Published
H.R. 10538 would pause AI training above 10^25 FLOP and ban superintelligence. What it says, why it can't pass soon, and the exposure for $NVDA, $MU, $CRWV.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
The Ban Artificial Superintelligence Act (H.R. 10538) is a September 2026 bill from Rep. Greg Casar and Sen. Bernie Sanders. It would create a federal Department of Artificial Intelligence and pause training of every AI model above 10²⁵ FLOP until that department is fully staffed and has written its rules. It would also ban "artificial superintelligence" outright, with prison terms of up to 20 years. If it became law, frontier-model training in the US would stop the day it was signed. For investors, the companies hit hardest would be the ones that sell training compute, not the labs.
👉 Track the exposed layers live: the Custom Silicon basket (Nvidia, AMD, Broadcom and the ASIC designers) and the AI Cloud / Neoclouds basket are the two groups whose revenue depends most directly on frontier training runs.
The bill was a fringe story when it was introduced on September 23–24. It isn't anymore. On November 3, Democrats won both chambers of Congress. On November 4, Rep. Yassamin Ansari, one of the bill's House cosponsors, posted that the "top of the Democratic agenda on day 1 needs to be a pause on advanced AI development," and memory stocks extended their losses the next morning (24/7 Wall St.). We covered the broader sell-off in Why are tech stocks down today?. This post covers what the bill itself says, which companies it would hit, and how likely it is to pass.
We read the full text as introduced (govinfo.gov). The House version has eight cosponsors besides Casar, including Ro Khanna, whose district covers much of Silicon Valley. It was referred to the House Oversight Committee. Sanders introduced a Senate companion at the same time (Casar press release).
| Section | What it does |
|---|---|
| §3 | Defines an "advanced AI system" as any model trained on 10²⁵ or more FLOP. The Secretary must adjust it every year for efficiency gains, which in practice means lowering it. |
| §4–6 | Creates a cabinet-level Department of Artificial Intelligence with a Senate-confirmed Secretary and an advisory board. |
| §8 | Pause on enactment. No advanced system can be trained, fine-tuned or modified, and no unreleased one can be deployed. The pause ends only when the Secretary decides the department is "fully staffed" and its rules are final. There is no deadline. |
| §9 | Bans developing, deploying, possessing, funding, importing or transferring superintelligence or any system with "precursor characteristics." Every advanced model needs pre-deployment approval. |
| §10 | A model found to have precursor traits is cut off from the internet and must be rendered inoperative within 30 days unless it can be shown to be clean. |
| §12 | Building or distributing an advanced model requires a federal charter. Charter holders give the Secretary full access and agree to forfeit their IP and assets if the charter is revoked. |
| §13 | Up to 20 years in prison for lab policymakers and outside "rogue actors" who recklessly break the rules. Other employees get a 10-year ban from the AI industry. |
| §15 | Makes it US policy to prevent superintelligence worldwide, "which may include export controls on artificial intelligence computing infrastructure." |
There is no appropriation in the bill and no dollar figures anywhere in it.
The name suggests a ban on science-fiction AI. In practice, 10²⁵ FLOP is a low bar for 2026. Epoch AI estimates GPT-4 used about 2×10²⁵ FLOP in 2023. Meta's own paper puts Llama 3.1 405B at 3.8×10²⁵. Epoch's estimates put every current frontier model from OpenAI, Google, Anthropic, xAI and Meta above the line, most by a wide margin. For comparison, the EU AI Act uses the same 10²⁵ number, but only to trigger extra safety obligations. It doesn't pause anything.
The six "precursor characteristics" in §3(6) are also broader than they look. Two of them are:
Coding agents already speed up AI research, and published lab safety evaluations already report deceptive behaviour in controlled tests. A broad reading would put deployed products, not just future models, under the 30-day shutdown rule in §10. Section 9(b) also bans releasing any model that "may be foreseeably modified" to show those traits. That would effectively end open-weight releases above the threshold, and Meta ($META) is the most exposed listed company there.
The bill regulates model developers. The revenue loss would mostly land on their suppliers. Frontier training is the main reason hyperscalers and labs buy the newest accelerators in such volume, so stopping training removes the reason for much of the build-out.
| Layer | Example companies | What the bill does to them | Exposure |
|---|---|---|---|
| Training accelerators | Nvidia ($NVDA), AMD ($AMD), Broadcom ($AVGO), Marvell ($MRVL) | §8 halts the training runs these chips are bought for | Highest |
| HBM and memory | Micron ($MU), SK Hynix ($000660.KS), Samsung ($005930.KS), SanDisk ($SNDK) | HBM demand follows accelerator shipments | Highest |
| Foundry and packaging | TSMC ($TSM), ASML ($ASML) | Leading-edge and CoWoS capacity is sized for AI training demand | High |
| Neoclouds | CoreWeave ($CRWV), Nebius ($NBIS) | Customers are labs and hyperscalers training models; §12 charters gate new customers | High |
| Hyperscalers and labs | Microsoft ($MSFT), Alphabet ($GOOGL), Amazon ($AMZN), Meta ($META), SpaceX (xAI) | Can't train new frontier models, but capex falls and free cash flow rises | Mixed |
| Power and cooling | Vertiv ($VRT), GE Vernova ($GEV), Constellation ($CEG) | Inference and non-AI load keep growing, but the new-campus pipeline slows | Medium |
Three things stand out in that table.
