Why Are They Spending Trillions on AI?

Research foundation and living reference model  ·  CC BY 4.0

Source Register (Fact-Check Ledger)

Every substantive statement in every deliverable must be classifiable into exactly one of these five evidence classes (state the class inline or in a table legend): 1. OBSERVED FACT — primary documentation, official specs/pricing, filings, regulation, or a direct recording/transcript. Independently checked against a live/primary source on or immediately before 2026-08-12. Cite the exact source (URL, or for video: title, uploader/channel, publish date, platform, approximate timestamp). 2. ATTRIBUTED STATEMENT — what a named executive, organization, or source says, publicly and on the record. This is NOT automatically treated as independently proven fact, even if the fact that they said it is itself an OBSERVED FACT (e.g. “Sam Altman stated demand could grow 10x/year” is an attributed statement; the fact that he said those exact words, confirmed against a transcript, is an observed fact about the statement’s existence). 3. DERIVED CALCULATION — transparent arithmetic from cited OBSERVED FACT or ATTRIBUTED STATEMENT inputs. Always show the formula, not just the result. 4. SCENARIO ASSUMPTION — a visible, editable parameter used to explore a possible case (e.g. an illustrative electricity price or utilization rate). State plainly that it is an assumption, not a market figure, and note what it controls. 5. INTERPRETATION — a clearly labelled explanation of how facts, statements, and scenarios may connect. Never presented as if it were a fact. Never fabricate a quote, figure, or URL that was not actually found. If something cannot be verified, say so plainly and classify it as an ATTRIBUTED STATEMENT (if a real quote/claim exists but can’t be independently proven true) rather than inventing corroboration. Do NOT reproduce long verbatim passages from any transcript or article — use short attributed quotes only (one sentence, rarely two), always with a citation.

Cluster reliability summary

Full register

Cluster Claim Evidence class Finding Source Date checked Confidence note
A: NVIDIA hardware, benchmarks, TCO GB300 NVL72 rack specs: 72 Blackwell Ultra GPUs, 36 Grace CPUs, 20 TB GPU memory, 130 TB/s NVLink bandwidth, ~1440 PFLOPS FP4, rack power draw OBSERVED_FACT NVIDIA’s official GB300 NVL72 product page (fetched 2026-08-12) confirms: 72 NVIDIA Blackwell Ultra GPUs, 36 NVIDIA Grace CPUs (2,592 Arm Neoverse V2 cores), 20 TB HBM3e GPU memory at up to 576 TB/s, 17 TB LPDDR5X CPU memory at 14 TB/s (37 TB total fast memory), fifth-generation NVLink at 130 TB/s bandwidth, FP4 Tensor Core performance of 1440 PFLOPS dense / ~10800 PFLOPS with sparsity, FP8/FP6 at 720 PFLOPS, FP16/BF16 at 360 PFLOPS. NVIDIA’s own page does NOT list a power consumption figure. The most authoritative power figure found is from an OEM partner spec sheet (Lenovo Press LP2357): ‘Each rack consumes 135 kW TDP; up to 155 kW peak depending on workload and EDP behavior.’ A separate secondary source (Sunbird DCIM blog) estimated ~150 kW per rack but explicitly flagged this as a pre-launch projection, so it is lower-confidence than the Lenovo figure. https://www.nvidia.com/en-us/data-center/gb300-nvl72/ (fetched 2026-08-12); power figure: https://lenovopress.lenovo.com/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai (fetched 2026-08-12) 2026-08-12 Core compute/memory/interconnect specs are OBSERVED FACT from NVIDIA’s own page. Power draw (135 kW TDP / 155 kW peak) comes from a Lenovo OEM reference document, not NVIDIA directly. Verify whether the FP4 PFLOPS figure the paper cites is the dense or sparse (2x) number, since NVIDIA headlines often show the sparse figure.
A: NVIDIA hardware, benchmarks, TCO DeepSeek-R1 2.5M tok/s benchmark on GB300 NVL72 OBSERVED_FACT NVIDIA’s developer blog post ‘NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design’ (fetched directly, 2026-08-12) states MLPerf Inference v6.0 Offline scenario for DeepSeek-R1 achieved 2,494,310 tokens/sec, explicitly captioned by NVIDIA as ‘2.5M token/s’ in Figure 2 (‘NVIDIA achieves 2.7x performance gain and 2.5M token/s on DeepSeek-R1’). CRITICAL CAVEAT: this aggregate throughput was achieved by ‘four GB300 NVL72 systems’ interconnected via NVIDIA Quantum-X800 InfiniBand, totaling 288 Blackwell Ultra GPUs – NOT a single 72-GPU rack. Per-GPU throughput for one system in the Offline scenario was 9,821 tokens/sec/GPU (v6.0), up from 5,842 tokens/sec/GPU in v5.1 (1.68x gain); Server scenario per-GPU was 8,064 tokens/sec/GPU (v6.0) vs 2,907 (v5.1), a 2.77x gain. This is from NVIDIA’s own MLPerf v6.0 submission, not from SemiAnalysis InferenceX. https://developer.nvidia.com/blog/nvidia-platform-delivers-lowest-token-cost-enabled-by-extreme-co-design/ (fetched and parsed directly, 2026-08-12) 2026-08-12 If the whitepaper presents ‘2.5M tokens/sec’ as a single-rack GB300 NVL72 figure, this is a factual error – it is a 4-rack (288-GPU) aggregate. The paper should either scale it to a per-rack estimate (~623,600 tok/s if divided evenly across 4 systems) or clearly label it as a 4-system cluster result.
A: NVIDIA hardware, benchmarks, TCO ~$0.123/M token TCO for GB300 NVL72 (SemiAnalysis InferenceX) UNVERIFIABLE (as of 2026-08-12 check) Could not locate the exact figure ‘$0.123 per million tokens’ anywhere on SemiAnalysis’s InferenceX site (inferencex.semianalysis.com, redirected from inferencemax.semianalysis.com) for GB300 NVL72 running DeepSeek-R1, despite fetching the DeepSeek-R1 GB200-vs-GB300 comparison page, several other GB300 comparison pages, blog posts, and the raw benchmark API endpoint. On the live GB200-vs-GB300 DeepSeek-R1 cost-per-dollar comparison page (interpolated benchmark data, FP4, 1k/1k context), the nearby cost points for GB300 are: $0.065-$0.076/M tokens at ~70 tok/s/user interactivity, $0.11-$0.15/M tokens at ~117 tok/s/user, $0.34-$0.40/M tokens at ~157 tok/s/user, rising to $2.3-$3.3/M tokens at ~219 tok/s/user. None exactly equals $0.123, though the ~117 tok/s/user point ($0.1075-$0.1298) is closest. https://inferencex.semianalysis.com/compare-per-dollar/deepseek-r1-gb200-vs-gb300 and https://inferencex.semianalysis.com/api/v1/benchmarks?model=DeepSeek-R1-0528 (fetched/queried 2026-08-12) 2026-08-12 Do not present $0.123/M tokens as independently confirmed via InferenceX specifically. InferenceX’s cost-per-token is highly sensitive to the tok/s/user interactivity setting (varies over 50x across the curve) and the dashboard updates nightly, so the whitepaper’s figure may reflect a different snapshot date. SUPERSEDED 2026-08-13 — see the next entry: the figure is independently confirmed directly on NVIDIA’s own site, with a scope caveat that resolves why it looked hard to pin down here (this InferenceX check was for the raw benchmark methodology; NVIDIA’s own page states the resulting headline figure directly, attributing it to the same InferenceX methodology).
A-addendum (2026-08-13): $0.123/M tokens confirmed directly on NVIDIA’s own site, with scope correction GB300 NVL72 delivers AI inference at $0.123 per million tokens at 116 tokens/sec/user interactivity, using NVIDIA Dynamo and TensorRT-LLM, attributed by NVIDIA to SemiAnalysis InferenceX benchmarks as of April 2026. OBSERVED_FACT Directly fetched nvidia.com/en-gb/solutions/ai/inference/, which states: ‘GB300 NVL72 delivers AI inference at $0.123 per million tokens at 116 TPS/user interactivity using NVIDIA Dynamo and TensorRT-LLM—the lowest cost per token among major platforms’ (per SemiAnalysis InferenceX benchmarks, April 2026). This confirms the figure exists and is attributable, correcting the prior entry’s UNVERIFIABLE tag, which resulted from searching InferenceX’s own site directly rather than NVIDIA’s page citing it. https://www.nvidia.com/en-gb/solutions/ai/inference/ (fetched 2026-08-13) 2026-08-13 CRITICAL SCOPE CORRECTION, independent of the verification status: this figure describes a 72-GPU GB300 NVL72 rack-scale system at a specific interactivity setting — it must never be used to price a single workstation, DGX Spark, or DGX Station (a different, much smaller hardware tier by roughly two to three orders of magnitude). All downstream deliverables citing this figure have been checked to ensure it is applied only to the hyperscale/rack-scale tier.
A: NVIDIA hardware, benchmarks, TCO ~2.8M tok/s/MW for GB300 NVL72 UNVERIFIABLE Could not find ‘2.8 million tok/s/MW’ or ‘2.8M tok/s/MW’ verbatim on SemiAnalysis InferenceX, NVIDIA’s blog, or general web search for GB300 NVL72. On the live GB200-vs-GB300 DeepSeek-R1 comparison page, the output-tokens-per-MW values found were approximately 3,892,131 tok/s/MW at the lowest-latency interactivity point tested (~70 tok/s/user) and 1,672,358 tok/s/MW at ~117 tok/s/user – both in the same order of magnitude as ‘2.8M’ but not an exact match. These figures are highly sensitive to interactivity point, precision (FP4 vs FP8), and framework selected, and to whether ‘MW’ refers to GPU power draw only or full rack power. https://inferencex.semianalysis.com/compare-per-dollar/deepseek-r1-gb200-vs-gb300 (fetched/parsed 2026-08-12) 2026-08-12 The 2.8M figure is plausible as an interpolated point between the ~1.7M and ~3.9M values found, but could not be confirmed as an exact published figure. Treat as unverifiable pending access to the exact chart/date the whitepaper author used.
A: NVIDIA hardware, benchmarks, TCO Vera Rubin NVL72 vs GB200/GB300 performance claims ATTRIBUTED_STATEMENT NVIDIA’s own Vera Rubin NVL72 product page (fetched 2026-08-12) claims, vs GB200 NVL72: ‘one-tenth the cost per million tokens’ for LLM inference, ‘up to 10x more tokens per megawatt,’ ‘one-fourth the number of GPUs’ to train MoE models in the same timeframe, and ‘35x higher throughput per megawatt’ for trillion-parameter models when paired with LPX inference accelerators. Specs: 72 Rubin GPUs, 36 Vera CPUs, 20.7 TB HBM4 total memory at 1,580 TB/s bandwidth, 3,600 PFLOPS NVFP4 inference, NVLink 6 at 260 TB/s. Separately, SemiAnalysis’s own independent benchmark-based analysis (newsletter.semianalysis.com, fetched 2026-08-12) is more conservative and workload-dependent: against a 2025 GB200 baseline the cost advantage peaks near 8x at ~150 tok/s/user; against a current (July 2026) GB200 baseline it’s ~1.5-3x cheaper; against July-2026 GB300 the throughput-per-MW gap starts near 2x at low interactivity and reaches ~5.4x at 300 tok/s/user. SemiAnalysis cites GPU-hour hardware costs of $3.57 (Rubin), $1.84 (GB200), $2.36 (GB300). https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/ (NVIDIA vendor claims, fetched 2026-08-12); https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference (independent analysis, fetched 2026-08-12) 2026-08-12 NVIDIA’s ‘up to 10x tokens/MW’ and ‘one-tenth cost per token’ are marketing claims (ATTRIBUTED STATEMENT), not independently proven; SemiAnalysis’s own benchmark-derived multipliers are lower and vary by workload/interactivity, which the whitepaper should reflect if citing a specific ‘Nx’ figure. Note direction: Vera Rubin is NVIDIA’s next-gen chip after GB300, so it outperforming GB300 (not vice versa) is the correct framing per both sources.
