If your production workloads run on Microsoft Foundry or your developers live in GitHub Copilot, the silicon answering your requests is about to change, and you get no say in it. Microsoft has not said it will tell you when it happens. It said on October 12 that GitHub Copilot and Foundry will follow Microsoft Copilot onto its in-house Maia 200 accelerator "by end of year," and that the next chip, Maia 300, starts production deployments in 2027 built around low-precision MXFP4 numerics. Customers cannot rent Maia or select it. What you can do is record what your workloads score today, so you can tell afterwards whether anything moved.
The announcement came in a blog post from Scott Guthrie, Microsoft's executive vice president for Cloud + AI. It was written for investors and capacity planners, and for a Foundry buyer it works as a change notice for a dependency you never chose.
What Microsoft Actually Announced
Microsoft announced a second-generation inference chip already in service and a third generation that arrives in 2027. Per the Microsoft blog, Maia 200 is "now in production" and deploying across the United States, Europe and Australia. It runs Microsoft Copilot, the built-in AI features in Teams and Cowork, and model development for Microsoft's MAI team, and it serves OpenAI models, MAI models and open-source model families. GitHub Copilot and Foundry are next.
Maia 300 is the bigger chip. The same post lists eight HBM4 stacks, 18 TB/s of peak memory bandwidth, a 950-watt TDP and compute chiplets on TSMC's N2P node. A rack holds 96 accelerators inside a 167-kilowatt envelope. Microsoft's architecture deep dive adds 288 GB of HBM per accelerator and claims, against Maia 200, "up to 7 times greater rack-scale token throughput" and "more than 2.5 times the performance per watt." The planned Maia 300 workloads include multi-step agentic coding for GitHub, "cloud-scale model serving for Foundry" and latency-sensitive work for Dragon Copilot.
Scale is the other half of the news. TIKR's summary of Reuters' reporting puts the plan at potentially more than one million chips next year, which TIKR works out to roughly 10,400 racks and about 1.7 gigawatts. Earlier reporting summarised by FourWeekMBA described talks with TSMC for more than 300,000 units for 2027 delivery and called those numbers subject to supply and packaging negotiations, not committed orders.
Can You Rent or Choose Maia? No
Maia is internal infrastructure, and nothing in the announcement changes that. The National CIO Review reports that Maia remains "reserved for Microsoft's internal operations," so Azure customers cannot select it directly. Neither the Guthrie post nor the architecture paper mentions customer rental, pricing or a hardware selector.
That puts Microsoft behind its two largest rivals on access. Google says you can use TPUs through Compute Engine, Google Kubernetes Engine and its Gemini Enterprise Agent Platform. AWS lists Trn3 instances and Trainium3 UltraServers under EC2. On both clouds you can benchmark the custom chip yourself and decide whether its price-performance is real. On Azure, Microsoft makes that call for you and decides what happens to any savings.
The change is narrower than it first looks. Foundry's serverless path never let you pick hardware. Its deployment types documentation says a deployment type determines where your data is processed, how you pay and the performance characteristics. It says nothing about which accelerator serves the request, and Global deployments "dynamically route traffic to available datacenters." Even Foundry's managed compute preview, the path for open-weight models, says "You don't need to pick virtual machine SKUs, because Foundry chooses GPUs per instance," drawing from A100, H100 and MI300X families. Hardware abstraction is standard practice in hosted inference. What is new is that the abstraction now hides a different chip architecture with different low-precision arithmetic.
Why the Numerics Matter More Than the Chip
The risk worth managing is a change in numerical precision, which can shift model outputs in ways a status page never reports. MXFP4 is a 4-bit microscaling floating-point format: weights or activations are stored in small blocks that share one scale factor, which cuts memory and bandwidth. The Guthrie post says Maia 300's MXFP4 formats "reduce memory and bandwidth needs for compute and KV-cache," and the architecture paper goes further, describing "enhanced MXFP4 and sub-3-bit storage formats." Those formats are where both the efficiency gain and any accuracy loss would come from.
The research is not settled in Microsoft's favour or against it. An ICLR 2026 paper, Bridging the Gap Between Promise and Performance for Microscaling FP4 Quantization, finds that MXFP4's power-of-two scale factor induces large error that "severely degrades accuracy" under standard post-training quantization. The same authors show a fix, MR-GPTQ, that brings MXFP4 close to the accuracy of Nvidia's NVFP4 format. Read the two findings together: 4-bit serving can be close to lossless when the quantization is done carefully, and visibly worse when it is not. From outside Microsoft you cannot tell which kind you are getting. We compared these formats on open models in FP8 vs NVFP4 vs INT4-AWQ, and the accuracy ranking shifted from model to model.
There is a recent precedent for exactly this failure. In its postmortem of three issues from August and September 2025, Anthropic traced degraded Claude output partly to a TPU server misconfiguration and a latent bug in the XLA:TPU compiler. Anthropic serves Claude on Trainium, Nvidia GPUs and TPUs, and wrote that "Each hardware platform has different characteristics and requires specific optimizations." Anthropic also wrote: "The evaluations we ran simply didn't capture the degradation users were reporting." A frontier lab's own evals missed a hardware-specific regression, so a vendor telling you a migration is transparent is a promise you should check yourself.
