Frontier infrastructure coverage audit

Issue #9387 delta

• Deepened the existing operator-autoscaling-unit-2026 record without adding a duplicate entry, URL, or entity. Totals remain 272 entries, 265 unique source URLs, 224 entity labels, and 3 explicit rumors.
• Separated evidence modes: the 10-100 QPS, 128-64K-sequence M/M/R analysis is an offline analytical oracle; the 929K-request/1.5B-prompt-token-per-model evaluation is production-trace replay on a research prototype; the fixed-capacity TPS and minimum-capacity curves are SLO-constrained benchmark sweeps; and plan latency/model error are prototype microbenchmarks.
• Bound the dynamic replay to Qwen2-7B and Qwen2-57B-A14B on the common nano-vLLM data plane. The main cluster is 40 A100-80GB GPUs across five Azure VMs, with NVLink inside each eight-GPU VM and InfiniBand between VMs; the hardware/granularity sensitivity cluster is 24 GB200 GPUs in one NVLink/NVL domain.
• Added the one-second OpScale versus 20-second model-level control intervals, one-hour displayed windows, P99 TTFT targets of one second for Qwen2-7B and two seconds for Qwen2-57B-A14B, average GPU counts of 7.1 and 10.8, SLO attainment of 98.4% and 98.1%, the baseline GPU counts, and the full Qwen2-7B scale-up latency table.
• Bound result deltas and denominators to their envelopes: versus model-level provisioning, 20.1%/35.7% average GPU reductions at 1K sequence length for dense/MoE, 36.3% at 4K and 22.1% at 8K, and 14-28% lower cluster power at high load; versus the model-level static deployment on the same GPU budget, 3-38% higher dense input TPS and up to 44% higher MoE input TPS at 40 GPUs. These maxima are not one stackable result.
• Recorded the placement/control mechanism without inventing realized counts: monolithic base instances plus elastic operator replicas; interference-aware best fit; locality preference from same device to NVLink/NVL to InfiniBand; weighted shortest-queue dispatch; and new-GPU activation only when no existing placement remains SLO-feasible.
• Preserved exact absence. The paper reports no achieved active batch, numeric queue depth or pending-token count, queue-wait distribution, exact cluster GPU-utilization value, operator-replica count/placement time series, configured per-model tensor/pipeline parallelism, SM-allocation distribution, numeric TBT SLO, failures/retries, explicit goodput, or currency/GPU-hour cost. The B=1-256 range is only a typical profiling space.
• Correctly typed the separate granularity/topology figure: the reported 1.7x is maximum request throughput in RPS, not input TPS. Its A100 result reports up to 33% lower GPU demand versus Attn-FFN; at 40 QPS, moving from A100 to GB200 reduces GPU count by 52% for OpScale and 38% for the monolith. The figure does not disclose model, sequence length, trace/arrival law, or SLO.
• No official OpScale code/data artifact is linked from the arXiv v1 landing page or TeX source. nano-vLLM, kvcached, Dynamo, and Production Stack are referenced dependencies/baselines, not an OpScale repository. The paper is benchmark evidence, not provider production-prevalence evidence.

Issue #9384 delta

• Added five non-duplicate official Google Cloud records: Cluster Director network topology, current GKE TPU Multislice, the maintenance-only Cloud TPU API Multislice ceiling, current Compute Engine/DWS provisioning semantics, and the November 2023 XPK/GKE TPU v5e training run.
• Preserved the physical/control hierarchy: host → rack/sub-block → block → cluster for Cluster Director; and job → Multislice → slice → VM/host → chip, with ICI inside TPU slices and DCN between slices.
• Typed quantities by lifecycle: “thousands of GPUs” is a supported Cluster Director scale class; 256 slices is a configured maximum for the legacy Cloud TPU queued- resource API; Flex-start is best-effort admission; approved reservation-bound capacity has very high assurance; and 50,944 TPU v5e chips across 199 pods is an achieved provider-reported workload.
• Kept current and legacy control planes separate. The current GKE page does not disclose a maximum slices-per-JobSet value, so the legacy 256-slice ceiling is not transferred to GKE or treated as a prevalent topology.
• Totals rise to 272 entries, 265 unique source URLs, 224 entity labels, and 3 explicit rumors. ai_cloud rises to 11, accelerator_platform to 4, official_documentation to 9, official_engineering_release to 11, vendor_specification to 10, and official_statement to 100.
• None of the five sources discloses a production queue-wait distribution, utilization, provider-wide active/schedulable scale, failure/retry distribution, power, or total workload cost. Regional and preview/GA status remain source-specific: the XPK release announces Multislice Training GA, while the living configuration pages do not label every listed topology, machine family, or region with one universal stage.

Issue #9383 delta

• Added one market_signal rumor record for the NVIDIA / Hugging Face August 2026 acquisition-report lifecycle.
• Preserved the independently sourced conflict: Business Insider reported talks above $13B with no deal, while The Information later reported a $12.9B agreement.
• Checked NVIDIA’s official news archive and Hugging Face’s official blog through 2026-08-27; no party announcement confirming signing or closing was found.
• Kept signing or party announcement, transaction structure and terms, regulatory conditions, and close unresolved, with last_checked_at 2026-08-27 and expiry 2026-11-30.
• Totals rise to 267 entries, 260 unique source URLs, 224 entity labels, and 3 explicit rumors; market_signal rises to 25.

