SemiAnalysis argues Microsoft has a path to “out-AI” rivals in the next phase of AI demand, where inference could be a $100B-per-gigawatt market

AI competition is increasingly shifting from building large models to operating them. In a viewpoint circulated by Yahoo Finance, SemiAnalysis frames the next battleground as “inference,” the process of using trained AI systems to generate responses for users. The argument is that inference demand, as enterprises deploy assistants and automation tools, could become as strategically important as the original race to train frontier models.

The semi-analyst case rests on sizing the opportunity for compute-intensive inference. The headline claim points to an inference market that could reach roughly $100 billion per gigawatt of power. While that figure is presented as an analyst estimate rather than a disclosed company metric, it indicates how large the expected spending could be once AI systems move from demos to sustained usage.

SemiAnalysis’s thesis also implies a different kind of competitive advantage. Training is expensive and episodic, but inference can be continuous, with models serving requests around the clock across many customers. That makes procurement, hardware availability, and the ability to deliver stable capacity through data centers central to who captures revenue as adoption accelerates.

Microsoft’s position in this framing is tied to its role as a major supplier of cloud computing capacity for enterprise workloads, including AI. The market question, as posed in the Yahoo Finance post, is whether Microsoft can “out-AI” OpenAI and Anthropic in terms of where the money is made, namely the downstream economics of running and scaling models in production. The premise is not that Microsoft necessarily trains every model better than specialized labs, but that it can package and sell the compute and platform layers that make inference economical and dependable for customers.

For Microsoft, this kind of inference-heavy environment matters because it aligns revenue with utilization. If customers pay for cloud capacity and related services as they deploy AI features broadly, ongoing inference traffic can translate into durable demand for data center resources. The economic link is straightforward: the more prompts, calls, agents, and applications businesses deploy, the more compute cycles are required to respond, which can keep underlying infrastructure in use.

Still, the post provides limited detail on exactly how Microsoft would capture that value, and it does not lay out specific capacity commitments, pricing changes, or contracted power arrangements in the way investors often scrutinize. Inference economics also depend on model efficiency, caching, batching, and how vendors set usage-based billing. Those operational variables are not quantified in the Yahoo Finance headline framing, leaving room for debate over how quickly the $100B-per-gigawatt opportunity could be realized in practice.

Looking ahead, the most relevant question for the market is not only whether inference demand grows, but how quickly power and GPU supply translate into usable capacity for customers. For Microsoft, the next indicates likely include disclosures around data center buildout, cloud AI service scaling, and customer adoption metrics that reflect sustained inference usage. For the broader sector, it will also matter whether pricing competition compresses margins or whether efficiency improvements shift spend from raw compute to software and tooling.