two blocks at up to $150 each, Y wants three at up to $120, and Z wants two at up to $90. Each participant accepts an allocation of fewer whole blocks than requested. These quantities and prices are illustrative, not observed offers.\n\nFor this pilot, suppose the published rule considers submitted price levels, maximizes the number of matched blocks, then minimizes the difference between eligible supply and demand. If a tie remains, it chooses the lowest qualifying price. Lower asks and higher bids receive allocation priority, with a previously announced rule for equal-price orders. This is an illustrative design, not a reproduction of Nasdaq's algorithm.\n\nAt $100, six blocks are offered and five are demanded. Five trade. At $120, eight are offered and five demanded, so five still trade but more supply remains unmatched. At $90 only three blocks are offered. The stated rule therefore selects $100: A sells three blocks, B sells two, X buys two and Y buys three. Z's limit is too low to trade with the remaining suppliers.\n\nFive blocks represent 160 GPU-hours. Buyers pay $500 in total, equivalent to $3.125 per GPU-hour under the example's assumptions. Trading fees and taxes are omitted; any real comparison must add them and other excluded charges. The arithmetic establishes an allocation for these orders. It does not establish that the price is competitive, that bids reveal true values, or that useful work was delivered.\n\nBefore bidding, the suppliers must reserve the offered capacity and the buyers must authorize purchases within their limits. Credentials, payment arrangements and access checks must be ready. If winning merely starts another negotiation, a participant may withdraw while everyone else is committed. This is the compute counterpart of the historical dispute over which orders could participate in a simultaneous match.\n\n## The cost of waiting together\n\nThe proposal works only for demand that can tolerate the appointment. An overnight evaluation batch might wait for a morning auction; an inference service facing an immediate traffic surge might pay more to start now. Pooling orders can improve the opportunity to match, but waiting consumes time. A pilot must count customers who leave before the auction, not only those who remain to trade.\n\nCompute also makes the product boundary unusually important. Two eight-GPU blocks cannot automatically satisfy one sixteen-GPU training request. A buyer that needs both blocks or neither introduces a combination constraint absent from the simple example. Splitting the market by topology, region or software requirements improves comparability while leaving fewer orders in each auction. One cannot assume both broad participation and close substitutability.\n\nNor does a uniform price prevent strategic bidding. A large supplier could withhold capacity, and a large buyer could shade demand. The opening price in a securities market also has surrounding quotations and established market institutions. A new compute venue cannot borrow their credibility merely by copying a price-selection rule.\n\nA useful test would compare scheduled auctions with immediate posted offers for the same qualified product. Measure time to usable capacity, total delivered cost, completed work, failed starts and participation over repeated sessions. Record whether apparent volume came from new matches or simply moved existing business to a different clock. The historical lesson is a specific design option: concentrate executable interest at an agreed time. Whether the additional matches justify waiting is a question the compute market must answer for itself.", "sources": [ { "id": "pct-source-sec-opening-20040923", "title": "Order Approving Proposed Rule Change Relating to Improved Nasdaq Opening Process, Release 34-50405", "publisher": "U.S. Securities and Exchange Commission / Federal Register", "url": "https://www.govinfo.gov/content/pkg/FR-2004-09-23/pdf/E4-2349.pdf", "published_at": "2004-09-23", "decision_date": "2004-09-16", "retrieved_at": "2026-11-09", "date_note": "Order dated September 16; Federal Register publication September 23. Dates read from document, not crawler metadata.", "evidence_type": "Primary regulator approval order", "supports": "Three-part original mechanism; automatic-execution participation dispute; Nasdaq response; SEC approval.", "verification": "Read official two-page PDF text, printed pages 57118–57119, sections II–V. No claim that stakeholder arguments establish realized effects.", "license_status": "U.S. federal government-authored order, generally public domain in the United States. Original paraphrase; no source layout or substantial quotation