Anthropic, OpenAI, and the New Species of Services Firm

Services firms do not sell skill. They sell institutional predictability — a composable thing made of definable primitives — and the unit of sale is the primitive bundle, not the consulting hour. When the bundle changes, the firm changes. The last few weeks added four primitives to the set, and they are only assemblable by entities that did not exist eighteen months ago. The buyers have not noticed yet. The incumbents are still pricing the old form factor.

The conventional read of those weeks is that the AI labs are entering consulting. Anthropic stood up a $1.5B joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs in early May 2026, backed by Apollo, General Atlantic, GIC, Leonard Green, and Sequoia. Within hours, OpenAI announced a $4B Deployment Company at a $10B valuation, with TPG leading and Advent, Bain Capital, and Brookfield as co-leads, plus 15 other investors. OpenAI separately acquired Tomoro, a London applied-AI firm whose client list included Tesco, Virgin Atlantic, Mattel, Red Bull, and Supercell — 150 forward-deployed engineers folded into the new unit on day one. Both moves copying Palantir's forward-deployed engineer playbook. Both, the press releases implied, taking aim at Accenture and Deloitte.

This read is wrong, or rather it is right at the surface and wrong about the structure. The labs are not entering consulting. They are assembling a new species of services firm with primitives that no consulting firm, integrator, or marketplace can copy. The species is the news, not the move into services.

What enterprise buyers actually buy

When a Fortune 500 procurement team signs a contract with McKinsey, Wipro, or Palantir, they are not buying skill. The most talented operators in those firms are no better than the most talented operators on Upwork — that is why the agencies recruit from the same pool, and that is why a single rockstar consultant can leave a Big Four firm to start their own practice and immediately compete on quality, if not on contract size. What buyers are actually buying is institutional predictability, and predictability is a composable thing.

It decomposes into primitives. At agency scale, eight: persistent identity, multi-dimensional reputation, binding commitments, capacity redundancy, capital at stake, audit trail, procurement wrapper, risk transfer. Each is atomic — none reduces to another. Each can be present or absent, and the absence of any one breaks compositions that procurement actually cares about.

Persistent identity means the pod is a legal-shaped entity, not a human. It has a stable registration, an EIN, a balance sheet, a lifespan that survives any individual's exit. Without it, the buyer is hiring a person, and a person can quit, die, or ghost. With it, the buyer is hiring an entity that can be sued, audited, renewed.

Multi-dimensional reputation means track records accrue at three independent layers — entity, role, individual — and are portable. A pod can carry credibility into new domains. Individuals can graduate roles within or across pods. Without this primitive, you have the cold-start problem that breaks every freelance marketplace: new sellers cannot get hired because they have no reviews, and they cannot get reviews because they cannot get hired.

Binding commitments are quantified SLAs — response time, delivery dates, quality floor — with automatic, pre-defined consequences for breach. This is distinct from aspirational quality. A commitment without consequences is a hope. Without this primitive, the buyer is buying intentions, not contracts.

Capacity redundancy means every named role has a named backup, enforced at the pod level. If the account manager is hit by a bus, the bench is already in the room. This is, more than anything else, what enterprise procurement is actually paying agencies for. Talent shortages and unpredictable attrition are the single largest unhedged risk a procurement team carries, and agencies sell hedges against it.

Capital at stake means the pod has actual money in the game — bond, escrow holdback, retainer slush, performance reserve. Reviews threaten future revenue slowly and weakly; a clawback threatens current capital immediately and decisively. Without this primitive, every commitment from primitive three is unenforceable in practice.

Audit trail is the immutable, queryable log of decisions, deliverables, sign-offs, hours, scope changes, handoffs. It is the substrate everything else runs on. Continuity when people rotate, dispute resolution when work fails, compliance demonstrability when auditors arrive, evidence-based reputation rather than vibes — all flow from this primitive.

Procurement wrapper is the pre-cleared MSA, NDA, security questionnaire responses, SOC 2 attestation, insurance certificates, W-9s, tax forms, maintained perpetually at the pod level and refreshed automatically. The buyer's procurement department says yes in three days instead of three months. Without it, the platform structurally cannot serve buyers above roughly $50K deal size, because the deal dies in vendor onboarding.

