The Great Compression, we documented how six model providers have deployed $200B+ in acquisitions and internal builds since 2024, systematically absorbing every middleware function that stood between foundation models and enterprise workloads. This week, two news events confirmed that the same logic is now operating one layer higher — inside the professional services firms that have been quietly positioning themselves as the implementation partners the enterprise needs. The partner layer is the new middleware.
PwC’s US CEO Paul Griggs delivered a pointed warning to his firm’s partners this week: those not “paranoid about being AI-first” would not last long. The firm is moving away from the billable-hours model toward automated platforms — “PwC One” — that clients can access directly for tax advisory and M&A due diligence, priced on subscription rather than time. It is a significant structural shift. But the cultural framing — individual urgency, adapt or be replaced — misnames what is actually happening at the market level.
The more architecturally revealing signal arrived the same week from Axios and Reuters: both OpenAI and Anthropic are separately in advanced talks with major private equity firms to build what reporting is explicitly calling “enterprise AI consulting arms” — staffed by forward-deployed engineers embedded inside portfolio companies, both advising and implementing AI solutions directly.
Pre-money valuation of the proposed OpenAI enterprise JV — structured as a majority-owned subsidiary staffed by forward-deployed engineers who advise and implement directly inside PE portfolio companies. Anthropic’s parallel JV with Blackstone, Permira, and Hellman & Friedman carries a separate ~$1B in proposed PE equity. Both deals are ongoing; no final agreements have been reached.
That is not a distribution arrangement. That is vertical integration arriving at the partner layer — the same structural move we have been documenting at the middleware layer since mid-2024, now executing one level up in the stack.
The Same Thesis, Extended Up-Stack
In The Great Compression, we argued that the agent harness layer — orchestration, evaluation, memory, observability, tool integration — was being absorbed from above by platform-native capabilities and from below by open protocol standards. The middleware companies that had stepped into the gap between foundation models and enterprise workloads in 2023 faced a structural threat: every function they had commercialized was being replicated by the model providers themselves, free, deeply integrated, and carrying the distribution advantage of the underlying model relationship.
The PE JV structures announced this week follow the identical pattern, applied to the consulting and advisory layer. The model providers are not simply selling licenses to PwC or McKinsey or Accenture. They are building the implementation infrastructure themselves — and using private equity’s portfolio reach to distribute that infrastructure at enterprise scale, bypassing the traditional consulting channel entirely.
OpenAI already has a distribution deal with PwC. Now it is building a JV that competes with that same channel. That is not a contradiction. That is a declaration about where the durable margin lives.
— Tom M. Gomez, Luminity DigitalThe distinction matters architecturally. A distribution deal means the model provider sells through a consulting partner who owns the implementation relationship. A forward-deployed engineering JV means the model provider owns both the model and the implementation relationship — and the PE firm is a minority investor and initial customer, not a channel. The consulting partner has been disintermediated before the deal is even signed.
What the JV Structure Actually Tells You
Advisory as the Asset
Model providers sell licenses. Consulting firms own the implementation relationship — scoping, configuration, change management, ongoing optimization. The advisor’s judgment and methodology are the billable commodity. Expertise is the moat.
Value capture sits with the firm that translates model capability into operational outcomes. The model company collects a subscription. The consulting firm collects the margin.
Structural PressureImplementation Substrate as the Asset
Model providers deploy their own forward-deployed engineers directly into enterprise and PE portfolio companies. They own both model capability and the implementation relationship. The PE firm is a minority investor and first customer — not a channel.
Value capture consolidates with the model company. The consulting firm is upstream of the deal or not in it. The implementation substrate — not the advisory relationship — is the durable margin.
AcceleratingThe Avanade Precedent
The Axios piece invokes Avanade — the joint venture formed in 2000 between Microsoft and Accenture to implement Windows and other Microsoft solutions into large enterprises. The analogy is apt, and its conclusion is instructive: Accenture eventually acquired a controlling stake. Microsoft retained a minority position. The implementation partner became the durable asset in that cycle.
In the Avanade cycle, the implementation partner won because Microsoft needed distribution and Accenture had the enterprise relationships. The question for the current cycle is structurally different: do the model providers need a traditional consulting intermediary at all — or does the PE JV model give them direct access to the enterprise relationships that previously required a channel? If the answer is the latter, the consulting firm is not Accenture in this analogy. It is the partner the Avanade deal was designed to bypass.
What Enterprise Leaders Should Understand
For enterprise decision-makers, the strategic implication is not primarily about consulting spend. It is about the implementation infrastructure you are building and who controls it. An enterprise that allows its AI implementation layer — its agent harnesses, its workflow integrations, its evaluation infrastructure, its deployment patterns — to be owned and configured by a model provider’s forward-deployed team has not simplified its architecture. It has concentrated a governance risk.
The Implementation Substrate Risk
When a model provider controls both the model layer and the implementation layer — the harness, the integration patterns, the evaluation criteria, the deployment configuration — it also controls what counts as success for agents running inside your organization. Enterprises evaluating PE-backed AI deployment programs should assess not just the model capabilities being offered, but who owns the implementation substrate beneath them and what vendor portability looks like when the JV terms change.
The Great Compression created exactly these conditions at the middleware layer. The same structural logic is now arriving at the services layer. The response must be architectural, not reactive.
The partners the PwC CEO is warning are not just competing with AI tools automating analyst work. They are competing with OpenAI building a $10 billion JV to own the relationship those partners have historically claimed. The compression was never just about middleware. It was always about who controls the full stack from model to outcome — and that stack now includes the implementation relationship itself.
It is coming. Faster than most would like to believe.
