Microsoft is likely to be the biggest consumer of the new Maia 200 AI accelerator, which could have a domino effect on cloud AI infrastructure, according to industry analysts.
Maia 200, launched this week, is part of a new focus in the technology industry on AI inference, which is part of a generative AI workflow that applies a large-scale trained language model to a set of data to generate an output. Nvidia last month launched six new chips, including Vera CPUs and Rubin GPUs designed for AI inference, and rack-scale Vera Rubin hardware and software packages supported by enterprise IT vendors such as Red Hat.
Microsoft’s Maia 200 rivals AWS’ Trainium and Google’s TPU as an AI accelerator. This is a chip built specifically for specific AI processing tasks, rather than a GPU running general AI models. Both the Maia 200 and Nvidia’s chips are built with AI inference in mind, but they take different approaches, said Mike Leone, an analyst at Omdia, a division of Informa TechTarget.
“Vera Rubin is built for types of inference that require more complex inference, such as when a single query triggers a large inference chain to answer a multi-step problem,” said Leone. “Maia is focused on processing millions of queries with Copilot and other more standard chatbots. The goal is not necessarily deep inference, but processing at scale with the lowest possible margins.”
Potential ripple effects of Maia 200
mike leone
The Microsoft Superintelligence team will use the Maia 200 internally, according to a company blog post this week. Microsoft also has a software development kit (SDK) in preview for AI engineers who want to use it, but Leone said internal workloads will likely be the chip’s primary consumers early on.
But similar to Nvidia’s Vera Rubin, a multimillion-dollar system that most mainstream IT buyers don’t have access to, the arrival of the Maia 200 could have indirect benefits for Microsoft cloud customers, Leone said.
“If they move large internal workloads like Copilot to Maia, they will effectively stop competing with their own customers for access to Nvidia GPUs,” he said.
But switching from Nvidia to Maia is not an easy task, warns Forrester Research analyst Naveen Chhabra.
“You can think of Nvidia’s CUDA library and Microsoft’s Maia SDK as two railroad tracks that aren’t necessarily compatible. If you need to replace one freight coach with the other, you need to make sure the trolleys, aka apps, are compatible,” he said.
Microsoft has amassed a large stock of Nvidia AI chips and systems used to build and run its own AI applications as well as sell GPUs to Azure customers, Chhabra said. He added that it’s hard to say whether the introduction of Maia will free up enough Nvidia GPU inventory to sell more to customers, as the hyperscaler doesn’t disclose specific usage numbers.
Another analyst said he thinks there will be some appetite among Azure customers to run AI inference workloads on Maia because it is likely to be much cheaper than Nvidia GPUs. It’s a similar value proposition to other hyperscalar AI accelerators, HyperFrame Research CEO Steven Dickens said in an interview with Informa TechTarget this week.
In the long term, I think IT leaders will have to choose between portability and price.
mike leoneOmdia Analyst
“Mya is [Google] TPU, and that makes perfect sense. “Azure’s cheaper options also make sense for inference workloads,” he said.
Leone acknowledged the potential pain of moving from Nvidia to Maia, but predicted that some IT buyers would be willing to make the trade-off.
“Long term, I think IT leaders will have to choose between portability and price,” Leone says. “You can commit to Maia for significantly better Azure economics, or you can stick with Nvidia for cloud-wide flexibility. Of course, that comes with the tradeoff of being more locked into the ecosystem.”
Bit by bit of Nvidia’s “strangle”?
stephen dickens
Concerns about Nvidia’s dominant position in AI chips have grown over the past year among some industry experts. Financial analysts estimate Nvidia’s GPU market share at 94% in Q2 2025. Given how heavily the company’s chips are used, the Compute Unified Device Architecture (CUDA) parallel computing framework used to run them is also “squeezing” the enterprise technology industry, according to analysts who criticized cloud-native computing at last year’s KubeCon conference. Foundation leadership on whether the open source community can introduce competitors to CUDA.
Dickens, who participated in the conversation at KubeCon, predicted that Maia 200, Trainium, and TPUs will begin to weaken Nvidia’s lead in the long term.
“NVIDIA will continue to grow with TPUs, AMD like Maia, and eventually Intel,” Dickens said. “NVIDIA’s 90% market share will normalize over time, but there is still plenty of room to grow as the market expands.”
Chhabra said he would not agree for at least the next three to four years.
“We’re looking at several aspects to arrive at that number,” he said. “Customer mindshare and sentiment, demand, ecosystem support, modernity, product and launch maturity, and proven architecture.”
Beth Pariseau, senior news writer at Informa TechTarget, is an award-winning IT journalism veteran. Any tips? send an email to her.
mike leone
stephen dickens
