Hinton wants an FDA-style approval gate for AI models
A safety proposal with the field's most famous name behind it, a first hard look at who actually pays for consumer AI, and a new bet on squeezing more compute out of data centers that already exist.
Geoffrey Hinton wants AI models to clear a regulator the way drugs clear the FDA. The Nobel laureate made the case on Tuesday's "Smart Girl Dumb Questions" podcast: "You're not allowed to just make a new drug and release it on the market. You have to convince the FDA. And to do that, you have to do a lot of work, about a billion dollars worth of work. That seems like the very least we should have for AI." He tied the urgency to models improving themselves — "we're beginning to get recursive self-improvement" — and offered a timeline that is explicitly a guess: "a year or two before everything gets much worse than it is now." The timing is not accidental: the remarks land a week and a half after OpenAI scrapped the public launch of GPT-6.1 Astra for failing its own safety bar, the same lab whose head of safety systems said the model fell short on permissions and on communicating what it had done. The proposal is more credible than the analogy is precise — drug approval reviews a fixed artifact, while a deployed model changes under you via updates and tool access, so an FDA-style gate would have to review a moving target and define who is legally the "manufacturer." That is exactly the fight regulators are already having, and Hinton just gave it a slogan. He has spent two years arguing the risk is imminent; now he is naming a mechanism, which is a different kind of statement — we covered the lab-side version of this mood in Altman says the world must accept AI's 'bad things'.
Only 4.5% of US consumers pay for AI — and the top 1% of those payers spend $903 a month. a16z's seventh Top 100 Gen AI Consumer Apps report adds something the traffic leaderboard never had: observed spending on US consumer cards via YipitData. Just 4.5% of eligible consumers had an active paid ChatGPT, Gemini, or Claude subscription in August 2026, up from 2.1% a year earlier; the median payer spends about $25 a month, but the top 1% averages $903 and accounts for 19.5% of all observed consumer AI spending — more than the bottom half of spenders combined. The structural finding matters more than the number: 29 of the top 50 vendors by spending never appear on a16z's web or mobile traffic rankings at all, and only seven companies (ChatGPT, Claude, Suno, Perplexity, Photoroom, Canva, Notion) make all three lists. Read together, that says the consumer AI market has quietly split into a free-app popularity contest and a prosumer spending economy built on tools like n8n, fal, and Manus — usage charts are flattering the wrong companies. The caveat a16z itself flags: this is a US card panel, not company revenue.
Turba Labs emerged from stealth with nearly $52 million to double compute without building a single data center. The two-year-old Palo Alto company raised a combined seed and Series A — the $40 million Series A led by Creandum, with Cusp Capital also participating — to sell digital twins of data centers: software models of a facility that simulate where workloads should run before they run there. Co-founder Hans-Juergen Schmidtke, formerly an engineering director at Meta, and Patrick Jahnke, ex-SAP, have grown the company to about 65 people across Palo Alto and Heidelberg. The company claims a blended cost-per-token reduction of 9.4 times in a modeled five-workload example, while stating plainly that this is not a universal guarantee. The pitch lands because the alternative is getting harder: new data centers are bottlenecked on power, water, and permits rather than chips, so the cheapest compute still available is the capacity already wired to the grid — as we noted when Lambda raises up to $4B from Blackstone ahead of its IPO.
What to watch: whether any regulator picks up Hinton's FDA framing in hearings this fall — and whether OpenAI's Astra shelving gets cited as Exhibit A.
Should frontier labs have to prove safety to a regulator before shipping — and who should pay for the review? Tell us in the comments.
