Anthropic added $18 billion in annualized revenue in two months. The Claude maker is now running at a $65 billion annual rate, a jump that makes most hypergrowth companies look quaint by comparison.

That number matters because it shows what’s actually happening in AI right now. The story isn’t about new model architectures or research breakthroughs. It’s about deployment. Companies are spending real money to use these systems at scale, and that’s changing where the smart money is flowing.

The infrastructure play

Nvidia is putting $1.5 billion into SoftBank’s data center developer, the one building out OpenAI’s infrastructure. The deal guarantees Nvidia chips will power those facilities, which is a tidy way to lock in demand while helping finance the build-out. Nvidia also disclosed a $21 billion stake in SpaceX, following Elon Musk’s announcement of an exclusive arrangement to equip its data centers.

Then there’s Groq, which just raised $350 million at a $3.5 billion valuation. The company started as an AI chipmaker but is pivoting to what it calls “neocloud.” Translation: they’re expanding their Nvidia-powered data center footprint instead of trying to win the chip wars. When a chip company decides to become a cloud company running someone else’s chips, that tells you something about where the margins are.

What’s working, what’s not

Not everything is scaling up. Relay, an AI automation startup, shut down and its team joined Google’s Chrome group. Founder Jacob Bank said they have “ambitious plans to help you work with AI in Chrome,” which is the polite way to describe an acqui-hire. The automation layer that seemed promising a year ago is getting absorbed into the platforms themselves.

Meanwhile, Wispr raised $280 million at a $2 billion valuation as it pushes beyond dictation into meetings and note-taking. The company is betting that voice interfaces will be how people actually interact with AI tools, not chat boxes or command lines.

The Amazon problem

Amazon, which famously started by selling books online, is now destroying rare texts to train AI models. The company is literally pulping unique books to scan them, because rare texts are valuable training data. Everything available online has already been scraped, so the frontier is physical media that never made it to digital.

It’s a perfect illustration of how the AI arms race creates weird incentives. The company that democratized book buying is now destroying the books it can’t sell enough of, because the data is worth more than the object.

What it means

The pattern is clear. The companies seeing massive growth are the ones selling access to deployed models or the infrastructure to run them. Anthropic is growing because enterprises are actually using Claude. Nvidia is investing in data centers because that’s where the compute demand is. Groq is pivoting because selling cloud access is easier than selling chips.

The speculation phase is over. The deployment phase is here, and it requires different bets. Less about who builds the best model, more about who can deliver it reliably at scale. The $65 billion question is whether that market is big enough to support all the companies trying to grab a piece of it.