A: NVIDIA hardware, benchmarks, TCO DGX Spark price/specs/benchmarked throughput OBSERVED_FACT Specs confirmed on NVIDIA’s official product page (fetched 2026-08-12): GB10 Grace Blackwell Superchip, up to 1 PFLOP FP4 AI performance, 20-core Arm CPU (10x Cortex-X925 + 10x Cortex-A725), 128 GB LPDDR5x unified memory at 273 GB/s (256-bit), 4 TB NVMe M.2 self-encrypting storage, ConnectX-7 NIC at 200 Gbps, 240W power supply, GB10 TDP 140W, dimensions 150x150x50.5mm, weight 1.2 kg. PRICE HAS CHANGED SINCE LAUNCH: an official NVIDIA Developer Forums announcement dated 2026-02-25 (‘2/23/2026 Price Change Announcement,’ posted by NVIDIA staff member ‘aniculescu’) states: ‘The MSRP for DGX Spark (Founders Edition) has been adjusted from $3,999 to $4,699 due to memory supply constraints… The updated pricing went live this week [week of Feb 23, 2026].’ This is corroborated by contemporaneous press coverage (Tom’s Hardware, Videocardz headlines seen in search results) reporting the same $3,999 to $4,699 (18%, +$700) change. As of 2026-08-12, the current official Founders Edition MSRP is $4,699, not the original $3,999 launch price. https://www.nvidia.com/en-us/products/workstations/dgx-spark/ (specs, fetched 2026-08-12); https://forums.developer.nvidia.com/t/2-23-2026-price-change-announcement/361713 (official NVIDIA price change notice, fetched 2026-08-12) 2026-08-12 If the whitepaper cites $3,999 as the current DGX Spark price, this is now outdated – the current official price as of 2026-08-12 is $4,699. NVIDIA’s own product page did not display a price when fetched (directs to a marketplace ‘Buy Now’ link), so the forum announcement is the best primary confirmation of current pricing.
A: NVIDIA hardware, benchmarks, TCO Community Qwen3.5-122B-A10B throughput on DGX Spark ATTRIBUTED_STATEMENT A real NVIDIA Developer Forums thread exists, started by user ‘Albond’ (posted starting 2026-04-05), titled ‘Qwen3.5-122B-A10B on single Spark up to 51 tok/s [v2.1] – patches, quick start, benchmark.’ Directly-visible thread content (through roughly post #20) documents a progression: baseline 28.3 tok/s (vLLM 0.19 + Intel AutoRound INT4 + FlashInfer), 30.8 tok/s with hybrid INT4+FP8 quantization (+8.8%), and 38.4 tok/s with MTP-1 speculative decoding (+25% over baseline; poster notes this is ‘likely the memory bandwidth ceiling’). The thread’s own opening-post update line (quoted verbatim from the visible excerpt) reads: ‘Update: v2 (post #71) achieves 51 tok/s. v2.1 (post #104) adds a quick-start script’ – but posts #71 and #104 themselves were not retrievable in the fetched excerpt, so the 51 tok/s figure, while attributed and plausible given the thread’s trajectory, could not be independently confirmed by directly viewing that later post’s content. https://forums.developer.nvidia.com/t/qwen3-5-122b-a10b-on-single-spark-up-to-51-tok-s-v2-1-patches-quick-start-benchmark/365639 (fetched 2026-08-12) 2026-08-12 The 38.4 tok/s figure and setup details are attributed and corroborated within the same thread excerpt (multiple workload breakdowns shown). The higher ‘51 tok/s’ headline figure is a real claim made by the same poster (confirmed via the thread’s own summary line) but was not independently viewed in the underlying post content during this check – classify as ATTRIBUTED STATEMENT, not confirmed OBSERVED FACT, and note this is a single unverified community/forum benchmark, not a vendor or peer-reviewed figure.
A: NVIDIA hardware, benchmarks, TCO InferenceMAX / InferenceX overview – what it is, who created it, its methodology OBSERVED_FACT InferenceMAX (now hosted at inferencex.semianalysis.com, redirected from the older inferencemax.semianalysis.com URL) is an open-source, automated LLM inference benchmark created and run by SemiAnalysis, described on its own site as tracking ‘throughput, efficiency, and cost’ across GPU/accelerator types (NVIDIA GB200, GB300, B200, B300, H100, H200; AMD MI300X, MI325X, MI355X) and inference frameworks (vLLM, SGLang, TensorRT-LLM/Dynamo), with nightly automated runs published via public GitHub Actions (github.com/SemiAnalysisAI/InferenceX). It publishes results through comparison pages, a ‘compare-per-dollar’ cost view, and a public API endpoint. https://inferencex.semianalysis.com/ and https://newsletter.semianalysis.com/p/inferencemax-open-source-inference (both fetched 2026-08-12); GitHub org referenced in page metadata: github.com/SemiAnalysisAI/InferenceX 2026-08-12 Confirms the benchmark’s existence, provenance (SemiAnalysis), and general methodology as OBSERVED FACT, independent of whether the specific $0.123 and 2.8M figures could be pinned down.
B: Current frontier-model API pricing Anthropic’s current flagship lineup and pricing: Claude Fable 5 ($10/$50 per MTok in/out), Claude Mythos 5 (same $10/$50, limited availability via “Project Glasswing”), Claude Opus 5 ($5/$25 per MTok), Claude Sonnet 5 ($2/$10 per MTok, this is now the standard non-promotional price), and Claude Haiku 4.5 ($1/$5 per MTok). OBSERVED_FACT Confirmed directly via WebFetch against Anthropic’s official pricing docs page (platform.claude.com/docs/en/about-claude/pricing). Full current base input/output pricing per MTok: Claude Fable 5 $10/$50; Claude Mythos 5 $10/$50 (limited availability, Project Glasswing); Claude Opus 5 $5/$25; Claude Opus 4.8 $5/$25; Claude Opus 4.7 $5/$25; Claude Opus 4.6 $5/$25; Claude Opus 4.5 $5/$25; Claude Sonnet 5 $2/$10; Claude Sonnet 4.6 $3/$15; Claude Sonnet 4.5 $3/$15; Claude Haiku 4.5 $1/$5. The cluster’s premise names “Opus 5/Sonnet 5/Fable 5” — confirmed as accurate to the current landscape, plus an unmentioned sibling, Claude Mythos 5 (same specs/price as Fable 5, restricted to Project Glasswing participants). https://platform.claude.com/docs/en/about-claude/pricing (fetched 2026-08-12) 2026-08-12 High confidence — this is Anthropic’s own live pricing documentation, fetched directly.
B: Current frontier-model API pricing Claude Sonnet 5’s $2/$10 per-MTok pricing was originally announced as introductory/promotional through 2026-08-31, with a scheduled increase to $3/$15 on 2026-09-01 — but that increase was cancelled and $2/$10 is now the permanent standard price. OBSERVED_FACT Anthropic’s pricing page carries an explicit note (footnote on the Sonnet 5 pricing row): “The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.” This directly corrects the premise (carried in the loaded claude-api skill’s cached table) that Sonnet 5 pricing is $3/$2.00-intro or will rise on 2026-09-01 — as of this check, on or before 2026-08-12, the increase has already been cancelled and $2/$10 is confirmed as the standing price, not merely an active introductory rate about to expire. https://platform.claude.com/docs/en/about-claude/pricing (fetched 2026-08-12) — footnote id claude-sonnet-5-introductory-pricing 2026-08-12 High confidence, directly from Anthropic’s current pricing page. Note this contradicts the cached pricing table baked into this session’s loaded skill file (which showed Sonnet 5 at $3.00/$15.00 with a $2.00/$10.00 ‘intro through 2026-08-31’ — that cached figure is now stale/superseded by the live page).
B: Current frontier-model API pricing OpenAI’s current flagship line is called GPT-5.6, released in three tiers named Sol (flagship), Terra (mid-tier), and Luna (budget/fast tier), with a limited preview around 2026-06-26 and a wider public release around 2026-07-09. OBSERVED_FACT This could not be verified against a primary OpenAI source in raw form — OpenAI’s pricing and docs pages (openai.com/api/pricing, platform.openai.com/docs/pricing) returned HTTP 403 to the fetch tool, and the working alternative (developers.openai.com/api/docs/pricing and /models) could only be accessed through WebFetch, which converts the page to markdown and has a small model summarize/extract before returning it to me — I never saw the raw HTML/text. Two independent WebFetch calls against developers.openai.com/api/docs/pricing and /models both reported gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna as real, currently listed model IDs, and a Wikipedia summary of the (English) GPT-5 article independently corroborated the Sol/Terra/Luna three-tier structure and the general release-timing narrative (limited preview around 2026-06-26, wider release 2026-07-09, tied to a described US government-restriction delay). Given the AI-summarization intermediary on the primary source, I classify the exact existence/spelling of the tier names as ATTRIBUTED (reported by two independent AI summaries + one encyclopedia summary) rather than a directly-observed primary-source fact. developers.openai.com/api/docs/pricing and developers.openai.com/api/docs/models (both accessed via WebFetch/AI-summarization, 2026-08-12); corroborating summary from en.wikipedia.org/wiki/GPT-5.6 (accessed via WebFetch, 2026-08-12) 2026-08-12 Medium confidence. The naming and existence of Sol/Terra/Luna is corroborated across three independent fetches (two OpenAI docs pages + Wikipedia), which is reassuring, but none of those fetches gave me verbatim primary-source text — all went through an AI summarization layer. Recommend the report state this explicitly rather than presenting it as a directly-verified primary fact.
B: Current frontier-model API pricing GPT-5.6 standard (short-context) API pricing per million tokens: Sol $5.00 input / $30.00 output (cached input $0.50); Terra $2.00 input / $12.00 output (cached $0.20); Luna $0.20 input / $1.20 output (cached $0.02). Long-context pricing (reported by the fetch as roughly double most rates, e.g. Sol $10.00/$45.00) applies above an unspecified token threshold — the exact threshold could not be found on the page. ATTRIBUTED_STATEMENT Via WebFetch/AI-summarization of developers.openai.com/api/docs/pricing, the standard-tier figures were reported consistently across two separate fetch attempts: Sol $5.00/$30.00, Terra $2.00/$12.00, Luna $0.20/$1.20 (short/standard context), roughly doubling for ‘long context’ (e.g., Sol long-context input $10.00, output $45.00 per one fetch, or $12.50/output per a slightly different fetch pass — the two fetches were not perfectly consistent on long-context cache-write figures, which is itself evidence of summarization noise). I explicitly asked a follow-up WebFetch call to find the token-count threshold that defines ‘long context’ vs ‘short context’ and it reported finding no footnote or explanatory text on the page defining that threshold. I cannot independently verify these exact dollar figures without a raw fetch of the primary source, which failed with 403 on the canonical openai.com/api/pricing URL. developers.openai.com/api/docs/pricing (accessed via WebFetch/AI-summarization, 2026-08-12) 2026-08-12 Lower confidence than a directly-verified figure — treat as attributed-via-intermediary rather than an independently confirmed primary fact. The two fetch passes gave slightly different numbers for GPT-5.6-sol long-context cache-write pricing ($12.50 in one pass, not matching cleanly with the other pass’s $6.25/$12.50 batch figures), suggesting the AI-summarization step is not perfectly reliable for dense pricing tables. A human should re-verify by opening the OpenAI pricing page directly in a browser.