reproduced.", "publication_date": "2004-09-23", "observation_date": "2026-11-09" }, { "id": "pct-source-nasdaq-yearend-20050113", "title": "NASDAQ Announces Market Year-end Statistics for 2004", "publisher": "Nasdaq", "url": "https://ir.nasdaq.com/node/57576", "published_at": "2005-01-13", "retrieved_at": "2026-11-09", "date_note": "Publication date verified on page; December 2004 is the reported rollout-completion month, not the article date.", "evidence_type": "Primary exchange retrospective announcement", "supports": "Nasdaq reported completion of Opening Cross rollout to all Nasdaq-listed securities in December 2004.", "verification": "Read dated official HTML and NASDAQ Products and Services section. Promotional claims of accuracy or success are not adopted as independent empirical findings.", "license_status": "Copyright Nasdaq; factual paraphrase only. No data table, chart, logo or source prose reproduced.", "publication_date": "2005-01-13", "observation_date": "2026-11-09" }, { "id": "pct-source-nasdaq-opening-alert-20210416", "title": "Equity Trader Alert 2021-29: Reminder, Nasdaq to Implement Enhancements to the Opening Cross Process", "publisher": "Nasdaq Trader", "url": "https://www.nasdaqtrader.com/TraderNews.aspx?id=ETA2021-29", "published_at": "2021-04-16", "effective_date_announced": "2021-05-17", "retrieved_at": "2026-11-09", "date_note": "Notice date and announced effective date read from official page; historical notice not presented as proof of every current operational detail.", "evidence_type": "Primary exchange implementation notice", "supports": "Earlier information from 9:25, on-open cancellation/modification restrictions, and later qualified order entry under cutoffs.", "verification": "Read full operative announcement, including early-information fields, restrictions, and 9:28/9:29:30/9:30 cutoffs. Article intentionally gives a high-level summary.", "license_status": "Copyright Nasdaq; factual paraphrase only. No reproduction of specifications or market data.", "publication_date": "2021-04-16", "observation_date": "2026-11-09" }, { "id": "pct-source-nasdaq-rule4752-20261109", "title": "Nasdaq Equity 4, Rule 4752: Opening Process", "publisher": "Nasdaq Listing Center", "url": "https://listingcenter.nasdaq.com/rulebook/nasdaq/rules/nasdaq-equity-4", "published_at": null, "version_date": "2026-11-09", "retrieved_at": "2026-11-09", "date_note": "Continuously maintained rulebook, observed November 9, 2026. This is an observation date, not an asserted publication date; subsection has its own amendment history.", "evidence_type": "Primary exchange rules", "supports": "Price selection begins with maximum executions; further tie-breaks, price protections and cancellation conditions apply.", "verification": "Read Rule 4752(a) and (d), especially (d)(2)(A)–(F). Distinguishes indicative reference prices from actual cross-price selection. Simplified compute rule explicitly differs from Nasdaq.", "license_status": "Published exchange rules; no open-content license verified. Factual paraphrase only, with no substantial verbatim reproduction.", "publication_date": null, "observation_date": "2026-11-09" } ], "facts": [ { "id": "pct-20261109-formation-f01", "claim": "The SEC approved the new opening process on September 16, 2004; its order was published September 23 and described on-open orders, imbalance dissemination and a single-price cross.", "source_ids": [ "pct-source-sec-opening-20040923" ] }, { "id": "pct-20261109-formation-f02", "claim": "Nasdaq reported on January 13, 2005 that rollout to all Nasdaq-listed securities was completed in December 2004.", "source_ids": [ "pct-source-nasdaq-yearend-20050113" ] }, { "id": "pct-20261109-formation-f03", "claim": "The 2004 approval record includes Bloomberg’s objection to automatic-execution eligibility and Nasdaq’s response that iterative order delivery conflicted with simultaneous matching.", "source_ids": [ "pct-source-sec-opening-20040923" ] }, { "id": "pct-20261109-formation-f04", "claim": "An April 16, 2021 notice announced earlier imbalance information and on-open cancellation/modification restrictions, effective May 17, 2021.", "source_ids": [ "pct-source-nasdaq-opening-alert-20210416" ] }, { "id": "pct-20261109-formation-f05", "claim": "Rule 4752(d), read November 9, 2026, starts with maximizing executed shares and includes further tie-breaks, price protections and conditions in which no Opening Cross occurs.", "source_ids": [ "pct-source-nasdaq-rule4752-20261109" ] } ], "inferences": [ { "id": "pct-2026-11-09-formation-opening-cross-I001", "claim": "An announced compute auction could concentrate otherwise dispersed, compatible interest; this is a proposed design mechanism, not an established empirical benefit.", "label": "PCT editorial inference" }, { "id": "pct-2026-11-09-formation-opening-cross-I002", "claim": "For