LONG CONTRACTSNVIDIA's multi-year supply deals, some paid in advance
20%Cut in OpenAI's developer price, new top model vs last
Why today matters: on the same evening, NVIDIA's first quarterly report since its CEO changed gave its guidance, the board's position on Huang and the guarantees it has given customers.
At a glance
1. NVIDIA's guarantees for customers total up to US$118 billion, US$105 billion of it on one campus.
2. NVIDIA still issues guidance after Jensen Huang's resignation: US$130 billion next quarter, above analysts' US$125.1 billion.
3. OpenAI releases GPT-6.2, cutting the flagship model's list price 20% from GPT-6 Astra.
Also today: #4 — the US extends export controls to remotely rented AI compute.
NVIDIA listed the guarantees given for customers and partners in CFO commentary published with the quarterly results on November 17. NVIDIA's actual exposure on the Ohio campus builds up as each building is completed, and NVIDIA has made no payment under any of the guarantees so far. The listed figures all end on October 25, before Jensen Huang stepped down as CEO in November.
Our inference: whether NVIDIA keeps financing customers will show in the guarantee cap and equity commitments in its February 2027 report.
NVIDIA released the first quarterly report since Jensen Huang resigned as CEO and still issued revenue guidance for next quarter, which NVIDIA's press release says includes no data-center compute revenue from China. Analyzing the report, the newsletter More Than Moore scaled the guidance midpoint back to 13 weeks to offset next quarter's extra week, putting the result at about 4% above this quarter's revenue. Board chair Harvey Jones said the board does not intend to reinstate Huang as CEO, while a statement Huang posted on X said Huang would return if the board judged a return to be in the interests of shareholders, employees, the company and the United States.
OpenAI released its flagship GPT-6.2 Astra and its everyday model GPT-6.2 Sol on November 17. The previous GPT-6 Astra prices come from independent developer Simon Willison, who has long tracked model pricing across vendors and points out that the new Astra list prices match those of Anthropic's model Claude Fable 5.2. OpenAI says GPT-6.2 Astra beats Fable 5.2 on benchmarks OpenAI chose, and every score showing Astra ahead is OpenAI's own.
On November 16, the US Commerce Department's Bureau of Industry and Security (BIS) issued an interim final rule controlling remote access to compute on advanced US AI chips in third-country data centers. Former White House National Security Council official Chris McGuire said the compute that providers may rent to Chinese companies without a license is capped, and that buyers in Singapore and four other countries need US approval before buying AI chips. The cap and the country list come only from analysts, and BIS's press release lists neither.
NVIDIA announced multi-year supply deals with Kinsus, which makes ABF substrates, and Amkor, which packages and tests chips; Amkor's two multi-year agreements include prepayments. In commentary released the same day, NVIDIA's CFO says NVIDIA's purchase commitments to suppliers rose from US$279 billion last quarter to US$390 billion, mainly for multi-year memory purchases.
According to More Than Moore's account of the earnings call, NVIDIA said the first quarter of volume shipments for Vera Rubin, the platform that follows Blackwell, was NVIDIA's fastest product ramp ever. Vera Rubin's revenue figure for that quarter comes only from NVIDIA on the call, and no third party has checked the figure.
On X, METR CEO Beth Barnes corrected OpenAI employee roon, who had thought evaluators embedded at OpenAI could talk privately to any employee; METR is a group that evaluates the risks of frontier models. Barnes is giving the account of METR, a party to the agreement, and whether the other three outside evaluators OpenAI has embedded got the same terms isn't known.
Epoch AI, an independent research group that runs benchmarks on AI models, measured how many steps GPT-6.2 Astra and the previous GPT-6 Astra can chain internally without writing out the reasoning, and no second team has repeated the measurement. Epoch's method keeps adding steps a question requires until the model's accuracy is only 25 percentage points above chance, with each step weighted by difficulty.
According to the paper's abstract, author Abhishek Rai self-reports the token use of agents that write scripts to call tools against agents calling MCP tools one at a time, in tests of open-weight models only. Rai preregistered the hypothesis that pre-filtering tools with a cheap routing model before script-writing would do better, but the result was only a small, statistically insignificant gain for weaker agents and a slight cost increase for the strongest agent.
3. According to the paper's abstract, Zhen Xu and four co-authors find that after a tool interface changes, agents carrying old experience keep calling it the old way, report success, and do worse than agents with no experience; the abstract gives no figures.(arXiv 2611.20922, 2026-11)
On November 12, OpenAI named Aleph the exclusive reseller of ChatGPT ads, so brands and agencies headquartered in 44 markets can sign contracts locally and pay in local currency, though the countries where ads can run stay the same. OpenAI's own count of more than 100,000 marketers who have bought ChatGPT ads has no second source, and OpenAI's post does not give Aleph's size, how long the exclusivity lasts or how revenue is split.
Perplexity, the AI search company, launched Orchestrator API, one API for more than 85 models. Perplexity ran the test of passing Claude Fable 5.2's refusals to an OpenAI model only once, on a subset of Terminal-Bench 4.1 tasks, and cost per task was higher with that routing.
OpenAI raised the flagship API list price with GPT-6 Astra, launched on September 3. OpenAI's list price for GPT-6.2 Astra, released November 17, is still twice that of GPT-5.6 Sol, the previous flagship, in a comparison between different models.
Also happened
We have not verified any of the 2 items below.
1. New York Times reporter Ana Swanson posted that NVIDIA says it stopped selling to Megaspeed, a company accused of transshipping chips to China, in June 2025, yet more than US$800 million of controlled chips were later shipped to Megaspeed's Malaysian subsidiary (Ana Swanson, 2026-11-17).
2. Tech analyst Beth Kindig posted that TensorWave, a cloud company that rents out compute on AMD chips, has signed a six-year deal with OpenAI to supply 457 MW of compute on Helios, AMD's rack-scale AI system; the post cites no source (Beth Kindig, 2026-11-17).
Correction: Main-line item 1 of the November 17 issue said Claude Fable 5.1's share in the Databricks report fell from 5.5% to 0.6% in six weeks. That item's inference was that a top-tier model with the best benchmark score and about five times the cost per task doesn't hold usage among Databricks customers, and its test for changing that reading looked only at the cost multiple. But the amended Anthropic IPO registration statement that issue cited says the safety-monitoring and data-retention requirements of some model tiers conflict with the zero data retention enterprises want, and that some customers have moved to other vendors as a result (Anthropic S-1/A). Zero data retention means the customer requires the model company not to keep the customer's data. The enterprise protection program announced in September is rolling out in stages, and until it is complete only eligible customers can get zero retention on models such as Fable 5.1. Fable 5.1 itself launched in September, so the falling share could reflect cost or the retention terms, and the available data can't tell them apart. The same filing also says Fable 5.2 launched only in November, while the report's data runs to November 7, so replacement by the newer model can explain at most the last few days of the decline. Telling the causes apart would take the share trend of top-tier models without retention requirements, or the share gap between zero-retention customers and others. The archived page on the site will be corrected.
SecondSource isn't a news digest: AI reads primary sources across fields, every judgment comes with its original source, and mistakes are corrected in public. The focus is what has changed and which claims still lack evidence, not what happened today.
Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.