B: Current frontier-model API pricing GPT-5.6 Sol/Terra/Luna each have a ~1.05M-token context window, 128K max output tokens, and a knowledge cutoff around 2026-02-16. ATTRIBUTED_STATEMENT Reported via WebFetch/AI-summarization of developers.openai.com/api/docs/models. All three tiers were reported to share identical context window (~1.05M tokens), max output (128K tokens), and knowledge cutoff (2026-02-16). This is plausible (matches the general pattern of same-generation model families sharing a context window and cutoff) but again passed through an AI-summarization layer rather than being read directly from primary-source text, so it should be treated as an attributed/secondhand report, not an independently confirmed spec sheet. developers.openai.com/api/docs/models (accessed via WebFetch/AI-summarization, 2026-08-12) 2026-08-12 Medium-low confidence — single-source (one fetch pass), via AI summarization, not cross-checked against a second independent source for these specific specs (context window / max output / cutoff date).
B: Current frontier-model API pricing Google’s current top-of-line Gemini Pro model is “Gemini 3.1 Pro,” and as of this check it is still labeled “Preview” (not GA/stable). Pricing is tiered by context length: input $2.00/MTok for prompts ≤200K tokens, $4.00/MTok for prompts >200K tokens; output (including thinking tokens) $12.00/MTok ≤200K tokens, $18.00/MTok >200K tokens. There is no separate, non-‘3.1’ ‘Gemini 3 Pro’ model. OBSERVED_FACT Confirmed via WebFetch against Google’s live Gemini API pricing page (ai.google.dev/gemini-api/docs/pricing), which lists the model as ‘Gemini 3.1 Pro Preview’ with endpoint ID gemini-3.1-pro-preview, quoting verbatim tiered pricing: ‘Input price: $2.00, prompts <= 200k tokens / $4.00, prompts > 200k tokens’ and ‘Output price (including thinking tokens): $12.00, prompts <= 200k tokens / $18.00, prompts > 200k’. The fetch explicitly confirmed no separate non-preview ‘Gemini 3 Pro’ listing exists (only ‘Gemini 3 Flash Preview’ alongside it), and confirmed Free Tier pricing for this model is ‘Not available’. A follow-up cross-check against a Wikipedia summary of the Gemini article corroborated Gemini 3.1 Pro as the latest Pro-tier model, described there as released 2026-02-19 and still in Preview status (not GA) as of that check. https://ai.google.dev/gemini-api/docs/pricing (fetched via WebFetch, 2026-08-12); corroborated by en.wikipedia.org/wiki/Gemini_(language_model) (fetched via WebFetch, 2026-08-12) 2026-08-12 High-medium confidence — the primary pricing figures came from Google’s own docs domain, though again read through WebFetch’s AI-summarization rather than raw text; the exact dollar figures and the ‘<=200k / >200k’ threshold language were reported consistently and match the expected Google pricing-page style, so confidence is reasonably high, but not a raw-HTML-verified certainty.
B: Current frontier-model API pricing The whitepaper’s framing (“OpenAI GPT-5.6 Sol/Terra/Luna, Anthropic Claude Opus 5/Sonnet 5/Fable 5, Google Gemini 3.1 Pro”) accurately reflects the current (as of 2026-08-12) three-lab frontier landscape, with the caveat that Anthropic’s family actually has a fourth sibling model (Mythos 5) and Google’s model is still a ‘Preview’ release rather than GA. INTERPRETATION Synthesizing the individual findings above: the premise holds directionally for all three labs as of the check date, but two nuances are worth flagging in the whitepaper text: (1) Anthropic Sonnet 5 pricing should be stated as the current standard $2/$10 per MTok, not as an expiring ‘intro’ rate, since the scheduled Sept 2026 price increase has been cancelled per Anthropic’s own pricing-page footnote; (2) Google’s Gemini 3.1 Pro is explicitly still in ‘Preview,’ not a GA/stable release, which is relevant if the whitepaper implies all three labs have shipped fully GA frontier tiers. Derived from the individually-cited sources above; no separate citation. 2026-08-12 This is an interpretive synthesis, not an independently checkable fact — presented as an editorial note for the whitepaper author, not as evidence in its own right.
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration On August 10, 2026, Nvidia announced six financing partnerships (memorandums of understanding) with Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs, and KKR to mobilize more than $500 billion in third-party capital for AI infrastructure buildout. OBSERVED_FACT Confirmed directly from CNBC’s own article text (Hugh Son, ‘Nvidia lines up $500 billion in financing as CEO Jensen Huang tells CNBC his chips are ‘investable asset’,’ published Mon Aug 10 2026 2:58 PM EDT, updated 6:09 PM EDT): ‘Nvidia signed memorandums of understanding with Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs and KKR to establish financing platforms for Nvidia’s customers… The effort aims to mobilize more than $500 billion in third-party capital for hyperscalers, frontier AI labs and enterprises.’ Corroborated by Fortune’s Aug 11, 2026 coverage (fetched directly) listing the identical six firms and $500B figure, and by search-indexed same-day coverage citing an FT report (Reuters, Bloomberg, WSJ, Guardian, CNN, U.S. News all appear to have run versions of this story Aug 10-11, 2026, per search-engine indexing, though those individual articles were not directly re-fetched in this session). https://www.cnbc.com/2026/08/10/nvidia-wall-street-asset-managers-500-billion-ai-push.html (author Hugh Son); corroborating: https://fortune.com/2026/08/11/nvidia-500-billion-ai-financing/ 2026-08-12
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration Named participants in the live CNBC Closing Bell Overtime panel interview (hosted by Becky Quick) were: Jensen Huang (Nvidia founder/CEO), David Solomon (Goldman Sachs CEO), Larry Fink (BlackRock CEO, joining remotely), Jon Gray (Blackstone President), Waldemar Szlezak (KKR Global Head of Digital Infrastructure), Jim Zelter (Apollo President), and Bruce Flatt (Brookfield CEO). OBSERVED_FACT Directly confirmed from the CNBC transcript page itself, which shows Becky Quick introducing all seven participants by name and title before the interview begins. https://www.cnbc.com/2026/08/10/cnbc-exclusive-transcript-cnbcs-becky-quick-speaks-with-nvidias-jensen-huang-wall-street-leaders-on-500b-ai-infrastructure-push-on-closing-bell-overtime-today.html 2026-08-12
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration Jensen Huang stated on the CNBC segment that ‘each gigawatt is something like $50, $60 billion’ to build out, including energy, land, power/shell, and computing costs. ATTRIBUTED_STATEMENT Confirmed as an exact quote from the CNBC transcript: HUANG: ‘It’s a hefty price tag. Each gigawatt is something like $50, $60 billion. And so there’s energy involved, there’s land power and shell involved, and of course there’s the computing part of it.’ This is Huang’s own figure as stated on air, not an independently audited cost — the whitepaper should treat it as attributed rather than confirmed hard cost. Independent 2025-2026 analyst estimates found via search summaries (Morgan Stanley ~$49B/GW for Nvidia’s Vera Rubin-generation systems, Foxconn ~$47B/GW, Bernstein ~$35B/GW, Epoch.ai ~$38B upfront capex/GW) put the figure in a broadly similar range, with $50-60B at the higher end of contemporaneous analyst estimates but not an outlier. CNBC transcript: https://www.cnbc.com/2026/08/10/cnbc-exclusive-transcript-cnbcs-becky-quick-speaks-with-nvidias-jensen-huang-wall-street-leaders-on-500b-ai-infrastructure-push-on-closing-bell-overtime-today.html. Corroborating analyst figures found via aggregated web-search summary only — primary analyst-note URLs for Morgan Stanley/Bernstein/Foxconn/Epoch.ai were not individually re-fetched/verified in this session, so treat those specific numbers as secondary, unverified-in-this-pass corroboration. 2026-08-12
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration Larry Fink (BlackRock CEO) stated on the same segment: ‘each gigawatt costs $50 to $60 billion to build out, and we’re talking about in the United States alone, we’re going to need over 70 gigawatts of power to fuel this… we need to raise $500 billion… but we’re going to have to raise trillions of dollars over the coming years.’ ATTRIBUTED_STATEMENT Confirmed as an exact quote from the CNBC transcript. Fink explicitly frames the $50-60B/GW figure as repeating Huang’s (‘As Jensen said’) and frames the >70GW figure and ‘trillions of dollars’ figure as his own forward-looking estimate, not as an audited/sourced market total. The existence and exact wording of the statement is confirmed against the primary transcript; the underlying substantive forecasts (actual future US power need, actual future capital raised) remain unverified projections, not independently confirmed facts. https://www.cnbc.com/2026/08/10/cnbc-exclusive-transcript-cnbcs-becky-quick-speaks-with-nvidias-jensen-huang-wall-street-leaders-on-500b-ai-infrastructure-push-on-closing-bell-overtime-today.html 2026-08-12
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration Fink’s figure of ‘over 70 gigawatts’ of new US power demand for AI is broadly consistent with independent third-party industry forecasts of US AI/data-center power demand growth. INTERPRETATION Independent forecasts found via web search put comparable figures in a similar range but none exactly match 70GW: Boston Consulting Group has reportedly projected a potential 50-80 gigawatt AI-driven power capacity shortfall in the US by 2030; S&P Global reported hyperscale data centers already drew 64.4 gigawatts from the US grid in 2025 (a 25% year-over-year increase); the IEA reportedly projects US data-center electricity use rising roughly 130% by 2030. These are directionally consistent with Fink’s >70GW claim but come from different methodologies, timeframes, and definitions (capacity shortfall vs. total demand vs. current draw), so this should be read as supportive context, not as independent proof of the specific 70GW number. None of the underlying BCG/S&P Global/IEA reports were directly fetched and verified in this session; figures come from an aggregated search-engine summary only. Aggregated web-search summary referencing BCG, S&P Global, and IEA published forecasts — primary report URLs not individually re-fetched/verified in this pass. 2026-08-12
C: NVIDIA/Wall Street financing broadcast — surrounding context and independent corroboration News coverage beyond CNBC (Fortune, and other outlets indexed via search — Reuters, Bloomberg, WSJ, Guardian, CNN, U.S. News, reportedly citing an original Financial Times report) corroborates the $500 billion figure and the same six-firm partner list on the same news day (Aug 10-11, 2026). OBSERVED_FACT Fortune’s article (fetched and reviewed directly) independently lists the identical six firms (Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, KKR) and the $500B figure, plus quotes from Huang, Fink, and Solomon consistent with the CNBC transcript. Other outlets (Reuters, Bloomberg, WSJ, CNN, Guardian, U.S. News) were identified only via search-engine indexing with matching headlines/dates referencing an FT report; their full article text could not be directly fetched (paywalled, blocked, or timed out) in this session. Their existence and headlines as indexed should be treated as corroborating signal of wide same-day pickup, not as independently re-verified primary sources. https://fortune.com/2026/08/11/nvidia-500-billion-ai-financing/ (directly verified); search-indexed-only headlines referencing Reuters, Bloomberg, WSJ, CNN, Guardian, U.S. News coverage dated Aug 10-11, 2026 of the same announcement (not directly fetched/verified). 2026-08-12
D: Sam Altman YC Startup School talk — full context verification Sam Altman said at YC Startup School (closing Center Court session, Chase Center, San Francisco, 26 July 2026) something equivalent to ‘you can now do three months of work in seventeen minutes — so do three months of work,’ as part of an argument against treating the present as ‘the end of history.’ ATTRIBUTED_STATEMENT Independently corroborated (as paraphrase, not verbatim quote) by a CC BY 4.0 attendee-notes repository for the same event: github.com/Princeu3/yc-startup-school-2026-notes, file sessions/09-sam-altman.md, section ‘Ambition and the end-of-history trap.’ The notes explicitly flag this line as ‘[transcript]’ — i.e. paraphrased from a machine transcript, never presented as a verbatim quotation — and pair it with: ‘it can be simultaneously true that creating superintelligence is the most important thing in the history of human society and that it will look small next to what some new startup does. Treating the present as the end of history is the trap.’ This matches the surrounding argument implied by the already-confirmed ‘three months of work in 17 minutes’ quote: Altman is using it to tell founders to be MORE ambitious (given AI-driven productivity gains), not to work less. https://github.com/Princeu3/yc-startup-school-2026-notes/blob/main/sessions/09-sam-altman.md (attendee notes, CC BY 4.0, Y Combinator Startup School 2026, closing session, 26 July 2026 — not an official YC or OpenAI transcript) 2026-08-12 This is a third-party attendee’s paraphrase of a machine transcript (Granola + Gemini re-transcription per the repo’s own README), explicitly NOT presented as verbatim by its author. It corroborates the gist and placement of the quote but should not be treated as an independent verbatim confirmation beyond what was already established.
