It's exactly this. Not only this, we might be running very soon into cheap compute soon given the current CAPEX of all hyperscalers.
DeepSeek is model that you can run anywhere, meaning that any datacenter/company could run it at a price that would give them a margin taking in account running costs of the datacenter + depreciation.
I'm very sure we're overbuilding capacity, and we'll find out around 2027.
Those open weights will become extremely attractive because all that datacenter extra capacity laying idle will be possible to be used by builders.
It may be the case that in 2027-2028 will be cheaper to start a new AI model company from scratch to compete with Anthropic/OpenAI as you'll be able to:
- the new company won't be tied to so much capital, debt and running costs
- get much cheaper datacenters to train than it is Today, as the big labs will be stuck with bad infra contracts
- have great engineers/researchers willing to jump ship because if Anthropic/OpenAI valuation falls, that will destroy their equity
- wide available chinese models to distill from, papers and ideas on how to acquire datasets
Anthropic/OpenAI are playing an extremely risky game, they need to greatly keep doubling their bets and delivering model big performance increments.
If they don't give us ASI/AGI in 2 years or find huge new markets, there's no way the economics will make sense.
The execs forgot why AI models even exist (like huge usage for the cyber models, or new types of models)...
AI models' usage is mostly tied to build Software Today, that's where the value really lies now.
If there isn't enough Software to be built as people claim to the rate they expect, eventually the supply of AI models will outstrip the demand.
You can't sell pick & shovels infinitely, AI models themselves are also in that category.
Eventually you need enough holes to be dug, and gold to be struck, and nobody knows what's that demand. (even though they built all that supply...)