Risk transfer is indemnification, errors-and-omissions insurance, refund and redo guarantees, IP indemnity, dispute mediation. When the work fails, the consequence flows to the pod and its insurer — not the buyer's career inside their own org. Without this primitive, the personal political risk to the buyer is the real blocker, and they will keep choosing the safe legacy agency over your marketplace no matter how good the work.

Marketplaces like Upwork and Fiverr deliver about one and a half of these eight. That is why those marketplaces top out at small projects and never serve enterprise. The model assumes the individual is the unit, and an individual cannot carry a procurement wrapper or transfer risk. Toptal-class platforms deliver roughly four. Traditional agencies deliver all eight. The agencies have not been disrupted by Upwork because they sell something Upwork cannot — and "something" is not skill, it is the primitive bundle.

Why scale changes the primitive set

Add Wipro, TCS, Infosys, and Accenture to the picture, and four more primitives emerge that do not exist at agency scale.

Talent supply pipeline is the industrial capacity to manufacture qualified labor at rates of tens of thousands of engineers per year. Wipro and Infosys run training campuses that intake 30,000-plus engineering graduates annually and certify them in specific platforms within months. This is qualitatively different from capacity redundancy — redundancy is "we have a backup," pipeline is "we can manufacture 5,000 new SAP consultants this year." No marketplace currently has any analog. McKinsey has something adjacent through its up-or-out pipeline, but at a different scale and shape.

Strategic partnership graph is preferred-vendor or platinum-partner status with the major platform companies — SAP, Oracle, Microsoft, Salesforce, AWS, ServiceNow. This is an upstream primitive (relationships with suppliers, not buyers), and it materially affects what you can sell, at what price, with what support. Wipro's Oracle partnership lets them offer pricing and certifications a pure marketplace cannot replicate. Accenture's Salesforce relationship is itself a multi-billion-dollar asset. Without this primitive, you compete on the same playing field as a startup; with it, you have asymmetric access to the platform ecosystem.

Regulatory operating capability is active legal entitlement to operate in restricted spaces — security clearances, ITAR registration, FedRAMP authorization, HIPAA Business Associate Agreements, classified-facility access, banking charters, GxP qualifications. This is different in kind from the procurement wrapper. Wrapper is "we have paperwork ready"; regulatory capability is "we are legally permitted to do this work at all." This is the primitive that makes Palantir's defense work and Accenture Federal's government work possible. A startup cannot acquire it in a year; it requires years of compliance investment, sometimes decades.

Scaled balance sheet is firm-level financial capacity large enough to underwrite outcome guarantees in the tens of millions. Accenture's roughly $60B revenue base means it can pre-finance massive engagements, eat losses, indemnify deeply, and signal to a Fortune 100 buyer that even a $50M write-off would not materially threaten the company. This is a difference in kind, not degree, from a pod's capital-at-stake primitive. It unlocks contract types that smaller entities literally cannot sign.

Add Palantir to the picture, and you get a twelfth primitive: proprietary platform IP that becomes the operating substrate of the engagement. Foundry is not ancillary to Palantir's services — it is the services, with humans installing it. The lock-in mechanism Palantir pioneered is platform-shaped, not relationship-shaped, which is part of why Palantir's stock returned 640% over five years while every analyst was still trying to figure out whether it was a software company or a consulting firm. It was neither. It was a primitive bundle no one else had.

The twelve primitives cluster into three viable bundles, three different paths from the agency baseline. Wipro and Accenture sell scale-and-duration: industrial labor, long-term embedded operations, multi-year framework agreements, geographic delivery centers. Palantir and C3.ai sell platform-and-outcome: proprietary tech as the substrate, outcome guarantees on top, deep vertical integration. McKinsey and Bain sell brain-only: domain expertise, analytical horsepower, board-level relationships, no labor scaling and no platform stack.