Issue #9381 delta

• Revised four existing primary-trace records rather than duplicating them: BurstGPT/Azure OpenAI, Chutes Year-in-Serving, ServeGen, and Alibaba Cloud Model Studio KV-cache.
• Named each random/rank variable, observation window, granularity, method class, directly reported parameter/test, and denominator/stationarity limitation.
• Kept entry, category, and entity counts unchanged; coverage.entities remains exactly sorted.
• Retained only released traces, primary papers, and official analysis/generator code as substantiation.
• Preserved the bounded conclusion: ServeGen v3 directly supports tier-specific Gamma/Weibull/Exponential IAT fits, category-specific Pareto/lognormal-mixture input fits, an Exponential output fit, and 24h/12h aggregate seasonality; it does not support a universal law or unreported parameters. Later BurstGPT v1.1 synthetic Zipf material, qualitative heavy-tail language, and empirical CDFs were not upgraded into production fits.

Snapshot: 2026-08-27 Issue: #9306
Authority: index.json

Issue #9379 delta

Issue #9379 adds three official first-party records and refreshes two existing records without duplicating their IDs. Google reports Gemini app and Antigravity active-user populations, a monthly developer population, aggregate model-API token throughput, and Cloud token-threshold customer cohorts; Meta reports almost 1 billion Meta AI monthly actives; Microsoft reports separate Microsoft 365 Copilot seat, GitHub Copilot user, Foundry customer, Agent 365 registered-agent, and cumulative Purview-audited-interaction populations; OpenAI reports 1 billion people using ChatGPT weekly; Anthropic reports more than 100,000 customers running Claude on Amazon Bedrock. The slice improves provider-scale population coverage, but it does not fill the production-distribution gap: none of these counts supplies requests, queries, sessions, messages, tokens, concurrency, interarrival laws, geography, tenant concentration, market share, or Zipf evidence. No extra frontier provider was added because no additional robust official disclosure located in this pass supplied a comparably useful, clearly scoped denominator.

Verdict

The corpus has a strong initial spine but is incomplete. It covers the major architectural seams and several high-value production traces; it does not establish entity-complete global coverage, complete production demand distributions, normalized financial or site lifecycles, or a resolved rumor history. “Exhaustive” remains an operating method—explicit taxonomy, dated evidence, and visible gaps—not a claim that the open web has a finite or fully enumerated boundary.

Issue #9373 now adds four official-source records: the existing PUCT-granted Batch Zero classification extension, New York Executive Order 62, ERCOT’s June Batch Zero approval release, and ERCOT’s August transmission-planning response. The slice preserves the operative boundaries: New York’s state-level abeyance is limited to specified discretionary applications pending and not deemed complete before the order, excludes local permissions, and is not a blanket construction ban or simple fixed one-year expiry; ERCOT’s >438,000 MW tracked request nameplate and nearly 89% data-center share are not verified or livable demand; and the 150 GW and 159 GW planning assumptions, 125 GW sensitivity, and 110 GW econometric forecast retain their distinct years, populations, and forecast roles. Request, verified status, Batch classification, forecast inclusion, completed study, approval, construction, energization, and live load remain separate states. The New York GEIS/report processes, final Batch Zero outputs, Fall 2027 statewide transmission plan, and end-2026 forecast including Batch loads remain future work. The Texas-governor directive/~474 GW claim remains explicit coverage debt because no direct governor source was supplied. The current exact indexed totals after issue #9384 are 272 entries, 265 unique URLs, and 224 entity labels.

Issue #9325 adds an opened Modine/Airedale cooling factory, a >$4B 2027-2029 capacity reservation with $165M upfront funding, and a shipping Schneider prefabricated pod rated to 1MW+. Factory opening, reserved capacity, shipping module, site acceptance, commissioning, and live IT load remain distinct states.

Issue #9317 adds four foundational serving mechanisms—Orca iteration scheduling, vLLM paged KV memory, Sarathi chunked prefill, and DistServe phase disaggregation—while binding each maximum to its year, baseline, model, hardware, workload, SLO, and transfer/control overhead. The maxima are not stackable universal multipliers.

Issue #9318 adds completed lifecycle outcomes for Graphcore/SoftBank, OctoAI/NVIDIA, and Ampere/SoftBank, including Ampere’s disclosed pre-deal revenue decline and operating losses. Acquisition price, retained IP/team, shipped product, and useful goodput remain separate evidence states.

Issue #9315 adds the operative post-Diffusion advanced-computing timeline: the May 2025 withdrawal, January 2026 China case-by-case H200/MI325X review, and July 2026 UAE A:5 / approved-end-user treatment. License posture remains separate from shipment and capacity.

