Microsoft Is Preparing a Million In-House AI Chips | Future Intelligence #31

Microsoft wants to deploy more than a million of its own Maia 300 accelerators in 2027—the first time the volume has been large enough to shift the balance between in-house silicon and purchased GPUs. ASML raised the lower end of its outlook to 44 to 45 billion euros, not because demand is stronger, but because orders outnumber the machines it can produce; the bottleneck in the entire AI cycle is lithography. And while Meta launched its first frontier model, Muse Wonder, JPMorgan barred external AI agents from accessing client accounts—the boundaries of what an autonomous model may do in a regulated environment are taking shape faster than the models’ capabilities themselves.

Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically every Sunday morning by email—in full, including specific scenarios and market implications. You can get ongoing access to the report by joining Bulios Black.

Maia 300: Microsoft is preparing a million in-house chips and gigawatts beyond Nvidia

On October 12, Microsoft unveiled the third generation of its own AI accelerator at the Open Compute Project summit. According to Reuters, it plans to deploy more than a million Maia 300 units as early as 2027, when volume deployment begins. The company’s own announcement stated its goal in gigawatts, not chip counts.

The debate over hyperscalers’ in-house silicon is moving from presentations to numbers at the power-grid level. Microsoft also says Maia 300 is intended to be its main platform for inference across its products—that is, for running models, not training them.

What Maia 300 is, technically

Each accelerator has eight HBM4 memory modules with bandwidth of 18 TB/s and power consumption of 950 W; TSMC manufactures the compute chiplets using its N2P process. High memory bandwidth matters more than raw performance for inference—the model must reread its weights from memory every time it generates a token.

The basic building block is a rack with 96 accelerators, drawing 167 kW, with 27,6 TB of HBM memory and aggregate bandwidth of 1,7 PB/s. Microsoft claims up to seven times the token throughput of a Maia 200 rack, which the company unveiled this January and which powers Copilot. The technical document cites more than 2.5 times the performance per watt.

But the key metric is different: 40% more tokens per megawatt than leading GPU-based systems. Microsoft did not name the system it compared against. Based on the specifications provided, a million chips would amount to roughly 10 400 racks and approximately 1,7 gigawatts of power draw—the limiting factor for expansion today is not silicon, but available electricity.

Why hyperscalers are building their own silicon

The reason is financial. Microsoft is projecting capital expenditures of around 190 billion dollars for 2026, up 61% year over year, including an estimated 25 billion impact from more expensive components. Gross margin fell to 67,6% in the fiscal third quarter, its lowest since 2022, as depreciation on data centers increased.

Maia is not yet aimed at Azure offerings for external customers—unlike Google or AWS, it remains an internal product. Microsoft therefore continues to buy Nvidia hardware at scale, and the savings will not show up until 2027 at the earliest. Its own chip improves its bargaining power, but not its cost structure immediately.

So far, this has not dramatically changed the market split. In March this year, Morgan Stanley estimated Nvidia’s share of AI processor revenue at around 85%, custom ASICs at over 10%, and AMD at under 5%. Bloomberg Intelligence expects Nvidia to retain 70 to 75% of the accelerator market through 2030.

Intel, Broadcom, and the division of profits

On the same day at the same summit, Intel revealed the specifications for its D-series GPUs, formerly known as Crescent Island: 480 GB of LPDDR5X memory, 1,5 TB/s of bandwidth, and 1,3 petaflops of FP8 performance in a 350 W envelope; an air-cooled PCIe card for models with up to 300 billion parameters. Samples are being tested by customers. Intel is betting on cheaper memory and operating economics, not peak performance.

The suppliers of custom chips are making the most money so far. Broadcom reported AI semiconductor revenue of 16,7 billion dollars for the third quarter of fiscal 2026, up 221% year over year, and forecast 21,7 billion for the following quarter. Marvell holds an estimated 20 to 25% of the segment, with Maia and Trainium projects.

Nvidia’s share price fell about 1% on the day of the announcement, to around 226 dollars. On Thursday, October 15, it closed at 221,30 dollars after losing 3,2%, at a market capitalization of 5,51 trillion dollars—still near its 52-week range of 164,27 to 243,37 dollars.

Jensen Huang keeps repeating that demand for AI infrastructure will grow faster than the supply of components. From a valuation perspective, what matters is whether pressure on Nvidia affects volumes or margins.

As long as Maia handles only internal workloads, Microsoft remains a major GPU buyer, but with greater bargaining power. The manufacturing value still goes to TSMC, regardless of whose logo is on the chip.

Watch two things: whether Microsoft confirms its 2027 timeline with its January results, and whether Maia ever moves beyond internal workloads and appears in Azure offerings. Only the latter step would mean direct competition in what customers currently buy as GPU instances.

ASML raises outlook to 44-45 billion euros: lithography is the bottleneck in the AI cycle

The Dutch lithography machine manufacturer reported third-quarter results on October 14 and narrowed its full-year revenue outlook to 44 to 45 billion euros from 43 to 45 billion previously. The reason is not demand optimism, but the opposite of the usual logic: there are more orders than the company can manufacture machines to fulfill.