Why OpenAI, Anthropic, and Google Spent $6.25B on Consulting

Anthropic's announcement for its deployment joint venture, published May 4, included this description of how engagements would work: "An engagement might begin with the company's engineering team sitting down with clinicians and IT staff to build tools that fit into the workflows that staff already use. Engagements like this will run across mid-sized companies across industries, each shaped by the people closest to the work."

Read that last clause again. Shaped by the people closest to the work.

We published a manifesto pillar named closest-to-the-problem in mid-May. The argument is that the person closest to the problem should be closest to the build. Anthropic's $1.5 billion deployment JV is now describing its operating model in those exact words. That's a clue about what the labs are actually buying with the $6.25 billion they have committed in the last 30 days.

What did the AI labs actually buy?

In one month, three frontier labs put capital behind embedded-engineering firms.

OpenAI launched the OpenAI Deployment Company on May 12, with $4 billion in initial capital at a $10 billion valuation, per OpenAI's own announcement. TPG led the round. Eighteen co-investors joined. The company was founded simultaneously with the acquisition of Tomoro, the London-based AI consultancy that has 150 Forward-Deployed Engineers across six offices in London, Edinburgh, Manchester, Singapore, Sydney, and Melbourne. Tomoro was created in 2023 in alliance with OpenAI. It was built from day one on the Palantir Forward-Deployed Engineer model.

Anthropic closed its deployment JV on May 4 at $1.5 billion total, per TechCrunch. Anthropic, Blackstone, and Hellman & Friedman each committed $300 million. Goldman Sachs, Apollo, General Atlantic, GIC, Leonard Green, and Sequoia rounded out the cap table. The JV places Anthropic engineers inside mid-sized companies across industries.

Google chose a third path. Rather than build in-house or buy, Google committed $750 million to existing consulting partners who already operate the same embedded model. Different mechanism. Same conclusion.

Three labs. Three strategies. The same shape of bet.

The prevailing read of this cluster is that lab-as-consultancy is vertical integration. AI labs are mature enough that customer adoption is the binding constraint on revenue growth, and a $4 billion bet on deployment services is what TPG and Blackstone recognize as a margin opportunity. Forrester has written this. The trade press is writing it. It is a coherent reading. It would explain everything you observed if the labs had decided to enter a category with structural demand.

What that read can't quite absorb is the specific shape of what the labs bought.

OpenAI didn't acquire a generic systems integrator. They acquired Tomoro specifically, a firm built on the Palantir model where engineers sit inside client organizations and build production systems in-place. Anthropic's own announcement uses identical language. Google funded firms that operate this way. The labs aren't buying SI capacity. They are buying the model of humans embedded in client orgs, doing the work alongside the customer, shaped by the people closest to the problem.

Why did the AI labs leave the API-alone thesis?

The build-with-the-API thesis says: the lab ships the model, the customer integrates the API, the customer builds the product. The forward-deployed thesis says: the lab ships the model AND ships engineers AND those engineers sit inside your company AND they shape what gets built with input from the people who already know the work. Those are two different products.

The labs are no longer selling the first one. They are now selling the second one. Quietly, with $6.25 billion of capital commitment, in 30 days.

I wrote yesterday about the Sinch AI Production Paradox: 74% of enterprises have rolled back a deployed AI agent in production, and 81% rollback rate among the orgs with mature governance investments. Daniel Morris, Sinch's CPO, said the line directly: "if governance was the fix, the most mature teams would roll back less, not more." That is the empirical-data side of today's story. The capital-allocation side is the labs putting $6.25 billion behind closing the same gap. They read the same data. They reached the same conclusion. The model alone doesn't ship outcomes. Embedded human and agent collaboration does.

That sentence is the cleanest verbatim echo of a recent manifesto pillar I have seen this year. It's the fourth time in three weeks a major vendor move has surfaced language from a recently-published Generative Labs pillar. I'm not writing this as a victory lap. The honest framing is that multiple parties are arriving at the same description of how building actually works in this era. The pillar names a structural reality. The labs are now operating on that reality with billions of dollars of weight behind it.

What changes for builders?

When OpenAI Deployment Company, Anthropic's JV, and Google-funded firms are competing for the same engagements, the market for AI product building gets three-sided.

Domain expert + independent builder or studio + AI tools used to be the competitive landscape. Now it's domain expert + lab-funded FDE team + AI tools (the labs' offer) versus domain expert + independent AI-native studio + AI tools (the existing offer). What differentiates the independent option? Not capability. The labs have the same models. Not access. The labs have the same APIs. The differentiator is the domain-tier methodology that emerges from being the studio that has done the work across many client contexts and on the specific problems the domain expert is bringing in.

The labs are doing the model-tier version of forward deployment. Independent studios do the domain-tier version. Both are needed. They are not the same product.

If you are a domain expert building a product, you now have a real choice. Option A is to engage one of the lab arms: forward-deployed engineers, capital backing, vendor-aligned tooling, optimized for breadth across mid-sized companies across industries. Option B is to engage an independent AI-native studio: domain-specialized methodology, smaller engagements, more direct accountability, optimized for depth on your specific problem.

The right answer is contextual. For some products, the lab-FDE model is the right call. Regulated industries with deep infrastructure complexity. Enterprises with $5M+ engagement budgets. Programs that need vendor-aligned compliance posture. For others, the independent-studio model wins. Founders or domain experts whose problem is not a generic "deploy AI into the workflow" but a specific "ship this product that doesn't exist yet." Products where the client is the co-creator, not the source of requirements to be extracted upstream.

What's gone is the option of "just use the API and figure it out yourself." Even the labs say that doesn't work.

What the lab bet actually means

$6.25 billion is a structural concession. The labs did not commit that capital because consulting is a fun business. They committed it because they have read the same Sinch numbers and the same H1 2026 incident retrospective everyone else has read, and they have concluded that the rate at which their own customers ship working production deployments without help is unacceptable. The lab-as-consultancy pivot is not a flex. It's a correction.

The model layer alone is not the product. Humans embedded in the build are the product. That's what the human and agentic collaboration model has been pointing at since we started naming it. The labs now have the model tier and the capital to put humans inside their customers. That is one of the products. The other product, the one independent AI-native studios deliver, is the domain-tier version of the same model. Different scale. Different focus. Same underlying claim.

The labs have voted with billions on which side of the API-alone-versus-embedded-collaboration debate the future lives on. The next decade of building gets shaped by what comes next, on both sides.