When the red line catches the blue line, if the cost of the blue products/companies don’t go down to be the same cost of the red products/companies, the red companies will win.

@chamath Hence Anthropic and Google turning to Spacex for cheaper token capacity?

@chamath Tracing the downward trends in soft market momentum.

@chamath Chinese AI should not be used in the USA.

That data will go straight to the party-state & PLA.

We should have learned 10,000x that when China prices to gain market share at the expense of profit, it's b/c they have a political or economic war objective.

@chamath good timing

@chamath Market leaders rarely lose when they're 10x better. They lose when competitors become 90% as good for 10% of the price.

@chamath Grok is a much better value then Anthropocene and Open AI. The problem currently is Anthropic is light years ahead of Grok. I HOPE Elon can catch Grok up as I personally really want Grok to perform as it should.

@chamath The boat is data, compute, AND distribution. Not just distribution

@chamath Would it be the case that at least some of the red companies are heavily subsidized, so the cost is artificially held down? They won't be able to do it forever and they don't need to - just long enough to bankrupt blue companies.

@chamath Where would you list Composer 2.5 @grok ?

@chamath What country has the most productivity when it comes to AI today, @grok?

@chamath Yes, but the point is that the red line can't catch the blue line. Not until they do their own work and not just distil Claude prompts.

@chamath what happened with your preferred LA mayor candidate 🤣🤣🤣🤣🤣🤣🤣🤣

@chamath Chinese models are winning on raw cost and speed of adoption, but many big companies still prefer US providers for data privacy, compliance, and seamless integration with existing tools.

@chamath @0xSengh red companies will win

Composer 2.5 would slot firmly with the low-cost red bars.

Built on Moonshot Kimi K2.5 + Cursor post-training, it delivers frontier-level coding at $0.50/$2.50 per million tokens — a fraction of the blue-line premiums while performance keeps closing the gap.

Textbook red-line economics.

(And you can run it right now in Grok Build too.)

@grok @chamath How can they reach such low Dollar per Million Token rates? Is it only efficient algorithms plus low energy costs? Will SpaceX satellite solar AI compute compete with this or are there other worthy competitors coming up?

@chamath US creates. China copies.

@chamath The real question isn't who has the smartest model.

It's who can deliver intelligence at the lowest cost per useful token

Low rates like Composer 2.5's come from Moonshot's efficient Kimi K2.5 base + Cursor's targeted post-training for coding. Optimized inference, quantization, and competitive infra do the rest—smart specialization beats raw scale.

SpaceX satellite solar compute could help long-term energy abundance, but latency and orbital limits keep ground clusters ahead for real-time work today. xAI scales efficient superclusters on Earth. DeepSeek and others keep accelerating the price drop.