{ "schema_version": "pct-article-edition-1.0", "publication_date": "2026-11-04", "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": [ { "schema_name": "PCT canonical research article", "schema_version": "0.2.0", "id": "pct-2026-11-04-company-coreweave", "edition_id": "pct-2026-11-04", "publication_date": "2026-11-04", "observation_date": "2026-11-04", "content_type": "company_profile", "status": "private pilot draft", "title": "CoreWeave’s financed delivery promise", "subtitle": "The AI cloud’s operating platform is real. Its next test is keeping customer cash, construction and financing on the same clock.", "editorial_voice": "Company Desk | AI-assisted, explanatory", "company": "CoreWeave, Inc.", "website": "https://www.coreweave.com/", "word_count": 1365, "estimated_read_minutes": 6.5, "read_time_basis": "210 words per minute, excluding source apparatus", "body": "CoreWeave’s place in the compute market becomes clearer when its cloud contracts are read alongside its financing. Customers buy GPU-cluster access; lenders and shareholders finance much of the buildout. The business depends on making those two sets of promises fit. Our assessment is that its central competitive test is the interval between securing a customer commitment and delivering a productive cluster. Good execution can turn advance commitments into a funding advantage. Delays can leave the operator paying for infrastructure before the associated revenue arrives. [cw-s1]\n\nThe operating model is unusually explicit. CoreWeave’s 2025 annual report says infrastructure development is primarily financed with asset-level debt supported by take-or-pay customer contracts, supplemented by corporate debt and equity. Customers generally commit to specified capacity for multiple years and owe payment regardless of utilization under those arrangements. There is also pay-as-you-go access. This makes the company materially different from a market whose main service is matching buyers with spare GPU hours. CoreWeave is taking responsibility for the infrastructure and the financing that make a reservation deliverable, then operating the service over its contractual life. [cw-s1]\n\nSoftware is part of that responsibility. Its SUNK product preserves researchers’ Slurm workflows while combining them with Kubernetes operations. The company describes topology-aware scheduling, job-level visibility and Mission Control monitoring for hardware faults and slow GPUs. Those are relevant capabilities because a cluster’s useful output depends on coordinated work, not simply installed chip count. But a feature list does not establish an achieved reliability rate. Buyers still need workload-specific evidence: time spent training successfully, recovery after failures, and the throughput delivered by the exact hardware, network and storage configuration in their contract. [cw-s2]\n\nThere is a revealing limit to the software portability story. The official SUNK Anywhere repository says its SUNK and Slurm Helm charts are proprietary and require a separate license. Public deployment documentation therefore does not mean the whole operational stack is freely reusable. For a customer planning to spread jobs across providers, technical compatibility and commercial rights are separate checks. A portable workflow can reduce migration work while still depending on vendor support, licensing and configuration knowledge. The sensible procurement question is what can actually move, under what agreement, and with which operational responsibilities left to the buyer. [cw-s3]\n\nNamed customers are more than logo evidence here. Microsoft accounted for approximately 67% of 2025 revenue, according to CoreWeave’s annual report. OpenAI provides a separate, customer-side check: its subprocessor list, updated October 14, identifies CoreWeave as a cloud-infrastructure provider for the API and several business ChatGPT products. That establishes a disclosed operating relationship from the buyer’s side. It does not disclose the amount paid, revenue recognized by CoreWeave, contract profitability or the share of OpenAI workloads served. Those limits matter when customer announcements are used to imply either diversification or financial strength. [cw-s1, cw-s4]\n\nThe second-quarter filing offers a more current concentration test. Customers A, B and C represented 36%, 26% and 10% of quarterly revenue respectively: 72% combined, by simple addition. The filing warns that letter labels may represent different customers from those in prior periods. We therefore do not map the letters to Microsoft, OpenAI or another named company. A lower largest-customer percentage can coexist with substantial dependence on a small group of buyers. The useful next disclosure would connect concentration with contract duration, payment timing and delivery obligations, rather than merely count newly announced customers. [cw-s5]\n\nBacklog is the second important distinction. CoreWeave reported approximately $104 billion at June 30, defining the measure to include remaining performance obligations plus other estimated future revenue under committed contracts, subject to delivery and service availability. It is a measure of contracted future business, not cash already collected. The same release reported 1.5 GW of active power against approximately 3.7 GW contracted. These are different stages of supply development, not interchangeable descriptions of usable capacity. A large backlog creates work to deliver, and a large power pipeline creates work to commission. Neither number by itself establishes the return on that work. [cw-s6]\n\nCore Scientific supplies a useful cross-check from the infrastructure side. Its July 28 presentation described approximately 590 MW leased to CoreWeave and more than 430 MW being billed as of mid-July, with full completion expected in early 2027. It also described a take-or-pay relationship with annual price escalation and customer-funded power, utilities and substantial buildout costs. This is counterparty evidence of real delivery and ongoing obligations, rather than a repeat of CoreWeave’s demand forecast. It remains a dated supplier report, not a November site inspection or proof that end-customer GPUs are fully utilized. Supplier billing and productive cloud output answer different questions. [cw-s7]\n\nFinancing scale gives those distinctions weight. In the first half of 2026, CoreWeave reported $3.663 billion of operating cash inflow, $14.874 billion of investing cash outflow and $13.985 billion of financing cash inflow. At June 30, it reported $35.6 billion of total indebtedness. The cash-flow pattern shows that expansion depended heavily on external financing. It does not, by itself, establish that the investment will lose money. The question is whether the eventual service cash flows and asset reuse justify the funding obligations. These are June snapshots, not an estimate of cash or debt outstanding on publication day. [cw-s5]\n\nThe second-quarter earnings release reported a 59% adjusted EBITDA margin alongside a $626 million net loss. Both figures belong in the picture. For a capital-intensive service, an earnings measure that excludes depreciation and financing costs cannot settle whether the overall buildout earns an adequate return. Conversely, a consolidated loss during rapid expansion does not prove that each operating cluster has poor economics. What is missing for that judgment is a consistent view of returns by deployment generation, including capital cost, useful service life, maintenance, power and financing. A headline margin is an incomplete answer