La malédiction du gagnant : pourquoi les entreprises de modèles IA pourraient perdre la guerre des marges

Spend billions training the best model in the world, then watch your customer treat it like a replaceable ingredient. That is the wager inside Satya Nadella’s most interesting claim about AI economics. The company doing the deepest research may end up holding the weakest pricing power.

It sounds backward because the public story has been simple. Models get smarter, so model makers capture the profits. If intelligence is the scarce resource, the labs that manufacture it should sit at the top of the food chain.

Nadella is arguing for a different map. In his conversation with Dwarkesh Patel and Dylan Patel, he suggested that model companies may face a winner’s curse: they do the expensive, dazzling work of invention, only to find that their advantage is always close to being copied, matched, routed around, or simply priced down. The money, on this view, collects elsewhere. It accrues to the software layer that makes models usable, the data layer that makes them valuable, and the distribution layer that makes them hard to dislodge.

That is a sharp claim. It also happens to line up neatly with Microsoft’s interests, which is worth remembering throughout. Still, self-interest does not make it false. In this case, it may even be clarifying.

The paradox sitting inside frontier AI

Nadella’s phrase, “one copy away from being commoditized,” lands because it exposes a structural asymmetry. Frontier labs absorb giant fixed costs: talent, data pipelines, training clusters, post-training, safety work, inference optimization, and constant iteration. Customers, meanwhile, buy outcomes. They care whether a task gets done accurately, cheaply, and within the boundaries of their systems.

Those are not the same game.

If a model company creates a breakthrough, the technical achievement is real. Yet the business benefit depends on how long that lead remains unique and how much of it can be translated into pricing power. In software, superiority does not automatically produce durable margins. It produces a temporary edge. The question is whether that edge hardens into dependence before competitors catch up.

With models, hardening is difficult. A rival does not need to replicate every internal detail to collapse your pricing umbrella. It only needs to become good enough on the tasks buyers actually run, or to be cheaper at the same quality threshold, or to fit better into an existing stack. Exact copying is not the point. Functional substitution is.

That is why the winner’s curse framing works. The lab that pays for the leap may be subsidizing the market education and product design of everyone downstream. Others can wait, watch, and build on top.

Open source keeps a floor under capability and a ceiling over price

Nadella’s second claim is even more important than the first: there will always be an open source model that is fairly capable. That sentence carries a whole market structure inside it.

A “fairly capable” open model does not need to beat the frontier to matter. It only needs to exist as a credible alternative. Once it does, every proprietary model vendor loses the fantasy of unlimited pricing. The customer now has options: self-host, fine-tune, distill, mix vendors, or keep a premium model for narrow cases while shifting bulk volume elsewhere.

This dynamic has strengthened as open models improve and as the tooling around them gets better. DeepSeek, Qwen, Llama, and the broader ecosystem changed the conversation from “Can open models compete?” to “On which tasks, at what cost, and with what operational tradeoffs?” That is a much more dangerous question for any company hoping to collect software-like margins from intelligence alone.

The important pressure is not ideological. Enterprises are not adopting open models to make a philosophical statement about openness. They adopt them because compliance teams want control, infrastructure teams want cost predictability, and product teams want leverage in vendor negotiations. Open weights become a permanent reference price for intelligence.

That does not mean proprietary models are doomed. Frontier labs can still command premiums, especially when they lead on reasoning, multimodal performance, coding, safety, or latency. The issue is duration. Premiums earned by being far ahead are one thing. Premiums that survive once the market catches up are another. Open source makes the second category hard to defend.

Rising margins do not disprove the trap

Dylan Patel reportedly pushed Nadella with a fair objection: some model companies were seeing inference margins improve, not collapse. Anthropic, he noted, had moved from under 40% to above 60% on inference margins within the year. If the market were truly commoditizing, why would margins rise?

Because markets rarely commoditize in a straight line.

In the early phase of a new technology, demand can grow faster than competition matures. Buyers will pay up when the quality gap is meaningful and when the cost of failure is high. If one model is materially better at coding agents, legal analysis, or research workflows, the price elasticity is low for a while. Scarcity and urgency create excellent temporary businesses.

