{ "schema_version": "pct-article-edition-1.0", "publication_date": "2026-11-03", "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": [ { "publication_date": "2026-11-03", "observation_date": "2026-11-03", "corrections": [], "correction_policy": "Append dated corrections; preserve previous versions.", "license_status": "PCT original text: unpublished, license not yet selected. Source material remains owned by its rights holders; no blanket republication permission assumed.", "id": "pct-2026-11-03-company-sfcompute", "content_type": "company_profile", "title": "SF Compute’s resale promise needs a denominator", "editorial_voice": "Company Desk | AI-assisted, explanatory", "body": "San Francisco Compute packages GPU rental with the ability to resell reservation hours that a workload no longer needs. The company says it builds and operates clusters, including infrastructure owned by partners, keeping physical delivery alongside its marketplace. [s1, s2]\n\nFinancing matters. On September 10, SFC announced $245 million of B300 capacity agreements with one unnamed AI company: two contracts lasting 36 months. That is announced contract value, not demonstrated revenue or cash collection. Capacity was scheduled to arrive through the fourth quarter and into 2027. [s3]\n\nIts September customer account describes Standard Intelligence reselling off-peak hours and Roboflow buying shorter reservations. This is the market: one team’s unused time becomes another team’s burst capacity. [s4]\n\nBuyers should interrogate the denominator behind SFC’s reported 25% average recovery through resale. Consider a simplified example: reselling one quarter of a reservation at its original hourly rate recovers one quarter of the bill, while surrendering one quarter of the hours. If the remaining hours are fully used, the price per used hour is unchanged before fees. Resale saves money relative to leaving those hours idle; it does not automatically make productive compute cheaper. [s5]\n\nThe current documentation supplies another qualification. Sell orders need a matching buyer. Proceeds arrive as platform credits after variable fees, and selling allocated time can terminate instances. Recovery is conditional, not a guaranteed cancellation payment. [s6, s7]\n\nPCT’s assessment: SFC is worth watching as an operator that packages capacity with resale flexibility. The missing evidence matters: disclosed customer concentration, realized resale volumes, fee schedules, and fill rates across hardware and delivery windows. The next useful proof is evidence that customers can repeatedly resell unwanted capacity, net of fees, without sacrificing the reliability their workloads require when demand shifts. Those measurements would reveal whether customers still recover meaningful value when many sellers compete for fewer buyers simultaneously. []", "research": { "id": "san-francisco-compute", "company": "San Francisco Compute", "short_name": "SFC", "as_of": "2026-11-03", "website": "https://sfcompute.com", "editorial_angle": "Resale flexibility can reduce stranded reservation cost, but a recovered share of purchase spend is not the same metric as a lower price per productive GPU-hour.", "identity": { "category": "GPU cloud operator and reservation-resale marketplace", "leadership": "Evan Conrad is identified as CEO in the September 9 announcement; Eric Park is identified as CTO on the cluster-build page.", "legal_entity_caution": "Public marketplace terms name San Francisco Compute Trading Company LLC. Do not assume that entity owns all infrastructure or signs every enterprise agreement.", "source_ids": [ "s2", "s5", "s8" ] }, "business_model": [ { "claim": "SFC combines technical operation with a market for time-bound GPU reservations. It also builds and operates clusters for capital providers who own the hardware.", "status": "Company-described", "source_ids": [ "s1", "s2" ] }, { "claim": "Asset owners can serve their own customers, sell through SFC, or combine both. The build-to-spec offer leaves hardware and residual value with the customer.", "status": "Product offer, not verified deployment economics", "source_ids": [ "s2" ] }, { "claim": "SFC charges sellers a platform fee on filled resale orders; current documentation describes a fixed amount per node-hour plus a percentage of sale value. Rates can differ by organization and change. Buy orders have no fee according to the Orders page.", "status": "Documented mechanics", "source_ids": [ "s6", "s7" ] }, { "claim": "The company describes providers receiving a share of resale fees in addition to underlying contract revenue. Public sources reviewed do not establish SFC’s consolidated take rate or gross margin.", "status": "Company-described revenue mechanism; financial outcome unknown", "source_ids": [ "s4" ] } ], "customers": [ { "name": "Standard Intelligence", "relationship": "SFC case study says it resells off-peak capacity to support peak-sized reservations.", "evidence_level": "Vendor-hosted named customer account", "source_ids": [ "s4" ] }, { "name": "Roboflow", "relationship": "SFC case study says it buys shorter GPU reservations for larger jobs.", "evidence_level": "Vendor-hosted named customer account", "source_ids": [ "s4" ] }, { "name": "Undisclosed AI company", "relationship": "SFC announced $245m across two 36-month dedicated B300 cluster agreements, $122.5m each.", "evidence_level": "Company announcement; contracts and customer identity not verified", "source_ids": [ "s3" ] } ], "financial_signals": [ { "date": "2025-11-26", "claim": "Announced $40m Series A at $300m valuation, co-led by DCVC and Wing, with Electric