The US AI order regulates the frontier and leaves apps open

Executive Order 14110 ties reporting duties to training runs above 10^26 operations. Frontier labs and clouds carry the new costs; most AI-native apps do not.
Any training run above 10^26 operations now carries a federal reporting duty. Executive Order 14110 on safe, secure and trustworthy AI, signed by President Biden on 30 October, uses the Defense Production Act to direct the Commerce Department to require, within 90 days, that companies developing powerful dual-use foundation models report their training activities, the ownership and protection of model weights, and the results of red-team safety tests.
Until Commerce sets its own definitions, the duty covers models trained with more than 10^26 operations, or 10^23 for models trained mainly on biological sequence data, and computing clusters with a capacity of 10^20 operations per second. Commerce must also propose rules requiring US cloud providers to report foreign customers who train large models on them. The order regulates by compute rather than by use, the lever behind last year's chip export rules, and that choice decides who pays.
The new reporting costs land on the few companies that can afford training runs at the threshold and on the clouds that host them, all of which already run compliance teams. The thousands of companies building products on top of models face no new federal license, and sector regulators keep authority over applications in health, finance and housing. That favours a two-layer market: a concentrated, supervised frontier and an open application layer where product decides competition. It also encourages the Federal Trade Commission to use its powers on competition, directs agencies to ease visas for AI talent and launches a pilot of a National AI Research Resource.
An open application layer does not mean quiet buyers. The order tasks NIST with standards for red-team testing and Commerce with guidance on watermarking and content authentication, and federal agencies will apply both to their own purchases. A startup selling to government, healthcare or finance should build evaluation records, provenance labels and documentation now, since they will become procurement requirements. Teams training their own large models should track their compute, because crossing the threshold brings reporting duties and a standing relationship with the Commerce Department.
The thresholds can still move. The Commerce reporting rules due within 90 days will show whether they do, and the way NIST red-team standards turn into terms in federal contracts will show how much of the frontier's burden reaches application companies through procurement. Two days after the order, 28 countries and the European Union, the US and China among them, signed the Bletchley Declaration on frontier AI risks at the UK's AI Safety Summit; the signatories meet again in Korea within six months and in France in a year, and the open question is whether they agree on shared testing of frontier models.
Washington has drawn its line at the frontier, which puts regulatory cost where capital is already concentrated and leaves the application layer open. Fanaura works on the open side of that line, with AI-native companies that turn evaluation, provenance and documentation into an advantage in regulated markets, the starting point of our approach to founders. Regulation by compute taxes the few who train at scale and hands everyone else a procurement checklist to win on.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building evaluation, provenance or documentation tools for AI in regulated markets? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.