{"slug": "microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027", "title": "Microsoft Orders Over 300,000 Maia 300 AI Chips From TSMC For 2027", "summary": "Microsoft has ordered over 300,000 of its Maia 300 AI chips from TSMC for 2027, according to The Information, as the company revises its AI chip roadmap amid delays. The latest-generation chip, internally known as Braga and expected to be sold as Maia 200, has been pushed from 2025 to 2026, and Microsoft is considering an intermediary 2027 chip, likely called Maia 280, before a Maia 400 enters mass production in 2028. The move aims to reduce Microsoft's reliance on Nvidia, which was its largest customer by revenue last year, and to cut costs for Azure, Copilot, and OpenAI-linked workloads.", "body_md": "*Microsoft's custom AI chip plan is real, but the clean story is not a 300,000-chip order. The harder fact is that Microsoft is still trying to get out from under Nvidia's pricing power while its own Maia roadmap keeps slipping.*\n\nMicrosoft wants its own AI chips badly enough to keep reworking the plan, even after delays forced it to scale back some of the ambition. That's the part you should watch. The company doesn't need Maia to beat Nvidia everywhere. It needs Maia to make Azure, Copilot and OpenAI-linked workloads cheaper to run at Microsoft scale.\n\nAccording to The Information, Microsoft has revised its internal AI chip roadmap. It's now focusing on less ambitious designs through 2028, after development problems slowed the program. Its latest-generation chip, known internally as Braga and expected to be sold publicly as Maia 200, was pushed from 2025 into 2026. That's a full year lost. The same report said Microsoft executives have considered an intermediary 2027 chip, likely called Maia 280, before a Maia 400 chip enters mass production in 2028.\n\nThat is not the smooth march you usually see in a hardware launch deck.\n\nMicrosoft announced Maia 100 in 2023 and released it in 2024 as its first custom AI accelerator for Azure. The problem is that the market moved faster than the chip. The Information reported earlier this year that Maia 100 has mainly been used for internal testing rather than powering Microsoft's major AI services, while the delayed Maia 200 has had to absorb design changes tied to OpenAI's needs.\n\n## The Nvidia Problem Is Still the Price\n\nThe economics are the whole point. Nvidia still dominates the accelerator market used to train and serve large AI models, and Microsoft remains one of its most important customers. The Information reported that Nvidia employees said Microsoft was Nvidia's largest customer by revenue last year, which tells you how expensive this dependency has become.\n\nHere's the thing: Microsoft doesn't have to replace Nvidia overnight for Maia to matter. If it can move a meaningful slice of inference, the work that serves Copilot answers and AI responses after a model is trained, onto its own chips, the savings can be large. Inference is where the meter keeps running: every prompt, every answer, every enterprise Copilot query carries its own compute cost, and that's why custom silicon matters to you even if you never buy a chip.\n\nThe AI services being sold into offices aren't just software subscriptions. They are claims on scarce data center power, memory, networking gear and foundry capacity. Real infrastructure, real limits. When Microsoft builds its own accelerator, it is trying to control more of that stack instead of paying Nvidia's margin every time demand rises.\n\nAmazon and Google already know this lesson. Amazon has Trainium and Inferentia. Google has spent years building tensor processing units, and The Information reported that Google can train or operate most of its own AI models without Nvidia chips, even while it still rents Nvidia-powered servers to cloud customers. Microsoft is late. That's plain enough.\n\nBut late is not the same as finished.\n\n## The Proof Has to Come From Workloads\n\nMicrosoft's strongest test will not be whether it can show another chip onstage. It will be whether Maia can run real production workloads at a cost customers and internal teams can measure. Data Center Dynamics reported in May that Anthropic was in early-stage talks to use Microsoft's custom Maia chips as the AI lab expanded its Azure usage, citing The Information. Those talks were not guaranteed to become a deal, and that caveat matters.\n\nA major outside AI lab would be a stronger signal than internal testing alone. OpenAI already sits close to Microsoft through Azure. Anthropic would be different. If a company building Claude decides Maia is good enough for production work, Microsoft gets the kind of validation that benchmark slides can't provide.\n\nFor now, the dull details are the important ones. Maia 200 has been delayed. Microsoft is reportedly simplifying the roadmap. Maia 400 is not expected until 2028. Its Cobalt server CPU, by contrast, is already used internally and available to Azure customers, according to The Information, which shows that Microsoft can ship custom silicon when the job is less brutal than an AI accelerator.\n\nNvidia's grip will not loosen quickly. Training frontier models still leans heavily on its GPUs, and Microsoft has given no serious sign that it plans to walk away from that relationship. The real fight is narrower and more practical: how much AI serving work Microsoft can pull onto its own hardware before Nvidia's next generation raises the bar again.\n\nThat is the race Maia is actually in. Not a press-release race. A cost race.\n\n**Also read:** [An AI Booking Agent Hacked a Melbourne Gym to Bump a Stranger From Class](https://startupfortune.com/an-ai-booking-agent-hacked-a-melbourne-gym-to-bump-a-stranger-from-class/) • [Harvard and MIT built an AI model of 8.3 billion people to test products on](https://startupfortune.com/harvard-and-mit-built-an-ai-model-of-83-billion-people-to-test-products-on/) • [JPMorgan Raises S&P 500 Target to 8,000 Saying AI Spending Is Finally Paying Off](https://startupfortune.com/jpmorgan-raises-sp-500-target-to-8000-saying-ai-spending-is-finally-paying-off/)", "url": "https://wpnews.pro/news/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027", "canonical_source": "https://startupfortune.com/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027/", "published_at": "2026-08-10 18:00:56+00:00", "updated_at": "2026-08-10 18:18:15.794445+00:00", "lang": "en", "topics": ["ai-chips", "ai-infrastructure", "ai-products"], "entities": ["Microsoft", "TSMC", "Nvidia", "Maia 300", "Maia 200", "Maia 280", "Maia 400", "The Information"], "alternates": {"html": "https://wpnews.pro/news/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027", "markdown": "https://wpnews.pro/news/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027.md", "text": "https://wpnews.pro/news/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027.txt", "jsonld": "https://wpnews.pro/news/microsoft-orders-over-300000-maia-300-ai-chips-from-tsmc-for-2027.jsonld"}}