Capital and Alt Capital participating.", "interpretation": "Historical financing announcement, not current enterprise value, revenue, cash balance, or total lifetime funding.", "source_ids": [ "s9", "s10" ] }, { "date": "2026-09-10", "claim": "The B300 announcement projected capacity arriving through Q4 2026 and into 2027.", "interpretation": "Forward delivery statement. Completion as of November 3 was not established in this research.", "source_ids": [ "s3" ] } ], "product_and_limits": [ { "claim": "VM, bare-metal and managed-Slurm options are publicly advertised.", "source_ids": [ "s12" ] }, { "claim": "Orders are pinned to one hardware SKU; documented node units contain eight GPUs. A market quote therefore needs hardware and a delivery window, not merely a GPU name.", "source_ids": [ "s7" ], "status": "Documented mechanics plus inference" }, { "claim": "Sell orders can remain unmatched. A fill removes allocation and may terminate instances; a sale is an operational change, not only a billing adjustment.", "source_ids": [ "s6", "s7" ], "status": "Documented mechanics" }, { "claim": "Resale produces platform credits after fees. Cash withdrawal rights or timing were not established in the reviewed sources.", "source_ids": [ "s6" ], "status": "Documented credit treatment; cash convertibility unknown" }, { "claim": "September 25 release places Elastic InfiniBand Partitions in public preview and requires physically colocated instances. Do not reuse older FAQ wording as a current blanket statement about VM InfiniBand availability.", "source_ids": [ "s11", "s12" ], "status": "Release state and source conflict" }, { "claim": "Published terms require prior company authorization for resale and written consent to copy or redistribute market data. Confirm the actual applicable agreement before building a commercial feed.", "source_ids": [ "s8" ], "status": "Contractual diligence point, not legal advice" } ], "facts_inferences_and_unknowns": { "company_claim": "September 9 claim: on-platform customers recover an average 25% of purchase spend through resale. Treat as company-reported, not independently audited or guaranteed.", "original_analysis": "If 25% of purchased hours are sold at the original rate, cost falls 25% but retained hours also fall 25%. With full use of retained time and no fees, price per used hour is unchanged. The benefit is avoided waste relative to retaining idle time.", "unknowns": [ "Customer concentration by recognized revenue, bookings or resale participation.", "Aggregate revenue, gross margin, debt exposure and actual take rate.", "Resale fill rates, volume, time to sale and bid depth by SKU and exact delivery window.", "Distribution, sample size, observation period and weighting behind the reported 25% average.", "Cash convertibility of resale credits and account-specific fee schedules.", "Whether announced B300 delivery milestones have been achieved." ], "source_ids": [ "s5", "s6", "s7" ] }, "editorial_controls": [ "Use 'SFC says' for contract values, customer usage and company performance statistics.", "Do not label $245m as revenue, cash received, annual recurring revenue, or total backlog.", "Do not equate resale rights with guaranteed liquidity, refundability, or risk elimination.", "Do not identify the unnamed buyer or infer its concentration percentage.", "Do not repeat 'first', 'only', near-100% utilization, or high-liquidity claims as established facts.", "No live price, orderbook depth, transaction execution, customer contact, account creation, or login was attempted." ], "article": { "title": "SF Compute’s resale promise needs a denominator", "byline": "PCT Research Desk | AI-assisted analysis", "date": "2026-11-03", "word_count": 300, "paragraphs": [ { "text": "San Francisco Compute packages GPU rental with the ability to resell reservation hours that a workload no longer needs. The company says it builds and operates clusters, including infrastructure owned by partners, keeping physical delivery alongside its marketplace.", "source_ids": [ "s1", "s2" ] }, { "text": "Financing matters. On September 10, SFC announced $245 million of B300 capacity agreements with one unnamed AI company: two contracts lasting 36 months. That is announced contract value, not demonstrated revenue or cash collection. Capacity was scheduled to arrive through the fourth quarter and into 2027.", "source_ids": [ "s3" ] }, { "text": "Its September customer account describes Standard Intelligence reselling off-peak hours and Roboflow buying shorter reservations. This is the market: one team’s unused time becomes another team’s burst capacity.", "source_ids": [ "s4" ] }, { "text": "Buyers should interrogate the denominator behind SFC’s reported 25% average recovery through resale. Consider a simplified example: reselling one quarter of a reservation at its original hourly rate recovers one quarter of the bill, while surrendering one quarter of the hours. If the remaining hours are fully used, the price per used hour is unchanged before fees. Resale saves money relative to leaving those hours idle; it does not automatically make productive compute cheaper.", "source_ids": [ "s5" ], "inference": "The example is original illustrative arithmetic, assumes equal purchase/resale hourly rates and zero fees, and is not a company performance estimate." }, { "text": "The current documentation supplies another qualification. Sell orders need a matching buyer. Proceeds arrive as platform credits after variable fees, and selling allocated time can terminate instances. Recovery is conditional, not a guaranteed cancellation payment.", "source_ids": [ "s6", "s7" ] }, { "text": "PCT’s assessment: SFC is worth watching as an operator that packages capacity with resale flexibility. The missing evidence matters: disclosed customer concentration, realized resale volumes, fee schedules, and fill rates across hardware and delivery windows. The next useful proof is evidence that customers can repeatedly resell unwanted capacity, net of fees, without sacrificing the reliability their workloads require when demand shifts. Those measurements would reveal whether customers still recover meaningful value when many sellers compete for fewer buyers simultaneously.", "source_ids": [], "inference": "Original desk assessment. Missing evidence means not established in the public sources reviewed, not a claim that it does not exist." } ], "body_markdown": "San Francisco Compute packages GPU rental with the ability to resell reservation hours that a workload no longer needs. The company says it builds and operates clusters, including infrastructure owned by partners, keeping physical delivery alongside its marketplace. [About San Francisco Compute](https://sfcompute.com/about), [Build your cluster](https://sfcompute.com/clusters-to-spec)\n\nFinancing matters. On September 10, SFC announced $245 million of B300 capacity agreements with one unnamed AI company: two contracts lasting 36 months. That is announced contract value, not demonstrated revenue or cash collection. Capacity was scheduled to arrive through the fourth quarter and into 2027. [San Francisco Compute sells $245M worth of compute](https://sfcompute.com/news/tcv-q3-2026)\n\nIts September customer account describes Standard Intelligence reselling off-peak hours and Roboflow buying shorter reservations. This is the market: one team’s unused time becomes another team’s burst capacity. [Standard Intelligence and Roboflow access larger GPU clusters through SF Compute’s resale market](https://sfcompute.com/news/scaling-with-gpu-resale)\n\nBuyers should interrogate the denominator behind SFC’s reported 25% average recovery through resale. Consider a simplified example: reselling one quarter of a reservation at its original hourly rate recovers one quarter of the bill, while surrendering one quarter of the hours. If the remaining hours are fully used, the price per used hour is unchanged before fees. Resale saves money relative to leaving those hours idle; it does not automatically make productive compute cheaper. [The average customer now saves 25% with SFC](https://sfcompute.com/news/realized-discount)\n\nThe current documentation supplies another qualification. Sell orders need a matching buyer. Proceeds arrive as platform credits after variable fees, and selling allocated time can terminate instances. Recovery is conditional, not a guaranteed cancellation payment. [Sell unused compute time](https://docs.sfcompute.com/preview/guides/reselling), [Orders](https://docs.sfcompute.com/preview/orders)\n\nPCT’s assessment: SFC is worth watching as an operator that packages capacity with resale flexibility. The missing evidence matters: disclosed customer concentration, realized resale volumes, fee schedules, and fill rates across hardware and delivery windows. The next useful proof is evidence that customers can repeatedly resell unwanted capacity, net of fees, without sacrificing the reliability their workloads require when demand shifts. Those measurements would reveal whether customers still recover meaningful value when many sellers compete for fewer buyers simultaneously." }, "sources": [ { "id": "s1", "title": "About San Francisco Compute", "url": "https://sfcompute.com/about", "published": null, "accessed": "2026-11-03", "type": "Company description", "supports": "Cluster operation and partner-owned infrastructure model." }, { "id": "s2", "title": "Build your cluster", "url": "https://sfcompute.com/clusters-to-spec", "published": null, "accessed": "2026-11-03", "type": "Company product page", "supports": "Customers own hardware and residual value; SFC designs, builds, operates, and offers routes to market." }, { "id": "s3", "title": "San Francisco Compute sells $245M worth of compute", "url": "https://sfcompute.com/news/tcv-q3-2026", "published": "2026-09-10", "accessed": "2026-11-03", "type": "Company announcement", "supports": "Two 36-month B300 agreements, unnamed customer, announced delivery timing. Not independent contract verification." }, { "id": "s4", "title": "Standard Intelligence and Roboflow access larger GPU clusters through SF Compute’s resale market", "url": "https://sfcompute.com/news/scaling-with-gpu-resale", "published": "2026-09-28", "accessed": "2026-11-03", "type": "Company case study with named customer quotations", "supports": "Reseller and short-term buyer examples; provider share of resale fees. Vendor-hosted evidence, not audited customer usage." }, { "id": "s5", "title": "The average customer now saves 25% with SFC", "url": "https://sfcompute.com/news/realized-discount", "published": "2026-09-09", "accessed": "2026-11-03", "type": "Company performance claim and illustrative guide", "supports": "25% average claimed realized discount; its example uses purchased GPU-hours. Does not establish a future guaranteed result." }, { "id": "s6", "title": "Sell unused compute time", "url": "https://docs.sfcompute.com/preview/guides/reselling", "published": null, "accessed": "2026-11-03", "type": "Product documentation", "supports": "Variable fixed-per-node-hour and percentage fees; proceeds as credits; no buyer guarantee; workload consequences." }, { "id": "s7", "title": "Orders", "url": "https://docs.sfcompute.com/preview/orders", "published": null, "accessed": "2026-11-03", "type": "Product