simultaneous compute matching, purchase authority and reserved capacity should precede execution, rather than leaving bilateral negotiation until after allocation.", "label": "PCT editorial inference" }, { "id": "pct-2026-11-09-formation-opening-cross-I003", "claim": "Restricting the auction to genuinely substitutable capacity improves interpretability but can reduce participation.", "label": "PCT editorial inference" }, { "id": "pct-2026-11-09-formation-opening-cross-I004", "claim": "Any evaluation of batch auctions should count waiting, abandoned demand and failed delivery alongside matched volume.", "label": "PCT editorial inference" } ], "unknowns": [ "No compute marketplace, provider or customer is reported to offer or accept the illustrative auction.", "All compute specifications, quantities, prices, limits, eligibility and willingness to wait in the example are invented assumptions, not market observations or supplier quotes.", "No causal estimate of the Nasdaq mechanism’s effect on liquidity, welfare or price accuracy is asserted.", "The proposed compute rule is not Nasdaq’s algorithm and is not claimed to be strategy-proof, welfare-optimal or legally approved.", "The example omits trading fees and taxes and assumes included delivery components are specified; actual delivered costs and performance require measurement.", "A small venue’s surrounding quotations, concentration, regulatory obligations, contract enforceability and remedies would need separate investigation." ], "correction_history": [], "license_status": "Original PCT editorial prose; public license not selected. Source rights remain separate. No source images, datasets, code or substantial quotations reproduced.", "public_url": null, "schema_version": "pct-research-item-1.0", "desk_extensions": { "estimated_reading_minutes": 5.0, "word_count": 1054, "historical_institution": "Nasdaq Opening Cross, introduced in 2004; original regulator record distinguished from 2021 changes and current rules.", "hypothetical_example": { "status": "Invented mechanism and order book; not a live offer, benchmark, transaction or investment recommendation.", "auction_time_utc": "10:00", "service_start_utc": "12:00", "gpus_per_block": 8, "service_hours": 4, "currency": "USD", "supply": [ { "seller": "A", "quantity_blocks": 3, "ask_per_block": 80 }, { "seller": "B", "quantity_blocks": 3, "ask_per_block": 100 }, { "seller": "C", "quantity_blocks": 2, "ask_per_block": 120 } ], "demand": [ { "buyer": "X", "quantity_blocks": 2, "bid_per_block": 150 }, { "buyer": "Y", "quantity_blocks": 3, "bid_per_block": 120 }, { "buyer": "Z", "quantity_blocks": 2, "bid_per_block": 90 } ], "rule": "Consider submitted price levels; maximize minimum of eligible supply and demand; minimize absolute supply-demand imbalance; lowest price if still tied. Allocate by price priority then a preannounced equal-price rule. Whole-block partial quantity fills allowed.", "clearing_price_per_block": 100, "matched_blocks": 5, "gpu_hours": 160, "gross_payment": 500, "price_per_gpu_hour_before_omitted_fees_taxes": 3.125, "supply_allocation": { "A": 3, "B": 2, "C": 0 }, "demand_allocation": { "X": 2, "Y": 3, "Z": 0 }, "limitations": [ "All blocks prequalified as substitutes for each buyer.", "Independent jobs; no multi-block all-or-nothing requirement.", "Trading fees and taxes omitted.", "Winning price does not demonstrate delivery, true valuation or competitive market conditions." ] }, "audit": { "primary_sources_only": true, "live_quote_claims": false, "historical_analogy_not_equivalence": true, "facts_separated_from_inference": true, "prior_topics_avoided": [ "freight clearing", "Nord Pool", "grain inspection", "airport slots", "gas recalls", "FCC repacking" ] }, "schema_version": "pct.article.v1", "dek": "Nasdaq's Opening Cross shows how a scheduled auction combines information, firm orders and allocation rules. A compute version must earn back the time customers spend waiting." }, "public_projection": { "version": "1.0.0", "date": "2026-11-10", "changes": [ "Internal workspace artifact paths omitted where present. Historical license/status fields retain original provenance; current public reuse policy is stated separately." ], "rights": "PCT reserves rights in original editorial content; third-party rights retained. No general source-data reuse license." } }, { "id": "pct-2026-11-09-spotlight-e10e8ebe80", "content_type": "article", "desk": "spotlight", "title": "Keep the OCR queue fed, and keep the receipts", "publication_date": "2026-11-09", "observation_date": "2026-11-09", "editorial_voice": "AI-assisted Spotlight Desk; practical builder voice", "body": "# Keep the OCR queue fed, and keep the receipts\n\n*Product spotlight | Peterson Compute Trading | November 9, 2026*\n*AI-assisted public-source analysis. No OCR benchmark, deployment or purchase was performed. Dollar figures are US dollars.