Korva is a scroll of the internet’s best ideas.
Withheld models stall revenue
Models kept back
The best model in the world was trained in February. OpenAI is not releasing Astra and stopped training for two weeks. Anthropic is not releasing what its safety assessment calls Model 2, widely believed to be the next version of Mythos. Anthropic also stopped giving Mythos to foreign employees for a time.
Why revenue stalls
If revenue per megawatt stalls, the labs lose the ability to outbid everyone else for compute. With no limits from safety, they could generate $100 million per megawatt, pay $50 million for compute, and every other owner would hand theirs over.
Rules add friction
Rules on data centers add to it. New York has banned data centers, Texas has moratoriums, and Ohio has tried to require developers to pay property tax for everyone within a certain radius. These cut supply and raise costs. Outwardly progress slows, while inside the labs the models keep improving, and a lab could hold its best model six months ahead of what the public sees, a gap that grows if progress speeds up.
Why do labs hold models back for testing?
Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework, both published in 2023, tie release decisions to evaluations of dangerous capabilities such as help with biological weapons. Models go through red-teaming and other tests that can take months. GPT-4, for example, finished training in August 2022 and was released in March 2023, according to OpenAI.
How do local rules limit data center growth?
Data centers need zoning approval, utility connections, and often water access, all of which fall under local control. A single large campus can draw as much power as a small city, so utilities and residents often object. Property tax rules and moratoriums also change what a developer can build.
Why does revenue per megawatt set bidding power?
A buyer can pay for capacity up to what that capacity earns, minus the buyer's other costs. A lab earning twice as much per megawatt as a rival can offer twice the price and still profit. When revenue stalls, that edge shrinks and other owners keep their capacity.