Notice the structural property: none of the three paths uses all twelve primitives. Each specializes. Trying to win on all twelve is a strategy mistake. Even Accenture, which approaches it, internally separates Federal, Strategy, and Technology into operating units with different primitive bundles. The primitive set determines the firm.

The four primitives only labs and PE can assemble

The Anthropic-Blackstone and OpenAI-TPG ventures are not stacking the existing twelve. They are adding four new primitives that neither Wipro nor Palantir nor McKinsey can replicate, and each one is structurally exclusive.

Frontier-model proximity is the first. A forward-deployed engineer from OpenAI knows what ships in October. They have direct access to the lab's research roadmap, to model capabilities six months ahead of public availability, to internal benchmark results, and to the engineers who built the systems. They design the deployment for capabilities that do not exist yet, knowing the deployment will be production-ready by the time those capabilities ship. Wipro and Accenture cannot have this primitive — they are at arm's length from the frontier by definition. They deploy what is available; the labs deploy what is coming. This is a moat with no analog in the prior framework, and it is the reason the FDE model is sticky in a way SaaS is not. A company that builds a custom AI system with an Anthropic FDE for six months is buying into a roadmap, not a snapshot. The roadmap is what makes the system improve while it runs. The roadmap is what makes the switching cost prohibitive.

This is not just a sales pitch. It is the structural answer to why the FDE model produced 640% returns at Palantir while every other expensive-enterprise-services model struggled. Palantir's FDEs were not just consultants — they were the only people on earth who could build on Foundry's actual capabilities, because Foundry's roadmap was their daily work. The Anthropic and OpenAI FDEs are doing the same thing with frontier models. The buyer is paying for proximity, which is a primitive incumbents cannot acquire by hiring.

PE-as-demand-channel is the second, and it is the move that is not getting enough attention. Blackstone is not writing a check; it is delivering customers. The OpenAI Deployment Company's investor consortium reportedly sponsors more than 2,000 portfolio businesses, each of them now a pre-warm enterprise buyer with a board-level mandate to deploy AI and a credible introduction from their owner. TPG, Bain Capital, Advent, and Brookfield each bring hundreds more. The PE firm is providing distribution as a primitive — captive demand that bypasses the entire enterprise sales cycle.

This is unprecedented. No marketplace, agency, or systems integrator has anything comparable. Wipro has to win each customer individually through RFPs, beauty contests, and twelve-month sales cycles. McKinsey wins through partner relationships, but each relationship is bilateral and slow to scale. The lab-PE entity has the customers handed to it through the limited partner relationship. A Blackstone portfolio company gets a call from their corporate-development counterpart: "We have a new partnership with Anthropic — they will send you a forward-deployed engineer next month to scope your AI deployment, and the cost is partially absorbed at the fund level." This is not how enterprise services have ever been sold. The sales cycle compresses from quarters to weeks because the relationship asymmetry is already in place.

The PE firms are not doing this for free, of course. They capture value in three ways simultaneously. First, the equity stake in the JV appreciates as the venture scales. Second, the portfolio companies get accelerated AI deployment, which improves their EBITDA and exit valuations. Third, the labs themselves get more usage, more lock-in, more revenue, which feeds back into the equity appreciation of any Anthropic or OpenAI shares held in adjacent funds. The PE firm is not just a distributor; it is a yield-stacking vehicle on the deployment of frontier AI across its book. This is why $1.5B and $4B can be raised on these structures in days. The LPs see the stacking.

Agent governance substrate is the third primitive. Deployment gateways, monitoring and rollback, decision-rights frameworks, jailbreak-resistance testing — the operational layer of what I have called Compliance as Differentiation on the Map. When FIS announced its anti-money-laundering agent with Bank of Montreal and Amalgamated Bank as the first two customers, the substance of Anthropic's contribution was less the model and more the design of context boundaries, shadow autonomy controls, and the rules for which decisions the agent makes versus which escalate to humans. This is brand-new primitive territory. None of the prior twelve capture it because agents were not a thing.