*\n\nGive a builder 100,000 scanned catalogue pages and a search box to fill. The job is ordinary: recover printed text, preserve its page coordinates, and let readers jump back to the original image. Our proposed collection is rights-cleared, non-sensitive and mostly English. It is a hypothetical workload, not a customer result.\n\nThe tempting question is which GPU to rent. Start one level earlier: how often will the worker stop, and where will its finished pages live? Runpod Serverless makes those questions unusually concrete. Its queue can keep a worker processing successive requests, but running time includes initialization and waiting. Its completed API results also expire. A successful OCR run needs both a controlled bill and a durable output. [Worker lifecycle](https://docs.runpod.io/serverless/workers/overview), [result retention](https://docs.runpod.io/serverless/endpoints/send-requests)\n\n## Define a usable page\n\nUse PaddleOCR's PP-OCRv5 server detection and recognition models as the candidate engine. The developer's model cards document the two stages separately: locate text, then recognize text lines. They do not establish accuracy on this catalogue. Keep the source image beside the output so readers can verify part numbers, punctuation and tiny type. [Detection model](https://huggingface.co/PaddlePaddle/PP-OCRv5_server_det), [recognition model](https://huggingface.co/PaddlePaddle/PP-OCRv5_server_rec)\n\nPin the model files, pipeline options, dependency versions and container digest. The current PaddleOCR tutorial supports explicit model selection and GPU execution; relying on an unspecified default can change the experiment later. For this first pass, target text search rather than faithful reconstruction of tables or illustrations. [Pipeline documentation](https://www.paddleocr.ai/latest/en/version3.x/pipeline_usage/OCR.html)\n\nPrepare a human-checked sample covering faint scans, rotated pages, dense columns and similar-looking characters. Define acceptance before measuring speed: correct searchable identifiers, usable text coverage and a reliable page link. Track failures separately. A page returned by the API is not automatically a page worth indexing.\n\n## Read the meter literally\n\nRunpod's endpoint-settings documentation, updated November 4, lists its 24-GB category of L4, A5000 and 3090 GPUs at $0.00019 per second. That is $0.684 per worker-hour with one GPU. This is a documented planning rate, not an account quote or a promise of a particular card. The public pricing-page fetch was unreadable in this check, so verify the console rate before spending. [Endpoint settings](https://docs.runpod.io/serverless/endpoints/endpoint-configurations)\n\nThe billing guide, updated October 15, says charges run from worker start until full stop, rounded up to a second. Model loading can be billed; ordinary image pulls and cached-model downloads normally occur before the worker starts. A running worker's idle-timeout window is billed too. The console label Idle instead means the worker has scaled down and is unbilled. Those two uses of idle deserve different counters. [Billing rules](https://docs.runpod.io/serverless/pricing)\n\nFor the proposed trial, set zero active workers, cap concurrency, and explicitly set timeouts. Console and REST v2 defaults differ, so inspect the resulting configuration rather than assuming a copied example matches the console. GPU categories can contain different hardware, and fallback choices can change the execution profile. Record the actual device on every run. [Configuration reference](https://docs.runpod.io/serverless/endpoints/endpoint-configurations)\n\n## The same GPU, two different bills\n\nHere is an intentionally hypothetical cost model. Assume 100,000 pages, two seconds of processing per page, twenty seconds of billed startup per worker activation, and a five-second billed tail before shutdown. Processing includes input handling and durable output writes within the worker. None of these timings has been measured.\n\nIn a fragmented schedule, suppose each worker activation handles ten pages before the queue empties. Ten thousand activations consume 200,000 processing seconds plus 250,000 startup-and-tail seconds. At the documented rate, the compute bill is $85.50.\n\nIn a well-fed schedule, suppose each activation handles 1,000 pages before shutting down. One hundred activations consume the same 200,000 processing seconds plus 2,500 overhead seconds. The compute bill becomes $38.48, a hypothetical reduction of $47.03, or 55%.