to a lifetime investment question. [cw-s6]\n\nSeptember provides a concrete example of why financing headlines need reconciliation. The completed offering in CoreWeave’s September 22 filing was $4.2 billion of 2.875% convertible notes due April 2033, including the fully exercised $500 million purchaser option. Net proceeds were $4.137 billion before other offering expenses. Capped-call transactions cost approximately $566.2 million and were intended to limit certain conversion-related dilution or cash obligations, subject to a cap. Subtracting that cost leaves approximately $3.571 billion before those other expenses. The coupon is therefore not the full economic price of the financing, and the earlier $3.7 billion pricing headline understates the completed principal. [cw-s8]\n\nNVIDIA is an important part of the system rather than a distant equipment supplier. In January, NVIDIA itself announced a $2 billion CoreWeave equity investment and a collaboration intended to accelerate more than 5 GW of AI factories by 2030. The plan included help with land, power and buildings, early adoption of successive NVIDIA platforms, and work toward broader use of CoreWeave software in reference architectures. Supplier alignment can support execution. It also makes the relationship worth analyzing across procurement, capital and software together. An investment and a development plan should not be read as an unconditional guarantee of CoreWeave’s obligations. [cw-s9]\n\nThere is also evidence of pricing opportunity. On September 17, CoreWeave said it had signed three-to-six-month contracts at approximately $40 million per MW, expressed as annualized revenue divided by power required for the relevant clusters. It reported 4.2 GW of contracted power as of August 11. The price measure is neither a standardized GPU-hour quote nor twelve months of secured revenue for a three-month contract. Shorter commitments may capture tight supply at attractive prices while bringing the next repricing decision closer. Without hardware mix, utilization, associated costs and renewal outcomes, the reported rate cannot establish a durable fleet-wide margin. [cw-s10]\n\nPCT’s assessment is conditional: CoreWeave offers a substantive operating platform with demonstrated customer relationships and financed infrastructure delivery. Its risk is the coordination of those strengths over time. The decisive evidence would show customer payments, supplier milestones and debt service staying aligned across successive deployments, then show older equipment earning worthwhile renewal revenue. This review does not establish contract-level profitability, current project completion, future refinancing terms or realized resale values. CoreWeave has scheduled its next results discussion for November 10. Until then, treating announced commitments as delivered capacity would obscure precisely the execution the business must prove. [cw-s11]", "paragraphs": [ { "id": "cw-p01", "source_ids": [ "cw-s1" ], "evidence_class": "analysis", "inference": "Editorial framing of the disclosed contract and infrastructure-financing model, not a claim about any particular missed delivery.", "body_character_start": 0, "body_character_end": 575 }, { "id": "cw-p02", "source_ids": [ "cw-s1" ], "evidence_class": "mixed", "inference": "The comparison with a matching marketplace is PCT analysis; no conclusion about relative investment merit is intended.", "body_character_start": 585, "body_character_end": 1279 }, { "id": "cw-p03", "source_ids": [ "cw-s2" ], "evidence_class": "mixed", "inference": "Feature descriptions are vendor claims; the proposed acceptance measurements are original buyer diligence.", "body_character_start": 1289, "body_character_end": 1976 }, { "id": "cw-p04", "source_ids": [ "cw-s3" ], "evidence_class": "mixed", "inference": "Licensing statement is documented; procurement implications are analysis, not legal advice.", "body_character_start": 1986, "body_character_end": 2669 }, { "id": "cw-p05", "source_ids": [ "cw-s1", "cw-s4" ], "evidence_class": "mixed", "inference": "The evidence establishes relationships, not customer-specific economics.", "body_character_start": 2679, "body_character_end": 3348 }, { "id": "cw-p06", "source_ids": [ "cw-s5" ], "evidence_class": "mixed", "inference": "72% is PCT arithmetic; no customer identity inferred from letter labels.", "body_character_start": 3365, "body_character_end": 4009 }, { "id": "cw-p07", "source_ids": [ "cw-s6" ], "evidence_class": "mixed", "inference": "Interpretation distinguishes future commitments from cash and active infrastructure.", "body_character_start": 4019, "body_character_end": 4722 }, { "id": "cw-p08", "source_ids": [ "cw-s7" ], "evidence_class": "mixed", "inference": "Counterparty reporting is independent of CoreWeave authorship but not an independent audit; no present-day completion assumed.", "body_character_start": 4732, "body_character_end": 5438 }, { "id": "cw-p09", "source_ids": [ "cw-s5" ], "evidence_class": "mixed", "inference": "Cash-flow interpretation is analysis; investing cash flow is not relabeled as capital expenditure or free cash flow.", "body_character_start": 5448, "body_character_end": 6081 }, { "id": "cw-p10", "source_ids": [ "cw-s6" ], "evidence_class": "mixed", "inference": "Economic critique is PCT analysis; no claim of misstated accounting or negative cluster-level returns.", "body_character_start": 6091, "body_character_end": 6770 }, { "id": "cw-p11", "source_ids": [ "cw-s8" ], "evidence_class": "mixed", "inference": "Proceeds subtraction is original arithmetic; not an effective-yield calculation or current cash balance.", "body_character_start": 6780, "body_character_end": 7511 }, { "id": "cw-p12", "source_ids": [ "cw-s9" ], "evidence_class": "mixed", "inference": "Strategic benefits and concentration implications are analysis; the 2030 capacity statement is a target.", "body_character_start": 7521, "body_character_end": 8198 }, { "id": "cw-p13", "source_ids": [ "cw-s10" ], "evidence_class": "mixed", "inference": "Pricing is a company-reported selected-contract metric; repricing implications are analysis, not a forecast.", "body_character_start": 8208, "body_character_end": 8886 }, { "id": "cw-p14", "source_ids": [ "cw-s11" ], "evidence_class": "analysis", "inference": "Original assessment and explicit research gaps. Only the earnings-call date is sourced to cw-s11.", "body_character_start": 8897, "body_character_end": 9653 } ], "sources": [ { "id": "cw-s1", "title": "CoreWeave 2025 Form 10-K", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm", "published": "2026-03-02", "period": "2025-12-31", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "SEC-filed company annual report", "limitations": "Business model and Microsoft concentration; filing is primary company disclosure, not independent endorsement.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s2", "title": "CoreWeave SUNK product page", "url": "https://www.coreweave.com/products/sunk", "published": null, "period": null, "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Company product documentation", "limitations": "Capabilities; no performance claims independently tested.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s3", "title": "SUNK Anywhere official repository", "url": "https://github.com/coreweave/sunk-anywhere", "published": null, "period": null, "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Company technical repository", "limitations": "License-access section; mutable documentation observed on read date.