Compute suppliers take the hit, not the people who pay for it. That matches how the market traded the election. Hyperscalers rose on November 4 while chip and memory suppliers fell, because lower capex is cash saved for the buyer and revenue lost for the seller.
The neoclouds are concentrated. Macroplane's supply-chain graph for CoreWeave shows Microsoft made up about 67% of its 2025 revenue, Meta has a contract worth up to $14.2 billion through 2031, and the filing says "all of our GPUs are NVIDIA." Most of that demand is training or serving frontier models. A pause plus a charter regime would hit a company like CoreWeave from both sides: lab customers can't train, and new customers can't sign up without federal approval.
*The AI Cloud / Neoclouds basket on Macroplane, equal-weighted, two-year view. The basket has given back much of its 2026 gain since late October.*
The listed lab exposure sits inside giant companies. OpenAI and Anthropic are private. Public investors hold them indirectly through Microsoft, Alphabet and Amazon, and hold xAI through SpaceX. For those companies, a training freeze is a strategic problem but a small hit to earnings. Their cloud businesses would keep serving existing models.
*The Custom Silicon basket on Macroplane: GPU designers, custom-ASIC houses and the EDA duopoly. Every name in it sells into AI training clusters.*
The odds are poor, even after the election.
Critics come from both directions. The free-market R Street Institute called it "AI authoritarianism", pointing to the open-ended pause and the risk of ceding the lead to China. AI-safety commentator Gary Marcus, who has argued for a temporary pause, still opposes the bill, calling a permanent unilateral ban "a guarantee of leaving the US behind."
The full bill isn't the risk. Pieces of it are, because they could survive the political process on their own.
The thing to watch is whether a pause shows up in the House Democrats' formal agenda for January, not just in individual members' posts. If it does, expect the risk premium on training-compute suppliers to stay high well beyond this week, whatever the bill's chances of becoming law.
It's H.R. 10538, introduced in September 2026 by Rep. Greg Casar with a Senate companion from Sen. Bernie Sanders. It would create a Department of Artificial Intelligence, pause training of models above 10²⁵ FLOP until that department writes its rules, ban superintelligence permanently, and require federal charters to build or distribute advanced models.
Not directly. The pause blocks new training, fine-tuning and unreleased deployments, so existing products could arguably keep running. But §12 requires a charter to "distribute" an advanced model, and no charters can exist until the department is set up. A model found to have "precursor characteristics" would also face a 30-day shutdown deadline.
The suppliers of training compute: Nvidia ($NVDA), AMD ($AMD), Broadcom ($AVGO), HBM makers such as Micron ($MU) and SK Hynix ($000660.KS), TSMC ($TSM), and neoclouds such as CoreWeave ($CRWV) and Nebius ($NBIS). Hyperscalers would lose the ability to train new frontier models but would save on capex.
Not in its current form. It expires when the 119th Congress ends on January 3, 2027. Even if it's reintroduced with Democratic majorities, it would need 60 Senate votes and would face a veto from President Trump. The near-term risks are hearings, export controls on AI chips, and state-level permitting fights.
It's the total amount of computation used to train a model. GPT-4 was about 2×10²⁵ FLOP by Epoch AI's estimate, so any model at or above GPT-4 scale would count as "advanced" under the bill. The EU AI Act uses the same number to trigger extra obligations for general-purpose models.
No. This article is for research and education only. Bill status and stock prices change quickly. Do your own research before making any investment decision.
It's H.R. 10538, introduced in September 2026 by Rep. Greg Casar with a Senate companion from Sen. Bernie Sanders. It would create a Department of Artificial Intelligence, pause training of models above 10²⁵ FLOP until that department writes its rules, ban superintelligence permanently, and require federal charters to build or distribute advanced models.
Not directly. The pause blocks new training, fine-tuning and unreleased deployments, so existing products could arguably keep running. But §12 requires a charter to "distribute" an advanced model, and no charters can exist until the department is set up. A model found to have "precursor characteristics" would also face a 30-day shutdown deadline.
The suppliers of training compute: Nvidia ($NVDA), AMD ($AMD), Broadcom ($AVGO), HBM makers such as Micron ($MU) and SK Hynix ($000660.KS), TSMC ($TSM), and neoclouds such as CoreWeave ($CRWV) and Nebius ($NBIS). Hyperscalers would lose the ability to train new frontier models but would save on capex.
Not in its current form. It expires when the 119th Congress ends on January 3, 2027. Even if it's reintroduced with Democratic majorities, it would need 60 Senate votes and would face a veto from President Trump. The near-term risks are hearings, export controls on AI chips, and state-level permitting fights.
It's the total amount of computation used to train a model. GPT-4 was about 2×10²⁵ FLOP by Epoch AI's estimate, so any model at or above GPT-4 scale would count as "advanced" under the bill. The EU AI Act uses the same number to trigger extra obligations for general-purpose models.
No. This article is for research and education only. Bill status and stock prices change quickly. Do your own research before making any investment decision.