Issue #9330 adds Schneider Electric and Eaton manufacturing milestones plus PJM’s large-load forecast/verification process. It keeps announced investment, production target, qualified output, shipment, site acceptance, energized capacity, and metered load as distinct lifecycle states; PJM forecast/request MW is not treated as live capacity.

Issue #9323 adds the first direct regional-demand row: OpenRouter reports weekly continent-level spend over more than 100T production tokens, while SkyLB/SkyWalker is deepened with six-country diurnal and session-locality evidence. Spend remains distinct from requests, tokens, users, tenants, and serving-region load.

Issue #9340 deepened three production traces without changing corpus counts. FineServe, ServeGen, and the one-year Chutes trace now parameterize model/task arrivals, client concentration, multimodal prompt sizes, reasoning-budget tails, user/model cadence, prefix reuse, and load imbalance. This evidence remains anonymized, provider-specific, window-scoped, and partly figure-read; it does not establish a universal fitted law.

Issue #9360 adds two source-bounded Azure production traces and one Huawei Ascend hardware specification. BurstGPT contributes 10.31M requests over 213 days, daily/weekly periodicity, long-tailed tokens, separated GPT-4 failures, and multi-duration burst case studies; Splitwise contributes one-day Conversation and Coding token/rate distributions. The Atlas 900 A3 row is a 384-NPU, 96-cabinet vendor reference topology, not proof of a built, healthy, schedulable production system, and it does not disclose power or cooling.

Issue #9362 delta

Issue #9362 closes the dedicated Baichuan, iFLYTEK, and Meituan census gap with six source-bounded records: two disclosed training envelopes, typed ecosystem adoption, production-scale asynchronous RL, stateful recommendation caching, and 200M one-week training sequences. iFLYTEK hardware and neutral Ascend operating receipts remain missing. Corpus totals are 213 entries, 208 unique URLs, and 176 entity labels.

Issue #9363 delta

Issue #9363 deduplicates the existing Untether AI bankruptcy and adds five realized negative outcomes to counter announcement/build survivorship bias: Untether AI bankruptcy, Builder.ai insolvency, Equinix Metal wind-down, Intel fab project cancellation/delay, Google Franklin Township withdrawal, and a lower-bound 2025 U.S. cancellation cohort. Totals are 218 entries, 213 unique URLs, and 181 entity labels. More shutdowns, bankruptcies, power denials, and commissioned-capacity failures remain open.

Issue #9364 delta

Issue #9364 adds four regulator/utility records that join large-load rules to named or typed loads: one rejected co-location amendment, two approved large-load tariffs, and one named generation/transmission approval. Totals are 222 entries, 217 unique URLs, and 185 entity labels. Direct queue-position, construction, commissioning, delivered-energy, and site-level utilization receipts remain incomplete.

Issue #9365 delta

Issue #9365 adds five component-delivery receipts across HBM, Ethernet switching, advanced packaging, and leading-edge wafers. Totals are 227 entries, 222 unique URLs, and 190 entity labels. Exact shipped units, yields, allocations, good-package output, assembled systems, optics/cabling joins, and deployed capacity remain incomplete.

Issue #9366 delta

Issue #9366 adds five dedicated speculative-decoding mechanisms and bounded acceptance or speedup examples. Totals are 232 entries, 227 unique URLs, and 195 entity labels. Production request-level acceptance histograms, draft/rejection waste, fallback rates, client retry counts, and load-conditioned distributions remain missing.

Issue #9367 delta

Issue #9367 adds WebArena, WorkArena/BrowserGym, tau-bench, GAIA, and OSWorld as distinct non-coding agent workloads. Totals are 237 entries, 232 unique URLs, and 200 entity labels. Production browser/desktop trajectories, action and tool-call histograms, session durations, observation bytes, reset costs, escalation rates, and user populations remain missing.

Issue #9370 delta

Issue #9370 adds four source-population records with direct conversation, turn/message, user-ID, language, country, timestamp, preference-vote, tree, and annotation boundaries. Totals are 241 entries, 236 unique URLs, and 204 entity labels. Provider-wide tenant/user denominators, billing/API segmentation, stable identities, verified geography, and production session/arrival distributions remain missing.

Issue #9371 delta

Issue #9371 adds three current, pinned official-repository records for vLLM, SGLang, and NVIDIA TensorRT-LLM. Totals are 244 entries, 239 unique URLs, and 207 entity labels; serving_system rises from 37 to 40, and official_repository rises from 1 to 4. The slice verifies available accepted-length histograms, separate drafted/accepted/ emitted-or-output/iteration accounting, benchmark report fields, and adaptive-controller inputs and parameter outputs. It does not establish enabled or collected production telemetry, a representative request distribution, fleet aggregate acceptance/goodput, production prevalence, or a controller outcome. Retry/fallback/client-side production distributions remain missing.

Status vocabulary

Status	Meaning
Complete	The declared bounded requirement is enumerated, source-linked, normalized, and internally checked for this snapshot. It does not mean no future source can exist.
Partial	Useful evidence exists, but named entities, regions, variables, dates, lifecycle states, or source classes remain absent.
Missing	No entry currently answers the requirement directly.
Unverified	A claim is present, but its denominator, original dataset, independence, or later outcome is not strong enough to rely on.

