Startup, release, acquisition, failure, and rumor chronology
As of: 2026-08-26. Tracker: #9305. This chronology records events that can change AI infrastructure availability or design assumptions. It separates completed events, announced/future products, reported negotiations, and unconfirmed rumors.
Failure and cancellation chronology (#9363)
| Date | Event | Lifecycle interpretation |
|---|---|---|
| 2024 | Equinix stopped commercially offering Metal and wound down supporting operations. | Service-line withdrawal and impairment, not physical portfolio shutdown. |
| 2025-05-20 | Builder.ai entered insolvency proceedings. | Counterparty failure; workload and hardware consequences remain unquantified. |
| 2025-07-24 | Intel discontinued Germany/Poland and slowed Ohio. | Two planned expansions canceled; one build delayed to demand. |
| 2025-09-22 | Google withdrew the Franklin Township, Indianapolis rezoning proposal. | Pre-construction site withdrawal, not lost live MW. |
| 2025-10-14 | Untether AI entered bankruptcy, retained from the existing corpus record. | Accelerator-startup insolvency, not a chip-capacity decrement. |
| 2025 full year | Heatmap found 25 opposition-linked U.S. cancellations, 21 in the second half. | Bounded heterogeneous cohort, not homogeneous MW. |
Named grid chronology (#9364)
| Date | Outcome | Lifecycle state |
|---|---|---|
| 2024-11-01 | FERC rejected the Susquehanna 300-to-480 MW amended co-location agreement without prejudice. | Regulatory rejection of an amendment, not permanent site denial. |
| 2025-07-09 | PUCO approved AEP Ohio’s data-center tariff. | Financially stronger queue/contract terms, not energized load. |
| 2025-08-20 | LPSC approved the Meta/Laidley generation and transmission settlement path. | Named project power approval, before construction and commissioning. |
| 2025-11-25 | Virginia SCC approved Dominion GS-5 effective 2027. | Customer-class tariff and cost allocation, not one site interconnection. |
Component delivery chronology (#9365)
| Date | Milestone | Lifecycle state |
|---|---|---|
| 2024-02-26 | Micron HBM3E entered volume production for H200 timing. | Production start tied to a named platform, without units. |
| 2024-04-18 | TSMC said CoWoS-L was in production/customer qualification. | Process/design readiness, not good-package shipment totals. |
| 2024-Q4 | TSMC Arizona N4 entered high-volume production. | Qualified wafer output, upstream of packages and systems. |
| 2025-06-03 | Broadcom Tomahawk 6 shipped in production volume. | Switch silicon delivery, not completed fabrics. |
| 2026-06-24 | Micron reported HBM4 high-volume shipments. | Customer-platform shipment state, without unit disclosure. |
2026 chronology
2026-08-20 — ERCOT publishes 765-kV planning and forecast distinctions
- Planning evidence, not built capacity: ERCOT’s August 12 response says the 2024 RTP used about 150 GW in 2030, the 2025 RTP used about 159 GW for the 2031 summer peak, and a lower approximately 125 GW 2030 sensitivity still required the full STEP 765-kV additions.
- Forecast boundary: ERCOT’s most recent approximately 110 GW 2032 econometric base-plus- mid-size forecast excludes Batch-process large loads; a final forecast including those loads was expected by the end of 2026. Assumption, sensitivity, forecast, actual load, planned line, constructed line, and energized asset are not interchangeable states.
2026-08-20 — PUCT grants ERCOT Batch Zero classification extension
- Completed decision: the PUCT granted ERCOT’s good-cause request to move the Batch Zero first-classification deadline from 2026-09-01 to 2026-10-01; the 2027-10-01 batch-study deadline did not move.
- Stated reason: ERCOT needed the month for manual initial screening; duplicate, terminated, withdrawn, and completed-request cleanup; lower-priority identification; transmission-service- provider validation of project status/readiness; and the resulting Batch Zero update to the Long-Term Load Forecast due 2026-10-01.