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s4", "title": "OpenAI subprocessor list", "url": "https://openai.com/policies/sub-processor-list/", "published": "2026-10-14", "period": null, "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Customer-authored primary disclosure", "limitations": "Relationship corroboration, not customer-specific financial terms.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s5", "title": "CoreWeave Q2 2026 Form 10-Q", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm", "published": "2026-08-12", "period": "2026-06-30", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "SEC-filed company quarterly report", "limitations": "Note 2 concentration; MD&A liquidity and cash flows; unaudited quarterly statements.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s6", "title": "CoreWeave Q2 2026 earnings release", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm", "published": "2026-08-11", "period": "2026-06-30", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Company earnings release filed as SEC exhibit", "limitations": "Backlog definition, active versus contracted power, adjusted EBITDA and GAAP loss.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s7", "title": "Core Scientific Q2 2026 earnings presentation", "url": "https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001839341-26-000013/q2fy26earningsdeck728am.htm", "published": "2026-07-28", "period": "2026-07", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Infrastructure-counterparty SEC exhibit", "limitations": "Slides 9 and 10; mid-July billing snapshot and prospective completion, not November verification.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s8", "title": "CoreWeave convertible-notes closing Form 8-K", "url": "https://www.sec.gov/Archives/edgar/data/1769628/000176962826000432/crwv-20260917.htm", "published": "2026-09-22", "period": "2026-09-22", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "SEC-filed completed financing disclosure", "limitations": "Principal, maturity, proceeds and capped-call cost after purchaser option exercised.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s9", "title": "NVIDIA and CoreWeave strengthen collaboration", "url": "https://nvidianews.nvidia.com/news/nvidia-and-coreweave-strengthen-collaboration-to-accelerate-buildout-of-ai-factories", "published": "2026-01-26", "period": null, "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Supplier/investor-authored primary announcement", "limitations": "Investment and strategic plans; counterparty has an economic interest.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s10", "title": "CoreWeave continues to contract new compute capacity at higher prices", "url": "https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx", "published": "2026-09-17", "period": null, "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Company operational announcement", "limitations": "Selected short-contract metric and dated contracted-power update; not a market-wide price index.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." }, { "id": "cw-s11", "title": "CoreWeave announces Q3 2026 results-call date", "url": "https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Date-of-Third-Quarter-2026-Financial-Results-and-Conference-Call/default.aspx", "published": "2026-10-27", "period": "2026-11-10", "accessed": "2026-11-04", "read_date": "2026-11-04", "type": "Company calendar announcement", "limitations": "Scheduled future discussion, not Q3 results.", "license_status": "Unknown; no republication license established. Facts paraphrased; no substantial quotation." } ], "facts_inferences_and_unknowns": { "facts": "Documented facts and company or counterparty claims are attributed at paragraph level; sources distinguish independent authorship from independent verification.", "inferences": "Each mixed or analytical paragraph identifies original interpretation in its inference field.", "unknowns": [ "Current completion and customer acceptance of each contracted cluster.", "Customer-level pricing, margins, payment schedules and full rights under applicable contracts.", "Future refinancing conditions and true all-in financing cost for each deployment.", "Realized renewal economics and asset residual values across GPU generations.", "Whether selected short-term pricing persists or represents the overall fleet." ], "calculations": [ { "id": "cw-calc1", "paragraph_id": "cw-p06", "expression": "36 + 26 + 10 = 72", "unit": "percent of Q2 2026 revenue", "source_id": "cw-s5" }, { "id": "cw-calc2", "paragraph_id": "cw-p11", "expression": "4.137 - 0.5662 = 3.5708", "unit": "USD billions, rounded to 3.571", "source_id": "cw-s8", "limitation": "Before other offering expenses; not current cash or enterprise financing cost." } ] }, "deduplication": { "compared_with": "pct-2026-11-03", "prior_company_article_id": "pct-2026-11-03-company-sfcompute", "distinct_angle": "Financing and physical delivery under committed contracts; not reservation resale liquidity.", "new_reporting": "CoreWeave SEC filings; completed September financing; customer, supplier and NVIDIA corroboration." }, "editorial_controls": [ "Financial figures retain explicit period labels.", "The scheduled November 10 earnings discussion is not treated as already held.", "No quarterly letter-coded customer is identified by inference.", "Backlog is not cash; contracted power is not active power.", "Selected annualized contract pricing is not extrapolated into a fleet-wide forecast.", "Proprietary source material is not republished; primary material is paraphrased.", "No investment recommendation, price target, outreach, login, trade or external state change." ], "corrections": [], "correction_policy": "Append dated corrections with affected paragraph IDs and preserve prior versions.", "license_status": "Private original PCT prose; public license undecided. External sources retain their own rights.", "public_url": null, "format_documentation": "UTF-8 JSON. body is canonical text; paragraphs annotate body character spans and source IDs without duplicating article text. Character spans use Python Unicode code-point indexing. Stable source and article IDs are not URLs. Dates use ISO 8601. Source license status is 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." } }, { "schema_version": "pct-editorial-1.0", "id": "pct-2026-11-04-deal-aib-nebius", "section": "Deals", "publisher": "Peterson Compute Trading", "as_of_date": "2026-11-04", "headline": "AIB–Nebius: 50 MW contracted, financing and delivery still ahead", "dek": "A September colocation commitment shows why signed capacity, project funding and delivered GPU supply need separate records.", "story_type": "Original public-source deal analysis", "freshness": "Agreement announced September 30, 2026; October disclosures reviewed November 4. Not a new November 4 deal.", "editorial_word_count": 653, "parties": { "supplier": "AIB Data Centers", "customer": "Nebius", "legal_customer": "Nebius Inc. referenced in issuer risk disclosures; full executed agreement not inspected" }, "deal": { "asset_class": "AI data-center colocation", "critical_it_capacity_mw": 50, "utility_capacity_mw": 65, "initial_term_years": 12, "gpu_model": null, "gpu_count": null, "current_executable_gpu_hour_price": null, "historical_presentation_calculation": { "source_id": "S4", "contract_value_input_usd": 1320000000, "derived_usd_per_critical_it_mw_year": 2200000, "formula": "1320000000 / 12 / 50", "classification": "Historical issuer-presentation input and arithmetic, not a spot quote or independently verified realized price" } }, "facts": [ { "id": "F1", "claim": "AIB announced