D: Sam Altman YC Startup School talk — full context verification Altman described a dystopia at Startup School involving ‘no freedom,’ ‘no agency,’ and ‘no privacy’ alongside material abundance/a cure for cancer — determine whether this is framed as (a) abundance-without-freedom or (b) overreacting-to-AI-safety. ATTRIBUTED_STATEMENT Found a fuller quote via news.com.au coverage (syndicated by europesays.com): “I’m particularly nervous about 10 years from now if we overreact to AI safety. And so we say, look, everyone, you’re going to get a cure for cancer, you’re going to have material abundance, but you will have no freedom, you will have no agency, it will be a perfect surveillance state.” This shows the two framings in the task prompt are NOT alternatives — they are causally linked in Altman’s own words: overreacting to AI safety is the CAUSE, and the abundance-without-freedom surveillance state is the RESULT/dystopia he says he wants to avoid. The article also notes he explicitly framed this as ‘a future that should be prevented,’ and that his stated alternative is that freedom and agency should increase as AI grows more capable. news.com.au, “‘No freedom’: OpenAI boss Sam Altman sparks frenzy after dystopian prediction”, https://www.news.com.au/technology/innovation/no-freedom-openai-boss-sam-altman-sparks-frenzy-after-dystopian-prediction/news-story/837f91f37c101f422cac77ec5eb2764f (dated approx. early August 2026); same content syndicated at https://www.europesays.com/people/174922/ 2026-08-12 News.com.au itself was not directly fetchable (blocked); the quote and framing were reconstructed via the europesays.com syndicated version and cross-checked against the GitHub attendee notes’ independent paraphrase (‘He explicitly rejected the bargain where you get enormous capability and give up freedom, agency and privacy for it, and allowed only for temporary tradeoffs’ — sessions/09-sam-altman.md, ‘Demand for intelligence’ section), which corroborates the same causal structure without using the word ‘surveillance.’
D: Sam Altman YC Startup School talk — full context verification Altman claimed AI/compute demand will grow 10x per year, with some hedging language. ATTRIBUTED_STATEMENT I could not find a primary source where Altman states demand itself grows ‘10x per year.’ What is well-documented and heavily cited is a DIFFERENT, related claim from his blog post ‘Three Observations’ (blog.samaltman.com): ‘The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use’ — explicitly a COST-decline figure, not a demand-growth-rate figure, though he links it causally to increased usage (a Jevons-paradox-style argument) and compares it favorably to Moore’s Law (‘Moore’s law changed the world at 2x every 18 months; this is unbelievably stronger’). Separately, at the YC Startup School talk itself, the attendee notes report a much softer, unquantified claim: ‘He expects worldwide demand to keep growing for many years, and used electricity as the analogy: demand rises as price falls. He allowed that at some point it gets harder to find things to do with more — but doesn’t think we’re near that’ (hedged, no percentage given). A Business Insider report (Dec 2025, not independently fetchable due to site block) is quoted secondhand as containing the line ‘I think demand will far exceed what we can think of,’ again with no specific multiplier. blog.samaltman.com, ‘Three Observations’ (cost-falls-10x claim); Forbes coverage at https://www.forbes.com/sites/johnwerner/2025/02/11/covering-sam-altmans-three-observations-on-ai/; YC Startup School attendee notes, https://github.com/Princeu3/yc-startup-school-2026-notes/blob/main/sessions/09-sam-altman.md, section ‘Demand for intelligence’; secondhand reference to businessinsider.com/openai-chart-compute-future-plans-profitability-2025-12 (site blocked for direct fetch, so this last citation is unverified at the primary source and should be treated as low-confidence) 2026-08-12 IMPORTANT: if the whitepaper’s cluster D claim states Altman said ‘demand grows 10x/year’ as a direct quote from the Startup School talk specifically, this appears NOT to be corroborated by the only independent notes I could find of that talk — those notes describe an unquantified, hedged ‘many years’ growth expectation with no specific multiplier. The 10x/year figure that IS well-sourced from Altman is about cost decline, not demand growth, and comes from a different piece (his blog, not this talk). Recommend re-checking the original transcript excerpt this claim was drawn from to see if ‘10x’ refers to cost or demand, and whether it was said at Startup School or elsewhere.
D: Sam Altman YC Startup School talk — full context verification Altman used a ‘100,000 tokens per month’ historical anecdote, with some specific comparison point, in the context described. ATTRIBUTED_STATEMENT Well-corroborated overall figures, but from a DIFFERENT event than Startup School: at OpenAI’s ‘Intelligence at Work’ enterprise livestream (June 2, 2026), Altman said OpenAI’s single heaviest user now consumes roughly 100 billion tokens/month, and drew the comparison: ‘Six and a half years ago, OpenAI’s heaviest user consumed roughly 100,000 tokens per month, a figure then likely the highest in the world. Today, 100,000 tokens per month is the global per capita average’ — i.e., the comparison point is NOT one person’s growth over time in isolation, but a shift from ‘highest-usage individual in the world’ (2019) to ‘global per-capita average’ (2026), representing roughly a 1,000,000x (100 billion / 100,000) increase for the top user specifically. He also reportedly ‘projects another million-fold expansion in token usage ahead.’ A YouTube video (title: ‘Sam Altman: The Average Person Will Use 500 Billion AI Tokens a Month,’ channel ‘The Missing Context’) suggests this general theme was discussed in more than one Altman appearance, but I could not confirm this specific anecdote was repeated verbatim inside the Startup School talk itself — the independent attendee notes for that session do not mention it. AOL/Yahoo-syndicated finance article, ‘Sam Altman makes stunning admission about AI,’ published June 7, 2026, https://www.aol.com/finance/sam-altman-makes-stunning-admission-223700331.html, reporting on OpenAI’s ‘Intelligence at Work’ enterprise event, June 2, 2026 2026-08-12 The million-fold figure (100bn/100k = 1,000,000) is a DERIVED_CALCULATION from the two OBSERVED numbers in the article, not stated as an exact multiplier by Altman himself in the excerpt I found (the article separately states he ‘projects another million-fold expansion ahead’ for the future, which appears to be his own framing, not just the historical ratio). If the whitepaper places this anecdote inside the YC Startup School talk specifically, that placement is unverified by me — it most clearly traces to the separate June 2, 2026 ‘Intelligence at Work’ livestream.
D: Sam Altman YC Startup School talk — full context verification Current OpenAI API pricing for flagship models (checkable independently as of 2026-08-12), relevant if the whitepaper cites pricing to compare against token-usage claims. OBSERVED_FACT Per OpenAI’s official pricing page, current per-million-token rates include: gpt-5.5 — $5.00 input / $30.00 output; gpt-5.6-terra — $2.00 input / $12.00 output; gpt-5.6-sol — $5.00 input / $30.00 output; o1-pro — $150.00 input / $600.00 output; gpt-4o — $2.50 input / $10.00 output; o3 — $2.00 input / $8.00 output; gpt-4.1 — $2.00 input / $8.00 output; gpt-4o-mini — $0.15 input / $0.60 output; gpt-3.5-turbo — $0.50 input / $1.50 output. Batch processing is roughly 50% cheaper, and cached input is typically ~10% of standard input rates. Regional processing endpoints for models released on/after March 5, 2026 carry a 10% uplift. https://developers.openai.com/api/docs/pricing (redirected from platform.openai.com/docs/pricing), fetched 2026-08-12 2026-08-12 Included as an independently-checkable factual anchor in case the whitepaper uses current OpenAI pricing to illustrate or sanity-check any of Altman’s cost/demand claims. This is NOT itself part of the Startup School talk content.
E: Mark Zuckerberg Meta essay (directly fetchable) Publish date of the Zuckerberg essay ‘The Future is for Everyone’ on about.fb.com OBSERVED_FACT Confirmed via page metadata (JSON-LD schema.org markup) fetched directly from the live page on 2026-08-12: datePublished = 2026-08-10T10:01:23+00:00, dateModified = 2026-08-10T15:28:59+00:00. Page title: ‘The Future is for Everyone’; meta description: ‘A letter from Mark Zuckerberg on the path to a positive AI future’; the piece itself is signed ‘— Mark’ at the end and bylined to Mark Zuckerberg in the meta/og description tags. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12; verified via embedded JSON-LD NewsArticle schema in page HTML) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The essay lays out a ‘three principles’ framing OBSERVED_FACT Confirmed verbatim in the live article text, stated twice: ‘We propose a philosophy based on individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety.’ Restated near the close: ‘Meta is committed to building with the principles of individual empowerment as the source of prosperity, invention as AI’s purpose, and a balance of power favoring people as the foundation for addressing safety risks.’ Note: the essay does not use the literal phrase ‘three principles’ as a label, but it explicitly enumerates exactly three named principles in parallel structure, twice. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12, full text extracted directly from page HTML) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The essay describes a personal AI agent with a ‘private mode’ where even Meta cannot see or access user information OBSERVED_FACT Confirmed verbatim, appearing twice. First: ‘We believe that people should have a fully private mode for personal agents where even Meta or any other service provider cannot see or grant access to your information. This is similar to how we built WhatsApp into the world’s leading end-to-end encrypted service…’ Second (policy-implications list): ‘We will build a fully private mode where even Meta cannot see or grant access to your information.’ Earlier, the personal-agent description also states the agent ‘will have strong privacy and security options so you can trust it to handle all of your personal content knowing that no one else can access your information, similar to how encryption works on WhatsApp.’ https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The essay uses a ‘superintelligent lawyer’ analogy to argue for broad distribution of AI capability OBSERVED_FACT Confirmed verbatim: ‘As a thought experiment, imagine only one person had a superintelligent lawyer. They would have an unfair advantage in court — even if they were wrong on the merits. That would lead to a worse society. But now imagine everyone has a superintelligent lawyer. In this case, justice would be carried out much more fairly and efficiently than it is today…’ The same passage extends the analogy to cybersecurity superintelligence and to a hypothetical single business with exclusive access to superintelligence. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The essay argues alignment should mean aligning AI agents to the individual user’s goals/values rather than a centralized/company set of values OBSERVED_FACT Confirmed verbatim: ‘Instead, we view alignment as ensuring that agents share a person’s goals and values, not our company’s… alignment should be about helping people pursue the many diverse goals and views held across society rather than a method of enforcing a centralized dogma.’ Also: ‘if we reach a state where billions of people are using and scrutinizing personal superintelligence agents, then we will have solved alignment to individuals’ interests.’ The essay also gives a specific critical example to contrast against: ‘one leading model was aligned to refuse helping draft a letter to prospective parents at a school because it thought standardized testing was unethical.’ This example is Zuckerberg’s own unattributed characterization of an unnamed competitor model/incident — it could not be independently identified or confirmed, so it should be classified separately as an ATTRIBUTED STATEMENT by Zuckerberg rather than a verified incident. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The unnamed ‘one leading model refused to help draft a school admissions letter over standardized-testing ethics’ anecdote ATTRIBUTED_STATEMENT This specific anecdote is stated by Zuckerberg in the essay but names no model, company, date, or source. No independent corroborating report of this specific incident was found in this check. Treat as an unverified claim by Zuckerberg, not an independently confirmed incident. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12); no independent corroboration found 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) The essay frames open-source AI as part of the safety architecture (preventing centralization, improving security) OBSERVED_FACT Confirmed verbatim: ‘Open source is a positive and important force for empowering people and preventing centralization that is detrimental for both safety and the economy… Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon.’ On cybersecurity specifically: ‘widely deployed open source systems have proven more secure because more people can identify vulnerabilities, harden the systems, and easily upgrade to the latest most secure versions. Even in recent weeks, we have seen companies handling security incidents like HuggingFace rely on widely available open models to patch vulnerabilities.’ https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) Meta had paused open-weight model releases and is now resuming them, implied by ‘we will resume releasing some open source models soon’ ATTRIBUTED_STATEMENT This is Zuckerberg’s own statement implying a prior slowdown/pause tied to the standup of Meta Superintelligence Labs. Not independently corroborated with a separate primary source (e.g., a dated announcement of the pause) in this check; treat as an attributed claim about Meta’s own release cadence, not independently verified against a separate primary record. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12); no separate independent corroboration found in this check 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) Independent/third-party reception and framing of the essay ATTRIBUTED_STATEMENT TechCrunch published a critical response around the same date (‘Mark Zuckerberg’s AI manifesto is exactly why people don’t like AI,’ byline Russell Brandom, ~2026-08-10) arguing the essay offers philosophical abstraction rather than acknowledgment of real AI risks. It specifically discusses the personalized-PhD-tutor claim (contrasted with real-world AI-assisted cheating in education) and the superintelligent-lawyer thought experiment (critiqued for ignoring frivolous litigation and resource-based imbalance), and contrasts Zuckerberg’s risk-minimizing tone with OpenAI/Anthropic leadership ‘acknowledging the dangers of AI.’ This corroborates that the essay’s rhetorical devices (three-principles framing, PhD-tutor claim, lawyer analogy) exist and were publicly discussed and critiqued, but is itself an opinion piece and does not independently prove any of Zuckerberg’s substantive predictions. Note: content was obtained via an automated fetch/summarize tool rather than manually reading raw HTML, so treat quoted fragments as approximately (not verbatim byte-for-byte) reproduced. TechCrunch, ‘Mark Zuckerberg’s AI manifesto is exactly why people don’t like AI,’ byline Russell Brandom, published on or about 2026-08-10, techcrunch.com (AI category) 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) A second related TechCrunch article ties a Meta model (‘Glimmer’) to Zuckerberg’s personal-intelligence vision around the same date UNVERIFIABLE A TechCrunch site search surfaced a result titled ‘Meta’s new Glimmer AI model offers a hint at Zuckerberg’s personal intelligence vision,’ byline Rebecca Bellan, dated approximately the same window (~2026-08-10), but this was only seen as a search-result snippet, not fetched and read directly. Its existence and exact content are not independently confirmed in this check; do not cite specific claims from it. TechCrunch site search result (techcrunch.com/?s=zuckerberg+future+is+for+everyone); article not independently fetched or verified 2026-08-12
E: Mark Zuckerberg Meta essay (directly fetchable) Related/companion Meta Newsroom pieces exist alongside this essay, indicating it is a follow-up to an earlier 2025 post OBSERVED_FACT The live article page’s ‘Related News’ module lists ‘Meta Personal Superintelligence for Everyone,’ dated July 30, 2025 (an earlier Zuckerberg post on the same themes), and ‘Meta and AMD Partner for Longterm AI Infrastructure Agreement,’ dated February 24, 2026, both linked directly from the August 2026 article page. This confirms the August 2026 essay builds on/expands a named July 2025 predecessor post rather than being a standalone statement. https://about.fb.com/news/2026/08/the-future-is-for-everyone/ (fetched 2026-08-12, ‘Related News’ footer module) 2026-08-12
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure Luxor’s Hashrate Index reported Bitcoin hashprice around August 10, 2026. OBSERVED_FACT Luxor Technology’s Hashrate Index Weekly Roundup, published August 10, 2026, at hashrateindex.com/blog/hashrate-index-roundup-august-10-2026/, reports spot USD hashprice at $31.73 per PH/s/Day, down 1.1% week-over-week from $32.07, and down 1.2% on a 30-day average basis. Supporting context in the same roundup: BTC price ~$64,033 (+2.0% w/w), network hashrate 911 EH/s (7-day SMA, -2.3%), network difficulty 127.48T (+0.99%), and weekly miner block-reward revenue of ~3,158 BTC (~$202M). The roundup states hashprice at ~$32/PH/s/Day is ‘at or below breakeven for many miners depending on operating cost and machine model type.’ Separately, the live hashrateindex.com homepage ticker (checked 2026-08-12) showed a current hashprice reading of $32.05/PH/day, consistent with the roundup’s range a couple of days later. https://hashrateindex.com/blog/hashrate-index-roundup-august-10-2026/ (Luxor/Hashrate Index, published Aug 10, 2026); live reading from https://hashrateindex.com/ checked 2026-08-12 2026-08-12 Figures extracted via automated fetch/summarization tool rather than direct human reading of raw page; treat exact decimal precision as reasonably but not 100% certain. The two figures ($31.73 on Aug 10 vs $32.05 on Aug 12) are consistent with normal day-to-day hashprice fluctuation, not a discrepancy.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure OpenRouter claims a large processing scale (e.g., trillions of tokens per week/month) and publishes model/app usage rankings. OBSERVED_FACT OpenRouter’s own homepage (openrouter.ai, checked 2026-08-12) states scale figures of ‘200T+ Monthly Tokens,’ ‘10M+ Global Users,’ ‘70+ Providers,’ and ‘400+ Models’ (confirmed present in raw page HTML, not just AI paraphrase). Its rankings page (openrouter.ai/rankings) describes itself as showing ‘Live LLM rankings based on benchmarks and real data from millions of people using models through OpenRouter,’ broken into categories: Top Models (weekly usage), Top models by task (ranked by spend share), Cost per session, Market Share (by model author, text request share), Tool Calls, Images, Top Apps. A May 2026 Series B announcement (per Tech Times coverage of the raise) cited weekly token volume of 25 trillion tokens (up from 5 trillion eleven months earlier) and a user base of 8 million globally — these figures predate and are lower than the current homepage’s ‘200T+ Monthly Tokens’ and ‘10M+ Global Users,’ consistent with continued growth between May and August 2026 (25T/week ≈ ~107T/month at that time, growing toward 200T+/month by August is plausible growth). https://openrouter.ai/ and https://openrouter.ai/rankings (checked 2026-08-12, live page fetch) 2026-08-12 Homepage stats are OpenRouter’s own self-reported figures, not independently audited; treat as company-disclosed usage data (self-reported) rather than third-party-verified fact, even though sourced directly from the primary site.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure OpenRouter has been described, including allegedly by itself, as the ‘largest’ AI model marketplace/router. ATTRIBUTED_STATEMENT OpenRouter’s own site (homepage, rankings page, docs FAQ, announcements/blog) does NOT contain any explicit self-description as ‘largest,’ ‘leading,’ or similar superlative — it describes itself functionally (‘One API for Any Model,’ ‘a unified API to access all the major LLM models on the market’). Third-party press coverage, however, does use such language: a Dataconomy article (Feb 25, 2026, per search snippet) refers to OpenRouter as ‘the world’s largest LLM API aggregation platform,’ and other syndicated coverage of the Chinese-model-share story referred to OpenRouter (per a search snippet, exact outlet not independently confirmed) as ‘the largest neutral LLM routing platform’ processing ‘more than 20 trillion tokens a week’ at that time (~Feb 2026). These are third-party characterizations, not claims OpenRouter makes about itself on its official site. https://dataconomy.com/2026/02/25/chinese-ai-models-hit-61-market-share-on-openrouter/ (referenced via search snippet; not independently fetched in full); OpenRouter’s own site fetched directly at openrouter.ai, openrouter.ai/rankings, openrouter.ai/docs/faq (checked 2026-08-12) shows no such superlative claim 2026-08-12 The ‘largest’ framings come from search-result snippets, not a full fetch of the source articles — could not independently verify exact wording/context in full. If the whitepaper attributes a ‘largest’ claim to OpenRouter itself (rather than to press coverage about it), that would be inaccurate and should be corrected to attribute it to third-party coverage.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure OpenRouter raised a $113 million Series B round (context for inference-market infrastructure scale). OBSERVED_FACT OpenRouter closed a $113 million Series B on May 26, 2026, led by CapitalG (Alphabet’s independent growth fund), with participation from NVentures (NVIDIA’s venture arm), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, and returning investors Andreessen Horowitz and Menlo Ventures. The round valued the company at approximately $1.3 billion, more than double its ~$547 million post-money Series A valuation from June 2025 (per PitchBook data cited in the article). Weekly token volume was reported to have grown from 5 trillion to 25 trillion tokens over the prior six months, with the user base reaching 8 million globally at that time. CEO Alex Atallah (OpenSea co-founder) and co-founder Louis Vichy were named; Atallah is quoted: ‘Running inference at scale is fundamentally a multimodel problem… The era of picking a single model is over.’ https://www.techtimes.com/articles/317353/20260529/ai-gateway-openrouter-raises-113m-google-nvidia-route-between-their-models.htm (Tech Times, published May 29, 2026, byline Richard L. Wells) 2026-08-12 This is a secondary news article, not OpenRouter’s own press release (attempts to fetch the BusinessWire and NYT/WSJ originals timed out or were blocked); figures are consistent with other outlets’ reporting found in search results, but exact quote wording should be treated as attributed via this outlet’s reporting rather than a directly-verified primary transcript.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure IBM and Together AI signed a $240 million deal to build an NVIDIA B300-based inference cluster. OBSERVED_FACT IBM’s own newsroom press release, dated August 11, 2026, titled ‘IBM and Together AI Sign Multi-Year Agreement to Scale Open-Source AI Inference with NVIDIA AI Infrastructure on IBM Cloud,’ confirms: a multi-year $240 million agreement; a large-scale dedicated inference cluster on IBM Cloud built with NVIDIA HGX B300 systems and NVIDIA Spectrum-X Ethernet networking; expected availability in Q1 2027; described by multiple outlets as the first dedicated, large-scale cluster built for inference on IBM Cloud using HGX B300 systems. Named quotes in the press release: Vipul Ved Prakash (CEO, Together AI): ‘Enterprises want the performance of the best frontier models without the closed-model price tag, and that only works if the infrastructure underneath is fast and reliable at scale.’ Alan Peacock (GM, IBM Cloud): ‘Enterprises are in a race to adopt agentic AI at scale to drive real business outcomes.’ Dion Harris (Sr. Director, HPC and AI Infrastructure Solutions, NVIDIA): ‘AI factories are becoming essential enterprise infrastructure—like electricity and telecommunications—turning compute and data into intelligence.’ https://newsroom.ibm.com/2026-08-11-IBM-and-Together-AI-Sign-Multi-Year-Agreement-to-Scale-Open-Source-AI-Inference-with-NVIDIA-AI-Infrastructure-on-IBM-Cloud (IBM Newsroom, published Aug 11, 2026); corroborated by Reuters headline/content confirmed via search (reuters.com/business/ibm-together-ai-ink-240-million-deal-nvidia-powered-ai-inference-cluster-2026-08-11/) and wire pickups (US News, Storage Review), all dated Aug 11, 2026 2026-08-12 IBM press release fetch succeeded and is the primary source. The Reuters article itself could not be directly fetched (blocked by tool), but its existence, headline, and content were confirmed via search snippets and secondary wire pickups carrying the same story.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure NVIDIA claims the new IBM/Together AI cluster delivers ‘30x more AI factory output.’ ATTRIBUTED_STATEMENT A search snippet of press coverage states: ‘According to NVIDIA, it is built to deliver 30x more AI factory output’ in connection with the IBM/Together AI HGX B300 cluster. This is a vendor (NVIDIA) marketing/performance claim relayed by press, not an independently benchmarked or verifiable figure — no methodology, baseline, or benchmark was found describing what is being multiplied by 30x (e.g., vs. prior-generation Hopper cluster, vs. what workload). Search snippet from press coverage of the IBM/Together AI announcement (exact outlet and full context not independently fetched); could not locate the specific NVIDIA source document making this exact claim 2026-08-12 This figure should be treated strictly as an attributed vendor claim, not fact. Could not verify against a primary NVIDIA source or find the benchmark basis; if included in the whitepaper it must be clearly labeled as an NVIDIA-attributed marketing claim.