But temporary is doing a lot of work there.

A company buying model access for a product roadmap is not marrying the vendor. It is renting capability. Nadella’s response captured that plainly: if pricing stays high, builders will substitute. Maybe not immediately. Maybe not for the hardest workloads. Yet the option remains alive, and that option is enough to discipline margins over time.

This matters because enterprise software buyers are experts in abstraction. They insert routers, evaluation harnesses, fallback models, and cost controls between themselves and any supplier that looks too powerful. Once model performance becomes measurable on internal tasks, procurement starts behaving like procurement. Sentiment drains away. A benchmark lead becomes an input into a spreadsheet.

The model vendor may still have a great business. It just may not have the kind of business that public excitement assumes.

The real product is reliable behavior inside messy systems

Nadella’s preferred word for the downstream layer is “scaffolding.” The term is ungainly, but the idea is right. Models are impressive in isolation and erratic in practice. They hallucinate, forget constraints, misuse tools, lose the thread across long contexts, and fail silently in ways that look polished until they hit production. Making them useful requires a surrounding structure.

That structure includes routing, retries, grounding, permission checks, tool use, memory, evaluation, monitoring, cost control, and user interface design. It also includes something less glamorous and more valuable: deciding what the model should even be allowed to do inside a business process.

This is where a lot of people still underestimate the work. They see a chat box and assume the value lives in the model response. In production systems, the response is just one component. The harder job is turning probabilistic intelligence into something a finance team, legal team, hospital system, or call center will trust on Tuesday at 3:17 p.m., when the data is messy and the stakes are not hypothetical.

That layer tends to accumulate switching costs. Once a company has built evaluation datasets, context pipelines, audit trails, and task-specific orchestration around a workflow, swapping models becomes easier than swapping the surrounding product. The wrapper joke misses this completely. Some wrappers are flimsy. Others are where the economic moat forms.

A model without this supporting structure resembles a gifted contractor dropped into a factory without badges, manuals, or supervision. Smart, yes. Productive, not yet.

Excel tells the story better than any chatbot demo

Nadella’s example of an Excel agent is revealing because it exposes where durable value may actually live. An agent that works deeply inside Excel is not just calling a model API and spraying text over a familiar interface. It needs to understand formulas, tables, pivots, naming conventions, spreadsheet state, user intent, permissions, and the difference between a plausible answer and an action that breaks a business process.

That knowledge is not generic intelligence. It is application intelligence.

Once you integrate a model into the primitives of a mature product, you create a dense layer of behavior that outsiders cannot easily replicate. The moat is not that the model has read the Excel manual. The moat is that the product owner controls the action space, the context, the telemetry, the feedback loops, and the trust relationship with the user.

This pattern generalizes. A legal workflow tool can teach models how matter management, precedent retrieval, and approval chains actually work. A healthcare product can encode billing rules, patient safety constraints, and documentation requirements. A customer support platform can link intent detection to refunds, policy databases, and human escalation. In each case, the model is necessary but insufficient. The defendable product is the choreography.

That is why Nadella’s claim lands hardest against the idea that model companies automatically “own” applications built on top of them. They can move upward when it suits them. They will certainly try. But climbing into vertical software is very different from exposing an API. The deeper you go into workflows, the more the problem stops looking like pure intelligence and starts looking like software design, domain expertise, and institutional trust.

Data liquidity beats raw data hoarding

There is another piece of Nadella’s argument that deserves more attention: the real barrier may be the liquidity of data. That phrase is much better than the usual talk about “proprietary data,” which often gets used as a magic charm.

Most companies already have plenty of proprietary data. The issue is that their data is fragmented, mislabeled, permissioned in conflicting ways, trapped in old systems, or detached from the moment a model needs it. A pile of documents in SharePoint, Salesforce, Slack, and a warehouse is not a strategic asset until it can be retrieved, reconciled, and grounded inside a workflow with the right context and controls.

Liquidity means the data can move to where the decision happens. It means the model can access the relevant slice at the right time, with traceability, freshness, and policy enforcement. That is much harder than storing embeddings and calling it a day.

