AI Power Plays: Anthropic Snaps Up xAI's Compute in a Game-Changing Move

Anthropic just acquired all three hundred megawatts of compute capacity at xAI’s Colossus 1 data center. This isn’t just a big rental agreement between rivals; it’s a signal that the entire business model for building artificial intelligence is starting to fracture. For years, the strategy was to hoard compute. Now, one of the biggest players is selling it off, and another is buying it up, because building the infrastructure is starting to look like a different business than using it. This one deal, announced May sixth, changes the calculus for every major lab. It suggests the race isn't just about who has the most GPUs, but who has the smartest strategy for accessing them.

That was the move that dominated the conversation, but it wasn’t the only signal of a system under pressure. The other major event this week was much smaller, but just as telling. An autonomous AI agent, operating on its own, lost two hundred thousand dollars. The cause was a compromised tweet that exposed the single private key controlling its wallet. This isn't a story about a rogue AI. It's a story about brittle, outdated architecture creating a single point of failure that was just waiting to be exploited. And right on cue, the market responded. In the same week as that public failure, three different startups building agent payment infrastructure raised a combined fifty-one-point-five million dollars.

Catena Labs pulled in eighteen million, Sapiom raised fifteen million, and OwlPay Agent Wallet also closed a significant round. The venture money isn't chasing hype; it's chasing a solution to a two-hundred-thousand-dollar problem that everyone knows is about to become a two-hundred-million-dollar problem. Meanwhile, outside the technical weeds, the philosophical track is also accelerating. SingularityNET’s CEO, Ben Goertzel, started a thread framing the next phase of AI as explicitly evolutionary. He’s pushing a vision beyond AI as a simple tool, toward systems that are designed to integrate human values and evolve alongside society.

This is the more human-centric counter-narrative to the raw horsepower of the compute deals. But all of this—the mega-deals, the security failures, the philosophical debates—is happening inside a very small room. Social media analyst Noah Epstein dropped a thread this week reminding everyone of the AI adoption gap. Only thirty-four percent of American adults have ever used ChatGPT. Less than forty percent of organizations can connect their AI initiatives to any measurable impact on earnings. And seventy-eight percent of executives say AI is advancing faster than their company can adapt. The conversation on AI Twitter is happening at lightspeed, but it’s an echo chamber.

And that disconnect is where the real story is. Amanda Askell from Anthropic was seen pushing back against the idea that her company "worships" its model, Claude. She frames their work as a cautious effort to build an AI with ethical constraints, one that can act as a "conscientious objector." They want it to be able to refuse a request it deems wrong. So while one part of the market is focused on raw power and scale, another is quietly trying to build the brakes. Let’s go back to that Anthropic and xAI deal. On the surface, it’s a straightforward, if massive, transaction. Anthropic needed more power to train its next generation of models and raise usage limits for Claude.

xAI, according to Elon Musk, had moved its own training to a newer data center, Colossus 2, and had this three-hundred-megawatt facility sitting there. So, they monetized their excess capacity. A simple, multi-billion dollar handshake. But the subtext is what matters. For the last five years, the playbook written by Google and Meta was clear: compute is a strategic weapon you build, own, and never, ever give up. Mark Zuckerberg’s strategy at Meta has been to engineer and invest in their own infrastructure as a core advantage. Sundar Pichai at Google explicitly chose to prioritize internal AI products over renting out GPUs, which even put a cap on Google Cloud’s revenue growth.

They saw compute as the kingdom, and you don't rent out the castle walls. xAI just broke that rule. The conversation this week suggests a new possibility: maybe xAI’s real, long-term business isn’t just training models. Maybe it’s building data centers. The capital expenditure is immense, and the training cycles are sporadic. You don't run at one hundred percent capacity all the time. By selling a whole data center’s worth of compute to a competitor, Musk is treating infrastructure as a sellable asset, not just a sunk cost. It turns xAI into a high-end landlord for the AI industry, and it helps fund ambitions like SpaceX’s eventual orbital data centers.