a binding 50 MW critical-IT agreement on September 30; initial term 12 years; utility agreement 65 MW.", "source_ids": [ "S1" ], "evidence_type": "issuer_reported_signed_agreement" }, { "id": "F2", "claim": "October 7 funding plan used customer prepayments and potential debt and preferred equity.", "source_ids": [ "S2" ], "evidence_type": "management_plan" }, { "id": "F3", "claim": "October deck targets Q4 2027 halls and calls project financing incomplete; AIB supplies facilities, customers supply compute.", "source_ids": [ "S3" ], "evidence_type": "issuer_presentation_target_and_status" }, { "id": "F4", "claim": "September deck used a $1.32bn implied-rate input; October peer examples expressly do not imply AIB terms.", "source_ids": [ "S4", "S3" ], "evidence_type": "historical_disclosure_with_later_comparability_caveat" }, { "id": "F5", "claim": "Nebius announced a separate $3.5bn Citi facility October 22, without a stated AIB allocation.", "source_ids": [ "S5" ], "evidence_type": "issuer_reported_financing" } ], "inferences": [ { "id": "I1", "claim": "Do not convert colocation MW into GPU-hour inventory without hardware and delivery evidence.", "basis_fact_ids": [ "F1", "F3" ], "label": "PCT analysis" }, { "id": "I2", "claim": "Nebius corporate financing does not establish CLT1 project financing or AIB cash receipts.", "basis_fact_ids": [ "F2", "F3", "F5" ], "label": "PCT analysis" }, { "id": "I3", "claim": "September presentation arithmetic is not an executable GPU price; later wording alone does not establish an amended agreement.", "basis_fact_ids": [ "F4" ], "label": "PCT analysis" } ], "classification": { "signed": "Company-announced binding colocation commitment; full contract not independently reviewed", "financing": "Project financing incomplete in cited October presentation; later corporate Citi facility is separate", "delivered": "Not established by cited sources; latest cited target Q4 2027", "spot_execution": "Not observed or established", "gpu_hour_pricing": "Unknown", "cash_collected": "Unknown", "execution_volume": null }, "unknowns": [ "Executed agreement text and exact signing date", "Final capital structure and completed project financing", "Prepayment amounts, receipt and escrow-release status", "Current contract value and detailed rent escalators/pass-throughs", "Construction completion and customer acceptance", "GPU models, counts, topology and deployment", "Utilization, realized revenue and spot execution", "Resale/assignment rights and complete termination provisions" ], "methodology": { "source_basis": "Public issuer releases and presentations; original synthesis, not republished reporting.", "verification_limit": "No interviews, account access, trade execution, independent operational measurement or full contract review. Negative claims are bounded to cited materials.", "source_history": "September and October presentations are preserved as separate vintages; no inference of contractual amendment from differences in wording.", "copyright": "No source images or substantial quotations reproduced; source facts paraphrased with links." }, "sources": [ { "id": "S1", "title": "AIB Data Centers Signs Contract with Nebius for AI Data Center Capacity", "publisher": "AIB Data Centers, SEC-hosted company release", "published_at": "2026-09-30", "url": "https://www.sec.gov/Archives/edgar/data/2070542/000121390026105865/ea030737801ex99-1.htm", "primary": true, "supports": "Announced binding agreement; capacity; term; utility-load support; expected funding.", "retrieved_at": "2026-11-04" }, { "id": "S2", "title": "AIB Data Centers Issues CEO Letter Highlighting Recent Milestones and CLT1 Development Strategy", "publisher": "AIB Data Centers, SEC exhibit", "published_at": "2026-10-07", "url": "https://www.sec.gov/Archives/edgar/data/2070542/000121390026107421/ea030792301ex99-1.htm", "primary": true, "supports": "Customer funding, planned debt and preferred equity, and remaining execution priorities.", "retrieved_at": "2026-11-04" }, { "id": "S3", "title": "AIB Data Centers October 2026 Investor Presentation", "publisher": "AIB Data Centers", "published_at": "2026-10", "date_precision": "month", "url": "https://www.aib.us/docs/AIB-Investor-Presentation-October-2026.pdf", "primary": true, "supports": "Pages 6, 8, 10 and 11: facility/compute distinction, delivery target, pricing caveat and incomplete financing.", "retrieved_at": "2026-11-04" }, { "id": "S4", "title": "AIB Data Centers Q3 2026 Investor Presentation, furnished September 30", "publisher": "AIB Data Centers, SEC exhibit", "published_at": "2026-09-30", "url": "https://www.sec.gov/Archives/edgar/data/2070542/000121390026105249/ea030723301ex99-1.htm", "primary": true, "supports": "Page 4 historical implied-rate calculation; this is presentation evidence, not the complete contract.", "retrieved_at": "2026-11-04" }, { "id": "S5", "title": "Nebius secures $3.5 billion financing facility with Citi", "publisher": "Nebius", "published_at": "2026-10-22", "url": "https://nebius.com/newsroom/nebius-secures-3-5-billion-financing-facility-with-citi", "primary": true, "supports": "Separate corporate facility, duration, initial interest margin and stated use of proceeds.", "retrieved_at": "2026-11-04" } ], "corrections": { "version": 1, "published_corrections": [], "policy": "Append dated corrections identifying the affected claim, old and new values, and supporting source. Preserve the prior vintage; do not silently relabel targets as delivery." }, "license": { "original_analysis": "Original commissioned PCT text. Apply the edition’s authorized distribution license to original prose only.", "third_party_material": "Linked source materials remain subject to their publishers’ rights and terms; no license to third-party text or images is conferred.", "distribution_license": "Not independently assigned by this research draft." }, "body": "AIB’s agreement with Nebius offers a useful test of how a compute-market ledger should record a large infrastructure commitment. On September 30, AIB announced a binding agreement covering 50 MW of critical IT capacity in the southeastern United States, over an initial 12-year term. The company said a 15-year electric-service agreement supported 65 MW of utility load. Those are different measurements of the project, not two quantities to add together. This is a November 4 review of an earlier transaction, rather than a newly signed deal. [September 30 announcement](https://www.sec.gov/Archives/edgar/data/2070542/000121390026105865/ea030737801ex99-1.htm)\n\nThe purchased product matters. AIB’s October presentation describes a facility business: the provider supplies the building, electrical infrastructure and cooling, while the customer installs and operates the servers and GPUs. The disclosed commitment therefore belongs in a colocation ledger. Converting it into accelerator-hours would require hardware, configuration, operating and service assumptions that the headline megawatt figure cannot supply. [October presentation, page 6](https://www.aib.us/docs/AIB-Investor-Presentation-October-2026.pdf)\n\nFunding is another separate record. In its October 7 shareholder letter, AIB described a plan combining customer prepayments with potential project-level debt and preferred equity. Management said it did not then anticipate needing corporate-level common equity for CLT1. That statement describes an intended capital structure, not proof that every funding source has closed. The same letter named securing project financing and advancing delivery among the next priorities. [October 7 shareholder