Anthropic’s $100 billion Amazon bet: scaling fast and pinky-swearing safety
Anthropic is currently attempting the ultimate Silicon Valley tightrope walk: securing enough power to light up a small country while promising the world that its AI won't accidentally end it.
The company just announced a staggering $100 billion commitment to Amazon Web Services over the next decade. That's "fuck you" money even by Big Tech standards, designed to lock down up to 5 gigawatts of compute capacity for training and running Claude.
To put that in perspective, 5GW is roughly what it takes to power nearly four million homes, or, in this case, one very hungry large language model.
But because you can't just spend $100 billion on digital brainpower without looking a little like a Bond villain, Anthropic is also rolling out a "significantly updated" Responsible Scaling Policy (RSP).
Think of the RSP as a tiered warning system inspired by biosafety levels. It’s a graduated set of safety standards that get more intense as the models get smarter.
Right now, Anthropic says it’s operating at ASL-2—industry standard, basically. But once Claude starts independently conducting AI research or helping a technician with a basic background create a chemical weapon, the "ASL-3" alarms go off.
"We will not train or deploy models unless we have implemented safety and security measures that keep risks below acceptable levels," the company stated, essentially promising that they have a kill switch ready for when the math gets too spicy.
It’s a nice sentiment, but the pressure to scale is immense. Anthropic’s run-rate revenue has exploded from $9 billion at the end of 2025 to over $30 billion today.
That kind of growth creates what the company calls "inevitable strain" on infrastructure, leading to reliability issues for free and Pro users alike.
Amazon, for its part, is dumping another $5 billion into the pot today, with the potential for $20 billion more down the line. CEO Andy Jassy is betting big on custom silicon, moving Anthropic toward Trainium2 and the upcoming Trainium4 chips.
The tension here is palpable. Anthropic is hiring a "Head of Responsible Scaling" to keep things from going off the rails, even as they admit they’ve already missed a few procedural deadlines in the last year.
They claim these minor slips—like finishing evaluations three days late—posed "minimal risk." That may be true for now, but when you're playing with 5 gigawatts of power and $100 billion in hardware, the margin for "procedural issues" tends to get a lot thinner.
Let's plan for the long game: will the safety case be enough to hold back the tide of a $30 billion revenue stream that demands more power, more speed, and more Claude?
Sources: Announcing our updated Responsible Scaling Policy, Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute.