documentation", "supports": "SKU-specific orders; eight-GPU nodes; no buy-order fees; partial fills and termination mechanics." }, { "id": "s8", "title": "Terms of Service", "url": "https://sfcompute.com/legal/terms-of-service", "published": "2025-11-18", "accessed": "2026-11-03", "type": "Published contractual terms; last-updated date", "supports": "Company-authorized resale, funding requirements, market-data redistribution limitation. Actual enterprise contract may differ." }, { "id": "s9", "title": "SF Compute raised a $40M Series A", "url": "https://sfcompute.com/news/sf-compute-series-a", "published": "2025-11-26", "accessed": "2026-11-03", "type": "Company funding announcement", "supports": "$40m round, $300m valuation, DCVC and Wing co-leads; not current valuation or total funding." }, { "id": "s10", "title": "The San Francisco Compute Company", "url": "https://www.dcvc.com/companies/the-san-francisco-compute-company/", "published": null, "accessed": "2026-11-03", "type": "Investor portfolio description", "supports": "Investor confirms relationship; describes owner-operated marketplace and transaction fees. Interested-party source, not independent audit." }, { "id": "s11", "title": "Introducing Elastic InfiniBand Partitions", "url": "https://sfcompute.com/news/elastic-infiniband-partitions", "published": "2026-09-25", "accessed": "2026-11-03", "type": "Product release announcement", "supports": "Public-preview isolated RDMA partitions for physically colocated instances; feature availability is conditional." }, { "id": "s12", "title": "GPU clusters / Specs", "url": "https://sfcompute.com/specs", "published": null, "accessed": "2026-11-03", "type": "Company product page", "supports": "VMs, bare metal, managed Slurm and support. FAQ still contains a Q3 2026 future-tense InfiniBand statement." } ] }, "sources": [ { "id": "s1", "title": "About San Francisco Compute", "url": "https://sfcompute.com/about", "published": null, "accessed": "2026-11-03", "type": "Company description", "supports": "Cluster operation and partner-owned infrastructure model." }, { "id": "s2", "title": "Build your cluster", "url": "https://sfcompute.com/clusters-to-spec", "published": null, "accessed": "2026-11-03", "type": "Company product page", "supports": "Customers own hardware and residual value; SFC designs, builds, operates, and offers routes to market." }, { "id": "s3", "title": "San Francisco Compute sells $245M worth of compute", "url": "https://sfcompute.com/news/tcv-q3-2026", "published": "2026-09-10", "accessed": "2026-11-03", "type": "Company announcement", "supports": "Two 36-month B300 agreements, unnamed customer, announced delivery timing. Not independent contract verification." }, { "id": "s4", "title": "Standard Intelligence and Roboflow access larger GPU clusters through SF Compute’s resale market", "url": "https://sfcompute.com/news/scaling-with-gpu-resale", "published": "2026-09-28", "accessed": "2026-11-03", "type": "Company case study with named customer quotations", "supports": "Reseller and short-term buyer examples; provider share of resale fees. Vendor-hosted evidence, not audited customer usage." }, { "id": "s5", "title": "The average customer now saves 25% with SFC", "url": "https://sfcompute.com/news/realized-discount", "published": "2026-09-09", "accessed": "2026-11-03", "type": "Company performance claim and illustrative guide", "supports": "25% average claimed realized discount; its example uses purchased GPU-hours. Does not establish a future guaranteed result." }, { "id": "s6", "title": "Sell unused compute time", "url": "https://docs.sfcompute.com/preview/guides/reselling", "published": null, "accessed": "2026-11-03", "type": "Product documentation", "supports": "Variable fixed-per-node-hour and percentage fees; proceeds as credits; no buyer guarantee; workload consequences." }, { "id": "s7", "title": "Orders", "url": "https://docs.sfcompute.com/preview/orders", "published": null, "accessed": "2026-11-03", "type": "Product documentation", "supports": "SKU-specific orders; eight-GPU nodes; no buy-order fees; partial fills and termination mechanics." }, { "id": "s8", "title": "Terms of Service", "url": "https://sfcompute.com/legal/terms-of-service", "published": "2025-11-18", "accessed": "2026-11-03", "type": "Published contractual terms; last-updated date", "supports": "Company-authorized resale, funding requirements, market-data redistribution limitation. Actual enterprise contract may differ." }, { "id": "s9", "title": "SF Compute raised a $40M Series A", "url": "https://sfcompute.com/news/sf-compute-series-a", "published": "2025-11-26", "accessed": "2026-11-03", "type": "Company funding announcement", "supports": "$40m round, $300m valuation, DCVC and Wing co-leads; not current valuation or total funding." }, { "id": "s10", "title": "The San Francisco Compute Company", "url": "https://www.dcvc.com/companies/the-san-francisco-compute-company/", "published": null, "accessed": "2026-11-03", "type": "Investor portfolio description", "supports": "Investor confirms relationship; describes owner-operated marketplace and transaction fees. Interested-party source, not independent audit." }, { "id": "s11", "title": "Introducing Elastic InfiniBand Partitions", "url": "https://sfcompute.com/news/elastic-infiniband-partitions", "published": "2026-09-25", "accessed": "2026-11-03", "type": "Product release announcement", "supports": "Public-preview isolated RDMA partitions for physically colocated instances; feature availability is conditional." }, { "id": "s12", "title": "GPU clusters / Specs", "url": "https://sfcompute.com/specs", "published": null, "accessed": "2026-11-03", "type": "Company product page", "supports": "VMs, bare metal, managed Slurm and support. FAQ still contains a Q3 2026 future-tense InfiniBand statement." } ], "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." } }, { "publication_date": "2026-11-03", "observation_date": "2026-11-03", "corrections": [], "correction_policy": "Append dated corrections; preserve previous versions.", "license_status": "PCT original text: unpublished, license not yet selected. Source material remains owned by its rights holders; no blanket republication permission assumed.", "id": "pct-2026-11-03-transaction-qumulus", "content_type": "transaction_analysis", "title": "QumulusAI's $71.9 million compute deal reaches billing, while expansion awaits delivery", "editorial_voice": "Deals Desk | AI-assisted, evidence-first", "body": "QumulusAI's $71.9 million compute agreement has crossed an important boundary: the supplier says the contracted hardware is serving its customer and generating bills. That is company-reported delivery evidence, not independent verification of operational performance. It is also an October development, not a newly signed November transaction.\n\nThe company announced the three-year agreement on July 28 with an unnamed AI inference platform. It covered NVIDIA Blackwell B200 and B300 capacity for production workloads, with renewal options. At announcement, customer availability was expected during the third quarter of 2026. [S1]\n\nDeployment came in stages. A September 15 update said the B200 portion had been handed over on September 8, while B300 deployment remained unfinished. The distinction matters: completion of one hardware tranche did not establish delivery of the entire agreement. [S2]\n\nOn October 15, QumulusAI reported that the final B300 handoff occurred at its Kansas City site on October 1. It said every GPU under the agreement was then in production and billing. The release did not disclose the contract's GPU count or the customer's power allocation. [S3]\n\nThe same customer separately signed a $211.9 million, three-year take-or-pay expansion announced September 29. That capacity was targeted for early 2027. Two optional one-year extensions, each valued at $70.6 million, sit outside the initial commitment. The original contract and expansion total $283.8 million, but that sum does not describe compute already delivered. [S4; S3]\n\nFor buyers and lenders, the distinction is practical. Delivery can reduce construction and commissioning uncertainty; it does not establish utilization, payment collection or customer credit quality. Those judgments require information these announcements do not provide.\n\nA further qualification concerns timing. The final October handoff followed the original third-quarter availability target. That comparison does not prove a contractual breach: the public releases do not establish contractual deadlines, remedies or whether customer-approved changes altered the delivery timetable.", "research": { "as_of": "2026-11-03", "story_type": "Commercial compute contract: delivery follow-through", "headline": "QumulusAI's $71.9 million compute deal reaches billing, while expansion awaits delivery", "freshness": "Earlier reporting, not new on November 3: contract announced July 28; B200 delivery reported September 15; final B300 handoff occurred October 1 and was reported October 15, 2026.", "parties": { "supplier": "QumulusAI", "customer": "Undisclosed AI inference platform" }, "deal": { "currency": "USD", "initial_contract_value": 71900000, "initial_term_years": 3, "renewal_options": "Exist; detailed original renewal economics not established", "hardware": [ "NVIDIA Blackwell B200", "NVIDIA Blackwell B300" ], "workload": "Production AI inference", "geography": "United States; final B300 deployment at Kansas City" }, "status": { "announced": "Signed agreement disclosed July 28, 2026; exact signing date not separately established", "commitment": "$71.9 million over three years, not cash received or revenue already earned", "closing": "No separate financial closing described", "service_execution": "Supplier reports B200 customer handoff September 8 and final B300 handoff October 1; all contract GPUs in production and billing as reported October 15", "executed_gpu_volume": null, "verification_basis": "Company primary-source statements, not independent operational measurement" }, "related_expansion": { "announcement_date": "2026-09-29", "value_usd": 211900000, "term_years": 3, "structure": "Take-or-pay", "delivery_target": "Early 2027", "initial_terms_total_with_original_usd": 283800000, "counting_note": "Original and expansion are separate commitments. Their sum is not deployed capacity, realized revenue or a second total to add again.", "optional_extensions": "Two one-year options at $70.6 million each, excluded" }, "unknowns": [ "Customer legal identity", "Exact GPU count and B200/B300 split under original deal", "Customer-specific MW and utility energization date", "GPU-hour price", "Actual utilization, cash collections and recognized revenue", "Full contract termination rights and penalties", "Independent verification of service delivery" ], "inference": [ "Reported deployment is stronger execution evidence than a signed purchase commitment but does not establish workload utilization or collection of the full contract price.", "Final handoff on October 1 falls after the initially stated third-quarter availability window; this does not by itself establish contractual default or penalties.", "The separate expansion makes customer concentration