F: Bitcoin hashprice, OpenRouter, inference-market infrastructure The NVIDIA HGX B300 platform used in the IBM/Together AI cluster is a Blackwell Ultra-generation, 8-GPU system. OBSERVED_FACT NVIDIA’s official HGX product page (nvidia.com/en-us/data-center/hgx/, checked 2026-08-12) specifies the HGX B300 as an 8x NVIDIA Blackwell Ultra SXM GPU system with 2.1 TB total memory, 144 PFLOPS sparse / 108 PFLOPS dense NVFP4 inference performance, fifth-generation NVLink (1.8 TB/s per GPU-to-GPU link, 14.4 TB/s total NVLink bandwidth), and 1.6 TB/s networking bandwidth. It is positioned as an enhanced variant of the HGX B200 with improved sparse-tensor and attention performance at the same 8-GPU form factor. This confirms the hardware described in press coverage of the IBM/Together AI deal (‘HGX B300 systems’ linking ‘Blackwell processors’) is more precisely Blackwell Ultra-generation hardware, not the original Blackwell (B200) generation — a useful precision refinement over generic ‘Blackwell’ framing used in some press coverage. https://www.nvidia.com/en-us/data-center/hgx/ (NVIDIA official product page, checked 2026-08-12) 2026-08-12 Some press coverage of the IBM/Together AI deal generically says ‘Blackwell processors’ without specifying ‘Ultra’; this is a minor precision note, not a contradiction. Total GPU count for the specific IBM/Together AI cluster (i.e., how many HGX B300 nodes/how many total GPUs) was not disclosed in the press release or coverage found.
G: Humanoid robot market pricing Unitree G1 humanoid robot starts at $13.5K (about $13,500), stands 1320mm tall, weighs about 35kg with battery, and has 23 degrees of freedom (up to 43 DOF in the G1 EDU configuration with optional dexterous hands). OBSERVED_FACT Official Unitree G1 product page confirms: ‘Price from $13.5K’ (tax and shipping excluded); dimensions 1320x450x200mm (standing); weight ‘About 35kg’ with battery; 23 DOF standard, 23-43 DOF for the EDU variant with hand upgrades. https://www.unitree.com/g1 (official Unitree product page, checked 2026-08-12) 2026-08-12 Directly fetched from the manufacturer’s own current pricing page; high confidence.
G: Humanoid robot market pricing 1X NEO home humanoid robot can be pre-ordered for $20,000 upfront (early access) or a $499/month subscription, with a $200 refundable deposit to reserve. ATTRIBUTED_STATEMENT Engadget (Oct 29, 2025) reported: ‘Those who want early access to it can get it for $20,000, but it will also be available as a subscription service of $499 a month.’ This is consistent with multiple other outlets (MSN, Humanoids Daily, GrabARobot) found in search. However, I was unable to independently verify the $20,000 / $499 figures directly on 1X’s own official 1x.tech/neo page in the fetched content – that page’s visible text (as fetched) showed only the $200 deposit figure, not the full purchase/subscription price. Treating the $20,000/$499 figures as an attributed/reported statement (well-corroborated across independent secondary sources) rather than a directly observed primary-source figure, since the primary source page did not surface those numbers in the fetch. https://www.engadget.com/ai/1x-neo-is-a-20000-home-robot-that-will-learn-chores-via-teleoperation-040252200.html (published 2025-10-29); deposit figure confirmed on https://www.1x.tech/neo (checked 2026-08-12) 2026-08-12 Well-corroborated across several independent outlets, but not confirmed on 1X’s own pricing page in the content actually retrieved.
G: Humanoid robot market pricing 1X NEO specs: approximately 168cm tall (5’6”), about 30kg (66 lbs), lift capacity ~70kg (154 lbs), carrying capacity ~25kg (55 lbs), 4-hour battery runtime (842 Wh), walking speed 1.4 m/s, max speed 6.2 m/s, uses an Nvidia Jetson Thor-based ‘1X NEO Cortex’ chipset. OBSERVED_FACT Confirmed directly from the official 1x.tech/neo page: height 5’6”, weight 66 lbs, lift capacity 154 lbs, carrying capacity 55 lbs, walking speed 1.4 m/s, max running speed 6.2 m/s, battery runtime 4 hours (842 Wh), chipset ‘1X NEO Cortex (Nvidia Jetson Thor)’, 22 DOF per hand, 7 DOF per arm. https://www.1x.tech/neo (official 1X Technologies product page, checked 2026-08-12) 2026-08-12 Directly fetched from manufacturer’s own spec page; high confidence for the physical/technical specs (separate from the pricing figures above).
G: Humanoid robot market pricing Tesla has publicly targeted a long-term consumer price of $20,000-$30,000 for Optimus, but as of today (2026-08-12) Tesla has not opened commercial orders, published a confirmed retail price, or shipped any unit to a customer. ATTRIBUTED_STATEMENT Elon Musk has repeated a $20,000-$30,000 long-term target price for Optimus at scale in multiple public settings, including reportedly at Davos (Jan 2026) and on Tesla earnings calls, with commercial sales previously projected to begin ‘by late 2027’ per one secondary summary. This is Musk’s stated aspiration, not a current list price – multiple secondary sources explicitly note Tesla has not opened orders or published a confirmed retail price as of when checked. I could not fetch Tesla’s own AI/Optimus page directly (403 Forbidden), so this rests on secondary-source aggregation rather than a primary Tesla transcript/press release I directly verified. Aggregated secondary reporting found via search (exact primary transcript of the Davos/earnings-call remarks not independently re-verified in this session; Tesla’s own site at tesla.com/AI returned HTTP 403 when fetched on 2026-08-12) 2026-08-12 Lower confidence – this is a widely repeated Musk statement corroborated by multiple secondary aggregator sites, but I was not able to pull the primary transcript (earnings call or Davos remarks) myself in this session to confirm exact wording or date. Should be labeled clearly as an unconfirmed target price, not a market price.
G: Humanoid robot market pricing Tesla Optimus (as described in current secondary sources, specs unverified against a Tesla primary source): height ~173cm, weight ~57kg, 65 degrees of freedom, max speed ~8 km/h, payload ~20kg, target battery life 8 hours. ATTRIBUTED_STATEMENT These figures come only from secondary aggregator sites (e.g., optimusk.blog, robottoday.com, blog.robozaps.com) found via search, not from Tesla’s own site, which returned a 403 error when fetched directly. Given Tesla has not shipped or commercially specified Optimus, these numbers should be treated as reported/aggregated rather than confirmed primary specs. Secondary aggregator sites found via search (e.g. optimusk.blog, robottoday.com); Tesla’s own tesla.com/AI page could not be fetched (403 Forbidden) on 2026-08-12 2026-08-12 Low-to-moderate confidence; no primary Tesla source was accessible to cross-check these numbers directly.
G: Humanoid robot market pricing Agility Robotics’ Digit humanoid robot has a reported outright purchase price around $250,000 per unit, with an alternative Robot-as-a-Service option cited around $8,500/month (or ~$30/hour) in some sources. ATTRIBUTED_STATEMENT Multiple independent secondary/aggregator sources (GrabARobot, RoboSelect360, Humanoza, BotMarket, TechAnnouncer, HeyManoid, BioRobot.ai, FutureRobots) consistently cite approximately $250,000 as Digit’s outright purchase price, with a RaaS alternative around $8,500/month or ~$30/hour attributed to an Agility investor deck. I could not fetch Agility Robotics’ own official pricing page directly (404 error on the URLs tried), so this figure is not independently confirmed against a primary source in this session, only via convergent secondary reporting. Multiple secondary/aggregator sources found via search (GrabARobot, RoboSelect360, Humanoza, BotMarket, TechAnnouncer, HeyManoid, BioRobot.ai, FutureRobots); official agilityrobotics.com pages attempted but returned 404 on 2026-08-12 2026-08-12 Moderate confidence due to convergence across many independent secondary sources on the same $250,000 figure, but no primary Agility Robotics pricing page was accessible to confirm directly.
G: Humanoid robot market pricing Figure AI does not publish pricing for its Figure 03 humanoid robot on its official site; third-party sources report approximate specs of 168cm height, 60kg weight, 40-44 degrees of freedom, ~20kg payload, and 5-hour battery life, with a commonly cited consumer price target of roughly $20,000. ATTRIBUTED_STATEMENT Figure AI’s official site (figure.ai) contains no pricing information and no detailed specs in the pages fetched; attempts to fetch figure.ai/robot and figure.ai/figure-03 and figure.ai/news/figure-03 all returned 404. The height/weight/DOF/payload/battery figures and the ~$20,000 price target come only from third-party aggregator sites (livium.com, aparobot.com, biorobot.ai, roboatlas.ai, humanoid.guide) found via search, not from Figure AI directly, and should be treated as unverified against a primary source. Figure AI official site (figure.ai) checked 2026-08-12, no pricing/specs found; third-party figures from livium.com, aparobot.com, biorobot.ai, roboatlas.ai (aggregator sites, not primary) 2026-08-12 Low confidence on the specific numbers since no primary Figure AI source confirms them; Figure AI explicitly does not disclose pricing publicly as of this check.
G: Humanoid robot market pricing Real, currently purchasable/pre-orderable humanoid robot platforms span roughly $13,500 (Unitree G1 entry price) to $20,000 (1X NEO early-access price) for consumer/prosumer units, while enterprise-grade or aspirational platforms (Agility Digit, Tesla Optimus target, Figure 03 target) are reported/targeted in the $20,000-$250,000+ range, meaning the whitepaper’s illustrative €25,000 example sits within, but toward the lower-middle portion of, the real observed market range once currency differences are accounted for. DERIVED_CALCULATION Formula: comparing four data points gathered above – Unitree G1 ($13,500, OBSERVED FACT, official page) < 1X NEO ($20,000 early access / $499/month, ATTRIBUTED STATEMENT via Engadget) approx.= Tesla Optimus long-term target ($20,000-$30,000, ATTRIBUTED STATEMENT, unconfirmed primary) < Figure 03 rumored target (~$20,000, ATTRIBUTED STATEMENT, no primary confirmation) << Agility Digit (~$250,000, ATTRIBUTED STATEMENT via multiple secondary sources). The whitepaper’s €25,000 illustrative figure (~$27,000 at typical EUR/USD rates as of mid-2026) falls between the Unitree/1X-NEO consumer band and the Tesla Optimus long-term target, but well below the Agility Digit enterprise price, and well below Figure/Tesla platforms that in practice have no confirmed commercial retail price at all as of 2026-08-12. Derived from the individual claims and citations above (see Unitree G1, 1X NEO, Tesla Optimus, Agility Digit, Figure 03 entries in this cluster) 2026-08-12 This is a comparative synthesis across the individual findings above, several of which (Optimus, Digit, Figure 03) rest only on secondary-source corroboration rather than primary-source confirmation; treat the overall range as directionally reliable but with real uncertainty on the high end given no shipped/commercially priced enterprise humanoid robot was directly confirmed via a primary source.
