Microsoft CEO Nadella: The compute price war is inevitable; in five years, enterprises' core asset will be proprietary evaluation benchmarks

On October 26, 2026, Microsoft Chairman and CEO Satya Nadella appeared on the venture capital podcast 20VC for a 24-minute conversation with host Harry Stebbings.

(Left: Microsoft Chairman and CEO Satya Nadella; Right: 20VC host Harry Stebbings)

The timing of this interview is worth noting. Five days earlier, on October 21, Nadella delivered the opening keynote at the Microsoft Frontier Tour at the ExCeL exhibition center in London, with 6,500 partners and customers in attendance. He proposed that enterprises deploying AI must solve three things: keep humans in the decision loop, control their own knowledge, and build trust. This podcast is an expanded, investor-facing version of the same set of judgments.

Addressing the computing cost and model competitive landscape that the market cares about most, Nadella said bluntly that, judging by the laws of technology, for the vast majority of task-specific business tasks, model capability will soon hit the saturation point of diminishing marginal returns. He emphasized:

Once capability saturates, open-source models or extremely cheap lightweight closed-source models will rapidly take over the market.

Computing power will be reduced to a commodity, and "the underlying layer will definitely go through an extremely brutal price war"

As model inference costs fall, the market has been constantly worried about the commercial moat of foundation models. Nadella did not mince words about this: he predicts that frontier models will belong only to a small, top-tier market, while everyday computing power will become thoroughly commoditized.

Nadella said:

So my prediction is very clear: the most cutting-edge top-tier models will always have their unique top-tier market, such as breakthrough scientific exploration, and the frontier baseline will keep being raised; but in the vast majority of everyday enterprise scenarios, the bulk of computing consumption will become commoditized, reduced to an extremely cheap basic raw material.

When asked whether the underlying traffic would be eaten up by open-source models, he gave an affirmative direction:

It could be open-source models, or it could be those lightweight closed-source models engaged in extremely fierce price competition. The underlying layer will definitely go through an extremely brutal price war.

The best evidence of this trend is the scheduling mechanism of Microsoft's own flagship products. Nadella admitted:

Just look at our own real scheduling mechanism inside Microsoft Copilot: every day we mix and match cheap closed-source models and open-source models in large quantities.

This statement in effect undermines the industry's basis for paying a premium for the scarcity of frontier models, and also reveals that Microsoft is transforming from a single-model provider into a model-invocation orchestrator and governor with more stable profit margins.

Overturning the "emergence theory" of safety alignment in favor of a radical "zero-trust architecture"

Over the past few years, the industry's default premise has been that the stronger the model capability, the more safety alignment will naturally keep up. Nadella said he was a believer in this premise back then, but has now abandoned it.

Nadella emphasized that traditional post-hoc reinforcement learning cannot achieve thorough alignment; alignment must be hard-coded directly into the model weights at the lowest level. He said:

I used to think that model 'alignment' would naturally emerge automatically as parameter scale expanded. Reality has proven that it does not.

He even rejected the path of post-hoc remediation:

I simply don't believe you can patch your way to thorough alignment through post-hoc reinforcement learning. Alignment must be hard-coded directly into the model weights at the lowest level.

Based on the current reality that underlying models cannot guarantee absolute trustworthiness, he proposed a brand-new engineering defense approach:

The safest posture for using advanced models today is to assume that the model has already been compromised.

He advocates:

Since the underlying model cannot guarantee absolute trustworthiness, we build an ironclad trust fence outside the model itself, being extremely aggressive in policy configuration and abnormal behavior monitoring.

In addition, regarding the recent practice by some model vendors in the industry of hiding the "chain of thought," he sharply criticized it, believing that "this makes the problem even more severe." He warned:

Frankly speaking, if closed-source model vendors do not recognize the necessity of this 'transparency toward trusted partners,' this opacity will instead become the biggest booster for the open-source ecosystem.

Microsoft has prepared Agent 365 and Purview for this layer: assigning each digital employee an independent identity, mandating the recording of all behavioral traces and the complete chain of thought, with chain-of-thought records required to be human-readable plain English, and tamper-proof logs handed over to multiple independent models for cross-review.

As a company that simultaneously sells models, orchestration frameworks, and governance tools, Microsoft has products to sell at all three layers. Defining security as external trust engineering is equivalent to directing enterprise budgets toward the governance layer where it has already staked out its position.

Beware the "reverse information paradox": in five years, the core of an enterprise will be "proprietary evaluation benchmarks"

On the business logic of AI empowering the enterprise B2B market, Nadella proposed the "reverse information paradox."

Arrow's classic information paradox is that when the technology owner fully discloses trade secrets to the buyer, without patent protection, the commercial value evaporates instantly; Nadella says this relationship has been reversed, and what now needs protecting is the enterprise customer as the buyer.

Nadella cautioned:

When enterprises interact with large models, even tiny business traces are providing reinforcement learning nourishment to large model companies for free. You are actually creating incremental value for the model vendors, rather than retaining the value within yourself.

Nadella said:

One thing I firmly believe is: in the future, every enterprise will build a set of proprietary evaluation benchmarks of its own, and this will become the enterprise's most core commercial intellectual property.

When asked what will become common sense in five years, he reiterated:

Unilever is a model case. They first used frontier models to run the process from sales orders to cash collection, then trained on business interaction traces. Claims processing has now been switched to the lower-cost MAI, while the business logic remains in their own hands. This case comes from his October 21 London speech.

The hard constraint of hundred-billion-dollar training costs: it must be paid for by "global GDP growth"

Given that the entire industry is currently expanding infrastructure at a frenzied pace, with the cost of a single model training run even being cited as high as one hundred billion dollars, the market is full of worries about the return on investment of this "arms race." Nadella offered the ultimate ledger from a macroeconomic perspective.

Nadella said:

This money must ultimately be paid for by an inflection point in real global GDP growth.

He explained that no financial income statement can defy gravity forever, and companies will ultimately face hard constraints on revenue and net profit. If a company can sustain hundred-billion-dollar R&D, "then the premise must be that the incremental value this company brings to the global economy is already large enough to legitimately earn it such a massive commercial return."

He offered the most core metric to watch:

Can AI actually cause the global GDP growth rate to make a substantial upward jump?

He is optimistic about this, specifically citing cost reduction and efficiency gains in the deep waters of public healthcare, as well as the endless potential demand elasticity of software development itself.

Finally, on the question of whether the industry will move toward "full-stack vertical integration monopolies" from chips to applications, he considers that highly unlikely.

English text is being prepared.

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.