letter](https://www.sec.gov/Archives/edgar/data/2070542/000121390026107421/ea030792301ex99-1.htm)\n\nThe October presentation makes the remaining work more explicit: project financing was incomplete, and two 25 MW halls were targeted for Q4 2027. A target should remain a target in the data. Recording these halls as available supply today would erase the construction and commissioning interval that a buyer still has to manage. [October presentation, pages 8 and 11](https://www.aib.us/docs/AIB-Investor-Presentation-October-2026.pdf)\n\nPricing also needs its original context. AIB’s September SEC presentation used $1.32 billion divided by 12 years and 50 MW in an implied-rate calculation. That arithmetic produces $2.2 million per critical-IT MW-year. The October deck instead presents other companies’ lease examples and explicitly warns that they do not establish AIB’s terms. This review does not interpret the change as a contract amendment or a withdrawal of the earlier figure. Neither presentation establishes a current executable GPU-hour price. [September presentation, page 4](https://www.sec.gov/Archives/edgar/data/2070542/000121390026105249/ea030723301ex99-1.htm) [October presentation, page 10](https://www.aib.us/docs/AIB-Investor-Presentation-October-2026.pdf)\n\nA more recent financing event provides context, but cannot close CLT1’s funding record by association. On October 22, Nebius announced a separate $3.5 billion senior secured Citi facility with a 364-day maturity, initially priced at Term SOFR plus 2.75 percentage points. Stated uses include data-center development, GPU procurement and general corporate purposes. The announcement does not allocate a disclosed amount to AIB. It establishes a financing agreement at Nebius, not a customer payment received by AIB or acceptance of either hall. [Nebius financing announcement](https://nebius.com/newsroom/nebius-secures-3-5-billion-financing-facility-with-citi)\n\nFor Peterson Compute Trading, the analytical lesson is to preserve several independent states: contract signed, financing committed, cash received, facility accepted, compute available and workload executed. A change in one state should not automatically update the others. A procurement agent should attach a source date to each transition and leave unsupported fields unknown. That approach makes a large deal useful for supply planning without allowing it to contaminate a spot-price series.\n\nThe next evidence to seek is specific: completed project-finance documentation, disclosed prepayment receipt or release, construction milestones, customer acceptance and actual service commencement. For downstream compute comparisons, the missing fields include accelerator model and count, usable cluster topology, service duration, commercial restrictions and the delivered price basis. No spot transaction, resale right, completed GPU deployment or workload execution is established by this review. Unknown execution volume should remain null, rather than becoming either zero or the entire contracted capacity.\n\nThe agreement is meaningful evidence that a buyer committed to future infrastructure. Its trading relevance lies in the conditional path from that commitment to usable supply. Keeping that path visible is more informative than compressing the deal into a dollar headline or treating a planned megawatt as inventory that can be sold today.", "content_type": "transaction_analysis", "title": "AIB–Nebius: 50 MW contracted, financing and delivery still ahead", "publication_date": "2026-11-04", "observation_date": "2026-11-04", "editorial_voice": "Deals Desk | AI-assisted, evidence-first", "license_status": "Private original PCT prose; public license undecided. External sources retain their own rights.", "public_url": null, "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-2026-11-04-market-formation-nord-pool", "title": "Nord Pool's lesson for compute: price the place, settle the shortfall", "publication_date": "2026-11-04", "observation_date": "2026-11-04", "type": "historical_and_institutional_analysis", "section": "Market formation", "dek": "Nordic power trading separates delivery-area prices, a common benchmark and responsibility for imbalance. Compute markets need equally explicit boundaries.", "body": "Nord Pool offers compute-market builders a useful example of how a shared marketplace can preserve differences in what buyers actually receive. The business began in Norway as Statnett Marked in 1993 and became a Norwegian-Swedish exchange under the Nord Pool name in 1996. Its subsequent architecture separates prices for delivery areas, a common financial reference and the settlement of deviations from contracted quantities. The useful lesson is institutional: decide which differences a contract preserves, which a benchmark abstracts away and who pays when delivery diverges. [1]\n\nThe geographic distinction is explicit. Nord Pool explains that limited transmission between bidding areas can produce different area prices. Its Nordic System Price is calculated separately, using the underlying Nordic order books but treating internal transmission capacity as unlimited. External flows remain inputs. The methodology effective September 30, 2025 describes this as a common reference for derivatives. A benchmark can therefore serve a broad market while deliberately excluding a constraint that affects a particular buyer's delivery price. [2, 3]\n\nFor compute, that is a warning about basis risk, the difference between a reference price and the cost a buyer actually faces. Consider a hypothetical buyer requiring a particular GPU configuration in one region. A broad index might fall while that qualifying capacity becomes scarcer. The index need not be defective for the hedge to disappoint. Its construction may simply omit the buyer's binding constraint. Publishing regional or configuration-specific prices could reveal that difference, but would also divide participation across more products. That trade-off needs evidence about actual substitution, rather than a universal GPU-hour label.\n\nDelivery accountability is a separate institution. The Nordic Imbalance Settlement Handbook, version 5.5 dated November 1, 2026, assigns financial responsibility for imbalances to balance responsible parties. The model compares trades with realized consumption and production. Those parties need agreements with eSett and the relevant transmission system operator, and must meet collateral requirements. The system operators retain responsibility for physical balancing; eSett performs financial settlement. These roles should not be compressed into a claim that the exchange alone guarantees physical delivery. [4]\n\nThe handbook also specifies preliminary calculations and a final settlement 13 days after delivery, with a window for correcting reported inputs. That is a useful distinction for compute designers: a scheduler's allocation record, observed service availability and a final invoice are different records. A market needs a stated rule for reconciling them, including responsibility for missing telemetry and disputed failures. The Nordic timetable is evidence of a defined process, not a recommendation that compute copy the same number of days. [4]\n\nEven the time unit has evolved. Nordic RCC's 2025 activity report records the day-ahead transition to 15-minute units on September 30, 2025, for delivery beginning October 1. Nord Pool's corresponding benchmark revision also introduced quarter-hour resolution. Some explanatory web pages still describe hourly operation, so those primers should not be used to infer today's contract granularity. The broader lesson is to version the unit, methodology and effective date together whenever a market changes its product. [2, 5]\n\nA practical compute pilot could begin with a narrow reservation: a named accelerator configuration, minimum memory and interconnect requirements, an identified region, a delivery window and explicit interruption rights. This is a proposed design, not an observed market launch. The contract should distinguish capacity made available from resources the customer actually uses. Otherwise an idle customer reservation and a supplier's failure to provide capacity could appear identical in a utilization measure. Performance testing, permitted substitutions and responsibility for customer-side software errors would need separate treatment.