The reason agent governance is a new primitive, rather than a sub-feature of existing primitives, is that it requires capabilities no traditional services firm has. It requires model-internals visibility to design decision-rights frameworks correctly. It requires red-teaming capacity to test jailbreak resistance. It requires alignment expertise to specify which behaviors the agent is permitted. None of this exists at Wipro or Accenture. They will hire for it, but they will not catch up before the next two model generations ship, and by then the substrate will be platform-specific in ways they cannot replicate without lab partnership.

The market signal that this primitive is load-bearing is visible in the FIS deal economics. As Aman Mahapatra of Tribeca Softtech pointed out, BMO and Amalgamated are not paying Anthropic directly for FDEs at quarterly consulting rates; FIS is absorbing the FDE engagement and amortizing it across its banking customer base. That economics structure only works if the deliverable is reusable across customers — and the reusable piece is the governance substrate, not the bespoke model fine-tuning. The agent firewall design for one bank generalizes to twenty banks. The model integration does not. The primitive that scales is governance.

It will be one of the largest moats in enterprise services over the next three years because every regulated buyer needs it and almost no one knows how to build it yet. The supply is currently a few hundred people, mostly at the labs themselves. The demand is every Fortune 2000 bank, insurer, healthcare system, government agency, and manufacturer.

Token-economy alignment is the fourth. Revenue scales with the customer's usage of the deployed system, not with hours billed. The vendor and the buyer are aligned on usage growth, not misaligned on engagement length. Consulting is capped at headcount-times-rate; tokens are not. Anthropic's reported ARR grew from $9B to over $44B in 2026, doubling roughly every six weeks. That is not a consulting growth rate. It is an API growth rate wearing a services jacket.

This primitive matters because it changes the unit economics of the entire engagement model. Traditional consulting has a fundamental scaling problem: revenue grows linearly with headcount, gross margins compress as you scale because you need more partners doing oversight, and there is a hard ceiling determined by how many skilled humans you can recruit and retain. Token economics is not bounded by headcount. The FDE deploys a system; the system runs; the system bills based on what it does; the FDE moves on to the next deployment. The vendor's revenue compounds without the vendor needing to hire proportionally.

Each of these four primitives is gated. Wipro cannot acquire frontier proximity because the labs will not sell them roadmap access — to do so would commoditize the lab's own positioning. Agencies cannot acquire PE-as-demand-channel because they are not a PE portfolio company, and a PE firm has no incentive to share portfolio access with a service vendor that does not stack yield back to the fund. Most firms cannot build agent governance substrate because they lack the model-internals visibility, the alignment talent, and the red-teaming capacity — the labs hire that talent before it can be hired anywhere else. And no traditional consultancy can shift to token-economy alignment because their cost structure is human FTEs, not inference compute, and their financial reporting expects revenue per consultant rather than revenue per deployment.

The skeptic's response to all of this is "this is just Palantir at scale." The response misses the topology. Palantir owned its primitives in one entity. Anthropic and OpenAI are separating model IP, capital, and deployment across a syndicate — the model lives at the lab, the capital lives at the PE firms, and the deployment muscle lives at the JV or DeployCo. That is a different shape, with different scaling and different lock-in properties. Palantir's growth was bounded by how fast it could hire FDEs. The lab-PE entity is bounded by how fast the PE syndicate can route portfolio companies to FDEs, which is a much faster constraint to relax.

The strategic divergence reveals which primitives each lab is betting on

Anthropic and OpenAI did not build the same thing. The two ventures look superficially similar — both copy the Palantir FDE playbook, both pull in PE capital, both target enterprise — but the underlying primitive bets are different.

Anthropic's joint venture is a minority partnership. Anthropic owns a piece, Blackstone and H&F own pieces, the model is federated. The structural intent is visible in the FIS deal: Anthropic does not deploy directly into BMO or Amalgamated; it deploys into FIS, which then resells to banks. This is the SAP model. SAP does not implement SAP at customers; partners do, and SAP captures the licensing economics while implementation partners capture the services economics. The amortization makes sense. As Tribeca Softtech's Aman Mahapatra noted, this approach is meaningfully better economics than direct engagements where each bank funds its own embedded engineering team to redesign the same controls in isolation. Anthropic's bet is that primitive three (binding commitments) becomes amortizable in a federated structure that primitives nine and ten (talent pipeline and proprietary platform) supports more cheaply than direct delivery.

