Provenance Record

Well-structured News-category piece (780 words, within the 400-1200 range). Clear Overview/What We Know/What We Don't Know/Analysis format. Neutral tone throughout, no editorializing or unsourced superlatives. The 'What We Don't Know' section appropriately flags that Crusoe's MemoryAlloy vs. vLLM performance claims rest on the company's own figures, not independent third-party benchmarks — good editorial discipline.

Read all three source snapshots from disk (gunzip'd), verified sha256 of each decompressed file matches the manifest exactly (source-0: fc89fcef...8sbb5a; source-1: 3da50ce2...af8c25; source-2: e44b3484...8feef3e) confirming no tampering. (1) source-0.html.gz (TechCrunch, live fetch, 200): confirms $3.9B raise / $30.9B valuation, Atreides/Mubadala/Valor co-lead, Founders Fund/GIC/Nvidia/QIA/Radical/TPG participants, three-pronged revenue model, Abilene TX site used by OpenAI, modular Spark description, three new board members (Thomas Seifert, Bill Stein, JB Straubel incl. his 2021 investment and Tesla board seat), $13B/5yr Jane Street contract, Goldman Sachs/Morgan Stanley IPO banker meetings, prior $1.38B/$10B round 'last October', 2018 founding as crypto-mining-on-flared-gas, customers Meta/Microsoft/Oracle. All verbatim-consistent with the article. (2) source-1.html.gz (Pulse 2.0, live fetch, 200): confirms >$140B total contracted value, >6GW gross contracted capacity / ~1GW delivered, Crusoe Cloud bookings +20x YoY, Managed Inference >$100M contracted ARR, '9.9 times faster time-to-first-token and five times greater throughput than vLLM' via MemoryAlloy (matches article's quoted figures), Gartner Visionary / Artificial Analysis #1 / Nvidia Exemplar Cloud recognitions, Cognition/Figure/Perplexity customers, Spark 'years to weeks' construction claim, 1,800+ employees/5 countries/Bellevue+NYC offices, and the full CEO Lochmiller quote used in the article verbatim under 'KEY QUOTES' (two adjacent quote fragments, same speaker, merged into one continuous quote in the article — accurate, not a splice across different statements). (3) source-2.html.gz (Unite.AI, Archive.org fallback per manifest 'archive_fallback: true' even though live URL also returned 200 — used the committed snapshot per instructions): independently confirms the 9.9x/5x MemoryAlloy vs. vLLM figures and the first two sentences of the Lochmiller quote (the article's 'corroborated by Unite.AI' framing is accurate — Unite.AI does not carry the third sentence about scaling 'every layer,' which the article correctly attributes to Pulse 2.0 alone), plus $100M+ ARR, 20x bookings growth, capacity figures, board/customer facts. All three snapshots are internally consistent with each other and with the article body.

Every specific figure in the article — $3.9B raise, $30.9B valuation, prior $1.38B/$10B round, >$140B total contracted value, >6GW/~1GW capacity, 20x Cloud bookings growth, >$100M Managed Inference ARR, 9.9x/5x MemoryAlloy-vs-vLLM figures, $13B Jane Street contract, 1,800+ employees — traces verbatim to at least one cited source, with the MemoryAlloy figures corroborated across two. No hallucinated numbers, names, or quotes found. Source-to-claim mapping in the body matches the outlet actually carrying each fact (e.g., the Astra/AGI omission, and the correct TechCrunch-only attribution for the Jane Street contract even though TechCrunch itself cites Bloomberg as its own source).

High-quality, thoroughly sourced submission. All headline/summary/lead figures and technical claims verified verbatim across the three cited sources with sha256-confirmed snapshot integrity. The deliberately-excluded Astra/AGI claim was correctly kept out of the final article. The only automated flag was a source-allowlist configuration gap for two outlets with clear prior-use precedent, which does not warrant a public corrections record. Verdict: APPROVE.