worth examining, but no credit assessment can be made without customer identity and payment terms." ], "sources": [ { "id": "S1", "date": "2026-07-28", "type": "Company announcement", "url": "https://www.qumulusai.com/articles/qumulusai-signs-719-million-three-year-agreement-with-ai-inference-platform-provider", "supports": "Original signed agreement, duration, hardware, workload, US footprint, Q3 target" }, { "id": "S2", "date": "2026-09-15", "type": "Company delivery update", "url": "https://www.qumulusai.com/articles/qumulusai-completes-b200-deployment-under-its-719-million-contracted-blackwell-agreement", "supports": "September 8 B200 handoff; B300 then incomplete" }, { "id": "S3", "date": "2026-10-15", "type": "Company investor release", "url": "https://investors.qumulusai.com/news/news-details/2026/QumulusAI-Puts-71-9-Million-Blackwell-Agreement-into-Production-Signed-Customer-Contracts-Surpass-500-Million/default.aspx", "supports": "October 1 final B300 handoff in Kansas City; production and billing; $283.8m customer commitments" }, { "id": "S4", "date": "2026-09-29", "type": "Company expansion announcement", "url": "https://www.qumulusai.com/articles/qumulusai-signs-2409-million-in-new-ai-inference-customer-contracts-for-nvidia-blackwell-b300-capacity", "supports": "Separate $211.9m three-year take-or-pay expansion, early-2027 target, optional extensions" } ] }, "sources": [ { "id": "S1", "date": "2026-07-28", "type": "Company announcement", "url": "https://www.qumulusai.com/articles/qumulusai-signs-719-million-three-year-agreement-with-ai-inference-platform-provider", "supports": "Original signed agreement, duration, hardware, workload, US footprint, Q3 target" }, { "id": "S2", "date": "2026-09-15", "type": "Company delivery update", "url": "https://www.qumulusai.com/articles/qumulusai-completes-b200-deployment-under-its-719-million-contracted-blackwell-agreement", "supports": "September 8 B200 handoff; B300 then incomplete" }, { "id": "S3", "date": "2026-10-15", "type": "Company investor release", "url": "https://investors.qumulusai.com/news/news-details/2026/QumulusAI-Puts-71-9-Million-Blackwell-Agreement-into-Production-Signed-Customer-Contracts-Surpass-500-Million/default.aspx", "supports": "October 1 final B300 handoff in Kansas City; production and billing; $283.8m customer commitments" }, { "id": "S4", "date": "2026-09-29", "type": "Company expansion announcement", "url": "https://www.qumulusai.com/articles/qumulusai-signs-2409-million-in-new-ai-inference-customer-contracts-for-nvidia-blackwell-b300-capacity", "supports": "Separate $211.9m three-year take-or-pay expansion, early-2027 target, optional extensions" } ], "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." } }, { "publication_date": "2026-11-03", "observation_date": "2026-11-03", "corrections": [], "correction_policy": "Append dated corrections; preserve previous versions.", "license_status": "PCT original text: unpublished, license not yet selected. Source material remains owned by its rights holders; no blanket republication permission assumed.", "id": "pct-2026-11-03-market-formation-freight-clearing", "content_type": "market_formation", "title": "Freight’s lesson for compute: follow the cleared trade", "editorial_voice": "Formation Desk | AI-assisted, historical comparison", "body": "Twenty years ago, Singapore Exchange’s freight expansion supplied a useful checklist for judging a new derivatives market. The record documents separate steps: securing reference prices, establishing clearing access and processing an actual trade. That sequence offers a more demanding test for compute-market development than an announcement alone.\n\nOn May 2, 2006, SGX announced an agreement with the Baltic Exchange to supply benchmarks for settling dry-bulk and dirty-tanker forward freight agreements. Baltic would also provide forward-curve assessments for margin calculations. The release therefore distinguished the price used to settle a contract from the information needed to manage exposure before settlement. [1]\n\nOn May 16, SGX announced AsiaClear’s launch and identified its first cleared freight transaction: a Panamax Time Charter Basket agreement between Noble Shipping and IMC Shipping. Freight Investor Services Physical brokered it; ABN AMRO Futures and UOB Bullion & Futures cleared it. Those named participants provide evidence of a transaction passing through the new facility, beyond a planned product description. The launch concerned clearing for over-the-counter agreements, so it should not be recast as proof of an exchange order book. [2]\n\nSGX’s 2006 annual report separately confirms the May launch and describes the facility’s purpose as mitigating participants’ credit risk and improving clearing. [3]\n\nFor compute, the editorial inference is practical: track the benchmark methodology, the source of margin inputs, access through clearing firms and verified transactions as separate milestones. Then ask whether the referenced GPU service matches the buyer’s actual exposure. A usable settlement price and a useful hedge require different evidence.