Microsoft deployed its Maia 200 AI chip in a data center this week, TechCrunch wrote. Cloud AI demand is outrunning hardware availability, forcing rollout pacing and capacity allocation constraints.
Microsoft chip self-supply push is paired with continued third-party buying because cloud demand is still outpacing capacity
Before Maia 200, cloud providers faced a supply crunch for advanced AI hardware and competed for Nvidia capacity. Hyperscalers could pilot custom AI silicon while still budgeting heavily for Nvidia GPUs. Hyperscalers pursued in-house chips due to cost and difficulty obtaining Nvidia hardware. Microsoft is working on its own frontier models to potentially reduce reliance on external model providers.
Microsoft deployed Maia 200 and plans additional rollouts in coming months. Satya Nadella stated Microsoft will continue buying AI chips from Nvidia and AMD. A reported $750 million Perplexity agreement would use Azure, attributed to Bloomberg via Reuters. Maia 200 is positioned as an inference optimized chip and expected to be used by Microsoft Superintelligence team first. Nadella framed vendor coexistence inside an TechCrunch exchange, "We have a great partnership with Nvidia, with AMD. They are innovating. We are innovating.", tying deployment to ongoing procurement.
The Maia 200 launch extends in-house silicon design while Nvidia and AMD supply remains in the stack. Business Insider’s earnings-call excerpt recorded Amy Hood allocating GPUs and CPUs, while Azure demand kept growing, Business Insider captured. Meta piloted an in-house training chip in March 2025 while forecasting up to $65B of 2025 capex, much of it for Nvidia GPUs. Microsoft’s inference positioning for Maia 200 still sits beside stated third-party buying for broader capacity coverage. The company also said Maia 200 will support OpenAI models running on Azure, widening workload compatibility pressure.
Custom silicon pilots can coexist with large external GPU budgets because internal chips do not cover all workloads. Hood’s allocation framing forces a queueing problem, while Maia 200 targets inference rather than universal acceleration. Meta’s in-house effort coexisted with continued large third-party purchases rather than replacing them near term. Microsoft’s plan to roll out more Maia 200 chips in coming months adds supply, while procurement stays active. The strategic logic inside a heterogenous fleet reframes chips as throughput multipliers, not vendor exits.
How many Maia 200 chips are deployed remains unclear. Unit economics and total cost of ownership stay undisclosed. Independent benchmarks constrain any claim that Maia 200 beats Nvidia or AMD. Report attribution limits what can be confirmed about the Perplexity agreement beyond its size. The short-term frame stays 6 to 18 months as Azure remains capacity constrained. Whether Maia 200 shifts a material share of workloads, plus Perplexity contract terms and in-house frontier model timing, will set the next constraint.