No broad topical slice below qualifies as complete.

Machine-derived inventory

The following counts are derived from index.json, not hand-maintained estimates.

Measure	Current value	Audit note
Entries	272	Every entry has an ID, entity, category, evidence class, confidence, published_at, event_at, source title, and source URL.
Unique source URLs	265	Repeated URLs represent distinct claims/events extracted from the same source; they are not independent corroboration.
Distinct entity labels	224	Joint labels such as “OpenAI / Oracle / SoftBank” are one ledger label, not three independently audited entities.
Categories	13	accelerator_platform 4; ai_cloud 11; datacenter_physical 22; frontier_lab 65; hardware_supply 1; hyperscaler 17; market_signal 25; policy_regulation 14; serving_system 45; standard 3; supply_chain 31; workload_model 13; workload_trace 21.
Source kinds used	54	Machine-derived leaders are official_release 55, research_paper 28, preprint 18, credible_reporting 16, official_earnings 16, official_model_release 15, official_repository 14, official_product_release 12, official_engineering_release 11, technical_report 11, and official_documentation 9; 67 entries use 43 other bounded source-kind labels.
Evidence classes used	48	Machine-derived counts now include official_statement 100, vendor_claim 31, benchmark_measurement 27, reported_observation 15, production_measurement 12, vendor_specification 10, production_observation 8, synthetic_experiment 6, and 63 entries across 40 other bounded lifecycle/benchmark/specification labels.
Confidence labels	16	Exact labels include high 189, medium_high 57, medium 11, low 3, and 12 source-bounded qualified labels used once each. Confidence describes evidentiary strength, not business likelihood.
Date fields	272/272 published; 272/272 event	Presence is complete. Date precision is source-bounded: the XPK run exposes only November 2023 for the event, while its publication date is exact.
Explicit rumors	3	All are low-confidence and carry state, last-check, expiry, corroboration, and fragment-level resolution metadata; final outcomes remain open.

Structural checks

Check	Status	Evidence
Required fields present	Complete	All 272 entries contain the schema’s required fields.
JSON parseability	Complete	python3 -m json.tool is the local validation command.
Unique-entry semantics	Partial	IDs are intended to be unique and URLs are counted, but no committed schema/link checker enforces the contract yet.
Source-class separation	Complete for current entries	The ledger keeps production, benchmark/synthetic, official, vendor, analyst/reported, and rumor classes distinct.
Claim contradiction handling	Partial	contradiction-matrix.md defines normalization; project-level reproductions and automated clustering remain open.
Refreshability	Partial	refresh-protocol.md defines manual refresh; scheduled validation and link checking are absent.

Requirement-by-requirement coverage

1. Frontier labs and regions — Partial

Present: OpenAI, Anthropic, xAI, Google DeepMind, Meta, Amazon Nova, Apple, Cohere, Ai2, Microsoft Phi, NVIDIA/Nemotron, DeepSeek, Alibaba/Qwen, Moonshot, MiniMax, ByteDance Seed, Baidu, Tencent, Z.ai/Zhipu, Huawei Pangu, 01.AI, StepFun, Shanghai AI Lab/InternLM, SenseTime, Xiaomi MiMo, Mistral, AI21, TII, MBZUAI/G42 IFM, SDAIA/HUMAIN, Sarvam, Sakana, NTT, LG AI Research, Samsung, SK Telecom, Kakao, NAVER Cloud, AI Singapore, and Sea AI Lab. The census spans the U.S./Canada, China, Europe, Israel, the Middle East, India, Japan, Korea, and Southeast Asia and records model/serving statements without treating plans or releases as delivered capacity.

Missing or shallow: Microsoft first-party traffic, Huawei Ascend physical evidence, iFLYTEK hardware/Ascend operations, G42/Inception Jais, NTT/Preferred Networks/Fujitsu/SoftBank, Grab/GoTo/SCB10X/VinAI, Europe beyond Mistral/DeepMind, and many Canadian/private labs. Most checked labs—including the newly added NVIDIA Nemotron, 01.AI, StepFun, Shanghai AI Lab, SenseTime, Xiaomi, Samsung, SKT, Kakao, MBZUAI, and Saudi rows—still lack physical serving fleets, traffic distributions, service health, and lifecycle resolution.

Proof needed for complete: a declared entity/region universe, at least one current primary source per entity, model and infrastructure lifecycle fields, and dated checks for launches, cancellations, partnerships, and regional constraints.

2. Hyperscalers, clouds, and AI clouds — Partial

Present: AWS, Google/Alphabet, Microsoft/Azure, Meta, Alibaba Cloud, CoreWeave, Nebius, and selected partner/cloud deployments. Official filings and earnings evidence for Alphabet, Microsoft, Meta, Amazon, Oracle, CoreWeave, Alibaba, Nebius, and Baidu are extracted in filings-ledger.md.