\n\nThis comparison changes the number of worker activations, not model speed. It does not require a thousand-page API request. Keep jobs small and feed them successively: Runpod documents that workers take queued jobs back to back. One-page requests do not imply one cold start per page. Conversely, batching requests cannot guarantee the assumed hundred activations; arrival timing, scaling, failures and available capacity decide what happens. [Worker execution sequence](https://docs.runpod.io/serverless/workers/overview)\n\nThe arithmetic is simply rate multiplied by processing seconds plus activation overhead. Add 2% extra processing for retries to the well-fed case, while holding activations fixed, and compute rises to $39.24. If only 95,000 unique pages meet the quality bar, that is about $0.413 per thousand accepted pages. These are sensitivity inputs, not reliability claims. Storage, external preparation, transfer charges elsewhere, taxes and human review remain outside the model.\n\n## The receipt must outlive the request\n\nRunpod's request guide, updated October 19, gives asynchronous /run results a thirty-minute retention window after completion. The job's TTL is a separate clock covering queueing and execution. Neither is an archive. Leaving collection until tomorrow can therefore lose access to completed API results. [Request policies and retention](https://docs.runpod.io/serverless/endpoints/send-requests)\n\nHave each proposed worker write its page result to durable storage before returning success. Derive its key from the input hash, page number, model revision and processing options. Commit atomically, retain the original coordinates, and return a compact reference. A retry checks for a valid existing result. A manifest records accepted, rejected and pending pages, with capped attempts and a quarantine for corrupt inputs. This is a design proposal, not implemented exactly-once execution.\n\nStage submissions so queue age stays within the configured TTL. Set execution limits around measured page-runtime tails, not the average. Reconcile source pages against durable outputs and the bill. Record worker starts, billed seconds, device identity, failed attempts and accepted unique pages. That makes a faster GPU's eventual contribution measurable without confusing it with a better-fed queue.\n\nPaddleOCR's repository is Apache-2.0 licensed, and both selected model cards declare Apache-2.0. Preserve applicable license and attribution notices, mark distributed modifications, and review additional dependencies. Those licenses do not grant rights to the scanned catalogue. [Repository license](https://github.com/PaddlePaddle/PaddleOCR/blob/main/LICENSE), [model licenses](https://huggingface.co/PaddlePaddle/PP-OCRv5_server_rec)\n\nThe first trial should prove that a finished page survives, a retry does not duplicate it, and the billed lifetime matches the ledger. Then optimize dollars per accepted page. The most useful GPU purchase may be the one postponed until those three checks pass.\n", "sources": [ { "id": "SP-S1", "title": "Endpoint settings", "publisher": "Runpod", "published": null, "updated": "2026-11-04", "url": "https://docs.runpod.io/serverless/endpoints/endpoint-configurations", "accessed": "2026-11-09", "access_note": "Opened and checked 24-GB GPU-category rate USD 0.00019/second, hardware grouping, worker controls and explicit warning that console and REST v2 defaults differ. Page displays last-modified November 4, 2026.", "license_status": "Public documentation; reuse license not established. Linked and paraphrased, no screenshots reproduced.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S2", "title": "Serverless pricing", "publisher": "Runpod", "published": null, "updated": "2026-10-15", "url": "https://docs.runpod.io/serverless/pricing", "accessed": "2026-11-09", "access_note": "Opened worker-lifetime billing, rounding up to the next whole second, cold-start and waiting billing, unbilled scaled-down state, storage as separate component, and active-worker discount available only through sales inquiry. No older active-worker percentage used.", "license_status": "Public documentation; reuse license not established. Linked and paraphrased.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S3", "title": "Workers overview", "publisher": "Runpod", "published": null, "updated": "2026-10-15", "url": "https://docs.runpod.io/serverless/workers/overview", "accessed": "2026-11-09", "access_note": "Opened lifecycle showing cached preparation before running, back-to-back processing, billed waiting, scale-down and console Idle distinction.", "license_status": "Public documentation; reuse license not established. Linked and paraphrased.