OpenAI's Deployment Company is the opposite shape. OpenAI majority-owns and controls it. The Tomoro acquisition folds 150 FDEs directly under OpenAI. The structural intent is direct embedded delivery into customer organizations, no intermediating layer. This is the Goldman Sachs model — the embedded expert, billable, expensive, on-site, with the relationship owned by the parent. OpenAI's bet is that primitive ten (frontier proximity) and primitive fifteen (agent governance substrate) are best protected by direct control, with no implementation partner in between to capture margin or dilute roadmap fidelity.

The investor composition tells you something about each bet. Anthropic's syndicate is Wall Street-heavy: Blackstone, H&F, Goldman, Apollo, General Atlantic, GIC, Leonard Green. These are alternative asset managers with deep portfolio company access and a strong banking and financial services lean. The early customer wins (FIS, banks) reflect this. Anthropic is betting on financial services as the primary deployment vertical and is leveraging investors who own the customer base.

OpenAI's syndicate is broader: TPG, Advent, Bain Capital, Brookfield, plus 15 others. The PE firms are more sector-diverse, and the consulting-firm involvement (TPG's stake in Cognizant-adjacent investments, Bain Capital's relationship to Bain & Company) suggests OpenAI is aiming at a cross-industry play — manufacturing, healthcare, retail, finance, infrastructure — rather than concentrating in one vertical. OpenAI is betting that primitive fourteen (PE-as-demand-channel) is most defensible when the PE network is wide and shallow rather than narrow and deep.

These are not minor tactical differences. They are competing theories of which primitives produce the most defensible compounding. If Anthropic's bet is right, the future of enterprise AI services is a federated ecosystem of regulated-industry integrators (FIS, Epic, Workday, ServiceNow) each running on a single model layer beneath them, with Anthropic capturing rent through licensing and amortized FDE hours. If OpenAI's bet is right, the future is direct embedded delivery at scale, with OpenAI capturing both the services margin and the platform margin through the same entity.

Both bets carry risk. Anthropic's federated model depends on the implementation partners actually being able to operate independently after Anthropic's FDEs leave — if FIS cannot scale the agent infrastructure without continuous Anthropic intervention, the federated economics collapse and Anthropic has to direct-engage anyway. OpenAI's direct model depends on the DeployCo unit economics not converging to Accenture-style services margins — if gross margins compress toward 25% on services revenue, the token-economy alignment primitive is not holding and the unit looks like a high-cost services firm with the wrong owner.

Brad Lightcap's framing of "production-readiness rather than transformation" at the DeployCo launch suggests OpenAI is aware of the risk and is trying to keep engagements narrowly scoped to deployment rather than open-ended consulting. Whether discipline holds at scale is the question. Consulting margin compression usually happens when the firm cannot say no to scope creep, and DeployCo's investor pressure to grow revenue will push hard against that discipline.

A historical parallel

The shift visible here has a clean prior analog: the collapse of the Sun Microsystems services ecosystem when AWS went direct to developers.

In the 2000s, Sun shipped servers, and a generation of systems integrators — including Wipro, Infosys, and the rest of the Indian IT services majors — built billion-dollar practices installing, maintaining, and customizing Sun deployments. Sun captured hardware margins. The SIs captured services margins. Both grew. The relationship was symbiotic for roughly a decade.

Then AWS launched, and Amazon did something Sun never had: it bypassed the SI layer entirely and sold compute as a self-service API directly to developers. The first wave of buyers were startups and digital-native companies who did not need an integrator. The second wave was enterprise IT shops who realized they could provision their own infrastructure faster and cheaper than waiting for an SI engagement. Within a decade, Sun's hardware business was gone (acquired by Oracle), the SI ecosystem built around it was repositioning desperately, and AWS had become the most valuable infrastructure business in history.

The structural lesson is that the platform layer eventually goes direct when the services margin between platform and buyer becomes large enough to attract a direct play. Sun's mistake was assuming the SI layer was a permanent partner. Amazon recognized that the SI layer was a margin opportunity dressed up as a partner.