Missing or shallow: IBM, Tencent, and sovereign-cloud financial normalization; full annual-report normalization remains incomplete for Oracle, CoreWeave, Nebius, Alibaba, and Baidu; customer-supplied hardware; partner prepayments; finance and operating leases; purchase obligations; depreciation; backlog quality; customer concentration; AI versus non-AI capex.

Proof needed for complete: issuer-by-issuer filings with identical field definitions, fiscal-calendar normalization, lifecycle linkage from obligation to installed capacity, and coverage of non-U.S. and sovereign clouds.

3. Datacenter power, grid, cooling, water, construction, and permitting — Partial

Present: IEA demand/grid evidence, U.S. project-pipeline and delay reporting, selected utilities/site announcements, NVIDIA DSX power architecture, GE Vernova/Crusoe ordered generation, a phased Siemens/Start Campus site record, and a supply-chain lifecycle framework. Power is treated as a gating resource rather than an afterthought.

Present but incomplete: Schneider/NVIDIA reference design, Trane/LiquidStack CDU availability and validation, Johnson Controls backlog, and earlier Vertiv/site records now separate design, product, order, and live-site states.

Present but incomplete: Eaton, Hitachi Energy, Siemens Energy, and GE Vernova now add dated transformer/grid manufacturing investments and backlog boundaries across the U.S. and India.

Missing or shallow: named operator/site census; utility interconnection queues; deliverable versus requested MW; product-specific factory output, transformer/switchgear shipments, substations; PUE and curtailment; water source/consumption; cooling-vendor and heat-rejection data; construction labor; permits; community opposition; cancellations; commissioning and acceptance dates. The disputed “half delayed” claim remains unverified at project level.

Proof needed for complete: site-level records from announcement through operation, with power-boundary labels, vendors, dates, MW, water/cooling method, permit status, and subsequent delay/cancellation resolution.

4. HBM, advanced packaging, optics, network, and storage — Partial

Present: selected NVIDIA/HBM, SK hynix HBM4/HBM4E, TSMC CoWoS, optical-interconnect, networking, and platform announcements; architectural recognition that accelerator availability alone does not set cluster capacity.

Present but incomplete: VAST/CoreWeave DPU data-plane deployment claims, WEKA storage/memory GA, Arista 1.6T product availability, and XPO specification/density evidence now bound storage and network lifecycle states.

Missing or shallow: supplier-by-supplier HBM output and allocation, CoWoS/advanced packaging capacity, substrates, optics/transceivers, switch silicon, cables, NICs, object/block storage shipments and production checkpoint paths, lead times, yields, shipment versus installed states, and China/regional supply chains. Electrical equipment is similarly sparse.

Proof needed for complete: component/vendor ledgers with physical units, time windows, order/shipment/install states, dependencies, and independent production or shipment evidence rather than only vendor roadmaps.

5. Batching and scheduling — Partial

Present: serving-system papers and releases cover scheduling, continuous/dynamic batching assumptions, queueing, and several workload traces. The corpus recognizes that batch opportunities depend on arrival, lengths, SLOs, hardware, and tenant policy.

Present but incomplete: phase/operator autoscaling, token-work signals, hybrid aggregated/disaggregated scheduling, production-scale heterogeneous coordination, and Google DWS Flex-start versus reservation-bound admission now have explicit evidence envelopes. OpScale adds exact prototype control intervals, scale-up latency, average GPU counts, TTFT-bound SLO attainment, topology, and SLO-constrained power/input-TPS results, but not the production-relevant queue/batch/utilization/operator-count fields below.

Missing or shallow: comparable production distributions for batch size, queue wait, SLO class, cancellation, priority, fairness, admission, and cross-tenant interference; policy prevalence by lab/cloud; DWS admission probability and wait distributions; scheduler behavior during failures and regional bursts. OpScale specifically leaves achieved active batch, numeric queue depth/wait, achieved GPU utilization, realized operator-replica placement/counts, and failure/retry reactions undisclosed.

Proof needed for complete: production traces or operator measurements with batch and queue fields, stratified by model, tenant, hardware, region, and SLO.

6. Prefill/decode disaggregation — Partial

Present: llm-d/Google, NVIDIA Dynamo, FineServe, and related system evidence make prefill/decode separation, topology, and KV transfer explicit product surfaces.

Present but incomplete: Dynamo/llm-d product surfaces, TaiChi SLO-dependent architecture results, TokenScale burst response, and HeteroScale production-scale coordination.

Missing or shallow: installed production share, transfer sizes, topology-specific break-even curves, failure domains, decode imbalance, WAN/regional use, and matched end-to-end comparisons that count orchestration and transfer overhead.

Proof needed for complete: neutral production evidence over representative prompt and output distributions, with quality, SLO, utilization, transfer, and failure accounting.

7. KV cache, prefix reuse, routing, and offload — Partial

Present: Mnemosyne, CacheRoute, Tair KVCache/HiSim, multi-tenant admission studies, robust cache work, and the Copilot trace’s approximately 90% average prefix-cached token share within sessions. The corpus distinguishes benchmark/synthetic results from production traces.

