The API wasn't enough

Seventeen and a half per cent, guaranteed, over five years. That's what OpenAI promised the private equity consortium backing The Deployment Company, its new $10 billion enterprise venture. That number tells you everything about what this actually is. Technology companies don't guarantee returns. Structured credit products do.

On the same day OpenAI closed that deal with TPG and 18 other PE firms, Anthropic announced its own $1.5 billion services entity backed by Blackstone, Goldman Sachs, and Hellman & Friedman. Both ventures do the same thing: embed lab engineers directly inside companies to deploy AI into core operations. Both target the same customer base: portfolio companies owned by the investors writing the cheques.

The conventional reading is that AI labs are diversifying into services. The real story is blunter. The API, on its own, doesn't close enterprise deals. Two years of selling model access taught both labs the same lesson: most organisations cannot bridge the gap between "impressive demo" and "running in production" without someone building the bridge for them. So now the labs are selling the bridge.

What makes the PE structure clever is distribution, not capital. Neither OpenAI nor Anthropic needed more money. What they needed was a channel that bypasses the 12-to-18-month enterprise sales cycle. Private equity firms don't ask their portfolio companies whether they'd like to adopt new technology. They install it. When Bain Capital or Blackstone decides their holdings will integrate AI into operations, procurement doesn't get a vote. The investment thesis is the purchase order.

This mirrors how enterprise software matured in the 1990s. SAP and Oracle discovered that licence revenue stalled without an implementation layer, so they built consulting arms, then partner ecosystems, then entire services industries around configuration and deployment. The model layer is following the same arc in compressed time. Except here, the labs aren't outsourcing implementation to Accenture. They're keeping it in-house and using PE as the distribution mechanism.

The financial architecture of these deals reveals who holds the leverage. OpenAI retains strategic control through super-voting shares while handing PE investors the economics. Anthropic's backers, each anchor committing roughly $300 million, get access to a services business competing directly against the $300 billion consulting industry. In both cases, the labs trade margin for distribution. The PE firms trade capital for a guaranteed clip on AI adoption across their own holdings. It's vertically integrated demand.

I think the implication for independent software companies is uncomfortable. If the frontier labs conclude that models require bespoke implementation to generate enterprise revenue, they'll keep expanding these services arms. Every workflow they rebuild with embedded engineers becomes a moat that no third-party integrator can replicate, because the integrator doesn't control the model's roadmap. The labs do.

The question worth sitting with: if selling the model requires selling the implementation, and selling the implementation requires owning the customer relationship, how long before "AI lab" and "systems integrator" become the same thing?

---