There is exactly one verifiable fact in the Anthropic–Hg announcement: the two firms intend to build an AI consulting business in Europe.
Everything else — the valuation uplift, the improved IPO window — is a sentence someone typed. No number. No equity split. No headcount. No client list. No date. Four of the five claims in the story are speculation wearing a press release as a costume.
I'll take the other side of that trade. Not because the deal is bad. Because the deal is not what the headline says it is.
Context
Hg is one of Europe's largest software-focused private equity houses, with tens of billions of euros under management and a portfolio stuffed with B2B software companies — Visma, IRIS and dozens of smaller vertical SaaS assets. That portfolio is a pre-qualified customer list: companies with budget, with legacy codebases, and with a board-level mandate to look AI-native before a competitor does it for them. Hg's motive is defensive as much as offensive. AI-native challengers are repricing the assets it already owns. Pushing Claude into the portfolio is asset protection with a growth story stapled to it.
Anthropic has the models and none of the plumbing. Microsoft bundles into Office, Teams and Azure. OpenAI has consumer gravity plus a system integrator network. Google has cloud and Workspace. Anthropic has an API, a safety narrative, and a partner network it does not control. In enterprise AI the fight stopped being about benchmarks roughly eighteen months ago. It is now distribution × delivery × compliance trust. Anthropic is first-tier on models, second-tier on the other two.
So it rents the third one from a PE firm. That is the whole deal. Everything else is commentary.
Core
Run the arithmetic. API and subscription revenue carries gross margins in the 50–70% band. Human-heavy consulting runs 20–40%. Public markets pay software 3–10x the revenue multiple they pay for services. Mix the two and you do not add a growth story — you dilute a multiple. Palantir spent years fighting the "that's not a software company" accusation for exactly this reason.
The consulting entity is not a revenue line. It is a customer acquisition cost with an invoice attached. The real KPI is seat penetration and token burn inside Hg's portfolio, not billable hours. If Anthropic is smart, the JV stays off its P&L: it takes licensing and a revenue share, avoids European labor law and headcount rigidity, and keeps its balance sheet as light as its model weights. Capped upside, controlled risk. That is the structure I would build.
I have run this experiment from the other chair. Last year I supplied the risk parameters for an LLM-driven options book in Paris, five hundred thousand euros of notional, and I overrode three hallucinated executions by hand. The model was never the failure point. Delivery was. Anyone selling enterprise AI consulting is selling the part where the hallucinations live, and that part does not scale with a better checkpoint.
Which raises the question nobody asked: exclusivity. If Hg portfolio companies default to Claude, Anthropic gets a captive distribution channel — and those companies get vendor lock-in, signed off by an owner with an incentive to prefer one supplier. The management teams and the LPs will have opinions about that. The release does not.
Nor does it mention MCP maturity, which is what actually determines whether an agent can be deployed into a customer's workflow rather than demoed at a customer's offsite. Nor the EU AI Act's provider/deployer line, which decides who carries the technical documentation and human-oversight obligations once the consultant starts operating the system instead of merely recommending it. Nor a "your data is not used for training" clause. Those absences are the story.
I have watched this movie on-chain. A protocol pays a growth partner to integrate; the integration is real, the revenue is real, and the token still bleeds — because the revenue is low-margin, non-recurring and controlled by the counterparty. Terra's code was poetry; Luna's exit was prose. The delivery layer is always where the elegance dies.
Contrarian
Retail reads "AI consulting" and hears "new revenue." Smart money reads "services mix" and hears "compressed multiple." Both are looking at the same press release, and only one of them is modelling the cap table. This is a channel trade dressed as a financial event.
Then the deeper blind spot. A model vendor selling advice about which model to buy has a conflict of interest baked into its incorporation documents. Neutrality is the product being sold and the first asset being spent. I flag the same structural flaw when I look at compliance-first stablecoins: an instrument whose issuer can freeze an address inside twenty-four hours is not a settlement layer, it is a permissions layer with better marketing. An advisor whose recommendations can be switched off by its own shareholder is not an advisor. It is a sales channel with a conference room.
And watch the vocabulary. "Partners with." "Launch." Those two words cannot distinguish a jointly funded legal entity from a memorandum of understanding from a purchase order. Three completely different assets. The market is pricing the first and the text may describe the third.
Takeaway
Watch four things. The entity's legal name, equity split and first named clients — that separates a business from a byline. Then whether Accenture or Deloitte object publicly; a silent partner network is a quiet channel, and quiet channels become loud conflicts. Then whether Hg portfolio companies disclose Claude as a standard stack component. And then whether "services revenue" appears as a line item in Anthropic's next disclosure — if it does, model the mix before you model the growth.
Risk isn't the drawdown; it's the gap between belief and reality. Right now that gap is four sentences wide.
When the consultant and the model share a cap table, who exactly is advising whom?