Missing or shallow: cross-tenant production prefix-popularity fits, reuse-distance and object-lifetime distributions, cache-hit opportunity by application, privacy boundaries, invalidation, fragmentation, offload bandwidth, routing overhead, and drift.

Proof needed for complete: raw production distributions with tenant weighting and cache policy, plus end-to-end hit/goodput measurements under realistic churn.

8. Autoscaling, placement, resilience, and fairness — Partial

Present: cluster-scale reliability research, routing/scheduling papers, and system releases expose topology-aware placement and failure/retry costs. Cluster Director now adds a host/rack-sub-block/block/cluster GPU hierarchy, current GKE Multislice adds atomic homogeneous multi-host slice scheduling, and the Copilot trace reports 9% of turns with tool failure and retry loops reaching 4× compute in the captured population.

Missing or shallow: production cold-start and scale-up times, capacity headroom, regional failover, heterogeneous accelerator placement, maintenance/health attrition, preemption, tenant fairness, quota enforcement, noisy-neighbor distributions, and installed-to-schedulable-to-goodput conversion. Google’s current pages also leave the maximum GKE slices per JobSet, exhaustive region/stage matrix, and production queue/utilization/failure/cost envelopes undisclosed.

Present but incomplete: workflow-DAG, agentic-OS, and tool-result-cache research now bounds request-level assumptions; Copilot/TraceLab supply production session, idle, tool, and retry evidence.

Proof needed for complete: operator traces linking demand, workflow DAGs, scaling decisions, health, placement, sandbox/tool queues, retries, SLOs, and per-tenant outcomes over failures and seasonal peaks.

9. User, tenant, geography, and seasonality distributions — Partial to missing

Present: the GitHub Copilot production corpus provides 3.2M users, 13.5M sessions, 95.1M turns, 760.5M LLM calls, 774.7M tool calls, 44.9T prompt tokens, 39.3B completion tokens, five user archetypes, and compaction/failure statistics. Other trace papers provide request/length/burst observations.

Present but incomplete: ServeGen identifies 29 dynamic top clients among 2,412 profiled clients; FineServe quantifies second-scale burst concentration; Chutes tracks one-year model/user evolution; Aliyun, Chutes, and Copilot expose distinct cache-reuse shapes. These are provider-specific observations, not universal tenant laws.

Present but incomplete: Azure OpenAI / BurstGPT adds 10.31M requests over 213 days, daily and weekly periodicity, long-tailed tokens, and case-study bursts at 400 req/s for 10 seconds, 100 req/s for 60 seconds, and 15 req/s for 600 seconds. Splitwise adds one day of empirical Conversation and Coding token/rate distributions. Four services are not all Azure traffic, the burst examples are not quantiles, and one-day samples cannot establish universal tenant, geography, session, or drift parameters.

Present but incomplete: SkyWalker adds six-country diurnal phase evidence; SageServe adds a >8M-request, 3-region mixed-SLO envelope; SMetric and Continuum add session-locality and tool-gap state. These are study-specific, not universal provider distributions.

Missing: exact tenant traffic shares; provider production geography/timezone weights; diurnal and weekly coefficients; launch/event burst parameters; paid/API/consumer segmentation; reasoning-mode selection; retry distributions beyond coding agents; speculative-token acceptance; cancellation; and cache opportunity by user cohort. User/adoption totals from labs do not fill these gaps.

Proof needed for complete: anonymized multi-product production traces with explicit population, interval, geography, tenant weighting, concurrency, and token accounting.

10. Distribution-family and parameter claims — Partial

Present: output-length evidence reports skewness 3.10, mean coefficient of variation 1.09, CV above 1 for 78.6% of examined workloads, top-decile share 35.7%, P90/P50 4.62, and P99/P50 10.77. BurstGPT and other studies cover burstiness/arrival modeling. workload-parameters.md records exact parameters available from selected sources.

Unverified: any universal distribution law. Current evidence supports category-dependent, heavy-tailed, bursty, multimodal, and nonstationary behavior. It does not prove one Zipf, lognormal, Pareto, Poisson, Hawkes, or MMPP law for users, tenants, prompts, prefixes, outputs, and arrivals collectively.

Missing: fitted parameters, estimator, goodness-of-fit, confidence interval, sample population, and drift window for every cited workload paper and every random variable.

Proof needed for complete: variable-specific parameter extraction and reproducible fit checks; “unknown” must remain a valid result.

11. Startups, launches, releases, and partnerships — Partial

Present: inference/serving startups, AI clouds, accelerator partnerships, financing, product launches, and selected acquisitions/partnerships appear in startups-landscape.md and market-chronology.md.

Present: Untether AI supplies a shutdown, engineering-team transfer, support termination, and bankruptcy case. Replicate, SchedMD, Groq, and Celestial AI supply platform acquisition, scheduler acquisition, IP-license/talent-transfer, and photonic-startup acquisition cases. Lambda, Applied Digital, Lightmatter, and Fireworks add neocloud, power-first site, optical-interconnect, and inference-platform lifecycle evidence.