- Lifecycle boundary: this is a classification/forecast-input delay, not cancellation, denial, a construction stop, completed study, approved interconnection, energization, or live load. Requested nameplate MW is not verified capacity, coincident peak, annual energy, or IT MW.
- Primary witness: PUCT Project No. 59142 ERCOT request and order granting the exception, checked 2026-08-27.
- Omitted candidates: no record was added for the suggested New York permit pause, Texas- governor directive/~474 GW quantity, or ERCOT planning sensitivity/transmission dependency; this pass did not obtain direct official witnesses establishing their exact scope and limits.
| Date | Entity/event | Type and state | Infrastructure expectation | Evidence caveat |
|---|---|---|---|---|
| 2026-01-05 | NVIDIA Vera Rubin / Rubin NVL72 | Product launch; announced production ramp | Rack-scale co-design, liquid cooling, RAS, and MoE fabrics become the integration boundary | Vendor peak claims require production goodput. |
| 2026-01-06 | xAI $20B Series E | Funding completed; company-reported scale | Million-H100-equivalent estates need continuous financing, power, and consumer distribution | Equivalent count and monthly users are not physical inventory or token distribution. |
| 2026-02-11 | Anthropic electricity-cost proposal | Policy release | Ratepayer allocation and grid investment become datacenter product constraints | Sector power forecast is company-authored. |
| 2026-03-02 | NVIDIA investments in Lumentum and Coherent | Strategic investments/purchase commitments | Lasers, optics, packaging, and manufacturing rights are strategic capacity inputs | Investments do not prove delivered component volume. |
| 2026-03-13 | AWS–Cerebras phase-split inference | Partnership/product announcement; future launch | Prefill and decode may run on different accelerator architectures | No matched public production economics yet. |
| 2026-03-16 | NVIDIA Dynamo 1.0 and DSX | Software/reference release | Topology-aware scheduling and facility digital twins become standard product surfaces | Reference architecture and vendor results need site/workload validation. |
| 2026-04-01 | Google TPU 8t/8i | Product announcement | Training and latency-oriented inference diverge enough for separate chips | Workload shares and independent net value remain undisclosed. |
| 2026-04-06 | Anthropic–Google–Broadcom future TPUs | Multi-gigawatt contract; future capacity | Labs commit across several silicon generations and years | 2027 capacity is not present supply. |
| 2026-04-20 | Anthropic–AWS up to 5 GW | Ten-year capacity commitment | Multi-cloud fleets and custom silicon are enduring, not temporary bridges | “Up to” and decade totals need delivery states. |
| 2026-05-19 | Google multi-site TPU / agent-token disclosure | Operational observation | Training spans sites; agentic enterprise demand can multiply within weeks | Aggregate tokens hide requests, cache, and tenant skew. |
| 2026-06-18 | Delay-rate rebuttal | Industry-analysis contradiction | Announcement totals require project/state denominators before aggregation | Proprietary project tracking limits reproduction. |
| 2026-06-24 | AWS Project Rainier operational | Cluster brought online | Multi-datacenter clusters can reach ~500K custom accelerators | Healthy fraction, topology efficiency, and utilization remain private. |
| 2026-07-13 | Meta Richland expansion to 5 GW | Site expansion announcement | One campus becomes a regional generation/grid/community program | Planned “compute capacity” needs IT/facility boundary. |
| 2026-07-23 | AMD–Cerebras integration | Partnership announced; not yet generally available | Heterogeneous inference phase placement can cross vendor systems | Transfer cost and availability not yet witnessed. |
| 2026-07-29 | Microsoft and Meta Q2 capex disclosures | Official financial results | Financing and leases are as important as chip count | Accounting boundaries differ materially. |