\n\nThe analogy has limits. This historical record establishes infrastructure and an initial cleared deal; it does not establish sustained liquidity, hedge performance or a timetable for compute adoption. Nor does clearing remove the mismatch between a benchmark and an individual customer’s bill. The next proof to seek is repeat participation under transparent terms, not simply another product name.", "research": { "id": "pct-2026-11-03-market-formation-freight-clearing", "title": "Freight’s lesson for compute: follow the cleared trade", "publication_date": "2026-11-03", "observation_date": "2026-11-03", "type": "historical_analysis", "section": "Market formation", "dek": "SGX’s 2006 freight rollout separated settlement data, margin assessment and an executed transaction. Compute deserves the same scrutiny.", "confirmed_facts": [ { "claim": "The May 2 agreement supplied both final-settlement benchmarks and forward assessments for margin calculations.", "source_ids": [ "1" ] }, { "claim": "SGX reported the May 16 launch and a first cleared Panamax freight agreement with named counterparties and intermediaries.", "source_ids": [ "2" ] }, { "claim": "The issuer’s annual report corroborates the launch and credit-risk purpose.", "source_ids": [ "3" ] } ], "inferences": [ "Market formation should be assessed separately at the benchmark, margin, clearing-access and executed-transaction layers.", "Compute-market readers should ask whether a referenced service matches the exposure being hedged." ], "limitations": [ "Historical analysis, not a November 2026 news event or a new compute launch report.", "Two contemporaneous issuer releases are accessible on a third-party archive; the independent issuer-hosted annual report corroborates launch, but not every named transaction detail.", "The source record describes OTC clearing, not exchange order-book activity.", "No claim of durable liquidity, hedge effectiveness, transaction price, volume or compute-adoption timetable is made.", "The annual report’s exact publication date was not verified." ], "uncertainty": "High confidence in the documented 2006 SGX launch and what its releases reported. The applicability to compute is explicitly editorial inference, not an observed empirical result.", "copyright_licensing_status": "unknown for cited sources; article text is original; no source text, image or dataset republished", "corrections": [] }, "sources": [ { "id": "1", "title": "SGX Partners With Baltic Exchange For Forward Freight Agreements", "author_or_issuer": "Singapore Exchange", "source_publish_date": "2006-05-02", "observation_date": "2026-11-03", "url": "https://mondovisione.com/media-and-resources/news/sgx-partners-with-baltic-exchange-for-forward-freight-agreements-2011124/", "provenance": "Original dated issuer announcement preserved by Mondo Visione; third-party archive, not the issuer’s own host.", "copyright_licensing_status": "unknown" }, { "id": "2", "title": "SGX Launches SGX AsiaClear Facility For OTC Clearing And Clears First OTC Trades", "author_or_issuer": "Singapore Exchange", "source_publish_date": "2006-05-16", "observation_date": "2026-11-03", "url": "https://mondovisione.com/media-and-resources/news/sgx-launches-sgx-asiaclearsmallsuptm-small-sup-facility-for-otc-clearing-and-cle-2011124/", "provenance": "Original dated issuer announcement preserved by Mondo Visione; third-party archive, not the issuer’s own host.", "copyright_licensing_status": "unknown" }, { "id": "3", "title": "SGX Annual Report 2006", "author_or_issuer": "Singapore Exchange", "source_publish_date": "2006", "source_publish_date_precision": "year; exact release date not verified", "observation_date": "2026-11-03", "url": "https://investorrelations.sgx.com/static-files/b8ea03f4-b2be-46a0-8627-f6863db2793c", "locator": "PDF page 13 for 16 May launch; PDF page 50 for purpose. The document also contains an AGM notice dated 30 August 2006, which is not treated as the publication date.", "provenance": "Issuer-hosted annual report.", "copyright_licensing_status": "unknown" } ], "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." } }, { "publication_date": "2026-11-03", "observation_date": "2026-11-03", "corrections": [], "correction_policy": "Append dated corrections; preserve previous versions.", "license_status": "PCT original text: unpublished, license not yet selected. Source material remains owned by its rights holders; no blanket republication permission assumed.", "id": "pct-spotlight-aws-capacity-blocks", "content_type": "product_spotlight", "title": "What a GPU reservation actually buys", "editorial_voice": "Spotlight Desk | AI-assisted, buyer-focused", "body": "A low hourly rate is useful only if the machines are there when the job needs them. Amazon EC2 Capacity Blocks for ML turn that scheduling problem into a purchasable reservation. AWS’s May 7, 2026 guide describes the product as a way to secure capacity for testing, model validation, workshops and planned inference demand. [S1]\n\nThe buyer specifies an instance type, quantity, date range and duration. Current documentation permits booking up to eight weeks ahead, with durations of one to 14 days or weekly multiples up to 182 days. The offer identifies an Availability Zone and reservation price. Once purchased, that price is fixed. [S2]\n\nThe commitment has edges. The reservation fee is charged upfront, and cancellation is prohibited. Operating-system charges can apply separately. Ordinary instance blocks begin shutting down 30 minutes before the reservation’s stated end; UltraServer blocks have a 60-minute cutoff. Extensions depend on available capacity, so an overrun needs a contingency plan. [S2][S3][S4][S5]\n\nFor a buyer, the question is whether a dependable work window outweighs the risk of paying for idle machines. A simple illustration: if only half the purchased GPU-hours are used, the reservation cost per used GPU-hour doubles. That arithmetic excludes other costs and says nothing about model throughput. Prepare data, software and checkpoints before the paid window opens.