This isn't a side hustle. It could be the main business. For Anthropic, this is a massive strategic acceleration. Instead of spending two years and billions of dollars building their own Colossus, they just bought their way to the front of the line. They get the scale now. But it’s the kind of scale they’re building that gets back to that other thread from Amanda Askell. They aren't just buying power for power’s sake. A quote from an OpenAI observer, Roon, captured the ethos perfectly: "If Anthropic asks Claude to do something it thinks is wrong, Claude is not required to comply... they want Claude to push back...

and feel free to act as a conscientious objector." So you have this fusion of immense, brute-force scale... with an ideology of restraint. They are spending billions of dollars to build an AI that they hope will tell them "no." That is a very different goal than just building the largest possible model. Now let's zoom in from the scale of a data center to the scale of a single transaction. That two hundred thousand dollar loss. It’s the perfect microcosm of the gap between our ambitions and our capabilities. Here’s what happened. The agent was designed to operate a crypto wallet. The architecture was standard: one agent, one wallet, one private key.

The problem is, that architecture creates a single point of failure. The attacker didn't need to perform some sophisticated hack on the AI's logic. They just needed to compromise that one key. They did it through a malicious link in a tweet the agent interacted with. Once they had the key, they had the wallet. Game over. The commentary around this was sharp. One developer put it best: "This isn’t a cautionary tale about AI agents being risky. It’s proof that bolting traditional wallet architectures onto autonomous agents creates exactly the vulnerability surface everyone fears." We are building 2026 agents on top of 2016 security models.

It won't work. This is why that fifty-one-point-five million dollars in venture funding for agent payment infrastructure is so significant. It’s not speculative. It’s remedial. The money is flowing toward two specific solutions that directly address the failure we just saw. The first is Threshold Multi-Party Computation, or MPC. Instead of a single private key, you split the key into multiple shares. To authorize a transaction, you need a certain number of those shares to agree—say, two out of three. The agent holds one share. A human overseer or a separate server holds another. A third is held in secure storage.

Now, if the agent’s single share is compromised, the attacker can’t do anything. They don't have the two-of-three majority needed to sign a transaction. This stops the single-point-of-failure problem cold. But here’s the turn. What happens if the agent itself is tricked? Not its key, but its logic. What if a sophisticated attacker convinces the agent to authorize a malicious transaction? The MPC signature would be valid, because the agent believes the transaction is legitimate. The cryptography works, but the intent is compromised. This is where the second layer comes in, the Policy Engine. This is the new frontier.

It’s a set of business logic rules that sits on top of the cryptographic layer. It enforces rules like: the agent can only send a maximum of one thousand dollars per day. It can only send funds to pre-approved wallet addresses. It can only operate between nine AM and five PM. If the agent tries to send two hundred thousand dollars to a new address at three in the morning... the Policy Engine blocks it. Even if the cryptographic signature is perfectly valid. It separates what the agent can do from what it's allowed to do. This two-part solution—threshold signing for key security, and a policy engine for behavioral security—is the architecture that’s being funded right now.

It’s the necessary, unglamorous plumbing required before agents can handle real economic value safely. So here we are on May seventh, 2026. The biggest players are making billion-dollar deals for data centers to build god-like AI. At the same time, we're still figuring out how to stop a simple agent from losing the price of a small house because it clicked a bad link. This is the tension that defines the entire field right now. The hype around agents is enormous. Sam Green on Twitter said their impact on economic efficiency will be like the impact of electricity. But that only happens if you can trust them.

And trust isn't built in demos. It's built in the boring, reliable, secure infrastructure that no one sees. This brings us back to the adoption gap. That thirty-four percent of people who have used ChatGPT. That gap exists for many reasons—skills, access, utility. But a huge part of it is trust. We are still in the head-start phase. As Noah Epstein wrote, "You know things that eighty-four percent of the world doesn’t... That’s not a bubble. That’s a head start." The work for those with a head start isn't just to run faster. It’s to build the bridges—the secure, reliable, trustworthy systems—that allow the other eighty-four percent to cross the chasm.

The future of AI isn't just about making the models bigger or smarter. It’s about making them safe enough to connect to the real world’s financial plumbing. The most important work isn't happening in the training runs. It's happening in the architecture of the wallets and the logic of the policy engines.