| 2026-07-30 | Alphabet/Amazon Q2 disclosures and reporting | Official results plus reported call update | Hyperscalers continue record investment while capacity is constrained | Amazon’s $220B update needs primary filing/transcript extraction. |
| 2026-08-11 | IBM–Together AI B300 cluster | Multi-year agreement | Specialized inference software attaches to incumbent enterprise cloud capacity | Scale and economics undisclosed. |
| 2026-08-12 | Claude through Mistral on Azure | Distribution partnership | A request can cross model provider, reseller, cloud, and region boundaries | Traffic share/routing not public. |
| 2026-08-13 | Reported Anthropic–Decart talks | Rumor/unconfirmed acquisition talks | Labs may buy infrastructure optimization capability | Bloomberg-origin report; no completed or official deal in corpus. |
| 2026-08-17 | Groq $350M financing after June $650M | Funding completed | Specialized inference clouds remain capital intensive and ecosystem-dependent | Financing is not deployment or profitability evidence. |
| 2026-08-24 | NVIDIA Groq 3 LPX production | Product reported in production | Ultra-low-latency decode is integrated with rack-scale context processing | Vendor-selected benchmark; cloud availability and net economics pending. |
| 2026-08-25 | Reported OpenAI datacenter leadership change | Rumor/reported personnel and strategy signal | Build-versus-lease strategy and organizational execution may change delivery | Official account and full strategic interpretation absent. |
| 2026-08-26 | AWS–NVIDIA two-million-GPU plan | Future 2027–2028 product/capacity plan | Fleets mix Blackwell/Rubin generations, custom silicon, and secure government capacity | Manufacturing, power, delivery, and demand can change planned counts. |
| 2026-08-26 | Anthropic–Nscale reported $45B lease | Reported contract/future project | Long-duration neocloud leases couple startup finance with dedicated generation | Planned West Virginia facility and 2027 start are prospective. |
| 2026-08-26 | NVIDIA–Hugging Face acquisition reports | Rumor/reported agreement; unconfirmed by parties | Ownership of a central model-distribution platform could affect neutrality, portability, and hardware routing | Business Insider reported talks above $13B with no deal; The Information later reported a $12.9B agreement. Signing/announcement, terms, regulatory conditions, and close are unresolved. |
2026-07-14 — New York orders a qualified state-permit abeyance for large data centers
- Operative scope: Executive Order 62 directs DEC to hold specified discretionary state applications in abeyance when they were pending and not deemed complete before the order; local-government permissions are excluded. Covered data centers consume or are designed to consume at least 50 MW, with manufacturing, research, education, and medical-care exclusions.
- Queue and future-work boundaries: the order cites nearly 12 GW of NYISO data-center load requests as of May 2026, more than 8 GW entered in 2025. Those are requests, not live demand. The report/GEIS and 60-day, 90-day, and twelve-month processes remain future work, so the order is neither a blanket construction ban nor a simple fixed one-year expiry.
2026-06-18 — PUCT approves ERCOT’s transition Batch Zero process
- Tracked queue: ERCOT reports more than 438,000 MW of tracked large-load requests, nearly 89% from data centers, and groups qualified projects of at least 75 MW into Batch Zero. ERCOT cautions that not all requests become built projects; tracked nameplate is not verified or livable demand.
- Announced milestones: classification was then expected in August 2026, later superseded by the October 1 deadline record above, and a final statewide transmission plan was expected in Fall 2027. Neither announcement proves a completed plan, constructed project, or energized load.
