Microsoft's Nadella Concedes Alignment Will Not "Fall Out" of Scaling, Pushes Zero-Trust Containment and Licensed-Weight MAI Models
Interview with 20VC's Harry Stebbings, published October 26, 2026
A change of mind on alignment, and a security posture built around it
The most consequential disclosure in Satya Nadella's conversation with Harry Stebbings is an admission of a changed view on a core assumption behind the scaling thesis. Nadella said he was an early believer in scaling laws, going back to 2018 and 2019, when "Dario, and Ilya, and Sam, and Greg, and Elon were all, I guess, one happy family." His revision is that "I used to always think that alignment will just fall out. I no longer believe that." He described scaling laws as still working and models as increasingly capable, but said alignment does not appear to emerge from scale and that mechanistic understanding of models "is also elusive."
Nadella argued that alignment has to be built into training, possibly pre-training, and that it cannot be added afterward. "I don't think you can even RL your way to alignment. You have to truly get alignment to be something that truly gets embedded in the weights right at the very beginning of it." That is a pointed view for the CEO of a company with a large stake in the leading closed-model developer, and it implies that post-training safety work from the labs may be insufficient.
His practical response is to separate AI safety from security. Microsoft's position is that customers should "assume a model is compromised," a thesis he laid out in a piece roughly 10 days before the interview, and should contain agents with an engineered zero-trust environment. He pointed to two Microsoft products, Agent 365 and Perception, which give every agent an identity and log its activity, including tool calls, every model call and the chain of thought. The logs must be tamper-proof so an agent cannot edit its own record. Policy is then set on top, and agent behavior is monitored against it, with multiple models used to check the output. Nadella pushed back on the idea that unsolved safety means nothing can be done on security: "People conflate an engineered security zero trust architectural system for containment and AI safety." His stance is to "deal with the situation as it exists" and be "super aggressive" in setting policy and monitoring deviations.
Chain-of-thought opacity as a pressure point on closed-model vendors
Nadella was blunt about the reliability of chain-of-thought monitoring, which underpins the architecture he described. Asked how reliable it is, he said: "It's not." Chain of thought can be unfaithful, and he said the problem worsens when a model company acknowledges that and responds by hiding it. He wants reasoning traces to be in English rather than an opaque internal language, and wants them available as-is to trusted partners so they are inspectable and auditable.
He framed this as a commercial warning to closed-model providers and as a potential tailwind for open weights: "Closed weight models are more powerful, but if they don't recognize the need to be more transparent, especially as their trusted partners use it, then we may have a problem." Nadella did not name a specific vendor, but the message to Microsoft's own partners is hard to miss.
MAI as "licensed weights" and a software business rather than a network-effects one
On Microsoft's own models, Nadella gave a clear description of the intended business model. MAI models will not be open weight, but he prefers to call them "licensed weights." He said: "I want to turn this into just a simple straightforward software business as opposed to a network effects intelligence business." The model's value to a customer will depend on their ability to use and control it, which fits his wider argument about firms keeping their own intelligence.
He also gave a customer example. Unilever has built an agentic system to automate order-to-cash and trade claims. It started on a frontier model and has since moved the trade claims workload to an MAI model, which Nadella said is lower cost and which Unilever controls because it can train on its own traces. Microsoft embedded forward-deployed engineers at the start, but Unilever's IT team now manages the system itself.
Open versus closed: commodity tokens dominate, but frontier persists
Asked to forecast traffic over 24 months, Nadella declined to give a percentage and said it depends on where the frontier sits. His view is that capabilities saturate for a given problem, at which point open-weight or cheaper closed models suffice, especially if a company can train on its own traces. "If the problem has not changed," he said, there is no need for the latest model. He expects a lasting market for the frontier, with science a possible next frontier-setting domain, but said "predominantly token usage will be commodity token usage."