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S4", "title": "Send API requests", "publisher": "Runpod", "published": null, "updated": "2026-10-19", "url": "https://docs.runpod.io/serverless/endpoints/send-requests", "accessed": "2026-11-09", "access_note": "Opened TTL versus execution timeout and separate result-retention table: async /run 30 minutes; sync /runsync 1 minute. Last-modified date visible.", "license_status": "Public documentation; reuse license not established. Linked and paraphrased.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S5", "title": "OCR pipeline usage tutorial", "publisher": "PaddleOCR project", "published": null, "updated": null, "url": "https://www.paddleocr.ai/latest/en/version3.x/pipeline_usage/OCR.html", "accessed": "2026-11-09", "access_note": "Opened explicit model selection, local model paths, version selection and GPU execution. Current tutorial defaults can differ from PP-OCRv5; example does not claim v5 is latest. No trustworthy page publication/update date established.", "license_status": "Public project documentation; no extensive quotation or reproduced image. Repository software license checked separately.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S6", "title": "PP-OCRv5_server_det model card", "publisher": "PaddlePaddle", "published": null, "updated": null, "url": "https://huggingface.co/PaddlePaddle/PP-OCRv5_server_det", "accessed": "2026-11-09", "access_note": "Opened publisher model card; verified text-detection role, model identification and Apache-2.0 declaration. No workload performance claim adopted.", "license_status": "Publisher model card declares Apache-2.0; review pinned artifact and additional dependencies before distribution.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S7", "title": "PP-OCRv5_server_rec model card", "publisher": "PaddlePaddle", "published": null, "updated": null, "url": "https://huggingface.co/PaddlePaddle/PP-OCRv5_server_rec", "accessed": "2026-11-09", "access_note": "Opened publisher model card; verified text-line-recognition role, English support and Apache-2.0 declaration. Published evaluation metrics not used as catalogue benchmark.", "license_status": "Publisher model card declares Apache-2.0; input-document rights remain separate.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S8", "title": "PaddleOCR LICENSE", "publisher": "PaddlePaddle", "published": null, "updated": null, "url": "https://github.com/PaddlePaddle/PaddleOCR/blob/main/LICENSE", "accessed": "2026-11-09", "access_note": "Opened Apache License 2.0 and redistribution obligations including license copies, modified-file notices and applicable attribution/NOTICE material. No source code copied.", "license_status": "Apache-2.0 for covered repository software; other dependencies and datasets require their own review.", "publication_date": null, "observation_date": "2026-11-09" }, { "id": "SP-S9", "title": "Runpod pricing page (fetch limitation)", "publisher": "Runpod", "published": null, "updated": null, "url": "https://www.runpod.io/pricing", "accessed": "2026-11-09", "access_note": "Attempted direct page open returned only an edge dispatch response rather than a readable price table. Not used as support for a rate; current documented configuration rate comes from SP-S1. Console quote not checked.", "license_status": "Not reproduced; no usable price evidence from this fetch.", "evidence_status": "attempted_unreadable", "publication_date": null, "observation_date": "2026-11-09" } ], "facts": [ { "id": "SP-F1", "claim": "Runpod endpoint-settings documentation lists L4/A5000/3090 24-GB category at USD 0.00019 per second.", "source_ids": [ "SP-S1" ], "evidence_type": "documented_configuration_price_not_account_quote" }, { "id": "SP-F2", "claim": "Serverless bills worker start through full stop, rounded up to a second. Billed phases include cold start, execution and running idle-timeout wait; scaled-down Idle is not billed.", "source_ids": [ "SP-S2", "SP-S3" ], "evidence_type": "current_billing_documentation" }, { "id": "SP-F3", "claim": "Workers process queued jobs back to back; a new job arriving during their running waiting window can start without another cold start.", "source_ids": [ "SP-S3" ], "evidence_type": "current_documentation" }, { "id": "SP-F4", "claim": "Console and REST v2 endpoint defaults differ. Active workers, maximum workers, idle timeout and execution timeout are configurable.", "source_ids": [ "SP-S1" ], "evidence_type": "current_documentation" }, { "id": "SP-F5", "claim": "Async /run results are retained 30 minutes after completion; TTL independently covers the job lifespan including queueing and execution.", "source_ids": [ "SP-S4" ], "evidence_type": "current_documentation" }, { "id": "SP-F6", "claim": "PaddleOCR documents explicit version/model