A provocative framework for managing the most powerful technology in human history is drawing fire — but much of the criticism, according to a detailed analysis by Scott Alexander of Astral Codex Ten, rests on a fundamental misreading of what AI chip regulation actually involves. Plan A, a governance proposal designed to coordinate AI development between the United States and China, has been dismissed by some critics as a blueprint for an “Orwellian dystopia” or a “global panopticon.” The reality, Alexander argues, is considerably more mundane — and the comparison to how the US already regulates Xanax is more instructive than most critics want to admit.
Summary
Key takeaways
- Plan A requires AI chip factories, buyers, and data centers to register with governments and submit to inspections — comparable to existing controlled substance regulations.
- Chips deployed under Plan A would carry cryptographic software allowing either the US or China to halt their operations remotely at any time.
- Starting in 2030, training new open-weight AI models would be banned under Plan A, though companies must release the research behind training runs.
- Plan A’s direct economic impact is estimated at a few percent increase in AI chip prices, with no meaningful effect on consumer devices like phones or laptops.
- Several Trump administration measures enacted this January already mirror core elements of Plan A’s chip controls — largely without public notice or outcry.
What Plan A Actually Proposes on AI Chip Regulation
At its core, Plan A’s approach to AI chip regulation follows a familiar template. Factories that manufacture high-powered AI chips must register with the government and accept inspections. Companies purchasing those chips — Google is the example cited — face the same registration and inspection requirements, along with government permission requirements if they resell. Data centers hosting the chips must also register, submit to inspections, and demonstrate robust cybersecurity defenses.
Beyond that, the chips themselves would contain cryptographic software enabling either the US or China to halt their operations remotely. Data centers conducting AI training runs would be required to publish basic operational information — such as the scale of their training runs — to a public database, and to prove they are running the code they claim to be running.
The flagship chip referenced in the analysis, the H100, currently costs $40,000 per unit. That price point alone makes clear why Plan A’s authors believe consumer hardware — phones, laptops, tablets — sits safely outside the regulatory perimeter. No device priced below that threshold contains the relevant chips, and the economics of combining large numbers of smaller chips to approximate frontier AI compute remain deeply unfavorable due to latency and memory constraints.
How Burdensome Are These Requirements?
Alexander’s analysis benchmarks Plan A’s chip controls against an unexpected reference point: US controlled substance regulation. Xanax, a Schedule IV drug, costs $14 for a 30-day supply at market price despite decades of factory registration, inspection regimes, and resale tracking. The regulatory overhead hasn’t collapsed pharmaceutical innovation or created a surveillance apparatus that ordinary Americans feel in their daily lives.
The estimate offered is that Plan A would move the AI chip industry’s regulatory stringency from roughly the 50th to the 95th percentile among US industries — a genuine increase, but not a civilizational rupture. Price increases of a few percent are likely, set against a backdrop where AI inference costs fall roughly 98% every year regardless.
The US-China Trust Problem and Why It Shapes Everything
Plan A’s most strategically significant feature isn’t the inspections — it’s the architecture of trust they create. The framework is deliberately designed to make a joint US-China AI regulation deal “trustless” in the technical sense: neither side can defect even if it wants to, because both countries have full visibility into where all the chips are and cryptographic control over their operation.
This matters because the alternative — two superpowers racing toward superintelligence on the assumption the other might cheat — produces the worst possible outcomes regardless of who wins. Plan A attempts to replace that dynamic with a verifiable, enforceable structure that doesn’t require either government to simply take the other’s word for anything.
The broader ambition is to replace a winner-takes-all race between one or two countries with roughly three to five countries and ten to fifteen companies, all operating at approximately equal capability levels, with no single actor able to gain a runaway advantage. By 2035, conditional on Plan A’s adoption, the projection is approximately ten AI companies holding at least 25% as much compute as the leading lab, spread across three to six different countries.
The Open-Weight Model Ban: A Real Cost, Not a Surveillance Issue
Plan A’s most contested restriction is its ban on training new open-weight AI models after 2030. This is, Alexander acknowledges, a genuine cost to openness — but it has nothing to do with surveillance. No one would search your devices for existing open-weight models. The restriction operates upstream, at the level of whether large companies are permitted to train and release new ones.
The economics may render the ban largely moot anyway. The most recent open-weight models in 2026 already cost over $100 million to train. If 2030-era models carry $10 billion price tags, it becomes an open question whether companies would spend that and give the result away for free regardless of any regulatory requirement.