Pre-2026 anchor events
| Date | Event | Why it remains relevant |
|---|---|---|
| 2024-09-25 | Mnemosyne multi-million-context inference paper | Long context can require sequence/KV parallelism and adaptive chunking; prevalence remains unknown. |
| 2024-12-27 | DeepSeek-V3 technical report | A 2,048-H800 cluster and co-designed MoE/FP8 stack challenge assumptions that capability scales only with million-chip estates. |
| 2025-01-21 | Stargate launch | Started the public $500B/10 GW multi-party infrastructure commitment frame. |
| 2025-03-18 | NVIDIA Dynamo announced | Made disaggregation, KV-aware routing, and cache offload a vendor platform thesis. |
| 2025-05-22 | Stargate UAE | Sovereign/regional capacity became a first-class frontier-lab program. |
| 2025-06-03 | Aliyun KVCache production study | Public evidence that reuse is skewed and category-dependent, not one universal Zipf. |
| 2025-09-09 | Rubin CPX announced | Hardware roadmap began specializing explicitly for massive-context inference. |
| 2025-11-18 | Anthropic–Microsoft–NVIDIA partnership | Demonstrated simultaneous commitments to Azure/NVIDIA alongside AWS and Google. |
| 2025-12-24 | NVIDIA–Groq non-exclusive license | Specialized inference IP can be licensed/integrated without a full acquisition. |
Rumor register
| Rumor | Origin in corpus | Confidence | Corroboration | Resolution/expiry |
|---|---|---|---|---|
| NVIDIA warned major customers of >15% AI-server price increases due to memory/HBM pressure | Industry reporting relayed by Tom’s Hardware, 2026-08-23 | Low | Broad memory-price pressure exists; specific customer notice not independently confirmed here | Review on official pricing, customer filing, or 2026-10-31. |
| Anthropic was in talks to acquire Decart for about $6B | Bloomberg reporting/video, 2026-08-13 | Low | Later outlets repeated the report; independent sourcing not established | Resolve on signed/terminated transaction or 2026-11-13. |
| OpenAI datacenter chief left amid reorganization and more lease/cloud use | Wall Street Journal, 2026-08-25 | Low | No complete primary confirmation of strategic interpretation | Resolve on OpenAI statement/leadership update or 2026-10-25. |
| NVIDIA reportedly agreed to buy Hugging Face for $12.9B after talks above $13B | Business Insider and The Information, 2026-08-26 PDT | Low | Two independently sourced reports establish a report lifecycle but conflict on whether a deal had been reached; TechCrunch reported no party response | Resolve on party announcement/signing, regulatory evidence, termination, close, or 2026-11-30. |
A rumor expires from active planning relevance if its review date passes without new primary or independently sourced evidence. Keep the historical entry, mark its outcome, and never silently promote it to fact.
Acquisition, failure, and cancellation watchlist
The public ledger remains much stronger on launches and funding than on failures. Every refresh must search for the following states, not just positive announcements:
- acquired, acquihired, technology licensed, assets sold, merged, renamed;
- shutdown, bankruptcy, restructuring, layoffs, impaired assets, covenant breach;
- delayed, canceled, downsized, phased, customer terminated, grid interconnect lost;
- product preview vs general availability vs deprecated/end-of-life;
- customer concentration, take-or-pay renegotiation, unpaid capacity, or stranded site;
- security, reliability, or performance incident that changes real goodput.
Priority named watches from startups-landscape.md: CoreWeave,
Nebius, Crusoe, Lambda, Together AI, Fluidstack, Nscale, Groq, Cerebras, SambaNova, Etched,
d-Matrix, Tenstorrent, Positron, inference/runtime vendors, KV-cache/storage vendors, optical
and cooling startups, and generation/datacenter developers.
Completed acquisition outcomes
| Date | Company / acquirer | Outcome | Evidence preserved | Missing operating proof |
|---|---|---|---|---|
| 2024-07-11 | Graphcore / SoftBank | Completed; Graphcore became a wholly owned subsidiary retaining its name and UK headquarters | Completed ownership change and strategic accelerator option | Purchase price, shipped post-deal systems, customer adoption, useful goodput |
| 2024-09-30 | OctoAI / NVIDIA | Acquired team and serving technology later appear in NVIDIA personnel and FlashInfer provenance | Post-deal team/technology integration | Transaction terms, hosted-service migration/shutdown, retained customers/revenue |
| 2025-11-25 | Ampere / SoftBank | Completed $6.5B acquisition; the original filing disclosed steep revenue decline and three operating-loss years | Completion, price, and historical financial deterioration | Unit shipments, customer concentration, roadmap delivery, AI-workload share |
Acquired is neither active-independent nor failed-without-value. Record ownership,
retained team/IP/product, customer migration, shipped artifacts, and runtime evidence
separately.
Event-state rules
- Funding is evidence of financing, not technical capability, adoption, profitability, or deployed capacity.
- Partnership is evidence of intent/contract only to the disclosed boundary; name launch state.