\n\nFor an agent, the purchasing record should preserve the exact instance type, accelerator count, region, Availability Zone, start time, usable end time, currency, total fee and cancellation terms. A generic “GPU-hour” field loses the conditions that make two offers comparable.\n\n**PCT reading:** Capacity Blocks show how time and delivery conditions become part of the compute product. The published rate is one supplier’s offer, not a market-wide clearing price. This review did not query live account-specific inventory or purchase a block; it establishes product mechanics, not available volume.", "research": { "id": "pct-spotlight-aws-capacity-blocks", "title": "What a GPU reservation actually buys", "section": "Product spotlight", "publisher": "Peterson Compute Trading", "as_of_date": "2026-11-03", "status": "desk_research", "editorial_word_count": 298, "facts": [ { "id": "F1", "claim": "AWS published the cited short-term GPU capacity guide on May 7, 2026.", "source_ids": [ "S1" ], "evidence_type": "dated_primary_publication" }, { "id": "F2", "claim": "Capacity Block offers specify hardware type, quantity, start time, Availability Zone and reservation price.", "source_ids": [ "S2" ], "evidence_type": "current_vendor_documentation" }, { "id": "F3", "claim": "Reservations can start up to eight weeks in advance. Durations are one to 14 days or multiples of seven days up to 182 days.", "source_ids": [ "S2" ], "evidence_type": "current_vendor_documentation" }, { "id": "F4", "claim": "The reservation price is charged upfront and does not change after booking; operating-system charges can apply separately.", "source_ids": [ "S3" ], "evidence_type": "current_vendor_documentation" }, { "id": "F5", "claim": "Capacity Block cancellation is not allowed.", "source_ids": [ "S4" ], "evidence_type": "current_vendor_documentation" }, { "id": "F6", "claim": "Instance termination starts 30 minutes before block expiry for instance types, or 60 minutes before expiry for UltraServer types.", "source_ids": [ "S2" ], "evidence_type": "current_vendor_documentation" }, { "id": "F7", "claim": "Extensions require sufficient available capacity and are not guaranteed.", "source_ids": [ "S5" ], "evidence_type": "current_vendor_documentation" }, { "id": "F8", "claim": "Savings Plans and Reserved Instance discounts do not apply to Capacity Blocks.", "source_ids": [ "S3" ], "evidence_type": "current_vendor_documentation" } ], "inferences": [ { "id": "I1", "claim": "Buyers trade scheduling certainty against utilization risk.", "basis_fact_ids": [ "F2", "F4", "F5" ], "label": "PCT analysis" }, { "id": "I2", "claim": "At 50% purchased-hour utilization, reservation-only cost per used GPU-hour is twice the nominal reservation rate.", "basis_fact_ids": [ "F4" ], "label": "hypothetical arithmetic", "assumptions": [ "Fixed prepaid reservation cost.", "All purchased accelerator-hours are equivalent within the example.", "Half of purchased accelerator-hours are used.", "Other charges and workload throughput differences are excluded." ] }, { "id": "I3", "claim": "Compute offer comparisons should retain date, region, hardware, quantity, usable duration and commercial conditions rather than only price per GPU-hour.", "basis_fact_ids": [ "F2", "F3", "F4", "F5", "F6" ], "label": "PCT buyer guidance" } ], "limits": [ "No interviews were conducted.", "No authenticated account-specific offering or inventory query was performed.", "No purchase was performed.", "No independently traded contract, secondary-market transaction, market-wide clearing price, liquidity measure or transaction volume was established.", "Documentation is current as read on November 3, 2026, not a frozen historical contract. Exact offer terms and availability must be checked before buying.", "AWS general FAQ says blocks cannot be modified, but current extension documentation permits extensions subject to capacity. Avoid the blanket claim that a block can never be extended." ], "suggested_agent_fields": [ "instance_type", "accelerator_model", "accelerator_count", "instance_count", "region", "availability_zone", "reservation_start_utc", "reservation_end_utc", "usable_end_utc", "currency", "total_reservation_fee", "separate_os_fees", "cancellation_policy", "extension_availability", "source_url", "retrieved_at" ] }, "sources": [ { "id": "S1", "title": "Secure short-term GPU capacity for ML workloads with EC2 Capacity Blocks for ML and SageMaker training plans", "publisher": "AWS", "published_at": "2026-05-07", "retrieved_at": "2026-11-03", "url": "https://aws.amazon.com/blogs/machine-learning/secure-short-term-gpu-capacity-for-ml-workloads-with-ec2-capacity-blocks-for-ml-and-sagemaker-training-plans/" }, { "id": "S2", "title": "How Amazon EC2 Capacity Blocks work", "publisher": "AWS", "published_at": null, "retrieved_at": "2026-11-03", "url": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-blocks-how.html" }, { "id": "S3", "title": "Capacity Blocks pricing and billing", "publisher": "AWS", "published_at": null, "retrieved_at": "2026-11-03", "url": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-blocks-pricing-billing.html" }, { "id": "S4", "title": "Capacity Blocks for ML", "publisher": "AWS", "published_at": null, "retrieved_at": "2026-11-03", "url": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-capacity-blocks.html" }, { "id": "S5", "title": "Extend Capacity Blocks", "publisher": "AWS", "published_at": null, "retrieved_at": "2026-11-03", "url": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-blocks-extend.html" } ], "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": [], "omitted": "Separate benchmark, provider-price and market-signal datasets are not republished in this article archive." }