Key takeaways
- Microsoft shipped Maia 200 on 26 January 2026 — TSMC 3nm, over 140 billion transistors, 216GB of HBM3e at 7 TB/s, over 10 petaFLOPS FP4 in a 750W envelope.
- In the quarter that contained the launch, NVIDIA's GAAP gross margin rose to 75.0% from 73.0% a year earlier, and data centre revenue hit US$62.3bn.
- The quarter after the launch, NVIDIA's data centre revenue reached US$75.2bn, up 92% year on year, with GAAP gross margin at 74.9% and the next quarter guided to the same 74.9%.
- Microsoft's own property and equipment additions were US$30.9bn in the March 2026 quarter alone, against US$16.7bn a year earlier.
- Maia 200 is live in one datacentre region, near Des Moines, with Phoenix described as 'coming next'.
What Microsoft actually shipped, and when
On 26 January 2026 Microsoft announced Maia 200, its second-generation AI accelerator. The published specification is unusually complete for first-party silicon: TSMC's 3nm process, over 140 billion transistors, 216GB of HBM3e delivering 7 TB/s of bandwidth, 272MB of on-die SRAM, over 10 petaFLOPS of FP4 and over 5 petaFLOPS of FP8, inside a 750W SoC envelope, with an integrated on-die NIC carrying 2.8 TB/s of bidirectional bandwidth. Microsoft claims 30% better performance per dollar than the latest generation hardware already in its fleet, three times the FP4 throughput of third-generation Trainium, and FP8 above Google's seventh-generation TPU.
The question for anyone holding NVDA or MSFT is narrower: does a hyperscaler building its own inference silicon compress the pricing power of the company it is buying from?
Two quarters of evidence say not yet
NVIDIA's fiscal Q4 2026 ended on 25 January 2026, one day before the Maia 200 announcement. In that quarter NVIDIA reported record revenue of US$68.1bn, record data centre revenue of US$62.3bn (up 75% year on year), and a GAAP gross margin of 75.0% — against 73.0% in the same quarter a year earlier and 73.4% the quarter before. Full-year FY26 revenue was US$215.9bn, up 65%.
The following quarter, ended 26 April 2026, is the first full quarter in which Maia 200 was deployed. NVIDIA's data centre revenue was US$75.2bn, up 92% year on year. GAAP gross margin was 74.9%. The Q2 FY27 outlook guides revenue to US$91.0bn and GAAP gross margin to 74.9% again — explicitly excluding any data centre compute revenue from China.
The chart tracks that margin across five prints, from Q4 FY25 through the Q2 FY27 guide. If custom silicon were eroding NVIDIA's ability to price, the first symptom would appear here, and it has not. The margin rose through the launch and has held within 10 basis points of 75% since.
Microsoft's capex is the tell
The strongest evidence that Maia 200 is not yet a pricing threat comes from Microsoft's own cash flow statement. In the quarter ended 31 March 2026 — the quarter in which Maia 200 was running in production — Microsoft's additions to property and equipment were US$30.9bn, against US$16.7bn in the same quarter a year earlier. Across the first nine months of FY26 the figure was US$80.1bn, against US$47.5bn. Azure grew 40%, Microsoft Cloud revenue was US$54.5bn, commercial remaining performance obligation rose 99% to US$627bn, and Microsoft's AI business passed a US$37bn annual run rate.
A company that had solved its accelerator problem in-house would be moderating that line. Microsoft is doing the opposite, because the binding constraint on inference capacity is not chip design — it is how much silicon can be installed, powered and cooled per quarter, and on that constraint first-party and merchant silicon compete for the same substation.
The counter-case, stated properly
Three facts cut against the vertical-integration thesis, and they come from Microsoft's own material rather than from anyone's scepticism. Maia 200 is deployed in one region — US Central, near Des Moines, Iowa — with US West 3 near Phoenix described as "coming next". Microsoft's announcement describes the workloads it will carry, including OpenAI's GPT-5.2 models and Microsoft Foundry, in the future tense: Microsoft says Maia 200 "will serve multiple models, including the latest GPT-5.2 models from OpenAI". And the predecessor, Maia 100 — detailed by Microsoft in April 2024 on TSMC's 5nm node — is not named anywhere in the Maia 200 material as having powered a production AI service. Twenty-one months separated that Maia 100 post from Maia 200's first deployment.
Any of that could change quickly. But a one-region, forward-tense deployment is not a substitute for the merchant supply chain; it is a bet placed alongside one.
The figure that would settle it
NVIDIA's Q2 FY27 guide of US$91.0bn at a 74.9% GAAP gross margin implies roughly US$68.2bn of gross profit in a single quarter — more than twice Microsoft's US$30.9bn of property and equipment additions in the March quarter. Each point of gross margin against that revenue base is about US$910m a quarter. Pricing power here is measured in single margin points, and a single point is a rounding error against Microsoft's capex but a US$3.6bn annual swing to NVIDIA.
What would falsify this
The argument here is that first-party silicon has not yet reached NVIDIA's price. It fails on two specific observations, neither of which has occurred: NVIDIA printing a GAAP gross margin below 73% for two consecutive quarters while data centre revenue growth decelerates, or Microsoft offering a generally available Maia 200 SKU that Azure customers can rent — which would convert Maia from an internal cost centre into a merchant competitor. Trade-press accounts of a schedule slip in the Maia programme could not be confirmed against any Microsoft primary source and are not relied on here. Nor can these filings tell you the transfer price Microsoft pays NVIDIA; margin is a blended number across a customer base, and a concession to one hyperscaler could hide inside 75% for several quarters.
Watch the second region. Phoenix going live, with a named production workload attached and a rentable SKU, is the point at which Maia stops being a hedge and starts being a substitute — and it will appear in Microsoft's capex line before it ever appears in NVIDIA's. The discount rate applied to all of this has been moving too, for reasons that have nothing to do with either company, and it is doing more to the multiple than any chip is doing to the margin. Anyone sizing this exposure alongside other high-beta sleeves is underwriting the same factor twice.
This content is for informational purposes only and does not constitute financial advice. Always do your own research or consult a qualified financial adviser before making investment decisions.