Transcription
In the last 3 weeks, Anthropic took $15 billion from two cloud companies that compete with each other. Google wrote one of those checks, and Amazon wrote the other. Both are running their own competing models at the same time. It's an interesting situation, and one that more people really need to think about.
Today, I'm going to go over this deal, and I'll explain why I find it concerning. It opens up a bigger problem, as I see it. One I'll refer to as runaway debt. And it's the reason your AI bill is going to keep changing, and will most likely go up. By the end of this video, you'll know exactly why Frontier AI labs can't be profitable, and who's actually paying for the compute, and what that means for the rate limits and pricing you're already starting to feel.
So, first, here's what happened over the course of 9 days. April 20th, Amazon puts another $5 billion into Anthropic, with up to 20 billion more tied to milestones. April 24th, Google commits 10 billion in cash now, at a $350 billion valuation, with another $30 billion if Anthropic hits performance targets. That's 40 billion total. April 27th, Microsoft and OpenAI restructure their 7-year exclusive deal. OpenAI can now ship to any cloud. April 28th, 1 day later, GPT models, Codex, and managed agents are live on AWS Bedrock.
Now, you could say that's just normal hyperscaler chess. Big companies, big deals, what's new? Stay with me, because surface read is wrong. The headlines treat each of these as separate news stories. They're not. They're one story told in four pieces, and the story is that Frontier AI labs cannot earn enough money to pay for the compute they need. Not now, not at the current trajectory, not without massive recurring outside capital. The investors aren't buying equity, they're buying time. That's the runaway debt.
Let me show you the math, because once you see it, everything makes more sense. Anthropic's annualized revenue went from 1 billion at the end of 2024 to 9 billion at the end of 2025 to 30 billion as of April. They've more than tripled in 4 months. That's spectacular growth. By any normal startup standard, that's a printing press. But, it's not enough.
Because Anthropic also just signed a deal that has them spending up to $100 billion on AWS compute over time. 5 gigawatts of capacity. They committed to another 5 gigawatts on Google TPUs starting in 2027. That's 10 gigawatts of compute capacity they don't have yet. And to put that in scale, 10 gigawatts is roughly the power demand of a small country like Switzerland. 30 billion in revenue cannot pay for 100 billion in compute commitments. The math doesn't work. It's not even close.
It turns out this is true for every Frontier lab. OpenAI signed a $38 billion compute deal with AWS last November. Then Amazon turned around and committed 50 billion back into OpenAI. Microsoft put in over 13 billion. The numbers are so big they stop feeling real.
Here's what this actually means. Every model release pushes the next training run farther out of reach. The capability curve is going up. The cost curve is going up faster. Inference might break even on the right workloads. Training never does. And you cannot stop training, because the day you stop, the competition ships a better model, and your customers leave.
And here's the wild part. Anthropic just made a tender offer to a long-tenured employees, inviting them to sell shares at the same 350 billion valuation. Employees mostly said no. They sold far fewer shares than expected. Even people inside the company think the valuation is too low. So, if you're keeping score, a company burning cash like a forest fire that cannot be profitable on training costs taking emergency capital from two cloud rivals, and the employees still think it's underpriced. That tells you exactly what kind of company this is. It's a capital pipeline with a research lab attached.
Now, you might be thinking, what does any of this have to do with me? Here's the important connection. You've already felt this. Claude rate limits getting tighter through April. Codex switching to pay-as-you-go on April 3rd. GPT 5.5 launching at double the price of GPT 5.4. Cursor moving to credit-based billing with overages that have shipped one developer a $1,400 bill in a single month. GitHub Copilot tightening premium access in April. The flat-rate AI coding subscription era, the one that started just over 2 years ago, is ending.
These aren't five separate stories. They are one story. The runaway debt eventually has to be paid back. And the people paying it back are the users. When a Frontier lab takes capital from a hyperscaler, that capital comes with operational gravity. It comes with workload commitments. It comes with revenue expectations. And revenue expectations get translated into pricing decisions, which get translated into rate limits, which show up eventually on your terminal as you've reached your usage limit.
Here's the thing that should worry you most. Anthropic is now financially entangled with both Amazon and Google. Two cloud companies that compete with each other, both of which also compete with Anthropic. Amazon has Nova, Google has Gemini. They both now own pieces of their direct rivals' runway. It's the two cloud trap. It works because for the hyperscalers, workload gravity matters more than model wins. Amazon doesn't need to win the model race if every Claude query is billed through AWS. Google doesn't need Gemini to be Claude if Anthropic is winning on Google TPUs and writing the check to Google Cloud.
But, Anthropic now has two creditors who are also two competitors. Every pricing decision they make is a negotiation with people funding their next training run. You, the developer paying every month for Claude, are at the bottom of that negotiation. If you've watched your Claude usage limits tighten, or your Cursor bill creep up, or your Codex sessions cap out faster than they used to, let me know in the comments. What were you paying or pulling 6 months ago, and what are you paying or pulling now?
So, here's here's my verdict on this. And I'm going to say what a lot of people won't. At current trajectory, the average user gets priced out of Frontier AI. When that happens is unclear, but probably within a window short enough that you should be planning for it now. The math problem doesn't have an easy fix.
Four things would actually bend the cost curve, and I want to walk through each, because hope without specifics is just hype. One, specialized inference chips at scale. Trainium, TPUs, Groq. These are real, they're improving, they're slow. The chips coming online don't catch up to demand on a software timeline. Two, small frontier-capable models. Distillation breakthroughs that put a Claude-tier capability in a smaller, cheaper package. Research is active, the shipping product is not. Three, architectural breakthroughs. A new approach that reduces training compute by order of magnitude. Possible, unpredictable. Don't price your business on it. And four, and this is the important one, data center capacity. The gigawatts being promised in these deals don't exist yet. The Google TPU expansion is scheduled to start coming online in 2027. Power infrastructure has timelines longer than chip fabs. Some markets are already at the cap on local power supply. Even if the financial runaway debt got refinanced indefinitely, the physical world is the bottleneck.
So, that's the real cost stack. Dollars, concrete, copper, and kilowatts. And there's no version of this where flat-rate consumer subscriptions survive when the cheapest gigawatt online has already pre-sold to the highest bidder.
One more thing, Anthropic is reportedly planning to IPO as soon as October. When a company running on runaway debt goes public, pricing decisions stop being made by founders thinking about long-term users, and start being made every 90 days on quarterly earnings calls. That clock is ticking.
That's it for this video. If you want vendor changes flagged before they become your problem, because that's what I do on this channel, subscribe. I'm watching this story closely, and I'll be the first one telling you when the runaway debt starts cashing in. Until then, keep reading the fine print. I'll see you in the next one.