- “Production” requires a defined customer/workload and availability boundary, not a press title.
- Acquisition talks remain rumor until signed and remain incomplete until closing.
- License, talent transfer, and partnership are not synonymous with acquisition.
- Product retirement, cancellation, and failure remain in the chronology; do not delete them.
- Repeated articles based on one source are one rumor origin, not corroboration.
Acquisition, shutdown, and support-continuity additions
| Date | Entity / event | Lifecycle outcome | Infrastructure lesson | Open verification |
|---|---|---|---|---|
| 2025-06-05 | Untether AI team joins AMD; company shuts down | Engineering-team transfer; speedAI and imAIgine support discontinued | Team, IP, product, and support can separate in an acquihire. | Full IP/asset disposition and customer migration outcomes. |
| 2025-10-14 | Untether AI bankruptcy filing | Legal/financial failure after shutdown | Reported ~$25M assets versus >$128M unsecured liabilities; funding raised did not produce durable operations. | Lifetime revenue, deployed systems, and AMD consideration. |
| 2025-11-17 | Replicate joins Cloudflare | Platform acquisition with announced brand/API continuity | Model-hosting primitives are consolidating into network/developer platforms. | Transaction value, GPU/traffic economics, and observed integration milestones. |
| 2025-12-15 | NVIDIA acquires SchedMD | Scheduler vendor acquired; Slurm promised to remain open and vendor-neutral | Scheduling/governance is a strategic cluster control point. | Post-acquisition release governance and cross-hardware neutrality. |
| 2025-12 / disclosed 2026-02-25 | NVIDIA–Groq non-exclusive license | IP license plus reported talent transfer; Groq remains operational | Consolidation does not require buying the legal entity. | Price, IP scope, staff transfer, support, and product-roadmap effects. |
Rumor state at this snapshot
| Rumor | Current state | Last checked | Expiry | Resolution rule |
|---|---|---|---|---|
| NVIDIA AI-server price-hike notice | Open / unverified | 2026-08-26 | 2026-11-30 | Confirm only with NVIDIA/customer primary evidence or multiple genuinely independent reports. |
| Anthropic–Decart infrastructure talks | Open / unverified | 2026-08-26 | 2026-11-30 | A later Anthropic infrastructure deal is not confirmation unless Decart and the claim terms match. |
| OpenAI datacenter leadership/strategy report | Open / unverified | 2026-08-26 | 2026-11-30 | Separate personnel departure confirmation from inferred strategic impact. |
| NVIDIA–Hugging Face reported agreement | Open / reported agreement, party-unconfirmed | 2026-08-27 | 2026-11-30 | Resolve signing or announcement, transaction structure/terms, regulatory conditions, and close separately; do not treat the report as current ownership. |
Expiry means remove from active architectural assumptions, not delete history. A
later check must mark each claim confirmed, partially_confirmed, refuted, or
expired_unresolved, with exact claim fragments and sources.
Speculative-decoding observability update — 2026-08-27
- vLLM: the pinned V1 implementation records drafted, accepted, and emitted token totals, speculative iterations, draft acceptance rate, a per-request consecutive accepted-draft-length histogram, and acceptance by draft position. These are available instrumentation surfaces; the source does not show that any production operator enabled, scraped, retained, or published them.
- SGLang: the pinned metrics collector exposes accepted-length histogram and cumulative accepted-token/forward-iteration seams, while the adaptive implementation consumes runtime outcome statistics and chooses later speculative parameters. Controller input and selected parameters are not a published controller outcome or causal production speedup.
- NVIDIA TensorRT-LLM: the pinned benchmark report schema separates total draft tokens, total accepted draft tokens, acceptance rate/length/probability, request count, and output-token total. It is benchmark instrumentation, not a production request distribution or fleet aggregate rate.
- Bounded absence: this slice found no primary-source retry/fallback/client-side production distribution. It therefore makes no retry or fallback telemetry claim. Benchmark acceptance or speedup likewise does not establish production prevalence or goodput.
