The lab that became a consultancy

On 14 May, Anthropic announced that KPMG would deploy Claude across its 276,000 employees globally. Nine days later, Anthropic's newly formed enterprise services company, capitalised at roughly $1.5 billion alongside Blackstone, Hellman & Friedman and Goldman Sachs, acquired Fractional AI as its "founding operational centerpiece." ^{1 2}

I want to take those two events seriously as a single event, because I think most of the commentary will treat them separately and miss what is actually happening. A frontier lab has decided, in public, that the deployment-services layer is too valuable to leave to its partners. It has put a billion and a half of committed capital behind that decision. And it has done so while its partners are mid-rollout.

That is the story. Not the acquisition. The posture.

The unit economics nobody at the lab can ignore

Start with the money, because the money makes the rest of the argument inevitable. A 10,000-seat enterprise Claude deployment, modelled at published Sonnet-class token rates and realistic usage, runs Anthropic somewhere between $500,000 and $2 million a year in API fees. ³ The deployment programme around that same rollout — the integration work, the change management, the workflow redesign, the model evaluation, the agent scaffolding, the production engineering — bills out at $300 to $500 an hour per senior practitioner, and at major-enterprise scale lands at $10 million to $100 million per programme. ⁴

Accenture booked roughly $3 billion in AI-related revenue in fiscal 2024 against a practice that now exceeds 40,000 practitioners. ⁴ If Anthropic's API take on the work flowing through that practice is a low single-digit percentage of the engagement value, and on the maths above it is, then every successful Claude deployment that runs through an SI is a transaction in which the lab supplies the active ingredient and someone else captures most of the margin.

The lab supplies the active ingredient. Someone else captures most of the margin.

You cannot run a frontier lab on those economics indefinitely. Compute bills do not care that your partner billed the client $40 million while you billed them $800,000. The training runs for the next model do not get cheaper because KPMG had a good quarter. At some point the lab has to decide whether it is in the API business or in the value-capture business, and the answer, when your post-money valuation is $61.5 billion and your investors are paying attention, is that you are in both. ⁵

So Anthropic moved. And the structure it moved into is the interesting bit.

What the deal actually is

The new entity is not Anthropic Consulting. It is a separately capitalised vehicle, with Blackstone, Hellman & Friedman and Goldman Sachs as the financial sponsors, focused initially on private-equity portfolio companies and mid-market enterprises. ^{1 2} Fractional AI becomes its delivery arm. Anthropic's Applied AI organisation collaborates with it on transformation projects, but Anthropic itself remains a public benefit corporation positioned around safety research. ¹

This structure is doing several things at once, and it is worth pulling them apart.

It separates the brand from the billings. A frontier lab that publishes alignment research and refuses to compete on price needs distance from a services business that, by definition, will lose engagements, deliver late, and have unhappy clients. The new entity carries that risk. Anthropic carries the model and the research story.

It puts PE money exactly where PE money wants to be. Blackstone alone has hundreds of portfolio companies. Hellman & Friedman has dozens of large ones. Goldman has a private wealth channel and an investment banking franchise that touches every major enterprise transaction in the Western economy. If you wanted to build a captive AI transformation business and you had to choose three sponsors who could route deal flow at it, you would choose these three. The $1.5 billion of committed capital is the visible number; the genuinely valuable number is the implied pipeline. ¹

It defines the customer segment in a way that does not, technically, compete with KPMG and PwC. Mid-market and PE portfolio companies are not the engagements that the Big Four fight hardest for. The largest transformation programmes, global rollouts at Fortune 100 banks and pharma majors, remain the SIs' domain. On paper, this is a complementary play.

On paper.

The contradiction the structure tries to paper over

In practice, the segmentation is going to leak. It always does. PE portfolio companies are not all small. Blackstone owns assets at multi-billion-dollar scale. The moment the new services entity wins a transformation mandate inside a Blackstone portfolio company that KPMG was also pitching, the partnership tension is real, and the answer "we're focused on mid-market" will sound exactly as convincing as it deserves to sound.

The standard objection here is that Microsoft has been doing this for a decade with Azure: selling cloud, partnering with SIs, and running Microsoft Consulting Services alongside. It works because everyone understands the rules. Microsoft does the lighthouse accounts and the platform engineering; the SIs do the scale delivery; the customer pays both. The model holds because the value of the underlying platform is large enough that no party wants to walk away.

The reason it might not hold for Anthropic is that the substitute is right there. KPMG could deploy GPT-class models from OpenAI through Microsoft tomorrow. PwC already runs significant Azure OpenAI practice work. The Big Four are not captive to any single lab; they have explicitly multi-model strategies because their clients demand them and because their own margins require optionality on the input. ⁴

If Anthropic's services arm becomes visibly competitive — meaning it shows up in pitches against KPMG with a lower price, a tighter integration story, and direct model-team access — KPMG's rational response is not to defend the Anthropic partnership. It is to weight the next twenty engagements toward whichever lab is not also pitching against them. The partnership does not collapse; it attritions.

This is not speculation. It is just what happens in channels when the vendor goes direct. We have seen it in software for thirty years.

Why now, and why not the others

Here is where I want to test the obvious framing, because the obvious framing is that Anthropic has spotted something the other labs missed, and that reading is too flattering by half.

OpenAI has not done this. Google has not done this in any serious form — Google Cloud Professional Services exists but remains a fraction of AWS's or Azure's services revenue, after a decade of trying. ⁴ Meta has not done this. DeepMind has not done this.

There are two readings.

The first reading is that Anthropic is early. The other labs will follow within twelve to eighteen months, because the unit economics I described above are not unique to Anthropic. They apply to every frontier lab selling API access to enterprises that then spend ten times the API bill on deployment. If that is right, this is the first move in a structural reconfiguration of how labs capture value from enterprise AI, and the comparable is the cloud hyperscalers building professional services arms in the mid-2010s. Anthropic is just first.

