They haven't revealed what they are changing in the price, but it's probably cache hit prices. They subsidized theirs to 10x less than normal to drive adoption. That's almost certainly below the cost of electricity for them.
Chinese model providers are in a bind because charging a reasonable margin will put them in competition with all the neoclouds that host their model weights without the development cost. Their economics are far more challenging than OpenAI/Anthropic/Google, who already have thriving high revenue business and mountains of compute online or coming online.
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...)
I agree that the trend of token price should be going down but it should also consider the effect and gain per token so it varies among different models, mainly according to their capability and availability. There’s a long way to go before it meets the equilibrium state.
So I think we should not limit our field of view on software development but on realms where ai software has not penetrated yet. That’s the blue ocean for ai business. The forth industry revolution hasn’t begun yet. The new steam engine powered textile machine is on the stage of developing and experiment yet.
The question is whether they can penetrate those markets before going bankrupt
My take: no
They are still pretty much competing for software development, you see only tiny verticals being tried by OpenAI and Anthropic, with Claude Design et al
This can be better done by other companies, with a fresh balance sheet, don't you think?
There'll be plenty of people at Anthropic and OpenAI opening those companies
I agree you on these issues. When we look back to the first and second industry revolution, how many companies survive to today since then? It takes time and it’s a relay game, but more fast than past.
So I think there’re huge opportunities for everyone involving in.
Talking back to deepseek’s announcement, I think it is smart and funny but we needn’t to be panic on this forecasted price rising at all. Just sit and relax, and wait for big news announced by other model players. At least glm 5.3 is on its way.
They haven't revealed what they are changing in the price, but it's probably cache hit prices. They subsidized theirs to 10x less than normal to drive adoption. That's almost certainly below the cost of electricity for them.
Chinese model providers are in a bind because charging a reasonable margin will put them in competition with all the neoclouds that host their model weights without the development cost. Their economics are far more challenging than OpenAI/Anthropic/Google, who already have thriving high revenue business and mountains of compute online or coming online.
Interesting insights. I will consider these factors. AFAIK, the electricity price is lower in China than in US.
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...)
I agree that the trend of token price should be going down but it should also consider the effect and gain per token so it varies among different models, mainly according to their capability and availability. There’s a long way to go before it meets the equilibrium state.
So I think we should not limit our field of view on software development but on realms where ai software has not penetrated yet. That’s the blue ocean for ai business. The forth industry revolution hasn’t begun yet. The new steam engine powered textile machine is on the stage of developing and experiment yet.
The question is whether they can penetrate those markets before going bankrupt
My take: no
They are still pretty much competing for software development, you see only tiny verticals being tried by OpenAI and Anthropic, with Claude Design et al
This can be better done by other companies, with a fresh balance sheet, don't you think?
There'll be plenty of people at Anthropic and OpenAI opening those companies
I agree you on these issues. When we look back to the first and second industry revolution, how many companies survive to today since then? It takes time and it’s a relay game, but more fast than past.
So I think there’re huge opportunities for everyone involving in.
Talking back to deepseek’s announcement, I think it is smart and funny but we needn’t to be panic on this forecasted price rising at all. Just sit and relax, and wait for big news announced by other model players. At least glm 5.3 is on its way.
Is DeepSeek a DeepSeek or it is cloude sonnet 3.5?
Show me if you could. I am interested in your answer.