Missing or biased: the sample remains launch- and funding-heavy. More failures, shutdowns, acquihires, distressed financing, deployment cancellations, missed roadmaps, customer concentration, revenue/profitability, and category-complete coverage across routers, KV systems, power, cooling, networking, and modular datacenters are sparse.

Proof needed for complete: a declared startup universe with founded/launch/funding/ deployment/acquisition/failure states and periodic resolution checks.

12. Rumors and resolution history — Unverified

Present: 3 open rumor entries are explicitly labeled and kept out of factual capacity totals. The OpenAI personnel-departure entry moved to reported observation after primary spokesperson confirmation, while its strategic interpretation remains bounded.

Present: each open rumor has a state, last-check date, expiry, corroboration note, and fragment-level resolution. Anthropic–Decart talks are independently corroborated but unclosed; the NVIDIA price direction is partially corroborated while magnitude/scope remain unverified; and the NVIDIA–Hugging Face record preserves conflicting talks/agreement reports without inferring signing, terms, regulatory clearance, or close.

Missing: complete original-source lineage, circular-republication detection, independent corroboration graph, claim-fragment matching, and later confirmed/refuted/partially-confirmed outcomes.

Proof needed for complete: a rumor state machine and dated resolution history. Until then, rumors are watch items only.

13. News and market chronology — Partial

Present: completed events, future announcements, partnerships, funding, and rumors are separated in market-chronology.md. Publication and event dates are present on every current entry.

Missing: systematic global media/source watch, change detection, corrections, archived copies, cancellation/failure follow-up, and resolution of future plans after their target date.

Proof needed for complete: a refresh cadence with source inventories, link/archive health, plan-expiry checks, and additive snapshots rather than silent rewrites.

14. Standards, regulation, export controls, and sovereign AI — Missing to partial

Present: dated BIS rule-lifecycle evidence; IndiaAI and EU sovereign-compute programs; Singapore datacenter-capacity policy; EU AI Act application dates; and Ultra Ethernet specification history. policy-standards-ledger.md separates proposals, enforcement, tenders, allocations, delivery, and operating capacity.

Missing: the operative U.S. replacement-rule matrix; comprehensive sanctions and transaction-level controls; China domestic policy/substitution; sovereign procurement beyond the initial EU/India cases; energy/water/permitting rules by jurisdiction; processor/site awards and delivered capacity; and adopted serving/KV/benchmark standards.

Proof needed for complete: jurisdiction-by-jurisdiction primary sources with effective dates, affected hardware/services, implementation status, and later amendments.

Coverage by source strength

Evidence tier	Current condition	Consequence
Production measurement/observation	20 entries; valuable but narrow	Strongest demand evidence is concentrated in coding-agent and selected serving workloads. Do not universalize it.
Benchmark/synthetic	33 entries	Useful for mechanism and break-even hypotheses; not proof of installed production prevalence.
Official statements	99 entries	Strong for what an entity said or filed, not for future delivery or neutral performance.
Vendor claims	31 entries	Retain exact envelope and reproduce before using as a fak gain claim.
Analyst/reported evidence	21 entries	Useful for market/site visibility; denominators and original datasets require checking.
Rumor	3 entries	Watch-only until primary evidence resolves the claimed lifecycle state.

Remote-browser execution-envelope slice (#9375)

Bounded coverage added: five workload_trace records from the official Steel BrowserBench repository pinned at commit 847e4ed604764ce8b887265709ddb2d1c3c5f020 (2026-01-15). The records cover the committed 5,000-attempt samples for Steel, Kernel, Browserbase, Hyperbrowser, and AnchorBrowser, including exact successful-sample counts/rates, mean create, CDP-connect, DOMContentLoaded-navigation, release, and total lifecycle latency, plus total median/p95/p99. Provider SDK versions and actual result-file test windows remain separate.

Method boundary: the README is a benchmark summary reproducible from committed JSONL, not production telemetry. The runner executes sequential lifecycle attempts from AWS EC2 us-east-1, excludes 10 warm-ups/provider, retries Kernel create after 429s, and gives Browserbase an approximately three-second full-cycle floor. Consequently, run rate does not measure concurrency, queue depth, or throughput. Create/start is not browser ready; browser ready is not CDP connected; CDP connected is not page navigated; page navigated is not task complete; task complete is not release. Successful samples are not users, sessions, tasks, actions, observations, or tool calls. Mean stage values are not stage p50/p95/p99 or failure distributions.

Known data anomaly retained: the Kernel result file marks all 5,000 attempts successful, but one successful row has session_release_ms: null; only 4,999 successful rows contain all four stage values. The README’s rounded stage and total summaries are indexed as published rather than silently repairing that row.

Unretired gap: hosted remote-browser lifecycle data does not characterize desktop-computer trajectories. Production action/tool-call distributions, observation bytes, resets, full session duration, escalations, user populations, and retry/failure/tail distributions remain unavailable in the reviewed primary sources.