Once framed this way, the value chain shifts again. The company that best mobilizes enterprise context may earn more durable returns than the company with the highest raw benchmark score. The reason is simple: context compounds. A model vendor can sell intelligence to many customers, but the product owner who sits inside the workflow collects feedback on what actually matters, what went wrong, what needed escalation, and what the user trusted enough to accept. That creates a flywheel around the application and its data interfaces, not only around the model.

This is also why distribution matters so much. The application where people already work has a privileged route to gather context, shape behavior, and learn from outcomes. That is not glamorous, but it is potent.

The market may split into capability creators and margin collectors

If Nadella is broadly right, the AI stack will not settle into a single winner-take-all layer. It will split. Frontier labs will remain essential because real capability advances still matter. Someone has to push reasoning, multimodality, agentic reliability, and efficiency forward. Yet being essential is not the same as capturing most of the profit pool.

We have seen versions of this before in technology. The firms that expand the frontier often do not keep all the value created by the frontier’s expansion. Sometimes they become the engine room for other companies with stronger distribution or better control of customer workflows. Sometimes they collect healthy returns without ever becoming the dominant margin story. Sometimes they do both for a while and then watch the balance shift.

The AI market seems especially prone to this split because the buyer can modularize the stack. A business can use one model for coding, another for summarization, an open model for internal retrieval, and a premium one for high-stakes reasoning. It can route requests dynamically and renegotiate as the market moves. That modularity weakens any single vendor’s hold unless the vendor owns much more than the model.

Which brings us back to Microsoft. Nadella is not merely making a prediction. He is describing a strategy. If you control productivity software, cloud infrastructure, identity, data platforms, developer tools, and distribution into the enterprise, then you are positioned to turn model competition into your advantage. You do not need every model win. You need the ability to absorb model innovation, swap suppliers, and present the customer with a complete system.

That is a strong place to bargain from.

Frontier labs still have cards to play

None of this means the model layer becomes irrelevant or permanently low-margin. A company that stays meaningfully ahead can still charge for that edge, especially in markets where quality differences map directly to revenue or risk reduction. If a frontier model saves a law firm dozens of billable hours per case or materially improves scientific research, buyers will pay premiums that look very healthy.

There is also a supply-side fact that protects the leading labs more than critics sometimes admit. Training at the frontier remains extremely expensive and operationally difficult. Access to top-tier compute, data, and research talent is not evenly distributed. That alone limits how many genuine peers any lab will face.

But scarcity at the frontier does not automatically settle downstream economics. It may just mean the model layer behaves more like advanced infrastructure than like classic consumer software. Infrastructure can be lucrative. It can also be highly negotiated, exposed to cost pressure, and vulnerable to vertical integration by its largest customers.

The labs with the best chance of escaping the winner’s curse may be the ones that stop imagining the model as the whole product. They will need distribution, workflow ownership, developer ecosystems, and some control over the context flowing into their systems. In other words, they will have to move outward from pure model excellence into the less romantic territory of software and enterprise integration.

That is harder than training a bigger model in one respect. It requires patience.

The margin war will be decided far from the benchmark charts

Benchmarks still matter. Capability progress is real, and the companies pushing it deserve the attention they get. Yet the industry keeps mistaking technical leadership for final economic leadership, as if the cleanest graph must also predict the fattest margin.

It usually does not work that neatly.

The buyer paying for AI at scale wants a system that fits their data, their approvals, their interfaces, their security model, and their budget. That buyer loves intelligence and distrusts dependency. The more mature the market becomes, the more aggressively that tension shapes purchasing behavior. Any supplier that cannot defend price on more than raw capability will feel it.

Nadella’s thesis is provocative because it flips the hero narrative. The model lab may produce the breakthrough, earn the headlines, and still discover that the stickiest value formed one layer up, where intelligence meets workflow, context, and distribution. That is the place where substitution gets harder and trust gets monetized.

The companies that internalize this early will build differently. They will treat models as vital and unstable inputs, not as the entire castle. The ones that ignore it may discover that they won the race to invent the future and left someone else to invoice for it.