The AI Labs Are Now in the Consulting Business

Anthropic and OpenAI just spent more than $5 billion in ten days to get into Palantir's business. Here is what that does to your AI procurement playbook.

EXECUTIVE SUMMARY

The foundation model labs spent the first half of May quietly admitting something: selling API tokens is not the same as driving adoption. Inside a ten-day window, Anthropic and OpenAI each stood up dedicated services companies — capitalized by Wall Street, staffed with Palantir-style forward-deployed engineers, and pointed at the same buyers their consulting partners have been selling to for years. They went from cloud vendor to boots-on-the-ground vendor in under a month.

// Anthropic launched a $1.5 billion enterprise services venture with Blackstone, Hellman & Friedman, and Goldman Sachs on May 4 — each anchor putting in roughly $300 million, with Apollo, General Atlantic, GIC, Leonard Green, and Sequoia rounding out a structure that mirrors Palantir's forward-deployment model and embeds Anthropic engineers inside mid-market PE-owned companies.

// OpenAI launched the Deployment Company seven days later — more than $4 billion in initial capital, 19 founding partners led by TPG, and an acquisition of London consulting firm Tomoro that brings roughly 150 Forward Deployed Engineers on day one. Bain & Company, Capgemini, and McKinsey are on the investor list — the consulting firms that have run enterprise AI for a decade are now financing the entity that competes with them.

// Three days later, Anthropic doubled down on the SI track — expanding its PwC alliance to train and certify 30,000 professionals on Claude and stand up a Claude-anchored Office of the CFO inside the firm. The labs are running both tracks simultaneously, which means the channel itself is being restructured around the model vendor's interests rather than the integrator's.

The strategic question is not which model to standardize on. It is whether your AI services budget, your incumbent SI contracts, and your internal Center of Excellence are designed for a market where the model vendor now writes the implementation playbook — and gets paid on the outcome.

In 2003, a different version of this story played out. Red Hat was selling Linux, and the big systems integrators — IBM Global Services (funny how that turned out), Accenture, EDS — were selling the projects that ran on it. For years, the SIs assumed they owned the customer relationship and Red Hat owned the bits. Then Red Hat built a consulting practice, and the SIs had to decide whether to fight, partner, or retrain on Red Hat's terms. Most retrained. The OS vendor's playbook quietly became the rulebook for the entire delivery economy underneath it.

What's different now is the time compression. Red Hat took years to build out its services arm. Anthropic and OpenAI did it inside a ten-day window in May, with $5.5 billion in committed capital and 150 Forward Deployed Engineers already on the org chart. The labs are not waiting for the SI ecosystem to mature into the role — they're buying their way past it. And every CIO I talk to is about to make a procurement decision they don't yet realize they're making.

The conventional wisdom on enterprise AI has been that the labs are software companies, the SIs are services companies, and the two coexist in a clean division of labor. The lab ships the model; the SI deploys it; the customer pays for both. That works as long as the model layer is the constraint. What the last two weeks demonstrated is that the model layer is no longer the constraint — deployment is. And the labs have decided they won't wait for their partners to solve it.

What you're actually buying now. Until April, an Anthropic or OpenAI contract meant access to the API and a relationship manager. After May, you can also buy a small team of engineers who will sit inside your operations, build the harness that wraps the model around your specific workflow, and stay until the production system runs. OpenAI calls this team Forward Deployed Engineers — a title and operating model adapted almost line-for-line from Palantir's playbook. Existing SI relationships do not go away, but their economics shift: PwC's deal to roll out Claude Code and Claude Cowork to its 364,000-person workforce, Cognizant's 350,000-employee Claude rollout, and Google's $750 million fund for the Big Three SIs all point the same way — the lab is now selecting which SI gets which workflow, not the other way around.

Four overlapping forces have been compounding for 18 months and broke the surface in May:

Pilot failure rates made API-only sales unsustainable. MIT's NANDA initiative found that 95% of enterprise generative AI pilots deliver zero measurable financial return, and IDC puts the agent-pilot-to-production rate at 12% — only 4 of every 33 POCs reach wide-scale deployment. When 88% of your customers can't get the technology into production, you can either blame your customers or handle the deployment yourself. The labs chose the second option.

Implementation is the highest-margin layer. It's where value gets unlocked and where switching costs get manufactured. Blackstone's Jon Gray said it plainly when announcing the Anthropic JV: the venture exists to break the bottleneck of engineers capable of implementing frontier AI at speed. That bottleneck is also the most profitable line item on a typical AI engagement.

The labs hedged with the SIs simultaneously. Three days after launching its services arm, Anthropic announced that PwC would train 30,000 staff on Claude and establish a Claude-native Office of the CFO. Not a contradiction — a portfolio. Compete with SI partners on the highest-value engagements; lock in volume underneath.

SI economics are restructuring in real time. The Big Four are now customers of the labs (paying for Claude or GPT subscriptions for their own consultants), channel partners (delivering lab-driven engagements), and competitors (running their own AI practices). Three roles, one P&L. The first SI to decide which role to optimize for is the one that survives the next 18 months intact.

// Key Takeaways

AI labs are no longer just model vendors. They are infrastructure, services, and channel partners simultaneously. A vendor framework that treats them as one of those three will miss what the other two are doing to your contracts.

The model layer is no longer the constraint. Deployment is. Whoever owns the deployment owns the margin, the relationship, and the switching cost. The labs noticed before the SIs did.

Pilot failure stopped being a customer problem. It became a vendor revenue problem. When 88% of your buyers can't get the technology into production, you can either blame the buyers or deploy it yourself. $5.5B in May was the labs picking option two.

The Big Four are now customers, channel partners, and competitors of the labs all at the same time. Three roles, one P&L. That's structurally unstable, and the first integrator to pick a role wins the next 18 months.

The strategic question for the next budget cycle is not which model to standardize on — that's already commoditizing. The question is who deploys the model, who owns the playbook, and who pays the price when the deployment relationship has to change. Until May, the answer to all three was your SI or your internal team. After May, it's contested in a way your existing contracts were not built to handle.