Open Algorithms as the Alternative
Plan A’s answer to the freedom concern raised by restricting open weights is an open-algorithms mandate. Companies training AI systems would not need to release their final model weights, but would be required to release the underlying research. Any company with equivalent compute could then independently reproduce a comparable model. Given that Plan A distributes compute access across middle-power countries as a condition of joining the governance regime, this could realistically mean dozens of different companies operating under different regulatory jurisdictions — achieving the distributed resilience that open weights currently provide, through a different mechanism.
The Trump-Era Precedent Nobody Noticed
Perhaps the sharpest point in Alexander’s analysis concerns timing. Plan A’s critics warned of dystopian overreach — yet the Trump administration quietly enacted a range of substantially similar AI chip controls in January, months before Plan A was even published. These included requiring chipmakers to obtain bank-style KYC verification from customers, mandating that customers certify their own customers’ compliance, requiring physical security certifications, and submitting chips to third-party performance verification labs.
None of these measures generated significant public debate or warnings of imminent authoritarianism. The contrast is hard to ignore: quietly passed chip regulations attracted no alarm, while an explicit proposal by idealists trying to prevent catastrophic AI outcomes attracted heated accusations of building a surveillance state.
The monitoring that already exists around AI usage reinforces the point. When a mass shooter in Canada killed eight people in February, OpenAI revealed that the perpetrator had their ChatGPT account banned months earlier over troubling posts about gun violence. Approximately a dozen OpenAI staff had debated whether to alert authorities. Canada’s AI Minister Evan Solomon subsequently summoned OpenAI executives to Ottawa to discuss escalation thresholds for harmful content. This kind of monitoring — of consumer AI interactions, at scale — is already happening, entirely outside Plan A’s framework.
Debunking the Surveillance State Misconception
The surveillance state critique of Plan A collapses under scrutiny when measured against the actual regulatory environment. Financial transactions, pharmaceutical manufacturing, and consumer AI interactions are all already subject to extensive government oversight that most people accept without question. Banks report $9.99 sandwich purchases if the counterparty triggers a flag. AI companies monitor chat logs for potential threats.
Against that backdrop, requiring AI chip factories and large hyperscalers to file paperwork and accept audits sits comfortably within the range of standard government oversight — closer to how milk or eggs are regulated, in Alexander’s framing, than to anything approaching a panopticon. The essay pointedly notes that Plan A actually calls for “zero data retention” in consumer AI applications, which would represent a meaningful improvement over the monitoring currently in place.
The analytical tension at the heart of this debate is worth naming directly. Critics who accept existing chip export controls, KYC requirements, and AI chat monitoring as unremarkable, but treat Plan A as an existential threat to freedom, are applying an inconsistent standard. The question isn’t whether AI chip regulation carries costs — it does, and Plan A’s authors acknowledge them explicitly. The question is whether those costs are proportionate to the risks being managed, and whether the alternative to governance is actually freedom or simply ungoverned concentration of power in whoever wins the current race.
FAQ
What industries and entities does Plan A regulation target?
Plan A targets AI chip manufacturers, companies that purchase chips such as Google, and data centers that host them. All three categories face registration and inspection requirements. Data centers must also demonstrate cybersecurity standards and, for training runs, publish operational data to a public database.
Does Plan A create a surveillance state or mass monitoring of consumers?
No. The analysis argues that Plan A’s requirements — factory registration, customer inspections, and data center audits — are standard government oversight mechanisms comparable to how controlled substances or food products are regulated. The proposal actually calls for zero data retention in consumer AI, which would reduce some existing monitoring.
Why does Plan A ban training new open-weight AI models after 2030?
The ban aims to prevent unregulated proliferation of highly capable AI models that cannot be safeguarded once their weights are publicly available. In place of open weights, Plan A mandates open algorithms — companies must release the research behind their training runs, allowing any similarly resourced entity to independently reproduce comparable models.
Will Plan A regulations significantly raise consumer hardware prices like phones or laptops?
No. Plan A’s direct economic effect is estimated at a few percent increase in AI chip prices. Frontier AI chips like the H100 cost $40,000 per unit, putting them categorically apart from consumer hardware. Phones and laptops would not be meaningfully affected under the current framework, though a dramatic future explosion in consumer hardware compute capacity could prompt additional measures.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
