Microsoft plans to scale its own Maia AI chips to gigawatt levels by 2027, according to a report by The Decoder. The target marks the most aggressive public indication yet of how far the company intends to push its in-house silicon program beyond the first generation of accelerators already installed in its datacenter fleet.

The plan sets a hard number on a trajectory Microsoft has hinted at for some time. Gigawatt-scale capacity — measured in total power draw of the compute infrastructure running Maia chips — would place the company's proprietary accelerators among the largest single-vendor AI compute footprints in the industry. A one-gigawatt cluster can power on the order of a million GPU-class chips, depending on the accelerator's thermal design power and the surrounding networking and cooling overhead.

Microsoft introduced the Maia 100 accelerator in November 2023, its first custom AI chip designed for training and inference workloads in Azure datacenters. The company positioned the silicon as a complement to, rather than a replacement for, its supply of Nvidia GPUs — but the 2027 gigawatt target changes the framing. The scale implies that Maia silicon would carry a substantial share of Azure's AI compute demand, not a marginal slice.

What does the 2027 target change?

The commitment affects three parties directly:

• Microsoft's capex planning. In-house chips cost less per unit of compute than merchant silicon at scale, and the company controls the design cadence. A gigawatt-class Maia deployment would reduce exposure to Nvidia's pricing and allocation cycles.
• Nvidia's largest customer base. Microsoft remains one of Nvidia's biggest buyers, but every gigawatt shifted to Maia silicon is revenue that does not flow to the GPU supplier.
• TSMC and the supply chain. Maia chips are fabricated externally, so the ramp still competes for the same advanced packaging and leading-edge node capacity that constrains the entire AI hardware market.

The 2027 date matters because it aligns with the window in which analysts expect AI training clusters to move from hundreds of megawatts to multi-gigawatt campuses. Microsoft, like Google, Amazon and Meta, is designing infrastructure around power availability rather than chip supply — and owning the accelerator design gives the company tighter control over performance per watt, the metric that ultimately limits how much compute a given power envelope can host.

How does Maia fit into Azure's hardware stack?

Microsoft has framed its silicon strategy as portfolio management. Azure currently runs AI workloads on:

• Nvidia GPUs across multiple generations
• AMD Instinct accelerators, added in 2024
• Maia 100 for selected internal and customer workloads

Scaling Maia to gigawatt levels by 2027 would move the custom silicon from third position to a primary pillar of that stack. The Decoder's report does not specify which Maia generation would carry the ramp, nor the split between training and inference capacity.

The announcement also fits a broader industry pattern. Google has run its TPU program at scale for a decade, Amazon has iterated Trainium and Inferentia across multiple generations, and Meta has begun deploying its own MTIA inference chips. Microsoft's 2027 target brings its roadmap to comparable ambition, on a compressed timeline — the company went from first custom AI chip announcement to gigawatt-class deployment plan in roughly four years.

For Azure customers, the practical question is pricing and availability. In-house silicon typically reaches customers through internal services first, where the economics of vertical integration show up as lower cost per token for inference. Microsoft has not disclosed commercial terms for Maia-based capacity.

The 2027 deadline now serves as a measurable benchmark: either Maia silicon reaches gigawatt-class deployment on schedule, or the gap between Microsoft's custom-silicon ambition and its execution becomes visible to the market.