Microsoft has a reasonable case. It already serves OpenAI models to Microsoft Copilot users on Maia 200, at a scale that would surface gross regressions quickly, and it has every commercial reason to keep quality flat while it moves GitHub Copilot, a paid product with vocal users. Most workloads will probably see no difference. You still have to verify yours, and once the switch happens there is no before-state left to measure.
Does 40% More Tokens Per Megawatt Mean Lower Prices?
Not on any timeline you should budget for. Microsoft claims Maia 300 delivers 40% more tokens per megawatt than "leading GPU-based systems" and that Maia 200 already delivers up to 1.7x more tokens per watt than "leading GPU infrastructure" for built-in AI features, per the Guthrie post. Neither claim names the GPU, the model or the workload, and the 40% figure does not appear in the architecture paper at all, which benchmarks Maia 300 only against Maia 200. TIKR's analysis notes that the number is Microsoft's own estimate and has not been tested outside the company.
The January claim gives you a test case. The National CIO Review reports Satya Nadella saying Maia 200 delivers 30% better performance per dollar than the latest hardware in Microsoft's fleet. Microsoft's January Maia 200 announcement said the chip would accelerate Foundry and support Microsoft 365 Copilot. Check your own Foundry invoices since then and see whether any of that 30% reached you.
The financials explain why it may not. TIKR reports that Microsoft's cash capex hit $35.80 billion in the quarter ended June, against $17.08 billion a year earlier, and gross margin slid from 69.05% to 67.20% across three straight quarterly declines. A company in that position uses cheaper inference to rebuild margin first. Evercore ISI analysts, cited by the National CIO Review, estimate Maia could reach roughly 10% of Microsoft's AI chip spend by the first quarter of 2027. A 10% share of chip spend is unlikely to reset a price list. Nvidia and AMD chips remain in the fleet, mainly for training.
What to Do Before the Switch
Microsoft Ignite runs November 17-20, and that is when Microsoft will add detail. Ask your questions before then.
This Week:
- Freeze an eval baseline for every production Foundry deployment. Take 200 to 500 real prompts per workload, record outputs, scores and token counts against today's model version, and store them with the date. If you have no eval set, start with our guide to building one from hand-written questions. Tools like Langfuse or Braintrust can schedule the re-run.
- Do the same for GitHub Copilot. Pick 30 to 50 representative tasks from your own repos (a refactor, a test-generation job, a bug fix) and record acceptance and pass rates now. GitHub Copilot moves by year end, and when developers say it got worse you will want numbers to check that against.
- Schedule a weekly re-run. A baseline only helps if you compare against it. A weekly diff on the same prompts tells you when, give or take a week, something changed.
Before Ignite:
- Put three questions to your Microsoft account team in writing. Which of our deployments will be served on Maia, and when? What numerical precision will our model run at, and is it the same precision we get today? Will Microsoft notify customers when serving hardware or precision changes for a deployment? Written answers become evidence later.
- Ask whether Provisioned or Data Zone deployments get different treatment. Microsoft's own docs say it can add regions to a data zone without prior notice. Assume hardware changes work the same way unless you are told otherwise.
Before Your Next Renewal:
- Ask for a hardware and numerics change-notice clause. You probably will not get the right to choose the chip. You can reasonably ask for notice of a material change in serving precision and an eval window before it takes effect.
- Keep 2027 budgets flat on unit price. Model Maia savings as upside you have to negotiate, and price any Azure-versus-other-cloud comparison on today's published rates.
- If bit-exact reproducibility is a regulatory requirement, move that workload to infrastructure where you choose the hardware. On Azure that means self-managed GPU VMs, not the serverless path, and you should cost it before the renewal conversation.
The Bigger Picture
All three hyperscalers are moving inference onto their own silicon, and Microsoft is catching up to Google and Amazon, who already let you rent the chip and measure it yourself (we compared them in our look at Nvidia alternatives for inference). Microsoft treats Maia the way a SaaS vendor treats its database, as an internal choice the customer inherits. For a model that writes your code and answers your customers, that deserves more scrutiny than it gets, because a precision regression shows up as a slow drift in quality that looks like users getting pickier.
Handle it like a model deprecation that arrives without a notice window: pin model versions where Foundry lets you, and keep dated eval results for everything else.
Record the baseline this week, before GitHub Copilot and Foundry make the year-end move.
Continue Reading
- Microsoft Maia 200 Chips Cut AI Costs 30% Over Nvidia GPUs
- FP8 vs NVFP4 vs INT4-AWQ: AWQ Outran NVFP4 on 3 of 4 Models
- NVIDIA Alternatives for Inference: Only Trainium Pays Off
- Intel's Crescent Island Fits 300B Models in Old Racks, Slowly
- CISA Named 6 Distillers. Your Agent Fleet Fits the Spec.
- Braintrust vs Langfuse vs Promptfoo: Don't Pay for the Gate
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic