hunter-gplpcnThe two companies that control most of the plumbing for AI are now building the standard for AI agents together. At Microsoft’s Build developer conference this week, Nvidia and Microsoft unveiled a combined technology stack that spans Windows devices, cloud data centers and on-premises servers, designed specifically for deploying the autonomous AI programs known as agents.
The announcement amounts to a division of labor. Nvidia supplies the acceleration layer, the GPUs and the software that runs model inference at scale, while Microsoft supplies the operating system, the cloud and the developer tools where agents will live. Together, the two companies are betting they can define how the next wave of AI applications is built, deployed and managed, before anyone else sets the terms.
Agents are the industry’s next act. Unlike chatbots, which answer questions, agents are programs that take actions: booking travel, managing email, writing and shipping code, reconciling accounts. Every major AI company is racing to build them, and every technology vendor is racing to host them. The infrastructure question, whether agents run best on a laptop, in a cloud or inside a corporate data center, is exactly what the Nvidia-Microsoft stack is designed to answer.
The design covers all three surfaces. On Windows devices, the stack lets developers build agents that use local models for quick tasks and hand heavier work to the cloud when needed. In Azure, it provides managed services for agents that need to run continuously. And for companies that will not put sensitive data in a public cloud, it runs the same code on-premises. The pitch to developers is simple: write once, run anywhere Nvidia and Microsoft support.
The competitive implications are immediate. If the two companies succeed in establishing their stack as the default, third-party frameworks for building agents will find themselves squeezed between the platforms. Developers tend to use whatever infrastructure their cloud provider makes easiest, and Microsoft’s share of enterprise software gives the partnership enormous reach. Startups that built agent frameworks on top of the old stack will have to explain why their tools are worth the integration cost.
The alliance also binds the two companies more tightly to each other. Nvidia has built its data-center empire selling chips to every cloud provider, including Amazon and Google, and it continues to court them. But a closer partnership with Microsoft gives Nvidia a software ally with the enterprise relationships to push its hardware into thousands of companies that would not otherwise buy from a chip vendor. Microsoft, for its part, gains access to the acceleration technology its cloud rivals already offer.
There is a history here. Microsoft and Nvidia have collaborated for years, from gaming GPUs to the supercomputers that trained some of the largest AI models. The new stack formalizes that relationship at a moment when both companies are repositioning: Microsoft is building its own AI models after its split with OpenAI, and Nvidia is expanding from hardware into software and services. Each needs the other’s distribution.
Developers at the conference gave the announcement a mixed reception, according to people who attended. Some welcomed a standard that could simplify the fragmented agent tooling; others worried that a two-company standard is still a duopoly, and that building on it means accepting the roadmaps and pricing of two of the largest vendors in the industry. Neither company has said what the stack will cost, and pricing will shape how fast it spreads.
The announcement includes concrete developer tools, not just promises. The two companies demonstrated a reference architecture for building agents, with templates for common tasks, a runtime that moves between device and cloud automatically, and telemetry that shows what agents are doing at every step. The details matter because developers choose platforms on specifics, and the demonstration was designed to answer the first question any engineering team asks: what exactly do I build on this, and how hard is it?
The partnership also positions the two companies against the model providers. OpenAI, Anthropic and Google each offer their own ways to build agents, but they control the model layer rather than the operating system or the accelerator. By defining the runtime standard, Nvidia and Microsoft are making the case that agents should be built on neutral models running on their infrastructure, a pitch that gives developers flexibility while keeping the economics in the two companies’ favor.
The question of openness is unresolved. Developers who attended the sessions said they were impressed by the tooling but wary of the lock-in, and neither company has detailed how the stack will interoperate with rival clouds or non-Nvidia accelerators. The history of computing suggests the standard that wins is the one developers adopt because it works, not the one imposed by the largest vendors. For now, Nvidia and Microsoft have the largest installed base to push from, and the strongest incentive to make the standard genuinely good.
The deeper question is whether agents, as a category, live up to the investment. The infrastructure being built now assumes that agents will generate huge volumes of compute demand, which is what justifies the GPUs, the cloud services and the developer tools. If agents become a niche, the stack will be overbuilt; if they become the default way software is used, the partnership will have picked the winning side. For now, the two companies are placing the largest bets in the industry on the answer being yes.
