242 billion dollars. That's how much money poured into AI in the first three months of this year alone — roughly four times what the industry saw in the same period last year. And this week offered a pretty clear picture of where it's all going. Meta signed another $21 billion deal with cloud provider CoreWeave, on top of the $14 billion it already committed, essentially buying up a small country's worth of computing power through 2032. Microsoft announced a $10 billion push into Japan, tying AI infrastructure to cybersecurity in a way that makes you realize this isn't just a tech race anymore — it's a geopolitical one. Amazon, not to be outdone, quietly let slip that its cloud AI business is now pulling in $15 billion a year. The plumbing of the AI economy isn't being laid anymore. It's setting.

The new models kept coming too. Google released Gemma 4, an open model built for the kind of AI that doesn't just answer questions but actually does things — books flights, writes code across multiple files, runs through multi-step tasks on its own. Anthropic went a different direction entirely, unveiling a cybersecurity-focused model called Claude Mythos that's apparently already found thousands of software vulnerabilities nobody knew about, deployed across more than 40 companies. It's a weird moment when the most interesting thing an AI can do isn't write your essay — it's hunt for threats in your code. Google also shipped Gemini 3.1 Ultra, which can hold two million tokens of context across text, images, audio and video at once, which is the kind of spec that sounds made up until you see it work.

But maybe the bigger story is quieter. AI agents — programs that go off and do things for you without constant hand-holding — are starting to plug into real-world tools at scale. Anthropic's open protocol for connecting agents to apps and databases crossed 97 million installs last month, and now every major player supports it. That's not a trend piece anymore, that's infrastructure. The company also started testing an always-on agent called Conway that just runs in the background, doing work while you do other things. Meanwhile, UK regulators put out a paper trying to define what "autonomous AI" even means, sorting it into five levels. They're clearly trying to get ahead of something they can feel accelerating.

Outside the AI bubble, it was a big week too. NASA sent four astronauts toward the moon — the first crewed lunar trip in over fifty years. SpaceX quietly filed paperwork for what could be the biggest IPO ever, potentially valued at $1.75 trillion. And back on the ground, a supply-chain attack on an open-source AI tool hit Mercor, a $10 billion recruiting startup, possibly leaking customer data. A useful reminder that all this speed comes with a cost: the faster everything moves, the more things there are to break.

Signals

Week in Review: April 10