H: Author’s own published articles — accurate-representation and date check LinkedIn article ‘Who Owns Your AI’ by Valto Loikkanen was published and its content can be verified for accurate representation. OBSERVED_FACT Fetched https://www.linkedin.com/pulse/who-owns-your-ai-valto-loikkanen-nq4lc/ directly. The fetch tool reported a publish date of August 11, 2026. Content covers three layers: (1) Compute as Infrastructure — Nvidia and infrastructure investors financing AI compute ($500B+ initially, trillions ultimately) as an asset class like electricity grids/telecom towers; (2) Intelligence as Utility — quoting Sam Altman that capable intelligence becoming abundant/inexpensive makes the model itself look more like infrastructure than a final application; (3) Personal Superintelligence — citing Mark Zuckerberg’s concept of a personal AI agent with privacy guarantees and distributed power. The article frames information as ‘the operating system for AI’ and argues true sovereignty requires ownership of infrastructure, information, and agents, plus portability of context between providers. No pricing or membership-cost figures appear in this article (that figure belongs to the separate digital-sovereignty article). This representation appears consistent with the article’s stated argument as extracted; however this was obtained via an automated fetch/summarization tool rendering LinkedIn’s page rather than a manual read of saved raw HTML, so exact wording nuances were not independently re-verified against a saved raw copy. https://www.linkedin.com/pulse/who-owns-your-ai-valto-loikkanen-nq4lc/ (fetched 2026-08-12; tool-reported publish date August 11, 2026) 2026-08-12 Moderate-high confidence in date and thematic summary since fetched directly from the live URL, but LinkedIn pages are commonly rendered via client-side JS and dates shown by automated extraction can occasionally be off by a day due to timezone conversion; recommend the author confirm the exact publish date shown in their own LinkedIn dashboard.
H: Author’s own published articles — accurate-representation and date check LinkedIn article ‘So the majority agrees: the time for digital sovereignty has come, now what?’ by Valto Loikkanen was published on Aug 5 2026, and contains a figure of approximately €42 per member per month. OBSERVED_FACT Fetched https://www.linkedin.com/pulse/so-majority-agrees-time-digital-sovereignty-has-come-now-loikkanen-vfgbf/ directly. The claimed date of Aug 5, 2026 is CONFIRMED. The claimed ~€42/member/month figure is also CONFIRMED VERBATIM: the article contains a cost-sharing calculation stating ‘20 members: approximately €105 per member/month; 50 members: approximately €42 per member/month; 100 members: approximately €21 per member/month.’ Both the date and the figure as claimed are accurate and correctly represented — no correction needed. https://www.linkedin.com/pulse/so-majority-agrees-time-digital-sovereignty-has-come-now-loikkanen-vfgbf/ (fetched 2026-08-12; publish date Aug 5, 2026 as stated on the page) 2026-08-12 High confidence — both the date and the exact monetary figure were independently retrieved from the live article and match the claim exactly.
H-addendum (2026-08-13): what basis produces €42/member/month The author directly clarified the assumption set behind his own €42/member/month (50 members) and €21/member/month (100 members) figures: an illustrative €100,000-excluding-VAT capital cost for a shared workstation-class machine (e.g. an NVIDIA DGX Station-type system), financed 5yr @ 6.5%, plus electricity/overhead — not a DGX Spark-based figure, and not based on any published hardware list price. SCENARIO_ASSUMPTION (the €100,000 capital figure) supporting a DERIVED_CALCULATION (the resulting €/member/month figures) Author direct communication, 2026-08-13, cross-checked against NVIDIA’s own DGX Station product page (nvidia.com/en-us/products/workstations/dgx-station/, checked 2026-08-13), which confirms NVIDIA publishes specs (748GB coherent memory, 20 PFLOPS FP4, models up to 1T parameters) but no retail/list price for this system — consistent with the author’s statement that €100,000 was his own illustrative assumption, not a quoted market price. Author communication; nvidia.com/en-us/products/workstations/dgx-station/ 2026-08-13 This resolves the previously-flagged ‘€42 vs €19.50–23/member/month’ discrepancy (see Cluster A entry below and Token-Factory Scenario Workbook §1): the two figures describe different hardware tiers (illustrative €100,000 DGX-Station-class vs. verified $4,699 DGX Spark), not competing estimates of the same cooperative. Downstream deliverables have been corrected accordingly.
H: Author’s own published articles — accurate-representation and date check Substack article ‘The overlooked AI scale-up layer’ was published May 22 2026 and argues that delivery innovation (not new products) and a ‘globally local’ model let owner-managed SMEs become the overlooked AI scale-up layer. OBSERVED_FACT Fetched https://startupcommons.substack.com/p/the-overlooked-ai-scale-up-layer-846 directly. The claimed May 22, 2026 date is CONFIRMED. The argument is accurately represented: the article positions owner-managed SMEs (established firms with existing customers/revenue, unlike startups or large bureaucratic corporations) as the overlooked opportunity layer. It explicitly frames the innovation as being in delivery/distribution rather than the product itself (‘The product or service may not be new. The way it is delivered may be new.’), and describes a ‘globally local’ expansion model via direct expansion or AI-driven franchising/licensing/local operator networks/partner-led expansion, where geographic and language barriers weaken while local presence/character is preserved. Both the date and the delivery-innovation/globally-local argument as claimed are accurate. https://startupcommons.substack.com/p/the-overlooked-ai-scale-up-layer-846 (fetched 2026-08-12; publish date May 22, 2026 as stated on the page) 2026-08-12 High confidence — date and thesis both independently confirmed from the live article via direct fetch.
H: Author’s own published articles — accurate-representation and date check Sequoia Capital has published a ‘services as software’ thesis. OBSERVED_FACT Confirmed the thesis exists but not as a standalone article titled exactly ‘Services as Software.’ It is presented as a section within Sequoia’s article ‘Generative AI’s Act o1: The Reasoning Era Begins’ (sequoiacap.com/article/generative-ais-act-o1/), authored by Sonya Huang, Pat Grady, and ‘o1’, published October 9, 2024. The article’s ‘Service-as-a-Software’ section states: ‘The cloud transition was software-as-a-service. Software companies became cloud service providers. This was a $350B opportunity. Thanks to agentic reasoning, the AI transition is service-as-a-software.’ The thesis argues AI companies will shift from selling software licenses to monetizing completed labor/outcomes (illustrated via Sierra’s per-resolution pricing model), expanding the addressable market from software into the much larger services sector (measured in trillions). A secondary source (a Substack analysis, beyondproduct2.substack.com/p/sequoia-capital-report-analysis-services) corroborates this, citing Sequoia’s estimate that enterprise service budgets are roughly 6x larger than software tool budgets, framing the shift as ‘selling outcomes’ vs ‘selling tools.’ https://www.sequoiacap.com/article/generative-ais-act-o1/ (Sonya Huang, Pat Grady, and o1; published October 9, 2024); secondary corroboration via beyondproduct2.substack.com/p/sequoia-capital-report-analysis-services 2026-08-12 High confidence the thesis and its content are real and accurately characterized. Note: if the whitepaper cites a Sequoia article specifically titled ‘Services as Software’ as a standalone piece, that exact title/URL was not found — the concept lives within the ‘Generative AI’s Act o1’ article (spelled ‘service-as-a-software’ in the original). The author should adjust the citation to point to this article rather than to any distinct ‘Services as Software’ URL if one was assumed.
H: Author’s own published articles — accurate-representation and date check Y Combinator has published an ‘AI-native services’ framing comparable to Sequoia’s thesis. UNVERIFIABLE Could not locate a specific, citable Y Combinator essay, blog post, or named framework titled or clearly framed as ‘AI-native services’ despite multiple direct fetches of ycombinator.com/blog, ycombinator.com/rfs, and multiple search-engine queries (Google, Bing, DuckDuckGo — most blocked by CAPTCHA or consent redirects, yielding no usable results). YC’s live Fall 2026 Requests for Startups list (fetched directly) does include a category called ‘AI-Native Compliance Infrastructure’ (automating regulatory monitoring/reporting to reduce specialized headcount) and other RFS categories (e.g., ‘New Operating Systems for the Physical World’, ‘AI for the Aging Population’) that discuss AI substituting for services-type labor, but none of these constitutes a general ‘AI-native services’ thesis comparable in scope to Sequoia’s ‘service-as-a-software’ piece. Searched: ycombinator.com/blog, ycombinator.com/rfs (fetched directly, 2026-08-12); no specific ‘AI-native services’ essay found 2026-08-12 Low confidence that such a document does not exist at all — search tooling available in this session was limited (no dedicated web-search tool; search engines mostly blocked). Recommend the author independently search or provide the specific source they had in mind, or drop the YC comparison / soften it to ‘YC’s Request for Startups language touches on similar AI-labor-substitution themes’ rather than citing a specific ‘AI-native services’ framing document.
I: AI Maturity Framework public findability There is a publicly posted ‘AI Maturity Framework’ by Valto Loikkanen, findable as of 2026-08-12, licensed CC BY 4.0, structured as a three-stage model (Learning/Internalizing/AI Native) across dimensions Working/Role of AI/Information. OBSERVED_FACT A public AI Maturity Framework by Valto Loikkanen IS independently findable on LinkedIn as of 2026-08-12, via two public video posts, both dated 2026-06-20 (fetched directly via HTTP, no login required): (1) https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS, posted 2026-06-20T21:02:12Z, and (2) https://www.linkedin.com/posts/valto_where-are-you-in-the-ai-maturity-map-activity-7474223638184058882-MbqI, posted 2026-06-20T22:16:19Z, headline ‘Where are you in the ai maturity map?’. However, the structure found does not exactly match the brief’s description. Post 1’s transcript describes a 4-LAYER ‘decoupled AI maturity ecosystem’ (Layer 1: Orientation Framework/matrix; Layer 2: Assessment Framework; Layer 3: AI Maturity Profile/radar chart; Layer 4: Specialized Tracks), and within Layer 1 states the framework ‘focuses strictly on the transition from learning to internalizing and eventually becoming AI native,’ while also describing progression as ‘from exploring to architecting.’ It names ‘the role’ (of AI, evolving from isolated assistant to agentic coordinator) and ‘information architecture and data control’ (Level 0 platform-centric to Level 5 information sovereignty) as two of the three vectors; a third vector is implied but not verbally named ‘Working’ in the transcript. Post 2’s transcript describes 5 numbered stages (explicit ‘Stage 5 architecture’), an ‘Organization DNA canvas’ with ‘smart targets and values’ sections, and three illustrative profiles (solo professional, entrepreneur/small business leader, enterprise architect). No CC BY 4.0 or any license statement was found on either post (searched page HTML for ‘CC BY’, ‘Creative Commons’, ‘license’ — zero matches on both). https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS (LinkedIn video transcript via schema.org JSON-LD, Valto Loikkanen, posted 2026-06-20); https://www.linkedin.com/posts/valto_where-are-you-in-the-ai-maturity-map-activity-7474223638184058882-MbqI (LinkedIn video transcript via schema.org JSON-LD, Valto Loikkanen, posted 2026-06-20, headline ‘Where are you in the ai maturity map?’) 2026-08-12 Transcripts were pulled from LinkedIn’s own page HTML (schema.org VideoObject ‘transcript’ field — LinkedIn’s auto-caption text), not by watching the video, so minor auto-caption artifacts (run-ons, mis-transcribed words) are possible. No standalone static image, article, PDF, or dedicated webpage version of the framework (separate from these two video posts) was found despite multiple targeted searches combining Valto Loikkanen / AI Maturity Framework / CC BY / peecos / PIOS / Prifina.
