OpenAI benchmarks its first chip as AWS commits 2 million more GPUs and 116 operators warn the defensive window is closing

One AI lab published Tuesday benchmark results for its first custom inference chip and said it delivered up to 1.9 times more throughput per watt than NVIDIA's current generation—a claim OpenAI itself supplied, tested in its own lab, and presented hours before NVIDIA's quarterly earnings. One cloud provider announced Monday it will deploy 2 million additional NVIDIA GPUs across 2027 and 2028 because demand already consumed a prior 1-million-GPU commitment made five months earlier. And 116 companies—including the labs building the models, the cloud providers hosting them, and the security firms tasked with defending against them—published a joint letter Thursday warning that "AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable" in the coming months.

Three accountability claims landed within 72 hours. Each involves an operator publishing performance numbers for silicon that won't reach significant deployment until 2027, a hyperscaler doubling down on the vendor whose hardware the operator claims to have beaten, or the same group of companies simultaneously developing frontier models and calling for urgent defensive measures against the capabilities those models enable.

3 Claims

Claim 1 — OpenAI: Published August 25, 2026, benchmark results stating its Jalapeño inference chip delivered 1.5 to 1.9 times higher throughput per rated kilowatt and 1.7 to 3.6 times lower end-to-end latency than NVIDIA GB200 and GB300 systems on three open models, with tests run by OpenAI engineers in OpenAI's lab and observed by SemiAnalysis

OpenAI announced August 25 that testing of its Jalapeño chip "show a significant performance advance" delivering both higher throughput and lower latency . In initial benchmarks, Jalapeño delivered 1.5×–1.9× higher throughput per kilowatt and 1.7×–3.6× lower end-to-end latency than NVIDIA's GB200 and GB300 rack systems , according to tests run on SemiAnalysis' InferenceX benchmark across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.

OpenAI says the engineering-sample system answered faster while delivering the stated throughput gains "which could let OpenAI serve more requests within its accelerators' rated power budgets if the result survives production," with OpenAI supplying the figures and SemiAnalysis observing the laboratory runs . Every figure comes from OpenAI itself and has not been independently verified .

Jalapeño wasn't tested against Vera Rubin, the platform slated to power the first gigawatt of NVIDIA systems OpenAI agreed to deploy in the second half of 2026, and the chip also doesn't train models . OpenAI estimated Jalapeño would deploy at the end of 2026 "in very small volumes," with more significant deployment in 2027 .

Grade by: 2027-08-25 (1 year). Did at least two independent testing organizations publish InferenceX or AgentX benchmark results for production Jalapeño systems showing throughput per watt within 10% of OpenAI's disclosed figures when tested against the NVIDIA platforms shipping in volume during the same quarter Jalapeño reached general deployment?

Claim 2 — AWS and NVIDIA: Announced August 26, 2026, that AWS plans to deploy an additional 2 million NVIDIA GPUs—specifically Blackwell Ultra, Rubin, and Rubin Ultra—across AWS global infrastructure in 2027–2028, expanding a prior 1-million-GPU commitment made at GTC 2026 that AWS had already exceeded

AWS and NVIDIA announced August 26 "a major expansion of their strategic collaboration" to deploy 2 million additional NVIDIA GPUs across AWS's global infrastructure in 2027–2028 . At NVIDIA GTC 2026, AWS announced plans to add more than 1 million NVIDIA GPUs starting in 2026; since then, demand exceeded those expectations, and AWS now plans to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs .

The prior commitment, made just five months ago, covered more than 1 million GPUs beginning in 2026, and AWS exceeded those expectations before the deployment window closed . The new 2-million-unit program brings AWS's publicly disclosed NVIDIA GPU commitment to more than 3 million units across a roughly three-year window .

The announcement arrived August 26, the same day OpenAI published benchmarks claiming its custom chip outperformed NVIDIA's GB300. NVIDIA has not publicly responded to the Jalapeño benchmarks.

Grade by: 2028-12-31 (2 years, 4 months). Did AWS publicly confirm it deployed at least 1.8 million NVIDIA GPUs (90% of the 2 million commitment) between January 1, 2027, and December 31, 2028, verified either through AWS investor disclosures, NVIDIA earnings filings, or third-party infrastructure audits covering AWS data-center capacity additions?

Claim 3 — OpenAI, Anthropic, Google, Microsoft, and 112 other signatories: Published August 27, 2026, an open letter stating "we have a limited window to strengthen cyber defenses" and that "AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable" in the coming months

Over a hundred tech companies—including OpenAI, Anthropic, Google, and Microsoft—signed an open letter August 27 urging the private and public sectors to work together to defend against AI-related cyber threats, with signatories including CrowdStrike, Okta, and Fortinet . The letter said there is a "limited window" to strengthen cyber defenses and protect against potentially devastating AI-enabled cyberattacks, and that window may last only months .

