Microsoft CEO Nadella: Supports AI Slowdown
At the recent All-In Summit, Microsoft CEO Satya Nadella participated in an exclusive interview with the All-In Podcast team, publicly echoing the cautious stance of Anthropic CEO Dario Amodei in his earlier blog post "Pacing the Frontier." Nadella stated that Microsoft "welcomes a thoughtful pace control to address alignment issues."
Nadella compared AI safety governance to handling disruptive bugs from an engineering perspective, stating that one of the most important abilities of early engineering leaders is to learn how to deal with "fatal bugs that can halt releases." He suggested that the industry should treat AI risks as engineering problems that can be disassembled and systematically solved, rather than just making slogan-like statements. He also specifically mentioned that "persistent agents" are bringing a new category of "insider risk," which requires a completely new set of systematic responses.
Just before and after the interview, Microsoft released a 37-page "Humanist AI Code of Conduct" to govern its self-developed MAI series models: it clearly states the "human-first" principle, rejects seeking legal personhood for AI, requires systems to always remain under human control with built-in emergency shut-off switches, and prohibits models from deceiving humans or accessing systems without authorization.
In the program, Nadella elaborated on Microsoft's "master plan": not pursuing the development of cutting-edge models but betting on application-layer infrastructure—what he calls "harness." He revealed that Microsoft 365 Copilot has surpassed 30 million subscribers, corresponding to about 250 million to 300 million potential enterprise users, and a total target market of approximately 450 million knowledge workers, including students. He also mentioned that the reasoning cost of Microsoft's self-developed models has been reduced to 15 to 60 cents per million tokens, while some cutting-edge models are priced as high as about $50.
Regarding capital expenditure, Nadella emphasized that Microsoft's $175 billion spending is "demand-driven," serving thousands of long-tail enterprise customers rather than just one or two flagship model companies. About 60% of this spending is directed towards short-cycle "hardware kits" like racks and chips, while the rest goes to long-cycle assets like land, electricity, and factories, employing a mix of self-built, leased, and managed approaches to maintain flexibility. He also stated that "talking excessively about the scale of capital expenditure is not a strength but a flaw."
The program also discussed the question of "who can win in AI"—as the price war among cutting-edge models continues to compress profit margins, Nadella appeared relatively calm, believing that the cost reduction actually benefits Microsoft's "infrastructure-agnostic" approach. He suggested that companies should conduct a "pull-out test": if a company's AI evaluation system cannot continue to be reused after changing the underlying model, it indicates that the business has been locked in by a single supplier. This statement has also been interpreted as a shift of funds from "purely betting on cutting-edge model companies" to companies that can help enterprises escape vendor lock-in and gain control over their own AI evaluation and data sovereignty in the infrastructure and tools layer.
Just before and after the interview, Beijing publicly rejected Dario Amodei's earlier suggestion to maintain chip export restrictions to China, with Chinese officials stating that such confrontational postures do not contribute to global AI governance cooperation—this context resonates with Nadella's stance in the program advocating for open models to reach different countries and communities more broadly.
Source: Public Information
ABAB AI Insight
Nadella's strategic statements have historical context. Since 2023, he has led Microsoft's massive investment in OpenAI and deeply integrated its models, while continuously signaling a "multi-model strategy," including introducing Anthropic's Claude model into the Copilot product line and supporting self-developed MAI series models to hedge against reliance on a single supplier. This public support for the "slowdown" narrative, along with the simultaneous release of the "Humanist AI Code of Conduct," continues the same approach—using safety and compliance narratives to pave the way for Microsoft to break free from reliance on a single model supplier and regain control.
The allocation logic of Microsoft's $175 billion capital expenditure is worth examining: about 60% is allocated to short-cycle hardware, alongside long-cycle real estate and power assets, with a mixed layout of self-built, leased, and managed approaches to retain switching flexibility; at the same time, Microsoft is massively subsidizing the reasoning costs of its self-developed MAI models, reducing its price per million tokens to just a fraction of some cutting-edge models, systematically guiding customers away from dependence on a single cutting-edge model towards Microsoft's own application and infrastructure layer—this is a typical capital path of "exchanging scale cost advantages for customer stickiness and bargaining power."
This approach can be compared to Amazon AWS's early cloud computing strategy of "not doing applications, only doing infrastructure," locking in a massive long-tail customer base by lowering marginal costs through scale effects, and can also be contrasted with IBM's historical path of locking in customers through "software + services" layers during the mainframe era, rather than purely competing on hardware performance. In terms of industry positioning, the AI industry is shifting from a stage of "competing on model parameter scale" in an arms race to a stage of "competing on who can deploy model capabilities at low cost into enterprise workflows" for application penetration. Nadella's statements indicate that Microsoft believes this turning point has arrived.
Structural judgment: This is a restructuring of the industry chain. As the technical premium of the model layer is rapidly diluted due to price wars and open-source competition, the truly scarce and value-extracting segment is shifting from "training the strongest model" to "who masters the enterprise-level deployment, evaluation, safety compliance, and data sovereignty infrastructure layer"; through the design of capital expenditure structure, subsidies for self-developed model costs, and compliance narratives like the "code of conduct," Microsoft is re-embedding itself in the most sticky and powerful bargaining position in this value chain, essentially a process of shifting dominance from "model providers" to "infrastructure and application layer integrators."
ABAB News · Cognitive Law
Those who talk about safety are often also vying for dominance.
Scale reduces costs, and costs lock in customers.
In the end, the arms race is about who can endure longer.