Rumor resolution update — 2026-08-27
- NVIDIA memory-price direction corroborated, magnitude still open. NVIDIA’s August 26 fiscal-Q2-2027 materials/call corroborate extreme memory-cost pressure and planned product price increases. They do not confirm the reported >15% magnitude, customer notices, named Grace Blackwell/Vera Rubin configurations, or shipment scope. Status remains
partially_corroborated_open; expiry remains 2026-11-30. - Anthropic / Decart talks remain open. Bloomberg-origin reporting is independently echoed by Reuters, Axios, and Israeli technology reporting, but neither company has announced a signed or completed deal as of August 27. Status remains
independently_corroborated_open; expiry remains 2026-11-30. - NVIDIA / Hugging Face report lifecycle remains open. Business Insider reported detailed talks above $13B with no deal at 2026-08-27T00:34:46Z; The Information later reported a $12.9B agreement at 6:26 p.m. PDT on August 26. NVIDIA’s news archive and Hugging Face’s official blog contained no signing or acquisition announcement at the August 27 cutoff, and TechCrunch reported that neither party had responded. Status is
reported_agreement_open; expiry is 2026-11-30. Signing/announcement, transaction structure/terms, regulatory conditions, and close remain unresolved.
This check resolves none of these rumors into completed fact. It tightens the corroboration
graph and preserves the exact fields still missing. The NVIDIA / Hugging Face source lineage,
publication times, and independence limits are in source-register.md.
Rumor resolution update — 2026-08-26
| Claim | Resolution state | Confirmed or corroborated fragment | Still unverified |
|---|---|---|---|
| OpenAI data-center leader departure | Partially confirmed; no longer classified as rumor | OpenAI confirmed Chris Malone left. | Why he left and the causal/strategic meaning of the reorganization and leasing changes. |
| Anthropic–Decart acquisition talks | Independently corroborated, open | Reuters independently reported talks after the Bloomberg-origin report. | Final $6B price, terms, closing, staff/IP transfer, and integration. |
| NVIDIA >15% AI-server increase | Partially corroborated, open | NVIDIA earnings commentary corroborated extreme memory-cost pressure and planned product price increases. | Exact >15% magnitude, affected configurations/customers, and timing in the original report. |
Resolution discipline: confirming one claim fragment does not confirm the rest. The OpenAI entry now records a reported observation for the confirmed departure and keeps strategic interpretation bounded. The other two remain rumors/watch items until their transaction or price terms are confirmed, refuted, or expire.
Remote-browser lifecycle benchmark slice — 2026-08-27
- Steel BrowserBench commit
847e4ed604764ce8b887265709ddb2d1c3c5f020, authored 2026-01-15, is the pinned official repository source. Its README, sequential runner, package lock, and committed results provide five provider-bounded 5,000-attempt lifecycle samples. - The actual result windows differ: Steel 2025-11-06; Hyperbrowser and AnchorBrowser 2025-11-05 to 2025-11-06; Kernel 2026-01-10 to 2026-01-11; Browserbase 2026-01-12 to 2026-01-13. Those date and SDK-version boundaries are not merged into a timeless provider ranking.
- Steel, Kernel, Browserbase, and Hyperbrowser each report 5,000 successful samples out of 5,000 attempts. AnchorBrowser reports 4,867/5,000 (97.34%) with 133 session-create failures. The Kernel file also contains one successful row with a null release timing, so only 4,999 rows have a complete four-stage timing tuple.
- BrowserBench’s mean stages are create, Playwright CDP connection, DOMContentLoaded page navigation, and release. Its total mean/median/p95/p99 cover the create→connect→goto→release lifecycle, not browser-ready, task-complete, or full agent-session latency.
- The sample is benchmark evidence, not production agent telemetry. Attempt rate is not concurrency, queue depth, or throughput; successful samples are not user sessions, tasks, actions, observations, or tool calls; means are not stage-tail or failure distributions.
- The unresolved production browser/desktop trajectory gap remains action/tool-call distributions, observation bytes, resets, full session duration, escalations, user populations, and retry/failure/tail distributions.