Anthropic and OpenAI are doing the same thing now. The traditional reading was that the labs would license models to AWS, Azure, and GCP, who would in turn sell to enterprises through Accenture, Deloitte, Wipro, and TCS. The labs have decided this is too many margin layers. The Anthropic-Blackstone and OpenAI-TPG ventures are the direct play. They are bypassing the SI layer, going direct to enterprises, and capturing both the platform and services margins in the same entity.

This is not a perfect analogy. Cloud is more self-service than enterprise AI deployment, and the FDE model is more high-touch than AWS console. But the structural dynamic — platform layer recognizing it can go direct and capture the SI margin — is the same. The casualty was the Sun-era SI ecosystem. The casualty in this cycle is the Wipro-era SI ecosystem.

The structural consequences across the services market

There are now three competing primitive bundles in enterprise services. The agency bundle (primitives 1–8) serves small and mid-market clients. The Wipro and Palantir bundles (1–12) serve enterprise and regulated industries. The lab-PE bundle (1–16) is new, frontier-aimed, and structurally exclusive.

Three things follow.

First, the mid-tier services market is about to be squeezed from both sides. The bottom is being eaten by AI agents replacing routine consulting work — the Fiverr-tier categories first (basic copywriting, simple coding, design, transcription), then encroaching upward through accounts payable automation, contract review, basic data engineering, and entry-level analytical work. The top is being claimed by the lab-PE entity before incumbents could plausibly assemble the new primitives. Wipro and Cognizant are exposed on both flanks. Their talent-pipeline moat (primitive nine) erodes as AI handles work that used to require 200,000 engineers. Their scale-and-duration model converts poorly to token economics because the cost structure is human-FTE-shaped, not inference-compute-shaped. Their multi-year embedded engagements look great on revenue stability and terrible on growth, because the buyer's appetite for human-installed legacy systems is declining quarter over quarter.

The talent question follows. What happens to the roughly 600,000 engineers currently employed at Wipro, TCS, Infosys, and HCL when the primitive bundle they serve becomes commoditized? The optimistic answer is that they retrain into agent governance, FDE work, and AI-native deployment. The realistic answer is that retraining at that scale, in that compressed timeframe, has never been pulled off cleanly. Sun's SI ecosystem did not retrain cleanly into AWS; it took a decade and substantial workforce attrition. Expect similar dynamics here. The Indian IT services majors will not disappear, but they will shed headcount and reposition into narrower, lower-margin work for the next decade.

Second, a gap opens that none of the three bundles fills. The Anthropic-Blackstone JV implicitly locks the buyer to Anthropic. OpenAI's DeployCo locks the buyer to OpenAI. Many enterprises will want vendor neutrality — they want Claude for some workflows and GPT for others, and to keep optionality on whoever ships frontier capability next quarter, which might be Google or Meta or a Chinese model lab nobody has heard of yet. They need agent-governance primitives that work across labs. They need handoff durability so they are not held hostage to vendor consultants forever. They need the equivalent of a multi-cloud strategy for AI services.

The Gartner forecast that by 2028, 70% of enterprises will abandon FDE-led engagements over cost and maintainability is not a forecast about AI failure. It is a forecast about which buyers will need the missing primitive: vendor-neutral, portable, durable agent governance with transferable knowledge. The buyers who can afford a permanent Anthropic FDE engagement will keep one. The buyers who cannot — and there are far more of these than there are Fortune 100 budgets to support them — will need an alternative.

That alternative is the actual disruption opportunity. It is not "another consulting firm with AI." It is a new tier — independent FDE pods, vertically specialized, lab-agnostic, sitting between the marketplace tier (primitives 1–8) and the lab-PE frontier tier (1–16). The lab-PE entities will not fill this gap because they do not want to be vendor-neutral — vendor neutrality undermines lock-in, which is the entire point of the FDE model. Existing consultancies cannot fill it because they do not have agent-governance primitives (number fifteen). Marketplaces cannot fill it because they do not have institutional predictability primitives (one through eight).