Preserved explicit gaps

Production-distribution parameterization added in issue #9340

• Three existing records were deepened: FineServe (23 days, 29 models, 9 tasks), ServeGen (text, multimodal, and reasoning), and Chutes (365 days, 6.122B requests, 314,970 users, 9,174 models).
• The parameter matrix records population, window, estimator, empirical family or candidate families, parameters, selection quality, confidence-interval status, drift, and missing fields.
• Request share is separated from token share, exact-request hits from prefix-token hits, aggregate arrivals from per-client processes, and text from multimodal, reasoning, and agentic workloads.
• No universal Zipf, log-normal, Pareto, Poisson, Hawkes, or MMPP law is supported. FineServe’s selected arrival family varies by workload, while ServeGen and Chutes expose empirical heterogeneity and drift.

The machine-readable coverage.explicit_gaps remains authoritative. In plain language, the open work is:

1. entity-complete global frontier-lab and sovereign-program coverage;
2. direct provider geography/timezone and tenant concentration, plus retry/failure, speculative-acceptance, stop-cause, session-length, and non-coding tool-call distributions;
3. comparable denominators, fit diagnostics, population confidence intervals, and drift extraction from every workload source; figure-read values remain approximate;
4. normalized hyperscaler/cloud filings and contractual obligations;
5. named physical-infrastructure and component lifecycle ledgers;
6. startup failures, acquisitions, cancellations, and delivered economics—not launches alone;
7. China/export-control/regional supply-chain coverage;
8. rumor provenance and resolution graphs; and
9. committed schema/link/contradiction validation and scheduled refresh.

Next audit actions

1. Complete the frontier-lab census by region, beginning with the unchecked high-scale China and hyperscaler labs.
2. Expand filings to Oracle, CoreWeave, Nebius, IBM, Alibaba, Tencent, and Baidu with normalized obligations and leases.
3. Build a named site/vendor lifecycle ledger for grid, power equipment, cooling/water, HBM/packaging, optics/network, storage, construction, and permitting.
4. Extract production-distribution parameters and explicitly test—not assume—Zipf, lognormal, Pareto, and MMPP fits per variable.
5. Add failure/acquisition/cancellation and rumor-resolution histories.
6. Only after the schema stabilizes, consider a committed Go validation leaf for field, uniqueness, link, lifecycle, and coverage checks.

   Remote-browser operational concurrency and boundary slice (#9376)

This bounded refresh adds five official-documentation records to index.json:

• browserops-steel-plan-envelope-2026 — plan concurrency, request-rate, and max-session quantities;
• browserops-browserbase-admission-envelope-2026 — distinct plan concurrency and per-minute creation limits, HTTP 429 admission refusal, configurable session duration, six-hour maximum, and separate 10-minute CDP inactivity timeout;
• browserops-kernel-idle-and-pool-envelope-2026 — standby timeout range and pool long-poll timeout behavior;
• browserops-hyperbrowser-configuration-lifecycle-example-2026 — a 60-minute timeout configuration example, with no plan or production-population inference;
• browserops-anchor-timeout-envelope-2026 — independent idle and maximum-duration timers.

Boundary coverage

Boundary	Steel	Browserbase	Hyperbrowser	Kernel	Anchor Browser
numeric documented concurrency allowance	yes	Free 3; Developer 25; Startup 100; Scale 250+ active browsers	not found	organization/project mechanism documented, quantity not public in reviewed page	not found
explicit admission/queue behavior	not found	HTTP 429 when either active concurrency or per-minute creation limit is reached; over-limit creation is dropped, not queued	not found	pool acquire long-poll; HTTP 204 on poll timeout	not found
numeric request-rate boundary	60/min Launch; 600/min Scale	session creations/min: Free 5; Developer 25; Startup 50; Scale 150+	not found	not found	not found
numeric lifecycle timeout	15 min / 1 h / up to 24 h by plan	configurable project default; 6 h maximum session duration; separate 10 min CDP inactivity timeout	60 min configuration example only; default and maximum not found	60 s standby default; up to 72 h	5 min idle default; 180 min hard default

All pages were accessed 2026-08-27. Steel’s page states Last Edit: June 30th, 2026; no stable release or repository commit was exposed for the other mutable documentation pages.

Unretired gaps: no reviewed official source supplies observed provider concurrency, throughput, production occupancy, or task-duration distributions. Public numeric queue depth, queue wait distribution, and universal concurrency quantities remain absent for most providers. These absences are retained as gaps rather than inferred from pricing, API rate limits, timeout settings, or anecdotal usage.

Frontier serving-configuration slice added in issue #9382

Five primary-source MLPerf records were added across DeepSeek-R1, Llama 3.1 405B, and Qwen3-VL-235B-A22B; NVIDIA B200/B300, AMD MI355X, and Red Hat’s B200 vLLM submission; Server and Interactive scenarios; and tokens/s versus queries/s result units. The NVIDIA configuration files pin TP/PP/EP and configured batching where exposed. AMD and Red Hat records deliberately leave those fields unknown because their published result/config files do not expose them.

Remaining gaps: replica count is unstated in all five records; numeric latency/SLO thresholds are not present in the selected submission files; two records omit runtime batch, concurrency, and sequence envelopes; none is production topology or prevalence evidence; no achieved-active-batch distribution is available.