Microsoft Maia 300 Chip Targets Nvidia With a 2027 Inference Shift

What Microsoft announced

Microsoft is no longer presenting its own AI silicon as a side project. At the OCP Global Summit in San Jose on October 12, the company unveiled the Azure Maia 300. It said the Maia family will grow to multigigawatt deployments starting next year and will become the "primary AI inference system for Microsoft products and services."21 Satya Nadella called it the company's "biggest bet yet on our own silicon."

The main claim is about power efficiency, not raw speed. Microsoft says Maia 300 will produce 40% more tokens per megawatt than "leading GPU-based systems."22 It also says a full rack of 96 chips will deliver up to seven times the token throughput of a Maia 200 rack.21 That rack figure is partly a packaging story, because each Maia 300 rack holds twice as many chips as a Maia 200 rack.

The chip is built around memory, which reflects Microsoft's view of where inference gets stuck. Each Maia 300 has eight HBM4 stacks with a peak bandwidth of 18 TB per second within a 950-watt power limit. A full rack carries 27.6 TB of HBM and 1.7 PB per second of total memory bandwidth. It supports the low-precision MXFP4 format to reduce the memory traffic needed for computation and for the KV cache. It also uses a software-defined dataflow design meant to cut down on moving data around. Microsoft's own explanation is that moving data, rather than raw compute, has become the main limit on large inference systems.

Why the "customer-turned-competitor" framing stuck

Axios summed up the moment in one line: every major Nvidia customer now also competes with Nvidia, in search of lower costs, better performance and more control.11 Its roundup of the field makes the point. Google has years of TPU development, and its latest generation comes in separate training and inference versions. OpenAI is building its Jalapeño chip with Broadcom. Amazon's Trainium line is used by Anthropic and is also part of OpenAI's infrastructure plans. Anthropic is exploring its own chip designs, and Meta runs an internal chip program.12

The rest of the roster has filled in during 2026. Google announced TPU 8t for training and TPU 8i for inference in April.7 In March, Meta revealed four MTIA generations (300, 400, 450 and 500) on a cadence of roughly one every six months, with Broadcom co-developing them under a deal that runs to 2029.10 Amazon said in July that its custom-chip business had passed a $25 billion annual revenue run rate.7 One report says most inference on Amazon Bedrock already runs on Trainium rather than Nvidia hardware.4

The framing is accurate, but it can mislead. These companies are not walking away from Nvidia. They are moving the workloads they control, and know well, onto hardware they design themselves.

The fine print: inference, internal use, and a long ramp

The coverage agrees closely on the limits of the announcement. Microsoft's plan to make Maia its primary inference platform covers its own services. It does not necessarily extend to third-party Azure customers, OpenAI included.23 Rani Borkar, who heads Azure hardware systems, said that "it's not an either-or strategy." She also warned that reaching multigigawatt scale will take until after 2027.23 Scott Guthrie has confirmed that Microsoft's cloud will keep deploying Nvidia, AMD and Intel products.24

The scope also matters for buyers. Maia 200 cannot be rented, and there is no Maia instance type that outside developers can move to.1 Microsoft has not said whether Maia 300 capacity will ever be sold to third parties the way Nvidia GPU instances are.24 For now, Maia is a cost tool inside Microsoft, not a product Microsoft sells.

The outlets disagree on some details. Most report the efficiency claim per megawatt, while at least one quotes Nadella as saying "tokens per watt." Neither version comes with a named comparison system. Microsoft did not say which GPU platform it measured against. One analysis says the 7x rack figure should be read as a target rather than a published benchmark, since the measurement method has not been disclosed. Volume estimates also vary. The Information reported in August that Microsoft was in talks with TSMC for more than 300,000 Maia 300 chips for 2027. A figure of over a million chips has also circulated, but one review traces it to secondary reporting rather than anything Microsoft said. The 300,000 figure is the better-sourced number. The million-chip figure is a goal, not a commitment.

