By Rebecca Stone

Microsoft has committed to deploying multiple gigawatts of next-generation Azure Maia AI accelerator racks across its global data-center footprint, The Register reported this week — a power envelope that, if fully realized, would rank among the largest single-vendor AI infrastructure pledges ever disclosed publicly.

The figure, drawn from a Register piece on Microsoft's planning for the Maia successor family, dwarfs the capacity of most national AI programs. One gigawatt roughly matches the continuous output of a modern commercial nuclear reactor. A multi-gigawatt plan, by definition, means Microsoft is preparing to plug rack clusters the size of mid-sized industrial facilities directly into the grid.

What Microsoft has already said about Maia

The Maia program first surfaced in November 2023, when Microsoft unveiled the Maia 100, a 5-nanometer accelerator designed in-house and manufactured by TSMC. The chip was Microsoft's first homegrown AI silicon aimed at training and inferencing large language models for Azure customers, including OpenAI workloads. At Ignite 2024, Microsoft confirmed a second-generation Maia was in development, though it withheld the process node, HBM configuration and tape-out window.

The Register's latest report does not specify which generation the multi-gigawatt plan targets, and Microsoft's own public materials continue to use the phrase "next-generation" without committing to a launch date.

How big is multi-gigawatt in practice?

A 50-megawatt AI campus is now considered a flagship deployment. NVIDIA's disclosed contracts with Oracle, xAI and others routinely call for 1-gigawatt-plus clusters. A multi-gigawatt plan, however, implies several such sites running in parallel — or one campus scaled beyond anything currently announced.

That scale carries grid consequences. Dominion Energy, NRG Energy and other U.S. utilities have already filed interconnection paperwork for Microsoft campuses exceeding 1 gigawatt in Virginia, Arizona and Wisconsin. The Register's report suggests Microsoft is now planning well past that ceiling.

What the Register did not disclose

The Register's headline, as captured, gives no timeline, no dollar figure and no manufacturing-partner disclosure. Crucial details remain open:

• Which node? The Maia 100 used TSMC's N5. A next-gen chip could plausibly land on N3, N3P or a customized TSMC process — but the report does not say.
• Which packaging? CoWoS-L, the advanced 2.5D scheme used by NVIDIA's Blackwell, is supply-constrained industry-wide. Multi-gigawatt commitments depend on TSMC allocating CoWoS lines.
• Which HBM? HBM3e dominates 2024–2025 AI accelerators; HBM4 sampling begins later in 2025. Microsoft's HBM mix for the next Maia is undisclosed.
• Power and cooling? Direct liquid cooling, immersion and 800-volt DC distribution are all candidates. Multi-gigawatt campuses almost certainly require on-site generation or dedicated substation upgrades.

Where this fits competitively

Google's TPU v6 "Trillium" and follow-on silicon are scaling aggressively inside Google Cloud. Amazon's Trainium 2 and forthcoming Trainium 3 anchor a separate AWS buildout. Meta's MTIA chips now power internal recommendation and ranking workloads. Microsoft's Maia bet, by contrast, has leaned heavily on NVIDIA GPUs to date — making a multi-gigawatt custom-silicon push a notable strategic pivot.

Analysts have pegged Microsoft's 2025 AI capex north of $80 billion, the bulk of which still flows to NVIDIA systems. If the multi-gigawatt Maia plan materializes, it would mark the first time Microsoft's custom-silicon spend is large enough to materially shift the supplier mix.

What to watch next

The first concrete proof point will be a tape-out disclosure — most likely on a TSMC earnings call or a Microsoft Ignite keynote. Until then, "multi-gigawatt" remains a procurement intent, not a shipped product. The Register's reporting puts a number on Microsoft's ambition; the silicon, the schedule and the supply chain still need to follow.

Microsoft did not respond to a request for comment as of publication.