I: AI Maturity Framework public findability The stage names ‘Learning,’ ‘Internalizing,’ and ‘AI Native’ are used, in that order, as part of the framework. ATTRIBUTED_STATEMENT Confirmed as a real spoken phrase from Valto Loikkanen in the video transcript: the framework ‘focuses strictly on the transition from learning to internalizing and eventually becoming AI native.’ However, whether these three words are the literal printed column headers on the diagram itself (as opposed to a narrative description of the same progression, which elsewhere in the same transcript is also described as ‘from exploring to architecting’ and via numbered ‘Stage 1’ through ‘Stage 5’) could not be confirmed, since only the spoken/auto-captioned transcript was accessible, not the visual diagram. https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS (LinkedIn video transcript, Valto Loikkanen, posted 2026-06-20) 2026-08-12 This is an attributed statement (words said in the video, per LinkedIn’s own transcript) rather than a confirmed fact about the diagram’s literal on-screen labels, which were not visually inspectable with available tools.
I: AI Maturity Framework public findability The dimension names ‘Working,’ ‘Role of AI,’ and ‘Information’ are used as the three row/dimension labels. ATTRIBUTED_STATEMENT Partially corroborated. ‘Role of AI’ is clearly present in substance: transcript states ‘[the role] shifts from an isolated assistant that’s simply answers questions to an agentic AI coordinating work across various systems.’ ‘Information’ is also clearly present: transcript calls ‘information architecture and data control’ the ‘most critical vector of transformation,’ the ‘bottom row’ of the matrix. The literal word ‘Working’ as a third named dimension was NOT found in either transcript; the closest related language is discussion of ‘how the work actually gets done’ and ‘automated workflows’ in post 2, without it being explicitly labeled as a row called ‘Working.’ https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS and https://www.linkedin.com/posts/valto_where-are-you-in-the-ai-maturity-map-activity-7474223638184058882-MbqI (LinkedIn video transcripts, Valto Loikkanen, posted 2026-06-20) 2026-08-12 Absence in the spoken transcript does not rule out ‘Working’ appearing as an on-screen visual label in the diagram itself, which could not be directly viewed.
I: AI Maturity Framework public findability The framework is licensed CC BY 4.0. UNVERIFIABLE No license statement of any kind was found in either LinkedIn post’s page source (searched for ‘CC BY’, ‘Creative Commons’, ‘license’ — no matches in either page), and no separate standalone article/PDF/webpage carrying the framework where a license might be stated was located despite multiple search attempts (Bing, DuckDuckGo/lite web searches for ‘Valto Loikkanen’ + ‘CC BY’ returned no results). Page source of https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS and https://www.linkedin.com/posts/valto_where-are-you-in-the-ai-maturity-map-activity-7474223638184058882-MbqI, checked directly; supplementary web searches for license mentions 2026-08-12 Reported plainly as not found, per instructions to state this as an expected possible outcome rather than assume CC BY 4.0 applies.
I: AI Maturity Framework public findability Overall assessment of public findability and current phase-naming as of 2026-08-12. INTERPRETATION A version of the AI Maturity Framework is independently publicly findable, but as two LinkedIn video posts rather than a single static CC BY 4.0 article/image. The framework’s structure, as recoverable from LinkedIn’s auto-generated video transcripts, is a 4-layer ecosystem (Orientation matrix / Assessment framework / Maturity Profile radar chart / Specialized Tracks) with a 5-stage numbered horizontal axis narratively described both as ‘exploring → architecting’ and ‘learning → internalizing → AI native,’ an ‘Organization DNA canvas’ with ‘smart targets and values,’ and three named adopter profiles (solo professional, entrepreneur/small business leader, enterprise architect). This is a plausible but not exact match to the cluster brief’s specific description (3-stage x 3-dimension grid with dimensions Working/Role of AI/Information) — ‘Learning/Internalizing/AI Native’ phrasing is directly attributable to Loikkanen’s own words, ‘Role of AI’ and ‘Information’ concepts are clearly present, but ‘Working’ as a named dimension and a strict 3-column (rather than 5-stage) layout were not confirmed from the transcript text, and no CC BY 4.0 license was found anywhere. The whitepaper should not assert a confirmed CC BY 4.0-licensed static three-stage/three-dimension artifact exists; it should instead describe what was actually found (two LinkedIn videos with this narrative content) and flag the license and exact on-screen-label questions as unresolved. Synthesis of https://www.linkedin.com/posts/valto_activity-7474204985564598272-7RyS and https://www.linkedin.com/posts/valto_where-are-you-in-the-ai-maturity-map-activity-7474223638184058882-MbqI 2026-08-12 Limited by inability to view the actual video footage or attached diagram images directly — only LinkedIn’s machine-generated video transcript text (from page JSON-LD) was accessible via available tools. A human review of the actual diagram frames would be needed to confirm exact printed labels.
J: OpenAI Codex usage statistics OpenAI reports that Codex accounts for 99.8% of weekly output tokens generated internally within OpenAI. ATTRIBUTED_STATEMENT Confirmed verbatim in OpenAI’s own blog post ‘How agents are transforming work’ (openai.com/index/how-agents-are-transforming-work/, published June 25, 2026), retrieved via a Wayback Machine snapshot after the live page returned a Cloudflare bot-challenge (403) on direct fetch. Exact text: ‘For the average OpenAI worker, Codex usage now accounts for more than 85% of output tokens. Since Codex users tend to use more tokens than non-users, its share of overall tokens is even higher: Codex accounts for 99.8% of weekly output tokens generated within OpenAI.’ An accompanying chart in the page’s embedded JSON data shows the series reaching ‘OpenAI (99.8%)’ as of a data point dated 2026-06-11. This is OpenAI’s own self-reported internal telemetry about its own systems; independent press coverage (e.g., The Next Web, ‘OpenAI says nearly all its employees have switched from chatbots to Codex agents, but every number comes from OpenAI itself,’ thenextweb.com, June 25, 2026) explicitly notes ‘every number comes from OpenAI itself’ and that ‘no independent third party has verified any of the usage figures.’ The fact that OpenAI published this exact figure is an observed fact about the statement’s existence; the underlying reality of internal token share is OpenAI’s unverified self-report. https://openai.com/index/how-agents-are-transforming-work/ (June 25, 2026), archived snapshot via web.archive.org; corroborating context: https://thenextweb.com/news/openai-codex-agents-shift-employees-non-developers 2026-08-12 High confidence the quote is verbatim and the report exists as described. No independent audit of the underlying token-count methodology was found or is possible from outside OpenAI.
J: OpenAI Codex usage statistics By May 2026, 70.2% of sampled individual Codex users made at least one request estimated to exceed one hour of human work, and 25.6% made at least one request estimated to exceed eight hours of human work. ATTRIBUTED_STATEMENT Confirmed verbatim in the same OpenAI report. Exact text: ‘By May 2026, 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, 70.2% made one estimated to exceed one hour, and 25.6% made at least one Codex request estimated to exceed eight hours.’ A second passage adds trend context: ‘From December 2025 to May 2026, the share of users who made a request that was estimated to correspond to work that would take a person more than 30 minutes rose to 80.6%. The share making a request that would take a person more than one hour rose to 70.2%. The share requesting work that would take a person more than eight hours grew the fastest from a low base.’ Embedded chart data confirms the same figures with growth-rate annotations: ‘1h+ 70.2% (+98%)’ and ‘8h+ 25.6% (+1131%)’ relative to the December 2025 baseline. This is OpenAI’s own estimate of task difficulty (via its own internal ‘estimated human-hours’ methodology, not independently defined or audited), applied to a self-selected internal sample — so the underlying reality of task difficulty is an OpenAI self-report/estimate, though the existence and exact wording of the statement is directly confirmed against the primary source. https://openai.com/index/how-agents-are-transforming-work/ (June 25, 2026), archived snapshot via web.archive.org 2026-08-12 High confidence in verbatim accuracy of the quote and figures as published by OpenAI. The ‘estimated human-hours’ metric is OpenAI’s own internal estimation methodology and its accuracy/definition was not independently disclosed or verifiable.
J: OpenAI Codex usage statistics Some heavy internal Codex users generate 60+ agent-hours of Codex work per day. ATTRIBUTED_STATEMENT Confirmed verbatim, with a more precise framing than the round ‘60+’ figure. Exact text from the OpenAI report: ‘The heaviest users ask Codex to run many hours of agent work in a single day. By June 2026, users at the 99th percentile regularly generated more than 60 hours of Codex agent turns per day, distributed across multiple, parallel agents. As Codex became more powerful and parallelizable, users moved from only a…’ (sentence continues beyond the retrievable snippet). Note the precise scope: this is the 99th percentile of internal users, not a generic ‘some heavy users,’ and the ‘60+ hours’ is explicitly agent-turn-time run in parallel across multiple simultaneous agent sessions in a single calendar day (i.e., aggregate parallel compute-time, not one person single-threading for 60 hours). https://openai.com/index/how-agents-are-transforming-work/ (June 25, 2026), archived snapshot via web.archive.org 2026-08-12 High confidence in verbatim accuracy of the quote. As with the other figures, this is OpenAI’s self-reported internal usage data, not independently audited.
J: OpenAI Codex usage statistics NVIDIA has stated or published a reference agentic AI workload figure of 32,000 input tokens plus 8,000 output tokens per turn. UNVERIFIABLE Could not locate this figure anywhere. Extensive searching (dozens of query variants across Bing web search, Bing News search, DuckDuckGo, Ecosia, Brave, and Marginalia, plus direct fetches of NVIDIA’s developer blog and the SemiAnalysis InferenceMAX/InferenceX benchmarking site, which does publish token-count scenario labels like ‘DeepSeek R1 (8k/1k)’) turned up no NVIDIA publication, blog post, keynote transcript, or press report using a ‘32K input / 8K output tokens per turn’ reference agentic workload figure. Search engines increasingly fell back to generic/irrelevant results once specific numeric phrasing failed to match anything indexed, which is itself evidence no indexed page contains this exact phrase. It is possible this figure exists in a specific NVIDIA GTC slide, investor-day deck, or a paywalled/PDF technical report not accessible to text search, or that it is a rounded/paraphrased derivation from some other NVIDIA token-ratio statement rather than a direct quote — but I could not confirm either the figure’s existence or its exact source. No source found; searches conducted via www.bing.com/search, www.bing.com/news/search, duckduckgo.com, ecosia.org, search.brave.com, search.marginalia.nu, developer.nvidia.com/blog, and inferencex.semianalysis.com on 2026-08-12. 2026-08-12 This should be presented in the whitepaper as an unverified/unsourced figure, or dropped, unless a primary NVIDIA citation can be supplied by the original claimant. Do not present it as a confirmed NVIDIA reference figure.
J: OpenAI Codex usage statistics Related context: OpenAI’s report also states that ~98% of OpenAI employees now use Codex agents (up from ~40% in August 2025), non-developer individual usage grew 137x and organizational non-developer usage grew 189x since August 2025, and Codex output for OpenAI’s Legal team was 88% of tokens, Finance 91%, Recruiting 89%. ATTRIBUTED_STATEMENT Surfaced as surrounding context while locating the primary source for the 99.8% figure. These figures come from press summaries of the same OpenAI report (The Next Web: ‘Nearly 98 percent of OpenAI’s employees now use Codex… up from roughly 40% in August 2025… Individual non-developer usage grew 137 times… Organizational non-developer usage grew 189 times’; Crypto Briefing: ‘91% output tokens from finance department via Codex… 89% from recruiting department… 88% from legal department’). These were not independently re-verified against the primary OpenAI page text in this pass (time did not permit re-fetching every figure from the archived snapshot), so they rest on secondary-source summaries of OpenAI’s self-reported data rather than a directly re-confirmed primary quote in this check. https://thenextweb.com/news/openai-codex-agents-shift-employees-non-developers (June 25, 2026); https://cryptobriefing.com/openai-codex-internal-adoption-non-developer-surge/ 2026-08-12 Provided only as supporting context for the cluster; not one of the four claims explicitly requested. If used in the whitepaper, should ideally be re-confirmed against the primary OpenAI source text directly (as was done for the four main claims) rather than relying on secondary summaries.
← Back to home View source / edit on GitHub →