The economics: why inference is the battleground

The companies building their own chips have picked their target deliberately. Like Microsoft, most of them are going after inference, where margins are highest, while Nvidia still dominates training.12 One estimate puts inference at about two-thirds of AI compute spending.1 Electricity is the largest operating cost of an AI data center, so producing more tokens from the same power directly lowers the cost of each response.

Microsoft has some track record on this. Maia 200 launched in January. It already powers Microsoft 365 Copilot and serves OpenAI models, Microsoft's MAI models and open-source models. The company says it delivers 1.7 times more tokens per watt than other GPU-based systems on the AI features in Teams, Copilot and Cowork.21 Earlier, Microsoft claimed Maia 200 offered more than 30% better performance per dollar than the hardware it previously ran.19 GitHub Copilot and Microsoft Foundry are expected to move onto Maia 200 by the end of 2026.

The best argument for why this matters to Nvidia does not depend on Microsoft dropping it. Microsoft only needs to shift enough work onto Maia to reduce how many outside accelerators it needs for each dollar of AI revenue.19 Nvidia faces that kind of gradual erosion across every company on the list above.

Nvidia's position: still growing, but with less leverage

Nvidia's numbers do not show a company in trouble. Its revenue more than doubled to $96 billion in the second quarter, which suggests total demand is big enough for both Nvidia and its customers' chips.12 Supply-chain forecasts point the same way. One set of estimates has Google's TPU shipments rising from 4.1 million units in 2026 to 9 million in 2027, Trainium going from 2.5 million to 3.2 million, and OpenAI's chip from 150,000 to between 600,000 and 700,000. Over the same period, Nvidia's 2027 shipment forecast was raised, not cut.6 Meta, often cited as a defector, signed in February for millions of Nvidia Blackwell and Rubin GPUs, plus a 6-gigawatt AMD deal.10

Customers could still put pressure on Nvidia's margins even while it sells more chips. One analysis argues the damage may show up less in units sold and more in pricing, as large buyers use their own chip roadmaps as leverage in contract talks.20 TrendForce data cited in the same coverage projects 44.6% growth in custom ASIC shipments in 2026, compared with 16.1% for commercial GPUs.20 Axios adds a broader point. Nvidia has started using its balance sheet to help finance the AI buildout, so even a small loss of standing could weaken its central role in that growth.12

Nvidia does keep one durable advantage: software. Models tuned for CUDA generally do not run unmodified on TPUs, Trainium, Maia or Jalapeño. None of the rival software stacks has come close to CUDA's adoption by outside developers.9 That advantage is weaker for Microsoft than for a typical enterprise, because Microsoft controls the full stack for Copilot and can write its own kernels. It is still a strong advantage for everyone else.

The constraint nobody escapes

Custom chips change who Microsoft depends on more than how much it depends on suppliers. Companies designing their own chips still draw on the same limited manufacturing capacity, advanced packaging and high-bandwidth memory that Nvidia uses.12 Key Maia 300 components will be made by TSMC.21 Nvidia GPUs, Google TPUs and Amazon Trainium are all made there too.6 HBM4, which Maia 300 relies on, is already rising in price as AMD and Nvidia compete for limited supply.14

Broadcom may be the clearest winner. It co-designs Google's TPUs and OpenAI's Jalapeño, and it works with Meta and Anthropic on custom silicon.3 Its custom AI chip business reportedly brought in $8.4 billion in one quarter, with a backlog of about $73 billion.20

The takeaway

Maia 300 is significant, but it does not threaten Nvidia's existence. Microsoft's commitment is a real change in ambition, from treating Maia as extra capacity to making it the default system for its own inference.24 In practice, the change covers mostly inference, applies to Microsoft's own products, depends on efficiency figures nobody has independently checked, and arrives gradually from 2027. That still matters. The biggest buyers of AI compute no longer depend on a single supplier, and in a market this large, that can affect Nvidia's pricing power long before it affects its revenue. Maia 300's performance in Copilot traffic next year, not its launch, will show how much.