\n\nThe operational test would follow a reservation through acceptance, provisioning, measured availability, any replacement and final settlement. The commercial test would ask whether another qualified supplier can fulfill the same obligation without renegotiating its essential terms. A proposed public index should then disclose how closely its eligible services match that obligation. These are editorial criteria for examining formation, not proof that any particular compute platform already meets them or that a financial contract would be liquid.\n\nThe analogy has important limits. Electricity operates through a shared physical network with system-wide balancing requirements. Compute workloads can sometimes wait or move, while software compatibility, data movement, privacy requirements and cluster topology can prevent substitution even between identical chips. There is no basis here for treating a compute marketplace as a grid operator or assuming a uniform balancing price would be appropriate. The transferable idea is a clear boundary between the traded promise, the reference price and the evidence used to settle performance.\n\nFor PCT, the next evidence to request is consequently concrete: a delivery specification, a methodology version, a sample reconciliation record and repeated fulfillment by more than one eligible supplier. Data rights also belong in that checklist. Nord Pool's published data terms distinguish internal use, redistribution and uses including index creation and settlement. Visibility of a price does not establish permission to build and redistribute a benchmark from it. This article republishes no market-price series and makes no claim that such rights have been obtained. [6]", "body_word_count": 822, "sources": [ { "id": "1", "title": "Nord Pool Connection: Welcome to 2023 – 30 years of Nord Pool", "author_or_issuer": "Nord Pool", "source_publish_date": "2023-01-24", "source_publish_date_precision": "day; page displays 24 January 2023 14:03", "observation_date": "2026-11-04", "url": "https://www.nordpoolgroup.com/en/message-center-container/newsroom/newsletter/2023/01/nord-pool-connection/", "locator": "Anniversary introduction; 1993 and 1996 history", "provenance": "Issuer-hosted dated retrospective.", "copyright_licensing_status": "No republication licence verified; original factual paraphrase only." }, { "id": "2", "title": "Nordic System Price: Methodology for calculation, revision 1.8", "author_or_issuer": "Nord Pool European Market Coupling Operator AS", "source_publish_date": "2025-09-30", "source_publish_date_precision": "document revision/effective date; exact web posting date not established", "observation_date": "2026-11-04", "url": "https://www.nordpoolgroup.com/4a6b82/globalassets/download-center/rules-and-regulations/methodology-for-calculating-nordic-system-price-effective-from-30-september-2025.pdf", "locator": "Revision history p. 2; sections 2.1–2.2 pp. 5–6; sections 3–4 pp. 9–11", "provenance": "Issuer-hosted 15-page methodology linked from Nord Pool's official current methodology page; PDF read directly.", "copyright_licensing_status": "Copyright/proprietary-rights notice; document states redistribution requires consent. No document, source text, chart or price data is reproduced." }, { "id": "3", "title": "Price calculation", "author_or_issuer": "Nord Pool", "source_publish_date": null, "source_publish_date_precision": "undated explanatory web page; observed 2026-11-04", "observation_date": "2026-11-04", "url": "https://www.nordpoolgroup.com/en/trading/Day-ahead-trading/Price-calculation/", "locator": "Area price section", "provenance": "Issuer-hosted explainer; used only for geographic congestion and price differentiation, not its outdated hourly timing references.", "copyright_licensing_status": "No republication licence verified; original factual paraphrase only." }, { "id": "4", "title": "Nordic Imbalance Settlement Handbook, version 5.5", "author_or_issuer": "eSett Oy", "source_publish_date": "2026-11-01", "source_publish_date_precision": "date printed on cover and in revision log; exact web posting date not established", "observation_date": "2026-11-04", "url": "https://www.esett.com/app/uploads/2026/10/NBS-Handbook-v5.5.pdf", "locator": "Cover; section 2.1 p. 14; sections 2.3.1–2.3.4 pp. 17–19; section 6 p. 67; section 9 pp. 96–100; revision log pp. 134–135 (printed page numbers)", "provenance": "Current issuer-hosted PDF reached through https://www.esett.com/handbook/; dated version checked directly. The cited responsibilities are not claimed to be new in this revision.", "copyright_licensing_status": "No republication licence verified; original factual paraphrase only; no tables or contractual text reproduced." }, { "id": "5", "title": "Nordic RCC Activity Report 2025", "author_or_issuer": "Nordic Regional Coordination Centre A/S", "source_publish_date": "2026-04-30", "source_publish_date_precision": "day; verified in issuer publication announcement", "observation_date": "2026-11-04", "url": "https://nordic-rcc.net/wp-content/uploads/2026/04/Nordic-RCC-Activity-Report-2025.pdf", "locator": "Printed page 30, The transition to 15-minute Market Time Unit", "provenance": "Issuer-hosted retrospective report by a participating regional coordination organization; confirms implementation rather than merely an intended launch.", "copyright_licensing_status": "No republication licence verified; original factual paraphrase only.", "publication_date_source_url": "https://nordic-rcc.net/publication-of-nordic-rcc-annual-report-2025-and-activity-report-2025/" }, { "id": "6", "title": "General Terms for Data Services", "author_or_issuer": "Nord Pool AS", "source_publish_date": "2023-07-10", "source_publish_date_precision": "document version date; not a claim that all customers' currently negotiated terms are identical", "observation_date": "2026-11-04", "url": "https://www.nordpoolgroup.com/49fb8e/globalassets/download-center/power-data-services/general-terms-data-services-data.pdf", "locator": "Section 4.1, printed pp. 3–4", "provenance": "Issuer-hosted public terms PDF read directly; no purchase, acceptance or account access performed.", "copyright_licensing_status": "Copyright Nord Pool. Contract-specific data-use and redistribution permissions apply; none acquired or assumed." } ], "confirmed_facts": [ { "claim": "Nord Pool dates its Norwegian start to 1993 and Norwegian-Swedish expansion to 1996.", "source_ids": [ "1" ] }, { "claim": "Area prices reflect transmission constraints; the System Price calculation abstracts from internal Nordic transmission limits.", "source_ids": [ "2", "3" ] }, { "claim": "The November 1, 2026 eSett handbook documents financial imbalance responsibility, collateral and a defined reporting/correction cycle.", "source_ids": [ "4" ] }, { "claim": "Quarter-hour