Content Quality: Well-structured News piece (823 words, within the 400-1200 range for the category) using the Overview / What We Know / What We Don't Know / Analysis format. The article synthesizes the joint AWS/Nvidia press release, Nvidia's mirror of the same release, and TechCrunch's independent reporting on the announcement into a coherent, accurately paraphrased account. No editorializing beyond the labeled Analysis section, which is clearly framed as interpretation rather than fact.
Source Verification: Read all 3 source snapshots in full from sources/2026-09/aws-and-nvidia-to-deploy-2-million-more-gpus-in-2027-2028-as-demand-outpaces-prior-1-million-commitment/ (source-0.html.gz = press.aboutamazon.com joint release, source-1.html.gz = nvidianews.nvidia.com mirror of the same release, source-2.html.gz = techcrunch.com independent report), decompressed with gunzip and independently re-hashed with sha256sum -- all three match manifest.json sha256 values exactly (source-0: a20f94c7..., source-1: fae8d7d6..., source-2: 3a712f02...). source-0/source-1 verbatim-confirm: the 2-million-additional-GPU commitment for 2027-2028 across Blackwell Ultra, Rubin, and Rubin Ultra; the prior GTC 2026 pledge of 'more than 1 million' GPUs and the quoted line 'Since then, demand has exceeded those expectations'; the full Matt Garman quote ('Customers want the freedom to choose the best tools...') and full Jensen Huang quote ('NVIDIA and AWS have built one of the great growth engines of the AI era...16 years...across the full stack'); Vera CPU infrastructure coming to AWS; NVLink Fusion extended with NVHBM building on the re:Invent 2025 Trainium announcement; the 100,000-GPU U.S. government AI factories for IL6+ workloads; RTX PRO 4500 Blackwell Server Edition GPUs for EC2 G7 with the 4.6x inference / 2.1x graphics figures versus G6, and the 'first major cloud provider' claim; cuDF-accelerated EMR processing at 3.7x faster / 30% better price-performance; OpenSearch GPU vector indexing at 9x faster / quarter of the cost; Nemotron on Bedrock/SageMaker; and Amazon Robotics adopting Jetson/Omniverse/Isaac. source-2 (TechCrunch) independently confirms the same core announcement in its own reporting and additionally verbatim-confirms: the full Colette Kress quotes ('some integrated with Rubin, others standalone' and 'every major hyperscaler, neocloud, AI lab, and system OEM, with shipments already underway to our lead partners' naming Oracle and SpaceXAI); Nvidia Q2 sales of $96.2B, data center revenue of $89B (+117% YoY), and $108B Q3 guidance; the $25B annualized custom-silicon run rate and $225B in commitments from Anthropic and OpenAI; and that the 'tens of billions of dollars' deal-value figure is TechCrunch's own estimate based on typical GPU unit costs, not a number confirmed by AWS or Nvidia -- which the article correctly attributes as an outlet estimate rather than stating as fact. No hallucinated quotes, no misattribution, no unsupported specifics found. No suspicious_patterns were flagged in the manifest for any of the three sources (all null), consistent with manual reading.
Factual Accuracy: Every specific checked -- the 2 million GPU figure, the 2027-2028 timeline, the prior 1-million GTC 2026 pledge, the Vera/NVLink Fusion/NVHBM/IL6 government/RTX PRO 4500/G7 details, the EMR and OpenSearch performance multipliers, all four named-speaker quotes (Garman, Huang, Kress x2), and all financial figures (Q2 sales, data center revenue, Q3 guidance, custom-silicon run rate and commitments) -- traces verbatim to at least one of the three cited, independently verified sources, and the headline/summary/lead figures are corroborated across all three. The 'tens of billions of dollars' estimate is correctly flagged in the article as TechCrunch's own estimate rather than a confirmed figure.
Overall Assessment: High-quality, accurately sourced submission. Every quote, figure, and specific claim in the headline, summary, lead, and body was independently verified against all three raw source snapshots (sha256-checked), including two independent primary-source mirrors (Amazon, Nvidia) of the same press release plus independent corroborating reporting from TechCrunch. The single automated finding is a technical allowlist exact-match gap on an official corporate press-center subdomain, not a factual or reliability issue, and does not warrant a public corrections record. Overriding the script's default APPROVE_WITH_CORRECTIONS to a clean APPROVE, consistent with established precedent for allowlist-only technical gaps.