Microsoft urges 'useful yield' metric at SEMICON Taiwan

Microsoft used a keynote at SEMICON Taiwan to call for a new industry measure for AI infrastructure: “useful yield.” The company framed useful yield as a way to judge how much real intelligence a stack delivers, not just how many chips or racks it installs.

Microsoft’s Azure Hardware leadership said useful yield should combine performance, energy and cost into a single, practical measure — for example tokens or useful responses per dollar and per watt, plus throughput and latency metrics. The argument appears in blog and press material accompanying the SEMICON Taiwan appearance.

Rani Borkar, president of Azure Hardware Systems and Infrastructure, anchored the pitch around cross-layer engineering: designing silicon, memory, networking, power and software together to raise the fraction of compute that produces useful outputs. Microsoft described this as the next efficiency frontier after raw capacity expansion.

Microsoft cited concrete platform work to illustrate the idea, pointing to the Azure Maia accelerator and system-level choices such as a two-tier scale-up network with an integrated NIC and finer-grained power controls on custom CPUs to squeeze more useful work from each megawatt. Those examples were used as evidence that co-design across layers can increase useful yield.

The push for useful yield arrives as hyperscalers pour capital into AI infrastructure and customers press cloud providers for predictable cost-per-inference and affordability. Industry trackers note that AI capex and procurement choices are now central to cloud economics and semiconductor demand.

Microsoft framed useful yield not just as an engineering metric but as a procurement lever that could reshape vendor selection, service-level agreements and RFPs. The company argued procurement should reward systems that raise usable intelligence per dollar, rather than only raw FLOPS or peak throughput.

That framing is designed to recast procurement debates inside hyperscalers and among enterprise cloud customers, where choices about accelerators, memory topology and power delivery are often made in isolation. Microsoft’s keynote argued those siloed decisions can hide system-level waste and distort vendor incentives.

SEMICON Taiwan provided the backdrop because the conference gathers chipmakers, equipment suppliers and datacenter operators who would need to adopt cross-layer approaches to make useful yield meaningful in practice. The event ran Sept. 2–4, 2026 and hosted forums on hardware, materials and systems that Microsoft said must collaborate to raise yield.

Implementing useful yield will be technical and political. Microsoft’s blog highlights how memory is increasingly the limiting resource for many inference workloads and how system optimizations across compilers, placement and network topology change the amount of useful work produced from a given memory footprint. Those are engineering levers toward higher yield.

Beyond engineering, useful yield poses measurement challenges: different models and workloads produce different ‘useful’ outputs, and customers may disagree about which outputs matter most. Microsoft acknowledged these trade-offs while urging industry standards and shared benchmarks to make useful yield comparable.

Not everyone will welcome a new procurement metric. Vendors that sell components optimized for isolated metrics such as peak throughput or densest silicon might resist standards that favor system-level co-design and longer integration cycles. Still, Microsoft and other cloud operators argue the long-term prize is lower operating cost and broader access to intelligence.

If useful yield catches on, it could change contract language, datacenter designs and investment priorities — favoring chips and systems that deliver more usable responses per dollar and per watt. Microsoft presented the idea at SEMICON Taiwan as a practical reframe: move from building more capacity to building more usable intelligence.