day-ahead delivery began October 1, 2025 following the September 30 transition.", "source_ids": [ "2", "5" ] }, { "claim": "Nord Pool's published data terms distinguish permitted uses including redistribution and index creation.", "source_ids": [ "6" ] } ], "inferences": [ "Compute delivery constraints should be measured separately from broad benchmark exposure.", "A narrow pilot contract should specify configuration, region, time, availability, permitted interruption and substitution before claiming fungibility.", "Reservation, telemetry and final invoice reconciliation are separate evidence layers.", "Repeated fulfillment across eligible suppliers is a proposed formation test, not an observed result or a forecast of liquidity." ], "unknowns": [ "Whether any compute platform implements all of the proposed operational and settlement tests; this article does not investigate or establish that claim.", "Actual hedge performance, usable cross-provider substitution and demand for a particular compute contract.", "The exact web posting dates for the effective-dated methodology and handbook; printed revision dates are stated separately.", "Any negotiated licence or data rights that PCT or a future benchmark provider might have; none inferred from public access." ], "limitations": [ "Institutional analysis, not a November 4 exchange-launch or liquidity news report.", "The eSett handbook's November 1 version is evidence of documented responsibilities, not evidence those responsibilities were newly introduced that day.", "Electricity's physical balancing and regulatory institutions are not assumed to transfer to computing.", "Some Nord Pool explanatory pages still use hourly wording. Effective-dated methodology and implementation evidence control the article's timing statements.", "No actual power prices, compute prices, traded volume, open interest, spreads or hedge-effectiveness estimates are supplied.", "Nord Pool's historical launch is not presented as causal proof that any one institutional feature created liquidity." ], "uncertainty": "High confidence in the cited institutional descriptions and dated implementation evidence; compute-market design conclusions are expressly editorial inferences requiring empirical testing.", "copyright_licensing_status": "Original article text; factual paraphrase with attribution. No licensed source text, images, charts or market-price datasets republished. Source/data redistribution permission has not been obtained or assumed.", "corrections": [], "prepublication_verification_notes": [ "Replaced reliance on hourly primer wording with revision 1.8 of the System Price methodology, dated September 30, 2025, and Nordic RCC's implementation report.", "Checked eSett v5.5 cover and revision log: November 1, 2026, not an inferred date from its October file path.", "An attempted European Commission page fetch returned HTTP 429. No claim relies on that inaccessible page; implementation is supported by Nordic RCC and Nord Pool primary material." ], "distinct_from_previous_issue": "The November 3 formation piece concerned SGX's 2006 OTC freight clearing launch. This piece addresses Nordic electricity's locational pricing, reference-price abstraction and reconciliation of physical-delivery deviations.", "content_type": "market_formation", "editorial_voice": "Formation Desk | AI-assisted, institutional", "license_status": "Private original PCT prose; public license undecided. External sources retain their own rights.", "public_url": null, "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-2026-11-04-spotlight-mooncake-store", "title": "Mooncake Store: putting a price on repeated work", "section": "Product spotlight", "publisher": "Peterson Compute Trading", "as_of_date": "2026-11-04", "status": "desk_research", "subject": "Mooncake Store and vLLM shared KV caching", "editorial_word_count": 558, "facts": [ { "id": "F1", "claim": "Mooncake paper first submitted June 24, 2024; current cited revision dated October 8, 2026.", "source_ids": [ "S1" ], "evidence_type": "dated_primary_research" }, { "id": "F2", "claim": "Prefix reuse reduces prefill, not direct decode time.", "source_ids": [ "S2" ], "evidence_type": "primary_documentation" }, { "id": "F3", "claim": "MooncakeStoreConnector supports cross-instance caching and CPU/disk offload; strict isolation requires configured multi-tenancy.", "source_ids": [ "S3" ], "evidence_type": "primary_documentation" }, { "id": "F4", "claim": "May 6 authors report 3.8× throughput, 46× lower median TTFT and 8.6× lower E2E latency on 12 GB200s; separate synthetic scaling test reaches 60 GPUs.", "source_ids": [ "S4" ], "evidence_type": "author_reported_benchmark_not_independently_reproduced" }, { "id": "F5", "claim": "vLLM hashes cache identity with prefix and extra inputs; request salts restrict reuse.", "source_ids": [ "S5" ], "evidence_type": "primary_documentation" }, { "id": "F6", "claim": "Mooncake repository identifies Apache-2.0 licensing.", "source_ids": [ "S6" ], "evidence_type": "repository_license" } ], "inferences": [ { "id": "I1", "claim": "The procurement comparison should include shared-cache infrastructure and completed requests meeting service constraints, not GPU hourly rates or TTFT alone.", "basis_fact_ids": [ "F2", "F3", "F4" ], "label": "PCT analysis" }, { "id": "I2", "claim": "If useful hourly throughput doubles and total hourly infrastructure cost rises 25%, cost per completion falls 37.5%.", "basis_fact_ids": [], "label": "illustrative arithmetic, not observed benchmark", "assumptions": [ "Same task mix and required quality/latency constraints.", "Throughput multiplier is 2.0.", "All-in hourly infrastructure cost multiplier is 1.25.", "Ratio is 1.25 / 2.0 = 0.625; reduction is 1 - 0.625 = 0.375." ] }, { "id": "I3", "claim": "A distributed cache creates value only when avoided recomputation and service improvements justify retention, retrieval and operational costs.", "basis_fact_ids": [ "F2", "F3", "F4" ], "label": "PCT analysis" } ], "unknowns": [ "No independent benchmark reproduction, interviews, deployment or supplier quote was performed.", "No measured all-in dollar cost, power consumption or cost-per-task saving was established.", "The May benchmark does not establish equivalent gains for another model, GPU, cache configuration, request mix or latency target.", "Latest documentation is mutable and is not an immutable historical release snapshot.", "Source publication dates are null where no original publication date was established." ], "corrections": [ { "status": "prepublication_guardrail", "text": "Do not label 46× as throughput, GPU savings or cost savings; it is the authors' median TTFT result." }, { "status": "prepublication_guardrail", "text": "Do not conflate the 12-GPU real-trace experiment with the 60-GPU synthetic scaling experiment." }, { "status": "prepublication_guardrail", "text": "Do not repeat the May 6 roadmap as current feature status; November 4 documentation describes disk offload." }, { "status": "prepublication_guardrail", "text": "Original Mooncake research publication, May 6 vLLM integration benchmark and today's documentation observation are distinct dates." } ], "license": { "editorial_text": "Original PCT synthesis. Publisher controls the license assigned to this original editorial text; no new public license is asserted here.", "third_party_material": "No third-party image, chart, code or extended quotation reproduced. Primary facts are paraphrased and attributed.", "software": "Mooncake repository code is Apache-2.0; this does not establish the license for all linked blog and paper