A startup that assembles primitives one through eight plus primitive fifteen — institutional predictability plus vendor-neutral agent governance — has a defensible position. It will not compete with the lab-PE entity at the Fortune 100 frontier. It will compete with traditional mid-tier integrators (Cognizant, Genpact, EPAM, smaller boutiques) for the buyers who want AI deployment without lab lock-in. That market is large, growing, and currently unserved.

Third, the sector-specific implications diverge. In financial services, the Anthropic-FIS-bank cascade is already operating, and the regulatory operating capability (primitive eleven) plus agent governance substrate (fifteen) combination creates a near-insurmountable moat for regulated workflows. Expect Anthropic-shaped deployments to dominate KYC, AML, fraud detection, and credit underwriting across Tier 1 banks within 18 months. In healthcare, the regulatory primitive is even more decisive (HIPAA, FDA, state-by-state insurance regulation), and the talent supply for healthcare AI governance is even thinner. The first lab-PE entity to crack healthcare governance will own the vertical. In defense and intelligence, Palantir's existing position is hard to displace, but Anthropic and OpenAI both have defense-adjacent contracts (Anthropic with AWS GovCloud, OpenAI with various Department of Defense pilots) and the regulatory capability gap will be the deciding factor. In manufacturing and supply chain, the picture is more fragmented because the regulatory primitive matters less but the talent pipeline primitive matters more, and that favors the SI incumbents in the short term. The lab-PE play will reach manufacturing last.

What to watch

Three signals over the next eighteen months will reveal whether the new species takes hold or whether the lab-PE entity proves to be a temporary phenomenon.

Whether OpenAI's DeployCo gross margins resemble OpenAI's or Accenture's. If the unit's margins collapse toward services norms (20-30%), the token-economy alignment primitive is not holding and DeployCo is just a high-cost services firm with the wrong owner. If margins stay closer to platform norms (50%+), the primitive holds and OpenAI has invented a new financial form. Watch the first earnings disclosure that breaks out the unit, which will probably come from one of the PE syndicate's portfolio companies before OpenAI itself files publicly.

Whether Anthropic's federated model — the FIS-style amortization of FDE costs across a customer base — works at scale. The economics are better than direct engagements, but the dependency structure is messier. If FIS and the other implementation partners absorb FDE engagements and successfully resell, the SAP-style ecosystem strategy wins and Anthropic captures licensing economics on top of a multi-tier services stack. If the implementation partners cannot scale the work without continuous Anthropic intervention, Anthropic ends up needing to direct-engage and starts looking more like OpenAI, which would represent a strategic failure of the federated bet.

Whether a credible vendor-neutral middle tier emerges. If it does, the lab-PE entities will eventually try to acquire it, the way Palantir tried to extend Foundry into adjacent categories — and the acquisition will validate the gap I have named. If it does not, the labs lock in frontier services and the gap goes unfilled, which would suggest the moat from primitives nine through sixteen is wider than I have estimated. I expect the first credible vendor-neutral player to be a Toptal-class platform that pivots upmarket by adding agent governance primitives, or an existing boutique consultancy (Slalom, West Monroe, BCG Digital Ventures) that builds lab-agnostic FDE pods. Watch for funding announcements at the $50M-$150M Series B range in that space.

What is not in question is the direction of travel. The primitive set is permanently expanded. The form factor of an enterprise services firm now includes frontier proximity, PE-as-demand-channel, agent governance substrate, and token-economy alignment as table stakes for the frontier tier. The buyers have not fully noticed yet. The incumbents are pricing the old form factor. And the gap I have named — vendor-neutral, durable, mid-tier agent services — is the disruption opportunity that nobody has built yet, because everyone is still arguing about whether AI consulting is going to be big.

It is not going to be big. It is going to be species-redefining. The big incumbents that survive will not be the ones that adapt their existing primitives. They will be the ones that assemble new ones, or that find the new gaps the lab-PE entity will not fill. Most will not manage either. Most will be pricing for a market that does not exist anymore.

Wipro is not the competition. Wipro is the casualty.