The structure already in place is harder to exit than most acquisitions.

We've spent a decade arguing about equity stakes, antitrust, and acquisition premiums. That vocabulary was built for a different era. What's forming between Google and Anthropic doesn't fit any existing category. The question is whether we have language precise enough to describe what it actually is.

Google holds a 14% equity stake in Anthropic at a hard contractual ceiling of 15%. No board seats. No voting rights. No governance exposure. Nominally passive. But the actual dependency runs far deeper. Google supplies the custom TPUs Anthropic trains on, not commodity cloud compute but proprietary silicon whose next-generation roadmap Google alone controls. Google's hardware division shapes the version of Claude being built right now. Google is co-financing a $5B+ data center in Texas for Anthropic, not as a compute contract but as a stake in the physical substrate. Distribution runs through Vertex AI, meaning Google's enterprise salesforce is already selling Claude to its own customers. When the Pentagon moved to exclude Anthropic from federal contracts, Google stood alongside Amazon, Apple, and Microsoft in public legal solidarity. That's not an investor protecting an asset. That's an alliance protecting shared infrastructure.

Anthropic's most capable model is being trained on Google's silicon. Google just published the inference architecture that deploys it. Those two facts, arriving in the same week, are the integration thesis made visible in real time. Google Research published TurboQuant, a compression algorithm achieving up to 6x reduction in KV cache memory with zero accuracy loss, presented at ICLR 2026. That's not just an efficiency paper. It's the inference architecture that makes frontier-scale deployment economically viable. Then Claude Mythos, internally codenamed Capybara, surfaced two days ago. Anthropic confirmed it: a new tier above Opus that scores dramatically higher on coding, reasoning, and cybersecurity. The cybersecurity dimension is what Anthropic flagged most carefully. Anthropic warning about its own model. Training and deployment. One stack. Two logos.

Formal acquisition would trigger antitrust review, governance obligations, and the talent exodus that would hollow out what Google paid for. Anthropic's founders left OpenAI specifically because of concerns about deploying powerful AI inside large commercial entities with misaligned incentives. Acquisition also destroys the independent safety narrative that makes Anthropic credible to regulators and enterprises, a narrative that serves Google as much as Anthropic. The current arrangement delivers everything acquisition would: model access through TPU dependency, enterprise distribution through Vertex AI, political cover through demonstrated solidarity. With none of the cost. This isn't a second-best outcome. It's a better-designed one.

Google doesn't tell Anthropic what to build. It shapes what's possible to build, at what scale, at what cost, on what silicon, with what political cover when things come under pressure.

The right frame isn't corporate. It's geopolitical. Structural power: not the ability to direct another actor, but the ability to shape the environment they operate in. Hardware roadmaps. Compression architecture. Physical infrastructure. Capital access. Political solidarity. None of those are traditional control levers. Together they constitute the conditions within which Anthropic makes every consequential decision. Japanese keiretsu offer a partial analog: interlocking stakes, shared supply chains, deep coordination without formal consolidation. But keiretsu relationships are roughly symmetric and operate across fewer simultaneous dependency layers. This one is neither, and spans more than any keiretsu precedent did.

What's forming here is mutual non-extractability: a relationship so load-bearing for both parties that the separability assumption underlying every existing corporate category simply doesn't hold. Five simultaneous dependency layers, silicon, physical infrastructure, capital, distribution, political alliance. Exit costs don't add. They compound. We need new vocabulary before we can have an accurate conversation about who shapes the trajectory of AI development, and what ‘shapes’ even means when it operates through the environment rather than through direct instruction.

The category error isn't calling this a partnership. It's thinking that word still means what it used to. Mythos trains on Google's TPUs. TurboQuant deploys it. A Texas data center houses it. Google's lawyers back it in court. What word covers all of that?