Why hyperscalers are descending toward the result, and why the return waits for them where they control only a subset of the factors
On July 2, 2026, Microsoft announced Frontier Company: 2.5 billion dollars and roughly six thousand engineers and industry specialists embedded inside enterprise customers to design, build, and operate their AI systems. An earlier and more immediate reading treated the announcement as a displacement of the authorship of value proof. This piece takes the level above. It reads Frontier as the first observable case study of a structural move: the migration of software rent from the availability of a means toward the result itself. The thesis fits in one sentence. Enterprise AI value is multiplicative, a single factor reduced toward zero collapses the product, and the factor that most often sits near zero, the customer’s institutional capacity, is precisely the one no vendor can deliver from the outside.
The cause is economic before it is strategic. The layer that briefly held the rent, the model itself, is commoditizing: the Stanford AI Index puts the cost of a GPT-3.5-level query on MMLU near twenty dollars per million tokens in late 2022 and around seven cents by late 2024, while Epoch AI measures inference prices falling close to two hundred times a year at a fixed capability. When invoking intelligence collapses in price, the rent must leave the model for the layer that does not commoditize: integration, governance, the customer’s own outcome. Vendor convergence is the evidence, not imitation: Amazon committed a billion dollars two days before Frontier, and Anthropic and OpenAI had stood up embedded groups back in May. The point most analyses miss sits on the demand side. The customer accepts the transfer of responsibility because it is also in distress: AI skills have become the hardest category of role to fill according to ManpowerGroup (seventy-two percent of employers, France near the top at seventy-four), EBIT impact remains rare despite broad adoption per McKinsey, BCG apportions about seventy percent of AI success to people and process, and unstructured data grows faster than the ability to govern it. Two distresses meet. Neither is a position of strength.
The tempting reading has Microsoft acting on the wrong side of a single bottleneck. It is wrong for a precise reason: the correct object is not a constraint but a production function. The right form is not the fixed-proportion Leontief, too rigid, but the O-Ring theory Michael Kremer formalized (Quarterly Journal of Economics, 1993): output is the product of task qualities, limited substitution remains possible, and yet the weakest task dominates the outcome. Paul Milgrom and John Roberts gave the organizational counterpart: raising one activity returns only when its complements rise alongside it. The consequence is immediate. Microsoft can raise infrastructure and part of the skills factor by placing capable engineers on site; it does not control governance, adoption, or the arbitration of internal incentives, and no amount of external engineering lifts a factor the vendor has no standing to touch. This is why the now-famous MIT NANDA finding, ninety-five percent of generative AI pilots with no measurable profit-and-loss impact, reads not as a deficiency of the models but as the statistical signature of multiplicative collapse: the models work in the demonstration, the product fails in the business because at least one client-held factor is near zero.
The friction in outcome-based accountability has an ontological source, and it clears once four classes of resource are separated rather than two. An asset is a thing that is transferable, financeable, deliverable: Azure, Copilot, a retrieval pipeline, an agent. A capability is the localized competence to operate that asset; it can be scaffolded from outside for a time. A capacity is the broader institutionalized state, endogenous to the firm, cumulative over years, non-transferable by contract. An institution is the layer that distributes decision rights and manufactures baseline trust. A vendor can sell an asset and briefly rent a capability; it can sell neither a capacity nor an institution, for the same reason a gym cannot sell fitness. It sells access to equipment and the presence of a coach; fitness is what the member builds after the coach goes home. This is the paradox at the center of the proposition: Frontier deploys an asset and bills for the institution of a capacity it cannot itself institute. The cadence makes it concrete. Engagements are described as forty-five day sprints staffed by five or six people, a rhythm calibrated for the demo-to-deployment gap of lightly regulated workflows; yet the showcase customers, Novo Nordisk and the London Stock Exchange Group, operate where qualification, change control, and audit trails do not compress to forty-five days. A capability that lasts only while the vendor remains is not a capability. It is a dependency.
A trust platform exposes controls, and a control is only as useful as the human role that operates it on the other side. Five forms of accountability must exist on the client side, and they form a hierarchy rather than a list: a Data Steward answers for the truth and lineage of the sources, a Knowledge Steward for the currency of the corpus, an AI Steward for drift, the boundary of refusal, and inference cost, a Process Owner for the operational and legal liability of the workflow. Crowning them, the Decision Owner governs the threshold of automated reasoning itself: what confidence is sufficient to act, when a human returns to the loop, how the cost of an error is weighed against the velocity of autonomy. The first is sometimes staffed; the last is almost nowhere formalized, which is a precise way of saying that the platform’s most consequential control connects, on the customer’s side, to no one. The causal story then reverses: the absent steward is not the cause of failure but the symptom of a missing capacity. And the measurement question sharpens. The complaint that Microsoft measures its own return is true but shallow; the consequential fact is that the vendor progressively defines the instrument, the metrics, the observables, the traces. Marking your own homework is embarrassing; designing the rubric everyone else will use is structural power, and it is the quieter of the two.
A grid that cannot be refuted is not worth publishing. Two futures test the thesis. The first is capacity-as-a-service: a share of the market may decline to institutionalize and rent its cognitive production permanently, converting the accountability deficit into a standing fee; if the market settles there, Frontier is not overreaching, it is early. The second is the internal AI factory: Gartner projects that about eighty percent of large enterprise finance teams will run internally managed generative AI platforms by 2026, and the center of excellence has become the default answer to scaling. The irony sits inside Microsoft’s own material, whose Cloud Adoption Framework instructs customers to build an internal center of excellence while Frontier offers to be that center on their behalf. The likely settlement is a split by maturity: the mature firm repatriates the capability, the immature one slides into permanent rental. The falsification window is twelve to eighteen months, along two vectors: autonomy, whether the accountability structures survive the engineers’ withdrawal, and institutionalization, whether capacity-as-a-service proves durable indefinitely without internal maturity. The grid generalizes to Amazon Web Services, Google Cloud, Oracle, and Salesforce, because none controls more than a subset of the factors. The difficulty they will all meet there is not one of engineering. The result is not deployed. It is instituted.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.