Anthropic is setting the pace for the entire industry. The latest example: a deal with Musk's SpaceX, which happily shared the capacity of its new data center amid its legal battle with OpenAI (Anthropic's main rival).
For context, at the start of 2026 Dario Amodei's startup was valued at $380B. Back then that ceiling looked pretty solid.
I just looked at the freshest numbers, and I was floored. Anthropic's annual recurring revenue (ARR) has shown incredible growth and shot up to $45B+ — even though just a few months ago it was "only" $9B.
And that's when the real frenzy began. As the Financial Times reports, major funds like Lightspeed and Dragoneer are ready to hand over almost unlimited cash (we're talking sums of up to $50B) and are literally waiting for Amodei to give the green light for a new round at a valuation near $1T.
Investors are behaving exactly like Fry from Futurama (shut up and take my money), because the product is scaling at a frightening speed and there's a risk of missing the chance to invest in it.
The question remains: why do startups even need such astronomical budgets? It's a consequence of competition for a still-forming market and for compute:
The SpaceX deal. Anthropic struck a strategic agreement with Elon Musk's company, securing access to critical compute capacity.
OpenAI's counterstrike. The main rival isn't sleeping and is buying up infrastructure. According to Bloomberg, they contracted 10 gigawatts of power for their servers ahead of schedule, even though they originally set that goal for 2029. To do it, they even had to cancel some Stargate projects in Texas and the UK.
In just 90 days OpenAI grabbed 3 gigawatts of capacity, two of which will be provided by Amazon. For scale: a single gigawatt can supply electricity to roughly 750,000 homes.
What can I say — this looks like the final shift from a battle of algorithms and new concepts to a war of infrastructure (though Yann LeCun, who criticizes LLMs as a limited technology, would disagree with me).
Building AGI (human-level artificial intelligence) now comes down not so much to code or top talent, but to access to power plants and the ability to quickly fine-tune (narrowly retrain) new models on endless server racks.