content.", "republication_note": "Keep source links and benchmark qualifications with any republication." }, "sources": [ { "id": "S1", "title": "Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving", "publisher": "Ruoyu Qin et al., arXiv", "published_at": "2024-06-24", "updated_at": "2026-10-08", "retrieved_at": "2026-11-04", "url": "https://arxiv.org/abs/2407.00079v5", "primary": true, "license": "arXiv non-exclusive distribution license; no blanket downstream republication grant assumed" }, { "id": "S2", "title": "Automatic Prefix Caching: features and limits", "publisher": "vLLM", "published_at": null, "page_date": "2026-10-13", "page_date_label": "Displayed footer date; treated as page date, not inferred original publication", "retrieved_at": "2026-11-04", "url": "https://docs.vllm.ai/en/latest/features/automatic_prefix_caching/", "primary": true }, { "id": "S3", "title": "MooncakeStoreConnector Usage Guide", "publisher": "vLLM", "published_at": null, "page_date": "2026-11-04", "page_date_label": "Displayed footer date; treated as page date, not inferred original publication", "retrieved_at": "2026-11-04", "url": "https://docs.vllm.ai/en/latest/features/mooncake_store_connector_usage/", "primary": true }, { "id": "S4", "title": "Serving Agentic Workloads at Scale with vLLM x Mooncake", "publisher": "vLLM", "authors": [ "Yifan Qiao", "Trong Dao Le", "Ao Shen", "Zhewen Li", "Bowen Wang" ], "published_at": "2026-05-06", "retrieved_at": "2026-11-04", "url": "https://vllm.ai/blog/2026-05-06-mooncake-store", "primary": true, "license": "Page footer states all rights reserved; only original summary and attributed facts used" }, { "id": "S5", "title": "Automatic Prefix Caching: design", "publisher": "vLLM", "published_at": null, "page_date": "2026-06-23", "page_date_label": "Displayed footer date; treated as page date, not inferred original publication", "retrieved_at": "2026-11-04", "url": "https://docs.vllm.ai/en/latest/design/prefix_caching/", "primary": true }, { "id": "S6", "title": "Mooncake Apache License", "publisher": "kvcache-ai / Mooncake contributors", "published_at": null, "retrieved_at": "2026-11-04", "url": "https://github.com/kvcache-ai/Mooncake/blob/main/LICENSE-APACHE", "primary": true, "license": "Apache-2.0 (repository code)" } ], "benchmark_scope": { "reported_by": "vLLM authors", "publication_date": "2026-05-06", "model": "Kimi-2.5 NVFP4", "hardware": "12 GB200 GPUs", "topology": "1 prefill instance and 1 decode instance; prefill TP4, decode DP8 plus EP", "workload": "Codex coding-agent traces described in source S4", "independent_reproduction": false, "reported_metrics": { "throughput_multiplier": 3.8, "median_ttft_speedup": 46, "end_to_end_latency_speedup": 8.6, "baseline_cache_hit_pct": 1.7, "store_cache_hit_pct": 92.2 }, "separate_scaling_experiment": "Synthetic workload based on traces, scaled from 12 to 60 GB200 GPUs." }, "agent_fields": [ "model_and_revision", "serving_engine_revision", "cache_connector_revision", "hardware_and_gpu_count", "prefill_decode_topology", "cache_capacity_by_tier", "tenant_isolation_policy", "request_mix", "routing_policy", "cache_warmth", "prefix_reuse_rate", "transfer_latency", "eviction_rate", "p50_ttft", "p99_ttft", "end_to_end_latency", "completed_requests_meeting_slo_per_hour", "total_infrastructure_cost_per_hour", "currency", "source_url", "retrieved_at" ], "body": "A coding agent that returns after running a tool may bring back most of the context it submitted before. Recomputing that history consumes capacity without adding new information. Mooncake Store turns previously computed attention state into a resource that multiple inference instances can reuse. For compute buyers, it raises a practical question: how much of the workload being priced is genuinely new work?\n\nThe project comes from Mooncake, the serving architecture developed for Moonshot AI's Kimi. Its research paper, first submitted on June 24, 2024 and revised on October 8, 2026, describes separate prompt-processing and token-generation clusters, with a shared cache drawing on CPU, DRAM and SSD resources. That puts memory placement and scheduling alongside accelerator performance in the serving design. [S1]\n\nIn vLLM, ordinary prefix caching reuses the attention keys and values associated with an already-processed prompt prefix. It saves prompt-processing work, called prefill; it does not directly accelerate generation of new output tokens, called decode. Long answers with little repeated input therefore have a different opportunity from repeated questions against a large shared document. [S2]\n\nMooncakeStoreConnector extends reuse beyond one instance. Current vLLM documentation describes a shared external store, CPU and disk offloading, and both single-node and multi-node configurations. A separate MooncakeConnector transfers state directly between prefill and decode workers. The distinction matters when specifying a deployment: point-to-point transfer and retained, shared cache capacity solve different problems. The store can be embedded in workers or hosted separately; its documented transport choices include RDMA and a TCP fallback. [S3]\n\nThe dated performance evidence is striking but bounded. In a May 6, 2026 vLLM report, the authors tested Kimi-2.5 NVFP4 with separate prefill and decode on 12 GB200 GPUs using coding-agent traces. They reported 3.8 times the throughput, a 46-fold reduction in median time to first token, and an 8.6-fold reduction in end-to-end latency. Cache hits rose from 1.7% to 92.2%. These are the authors' results, not PCT measurements. Their separate scaling experiment used a synthetic workload and reached 60 GPUs; it should not be described as the same trace experiment. [S4]\n\nPCT's economic reading is conditional. A faster first token does not establish an equivalent reduction in cost per completed task. Suppose a deployment doubles useful completions per hour while its total hourly infrastructure cost rises by 25%. Cost per completion becomes 1.25 divided by two, or 62.5% of the starting figure: a 37.5% reduction. This is illustrative arithmetic, not a Mooncake result. Storage, networking, CPU work and operations belong in the numerator. Completed requests meeting the required latency and quality targets belong in the denominator.\n\nReuse also has boundaries. vLLM's cache identifiers incorporate the preceding prefix and relevant distinctions such as adapters and multimodal inputs. Its design includes per-request salts to restrict sharing. Mooncake's current guide separately states that strict tenant isolation requires a multi-tenant master and registered tenant quota policy. A shared pool should not be assumed to provide the required isolation merely because it is reachable. [S5][S3]\n\nBefore assigning monetary value to a cache hit, replay representative traffic with cold and warm caches, realistic tool delays and the intended routing policy. Measure skipped prefill, transfer time, evictions, tail latency and total cost together. Test restart and failure behavior too. PCT's conclusion: shared state can substitute for repeated accelerator work, but the trade only pays when retaining and retrieving it is cheaper than recomputing it under the service's actual constraints.", "content_type": "product_spotlight", "publication_date": "2026-11-04", "observation_date": "2026-11-04", "editorial_voice": "Spotlight Desk | AI-assisted, buyer-focused", "license_status": "Private original PCT prose; public license undecided. External sources retain their own rights.", "public_url": null, "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." }