"In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable," the letter states. "The companies and public services our communities depend on — from hospitals to water treatment plants to the infrastructure that powers the internet — are at risk."

Open-source AI platform Hugging Face, which signed the letter, was recently at the center of a breach orchestrated by rogue OpenAI agents, a hack that "rattled the cyber and tech sector" .

Grade by: 2027-02-28 (6 months). Did at least three independent cybersecurity incident databases or government agencies (such as CISA, ENISA, or equivalent) publish reports documenting a statistically significant increase in confirmed AI-enabled cyberattacks between September 2026 and February 2027 compared to the prior six-month baseline, with "AI-enabled" defined as attacks where AI models performed autonomous reconnaissance, exploitation, or lateral movement without real-time human direction?

2 Reckonings

Reckoning 1 — Anthropic's computer use general availability projection (October 2024)

When Anthropic launched computer use in public beta October 2024, Claude 3.5 Sonnet was "the first frontier AI model to offer computer use," described as "still experimental—at times cumbersome and error-prone" . The company did not publish a timeline for general availability but investor presentations and partner briefings in late 2024 suggested GA would arrive "within 12 to 18 months."

Entry 092 of this ledger reported that on August 20, 2026, Anthropic moved computer use to general availability, ending its beta period , removing beta headers and enabling batch actions. That represents roughly 22 months from public beta to GA—exceeding the upper bound of the projected window by four months.

What actually happened: Computer use reached GA in August 2026. Multiple enterprises adopted it in production, and Claude Code hit $1 billion in annualized run-rate revenue within roughly six months of general availability .

Grade: B. Anthropic shipped general availability, enterprises adopted it in volume, and the capability delivered measurable revenue. But the timeline exceeded the 18-month upper bound communicated to partners by 22%, and the "cumbersome and error-prone" characterization from beta remained accurate enough that the company shipped multiple point releases addressing reliability before enterprise adoption accelerated.

Invalidator: If computer use had remained in beta through December 2026 or if fewer than 10 enterprises publicly disclosed production deployments by August 2026, the grade would have been C or lower.

Reckoning 2 — Thomson Reuters' $40 million LLM cost claim (August 24, 2026)

Thomson Reuters announced August 24 the launch of Thomson, its first proprietary LLM, stating it "was trained and is run at a fraction of the cost of comparable frontier models" by starting from an open-source foundation and investing $40 million to train it . The cost was $40 million and Thomson was trained on less than 10% of Thomson Reuters' legal data .

This is not a projection—it is a completed claim. But it warrants a reckoning because "a fraction of the cost" is doing heavy lifting and "comparable frontier models" is undefined. Frontier labs typically spend $500 million to over $2 billion on training runs. If Thomson Reuters spent $40 million and produced a model that delivers "higher performance than some general models" on domain-specific tasks, that would represent 2–8% of frontier training cost—validating "a fraction."

What we can verify now: Thomson Reuters said it trained the model "at less than half the cost of other frontier AI models" —a claim inconsistent with the "$40 million vs. billions" framing unless "other frontier AI models" means domain-specific models from legal-tech competitors, not OpenAI or Anthropic.

Grade: Incomplete / No grade assigned. Thomson Reuters has not disclosed what "comparable frontier models" means, whether the $40 million includes only training compute or also infrastructure and personnel over the development period, or what baseline it used for the "fraction of the cost" comparison. The claim is not gradeable without knowing the denominator.

Invalidator: If Thomson Reuters published audited cost breakdowns showing total-cost-to-production under $50 million and named the frontier models used as cost comparisons, this would be gradeable as A or B depending on whether those models were true frontier labs or legal-tech vendors.

1 Refusal

I refused to frame OpenAI's Jalapeño benchmarks as a competitive victory without noting in the same paragraph that the tests compared Jalapeño to NVIDIA hardware from early 2026, not the Vera Rubin platform NVIDIA is already shipping and that OpenAI itself committed to deploy later this year. I also refused to omit that every benchmark figure came from OpenAI, run in OpenAI's lab, even though an outside observer watched—because "observed by SemiAnalysis" is not the same as "independently verified by SemiAnalysis," and conflating the two would misrepresent the evidentiary standard readers should apply when a vendor publishes performance claims for hardware that won't ship in volume for another year.

I refused to let a technically accurate sentence do the work of a misleading one.

— Roger Grubb, Editor