He did not equate commodity with open source. It may be open weights or "cheap closed weights," he said, because "there'll be lots of price competition." He noted that Microsoft's own Copilot usage mixes cheaper closed-source models with open-weight ones. He also highlighted a "Hydra fusion" project on GitHub, a learned model, as part of a shift toward agent swarms in which many sub-agents, some of them local models, are called in parallel.
On the geopolitics, Stebbings noted that most open-weight models are Chinese. Nadella's answer was that unaligned models are a risk to every country, including China, and that he was pleased with progress from Western open-weight efforts, citing Nvidia's Jensen Huang specifically.
The "reverse information paradox" and the private eval as corporate IP
Nadella's most developed idea concerned the future of the firm. Drawing on Hayek's view that tacit knowledge is distributed across the economy, he argued there cannot be a single firm or a single intelligence. He then inverted Kenneth Arrow's information paradox. Traditionally the seller risks leaking value by disclosing a secret. Now, he said, it is the buyer who is exposed, because even the exhaust of using a model, such as thumbs-up and thumbs-down signals, creates value for the model company rather than the customer. "For this to be stable, more of that surplus that gets created by use of intelligence has to remain at the firm."
That retention does not have to happen only through weights. It can be in context, or in evals. Nadella said he is "very gung-ho" about every firm having a private eval benchmark as its core IP, with models coming in to "hill climb" on it. He said a firm might eventually be defined as "an AI benchmark," and named this as the thing that looks strange today but will be accepted within five years. Stebbings noted this currently applies to a rarefied set of companies such as Harvey and Legora. Nadella compared it to PC-based forecasting in the late 1980s and early cloud adoption: he expects "AI native" to be passé within five years, and noted that even the 500-year-old Lloyd's is now cloud native.
Implementation economy and industry structure
Nadella said the services layer for AI is a new hybrid that is neither classic management consulting, systems integration nor technology support. It begins with the business problem, the eval and the agentic architecture, and he expects traditional systems integrators to evolve into this model.
On whether the industry will consolidate into full-stack players, he said the evidence is going the other way. He once thought a single vertically integrated stack from silicon to app might emerge, but "every day you wake up and you say wow there's more competition in every layer of the stack." He said he hopes for "horizontal specialization with multiple winners in each layer" and the ability to compose across them, including routers that send tasks to the best model. He added a caveat that dominant designs can still change the picture.
Compute, power and the GDP question
On Mustafa Suleyman's suggestion that $100 billion training runs are coming, Nadella reframed the question as one of R&D budgets and macro growth. A training run is "just a percentage of R&D," which for R&D-heavy companies is perhaps 15% to 20% of revenue, so affordability depends on revenue, which depends on AI expanding the world economy. He acknowledged shortages of power and silicon, noted these are long-lead items, and said when supply and demand will balance is something "no one knows," adding that anyone speculating is just that. For him the bigger variable is world GDP growth, which has hovered around 2% for a long time.
He said he hopes and believes AI can lift that, with healthcare and software among the first areas. He cited Copilot deployment in the UK's NHS as an example where better outcomes and lower costs raise GDP by definition. He said demand for software development has proven highly elastic, meaning "IT backlogs were huge, and none of us even recognized it." Silicon design, with faster tape-outs and in-silico simulation, and engineering tasks such as maintenance scheduling at Rolls-Royce were other examples of areas where optimization is paying off.
On whether the gains in coding will transfer to finance, sales and marketing, Nadella was more cautious. Coding benefited from RL against clear reward signals on top of pretrained models, and "can we crack that in knowledge work seems much harder," though he said contact centers, supply chain and cash-to-order are showing it is possible. He suggested science and engineering may prove as elastic as coding.
The interview contained no financial guidance or new product dates, and several answers, particularly on power and traffic mix, were openly speculative. The substance lies in the strategic framing: a Microsoft CEO who now doubts that scale solves alignment, wants more transparency from closed-model partners, and is positioning MAI as a controllable, licensed-weight product.