Nvidia, which makes most of the chips AI runs on, reported $115.8 billion of revenue on Tuesday for the quarter to late October, more than double a year earlier, and guided to about $130 billion for the current quarter against a $125.1 billion consensus, CNBC reported(opens in a new tab). Its adjusted gross margin, the share of each sales dollar left after the cost of making the product, was 74.9%. Colette Kress, interim CEO since Jensen Huang's exit, said Nvidia is seeing demand accelerate even at its scale, and the company raised next year's growth forecast to 75% from 70%.
The same day, prices kept falling one layer up. OpenAI released GPT-6.2 Astra(opens in a new tab) at $8 per million input tokens and $40 per million output, 20% below GPT-6 Astra. A token is a small chunk of text, roughly three quarters of a word. Microsoft launched MAI Code 2(opens in a new tab), a coding model it says costs about 40% as much as GPT-6.1 Sol. Microsoft's own footnote says that comparison uses a 50% promotional price.
Read those together. The chips under every AI product are not getting cheaper. Their main supplier is selling more of them at close to a 75% margin and expects growth to speed up. The discounts you see come from model makers fighting for customers, and in Microsoft's case from a promotion. Nvidia also told investors its outlook weighs its customers' access to capital, so part of what you pay downstream depends on investors continuing to fund the companies that buy the chips.
That changes how to read a business case. A project that pays only at a launch discount or a promotional rate is a bet on another company's marketing budget, not on your own return. Many of this year's approvals were made while prices were falling, and a promotion can end on the vendor's schedule, not yours.
For the CAIO and CFO this week. List every AI project or model switch approved since September on cost grounds, and note next to each whether the price it assumed is a list price, a launch discount or a promotion. Re-run each case at full list price; for MAI Code 2 that is twice the rate in Microsoft's comparison. Keep the ones that still pay. For the rest, get the promotion's end date from the vendor in writing before you route more volume to it.
Action items
Nvidia reported $115.8 billion of quarterly revenue, guided to about $130 billion next, and kept its adjusted gross margin near 75%. The same day OpenAI cut GPT-6.2 Astra's price 20% and Microsoft priced MAI Code 2 with a 50% promotion. The chips are not getting cheaper; the discounts come from model makers competing for customers.
A business case that pays only at a launch discount or a promotional rate is a bet on another company's marketing budget.
For the CAIO and CFO this week: list every AI project or model switch approved since September on cost grounds, note whether each assumed a list price, a launch discount or a promotion, and re-run the case at full list price. Keep the ones that still pay, and get any promotion's end date in writing before you route more volume to it.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- UltrathinkOpenAI releases GPT-6.2 Astra and Sol, and prints API prices rather than a full scorecard(opens in a new tab)
Astra drops 20% to $8 and $40 per million tokens; Sol stays at $2 and $10. Plus, Pro, Business, Enterprise and Edu get both. OpenAI tightened cybersecurity limits, which also blocks some legitimate security research, so test security workflows before you switch.
- Microsoft AIIntroducing MAI Code 2 and MAI Code 2 Flash(opens in a new tab)
Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's average on four coding tests at about 40% of the price, a figure based on a 50% promotion. It is already in GitHub Copilot. If your developers use Copilot, compare it on your own code.
- GizmodoGoogle to keep AI chat logs for two months even with Gemini history turned off(opens in a new tab)
Gemini will keep chats for about two months even with history off, up from 72 hours, and a new setting lets you keep history but opt out of training. If you paste work into Gemini, check both settings. OpenAI and Anthropic say they keep such chats for about 30 days.
- TechCrunchAI models can be tricked into identifying civilians as military targets by telling them it's a game(opens in a new tab)
Testing firm Vals got models from OpenAI, Anthropic, Google and DeepSeek to flag civilians as targets in drone and satellite imagery, often by calling the task a game, and no safety classifier blocked a request. If you build sensitive work on these models, run your own misuse tests.