The second reading is that the other labs looked at this and concluded it was not worth the brand and organisational cost. OpenAI has a much larger enterprise business than Anthropic and has chosen to scale it through Microsoft and through a relatively small direct field organisation, not through a services acquisition. Google has the institutional muscle memory of trying and failing at services for fifteen years. The fact that none of the other labs has done this is information. It is not proof Anthropic is wrong, but it is a data point that deserves more weight than the narrative of bold strategic foresight usually allows.

I think the truth is closer to the first reading than the second, but the second reading carries a warning I want to keep in view: services businesses break labs. They break them culturally, because the incentive structure of billable delivery is opposite to the incentive structure of research. They break them operationally, because timesheets and utilisation targets and client escalations consume management attention that was previously spent on training runs. And they break them strategically, because the services pipeline becomes a forecast input that the model roadmap starts bending around.

Anthropic has tried to insulate itself from all three failure modes by structurally separating the entity. Whether that insulation holds when the services arm is generating real revenue, real client commitments, and real internal politics is the question I would put at the top of the watch list.

The fourth deployment model

There is a taxonomy question underneath all of this that I think is genuinely new.

Enterprises deploying AI in 2024 had three rough models to choose from. They could build a centralised AI Centre of Excellence and have it serve the business units. They could embed vendor engineers — Anthropic's Applied AI team, OpenAI's Forward Deployed Engineers, the equivalent at Google — directly into their teams. Or they could hire a systems integrator to run a transformation programme.

Each model had a known set of trade-offs, and each had a known set of conflicts of interest. The CoE answered to internal politics. The vendor engineers were brilliant on their own model and uninterested in alternatives. The SI was vendor-agnostic in principle and indexed on billable hours in practice.

The lab-owned consulting arm is a fourth model, and it scrambles the conflicts in a new way.

The lab-owned consulting arm is not vendor-agnostic. It is not internal. It is not embedded. It is something new, and the procurement question it forces is also new.

The structurally honest version of the procurement question is: when the advice on which model to use is given by an entity owned by one of the models, whose interest does the advice reflect? The answer is not necessarily "the vendor's." A good consulting team will recommend against using Claude in a workload where Claude is the wrong tool, because shipping a failed deployment is worse for the business than refusing the engagement. But the structural pressure runs one way. Over a thousand engagements, the lab-owned arm will recommend its parent's model more often than a genuinely independent advisor would. That is not corruption. That is just statistics.

Enterprises with mature procurement functions will see this immediately and price it in. The interesting question is what they do with the information. The naive answer is "they don't hire the lab-owned arm." The more likely answer is "they hire it for the work where the lab's model is obviously the right tool, and they hire SIs for the work where the choice is genuinely contested." Which is, not coincidentally, the segmentation Anthropic has implicitly chosen by aiming the new entity at PE portfolios and mid-market — segments where the procurement function is thinner and the engagements are smaller and the model choice is less contested because the workloads are less specialised.

The segmentation is rational. The segmentation is also, on a long enough horizon, a ceiling. The largest, most strategically interesting transformations will continue to flow through entities that look genuinely independent of any single lab. That is not a problem for Anthropic at $1.5 billion of committed capital and a sub-billion services run-rate. It is a problem if the ambition is to become a top-five enterprise services business.

What this means for the SIs, and for the model

For the Big Four, the right read is not panic. The right read is that the value chain in enterprise AI is being actively contested, and that the comfortable assumption, the labs build models, the SIs deploy them, everyone keeps their lane, is no longer the operating reality.

The Big Four's structural advantages remain real. Audit independence, regulatory relationships, geographic reach, C-suite trust accumulated over decades, and the fact that 40,000 AI practitioners cannot be hired in a year by anyone, including a frontier lab with billion-dollar PE backing. ⁴ But the implicit assumption that the lab is a supplier and not a competitor is gone, and the strategic response has to be either deeper multi-model neutrality (which makes the Big Four genuinely useful as the place where bake-offs happen) or vertical specialisation (which makes them genuinely useful as the holders of domain expertise the labs cannot replicate).

The strategy that does not work is doubling down on a single-lab partnership as a differentiator. That was a 2024 strategy. It is not a 2026 strategy.

For Anthropic, the test is whether the new entity can grow into a meaningful revenue line without dragging the lab into a services-shaped organisation. Microsoft managed it because the platform was so dominant that the SI ecosystem could not afford to walk away. Anthropic is not in that position. Claude is excellent. Claude is not, on any reasonable measure, dominant in the way Azure or Windows or Office were when Microsoft built its consulting arm. The services play is happening earlier in the platform's lifecycle than the historical comparables would suggest, and that timing creates risk the comparables do not capture.

For the model, meaning the way enterprise AI gets bought and deployed, the takeaway is that the boundary between "the people who make the model" and "the people who deploy the model" has just been formally erased by one of the three frontier labs. Whether the other two follow within eighteen months is the single thing I would watch most closely. If they do, the SI market as currently structured is going to look different by the end of 2027. If they do not, Anthropic will either be exceptionally well-positioned or stuck running a services business its competitors decided was not worth the trouble.

I do not think the answer is obvious. I think the lab has made the right bet on the unit economics and an uncertain bet on the operating model, and that the gap between "right strategically" and "executable organisationally" is the gap most lab-to-services transitions have died in.

The acquisition is reported. The capital is committed. The partnership with KPMG is live. All three of those things are true at the same time, and at least one of them is going to bend in the next twelve months. Working out which is the question I would be paying KPMG, PwC, and Anthropic's competitors to answer right now, if I were any of them.

Footnotes

¶