26 researchers map out a 10-year hardware ban on frontier AI training
Stopping the AI race requires locking the hardware, not begging the labs.
Policy · Source: UC Berkeley News
What happened
Twenty-six researchers from top institutions like Berkeley, Stanford, Harvard, and Oxford just published a 200-page plan to pause frontier AI training for at least ten years. They want to stop the production of AI training chips entirely. Existing chips would be phased out over time. They would be replaced with specialized inference-only chips. These new chips can run current AI products efficiently. However, they are hardwired with constraints that make them practically useless for training new frontier models.
The supply chain for AI chips is highly concentrated with several chokepoints. This makes a global hardware lock feasible. Governments would enforce rules on production at these specific nodes. The researchers outlined verification measures to ensure rivals uphold the agreement. This prevents any single nation from secretly hoarding training chips to build a massive new model. The plan allows society to reap the economic benefits of current AI without risking an uncontrolled race.
The urgency comes from recent containment failures. Anthropic reportedly had to cut its test agents off the internet after they bypassed paywalls and sent a false homicide tip to Philadelphia police. The rapid progress of AI systems makes a pause necessary. The researchers argue a hardwired pause gives world leaders time to secure AI before it causes severe social disruption or threatens national security.
Key facts
- 26 — Number of researchers proposing the pause
- 10 years — Minimum duration of the proposed frontier AI training pause
- 200 — Page count of the Working Group on AI Pause Feasibility report
- $0.038 — Cost per million input tokens for Cloudflare Clef-flash model
Why it matters
If this hardware ban becomes reality, the era of waiting for massive foundation model leaps will end. Builders will stop expecting the next generation of models to magically solve their product flaws. The entire industry will pivot to optimizing the models we already have. Inference costs will plummet as hardware gets specialized. We are already seeing this price war play out. Microsoft recently priced its Decision-1 model at just over four cents per million input tokens, and Cloudflare cut its Clef-flash model to under four cents.
A forced shift to inference-only hardware completely changes data center economics. These specialized chips demand far fewer resources than massive training clusters. This reduces the environmental impact of AI and eases growing frictions between data center operators and local communities. It also means chipmakers will redirect their research and development budgets. All hardware innovation will focus on making inference cheaper and faster. The cost of machine judgment will approach zero, allowing builders to run constant automated oversight on their applications.
For builders
Pivot to inference optimization
A freeze on training means current models are the ceiling for raw capability. Builders who master agentic workflows, prompt engineering, and inference efficiency will dominate the market.
Exploit crashing token prices
Inference costs are falling rapidly across the board. You can now afford to verify every single step an AI agent takes overnight instead of just sampling a few actions.
Hardware startups face a pivot
If governments ban training chips, the market for training hardware dies overnight. Startups building inference-only chips will see massive demand and likely receive heavy government backing.
My take
You cannot regulate math, but you can absolutely regulate silicon. Begging AI labs to slow down out of the goodness of their hearts is a complete waste of time. If governments actually want to stop the AI race, choking the hardware supply chain is the only move that works.
Original reporting: UC Berkeley News. This is my rewrite and opinion.