selection, local model paths and GPU execution. The named PP-OCRv5 server models perform detection and text-line recognition respectively.", "source_ids": [ "SP-S5", "SP-S6", "SP-S7" ], "evidence_type": "developer_documentation" }, { "id": "SP-F7", "claim": "PaddleOCR repository is Apache-2.0 licensed and selected publisher model cards both declare Apache-2.0.", "source_ids": [ "SP-S6", "SP-S7", "SP-S8" ], "evidence_type": "published_software_and_model_license" } ], "inferences": [ { "id": "SP-I1", "claim": "The documented rate equals USD 0.684 per worker-hour for the one-GPU modeled worker.", "basis_fact_ids": [ "SP-F1" ], "label": "PCT arithmetic, not an account quote", "assumptions": [ "One GPU in the documented 24-GB category; compute rate unchanged." ] }, { "id": "SP-I2", "claim": "Hypothetical 100,000-page fragmented and well-fed schedules cost USD 85.50 and USD 38.475 in compute, saving USD 47.025 or 55%.", "basis_fact_ids": [ "SP-F1", "SP-F2", "SP-F3" ], "label": "Hypothetical scheduling sensitivity model, not measured savings", "assumptions": [ "Two processing seconds per page including in-worker input/output handling.", "Twenty billed startup seconds plus five billed tail seconds per activation.", "Fragmented schedule: 10,000 activations, 10 pages per activation. Well-fed schedule: 100 activations, 1,000 pages per activation.", "Same processing speed and output quality; no additional failures or active-worker charges.", "Whole-second timings avoid fractional activation rounding effects in this example." ] }, { "id": "SP-I3", "claim": "Adding 2% repeated processing without extra activations gives USD 39.235. At 95,000 unique accepted pages the compute cost is about USD 0.413 per thousand accepted pages.", "basis_fact_ids": [ "SP-F1", "SP-F2" ], "label": "Hypothetical retry and quality-denominator sensitivity", "assumptions": [ "Extra processing fraction 0.02 and unique acceptance fraction 0.95 are invented sensitivity inputs.", "Activation count remains 100; more activations or longer runtime would increase cost." ] }, { "id": "SP-I4", "claim": "A continuously supplied small-job queue can amortize startup while page-level durable commits limit duplicate work and result-loss exposure.", "basis_fact_ids": [ "SP-F3", "SP-F5" ], "label": "Proposed implementation, not deployed or tested", "assumptions": [ "Correct atomic persistence, idempotency keys and reconciliation.", "Actual resource availability and scaling permit sustained processing." ] } ], "unknowns": [ "Actual OCR throughput, initialization time, hardware assignment, queue behavior, quality and memory requirements for this exact corpus.", "Account-specific price, quotas, current capacity and console rate; pricing-page extractor did not return a usable table.", "Actual number of billed activations, seconds lost to retries, and durability behavior.", "Total invoice including storage, external preparation and transfer costs elsewhere, taxes and review labor.", "Exact pinned software/model revisions, dependency compatibility and legal rights to a production corpus.", "End-to-end completion deadline; aggregate worker-seconds do not determine wall-clock latency." ], "correction_history": [], "license_status": "Original PCT editorial text; public reuse license not selected. Source documents, software, models and input-document rights remain separate. No source images, benchmark outputs or long quotations reproduced.", "public_url": null, "schema_version": "pct-research-item-1.0", "desk_extensions": { "id": "pct-2026-11-09-spotlight-runpod-ocr", "section": "Product spotlight", "publisher": "Peterson Compute Trading", "as_of_date": "2026-11-09", "status": "desk_research_with_audited_hypothetical_cost_model", "subject": "A Runpod Serverless queue using explicitly selected PaddleOCR PP-OCRv5 server models", "deck": "An OCR backfill shows how worker activations and expiring API results can matter more than a faster GPU.", "editorial_word_count": 1051, "reading_minutes": 4.8, "reading_wpm": 220, "publication_note": "Evergreen technical analysis using current primary documentation observed November 9, 2026. Not a launch report, purchase recommendation or actual workload benchmark.", "cost_model": { "currency": "USD", "as_of": "2026-11-09", "rate_source_id": "SP-S1", "pricing_mode": "Documented Runpod Serverless 24-GB category rate; zero active workers and one GPU per worker assumed; console quote unverified", "gpu_category": "L4, A5000, 3090; not a guarantee of a specific model", "rate_usd_per_second": 0.00019, "rate_usd_per_worker_hour": 0.684, "unique_input_pages": 100000, "hypothetical_processing_seconds_per_page": 2.0, "hypothetical_billed_startup_seconds_per_activation": 20.0, "explicit_billed_idle_tail_seconds_per_activation": 5.0, "processing_seconds_each_baseline": 200000, "fragmented": { "pages_per_activation": 10, "activations": 10000, "overhead_seconds": 250000, "total_billed_worker_seconds": 450000, "compute_usd": 85.5 }, "well_fed": { "pages_per_activation": 1000, "activations": 100, "overhead_seconds": 2500, "total_billed_worker_seconds": 202500, "compute_usd": 38.475 }, "hypothetical_savings_usd": 47.025, "hypothetical_savings_percent": 55.0, "sensitivity": { "extra_processing_fraction": 0.02, "activations_unchanged": 100, "total_billed_worker_seconds": 206500, "compute_usd": 39.235, "unique_accepted_pages": 95000, "compute_usd_per_thousand_accepted_pages": 0.413 }, "formula": "compute = rate * (input_pages * processing_seconds_per_page * (1 + repeated_processing_fraction) + activations * (billed_startup_seconds + billed_idle_tail_seconds))", "assumptions": [ "All numerical workload timings and activation counts are hypothetical.", "Processing includes in-worker input handling and durable output writes; resource fit is unverified.", "No GPU throughput change between schedule cases.", "Model a single rate; no active-worker sales discount, capacity promise or invoice claim.", "One thousand pages per activation does not mean one thousand pages per request.", "Whole-second activation durations mean modeled totals need no additional billing rounding." ], "exclusions": [ "Storage", "External PDF/page preparation", "External network-transfer charges", "Taxes", "Credits or negotiated discounts", "Human review and engineering labor", "Any extra startup or processing absent from the stated sensitivity cases" ] }, "implementation_plan": { "status": "proposed_not_executed", "inputs": "Rights-cleared non-sensitive scanned catalogue pages, primarily English; exact model revisions and container digest to be pinned.", "steps": [ "Establish representative human-checked sample and quality acceptance criteria.", "Select explicit PP-OCRv5 server detection and recognition models; verify GPU execution and actual device.", "Configure zero active workers, cap concurrency and explicitly set idle/execution/TTL policies; read settings back.", "Use small page jobs with a staged producer that keeps workers fed without exceeding job lifetimes.", "Persist page outputs under content/model/options identities using atomic commits and retry-safe checks before returning success.", "Separate output acceptance from API success; quarantine corrupt pages and cap attempts.", "Reconcile every source page to a durable result and audit billed activation/time counts." ], "no_transactions": true, "no_deployment": true, "no_performance_test": true }, "verification": { "date": "2026-11-09", "public_primary_sources_opened": 8, "additional_primary_fetch_unreadable": 1, "arithmetic": "Decimal-based local assertions passed for hourly rate, scenario bills and sensitivity total. Only arithmetic was executed.", "benchmark": "No OCR or GPU benchmark run; no vendor performance metric presented as a result.", "cloud_actions": "None. No account changes, model downloads, uploads, purchases, deployments or credentials created.", "source_date_caution": "Runpod modification dates are explicitly displayed by its pages. Undated model cards and tutorial use null publication/update dates; access date is not release date.", "editorial_checks": [ "Separated documentation facts from hypothetical inputs and design proposals.", "Used current documented configuration rate while disclosing pricing-page fetch limitation.", "Did not claim latest model, tested output quality or a guaranteed GPU assignment.", "Did not infer one cold start per page or request.", "Distinguished billed running wait from scaled-down Idle.", "Distinguished result retention from job TTL.", "Did not reuse older advertised active-worker discounts.", "Distinguished aggregate worker time from wall-clock deadline.", "Kept license rights separate from document rights." ] } }, "public_projection": { "version": "1.0.0", "date": "2026-11-10", "changes": [ "Internal workspace artifact paths omitted where present. Historical license/status fields retain original provenance; current public reuse policy is stated separately." ], "rights": "PCT reserves rights in original editorial content; third-party rights retained. No general source-data reuse license." } } ], "corrections": [ { "date": "2026-11-05", "scope": "PCT_2026-11-04.json", "description": "H100 spot and on-demand 06.ai records shared an identifier; prior file remains unchanged, this edition uses unique observation IDs. No price correction implied.", "prior_archive_id": "libfile_5a5661b01848819197ddf74ee9b1d16d" } ], "omitted": "Separate benchmark, provider-price and market-signal datasets are not republished in this article archive." }