Each time you ask a chatbot a question or let an assistant draft an email, a chip in a distant data center does the work. It burns electricity the whole time. Repeat that for hundreds of millions of people every day, and the bill for chips and power becomes one of the largest costs in the technology business.
Most of that computing runs on chips from a handful of makers, led by graphics processing units, or GPUs, that can handle almost any AI model. Buyers pay for that flexibility, and the price carries a wide profit for the chip maker. When demand runs ahead of supply, every buyer also waits in the same line for delivery.
So the biggest cloud companies now design some of their chips in-house. Each chip is shaped around the work its owner runs most often, a contract factory makes it, and standard GPUs still handle everything else.
Microsoft is one of those companies. It designs Maia accelerators for AI and Cobalt processors for general cloud work, and it still buys chips from Nvidia and Advanced Micro Devices (AMD). On BloFin, crypto traders can follow Microsoft's share price through the MSFTUSDT Perpetual, which tracks the stock and settles in USDT.
What are Microsoft's Maia and Cobalt chips?
Microsoft's Maia and Cobalt chips are the two families of processors it designs for its own Azure data centers. Maia chips are AI accelerators built mainly to run trained models, while Cobalt chips are Arm-based processors for everyday cloud work. Maia 200 is in production, and Maia 300 is due in 2027 (source: Microsoft's Maia 300 announcement).
An accelerator is a processor made for one kind of math, done very fast, and that math is what AI models run on. It works beside a server's central processing unit, or CPU, which runs everything else on the machine. A GPU is an accelerator too. Nvidia and AMD sell theirs to every buyer, so each one has to cope with any model a customer brings.
AI work comes in two stages. Training teaches a model from vast amounts of text and images, and it can keep thousands of chips busy for weeks. Inference is the answering stage, when the finished model responds each time you send Copilot a question, and Maia is built mainly for that second stage.
Microsoft showed its first Maia and its first Cobalt at its Ignite conference on November 15, 2023. It said the chips would sit beside products from its chip partners. In the same news, it said it would add Nvidia's H200 GPU and AMD's MI300X accelerator to Azure (source: Microsoft Source on the 2023 chips). That mix of its own chips and bought ones has held ever since.
Maia generations from Maia 100 to Maia 300
Microsoft has announced three Maia chips. More than two years passed between the first and the second, and the third followed less than nine months after the second.
Chip | Announced | What Microsoft said about it | Status |
|---|---|---|---|
Maia 100 | November 15, 2023 | first Maia, built for AI and generative AI work | planned to roll out from early 2024 |
Maia 200 | January 26, 2026 | made by Taiwan Semiconductor Manufacturing Co (TSMC) on a 3-nanometer process, over 140 billion transistors, 216 GB of fast memory, 750 watts | in production, serving Microsoft Copilot |
Maia 300 | October 12, 2026 | computing chiplets made on TSMC's N2P process, eight memory stacks, 950 watts, 96 chips to a rack | production deployments from 2027 |
Microsoft announced Maia 200 in January as a chip built for inference, the answering stage (source: Microsoft's Maia 200 announcement). The memory is high-bandwidth memory, or HBM, stacked right beside the chip so data reaches it quickly. Large models spend much of their time waiting on memory, which is why each generation adds more of it.
Maia 300 is built from chiplets, smaller pieces of silicon joined in one package, with separate parts for computing, memory control and the links to other chips. Microsoft designed it for long, memory-heavy tasks, such as an AI agent reading a long document before it answers (source: Microsoft's Maia 300 announcement).
Cobalt CPUs for everyday cloud work
Cobalt is the other half of Microsoft's chip work: a CPU for the ordinary work every cloud server runs, from databases to websites. Cobalt 100 runs in 32 Azure regions, and Cobalt 200 offers up to 50% better performance than its predecessor. TSMC, the contract chip maker that also builds Maia, makes it on a 3-nanometer process, using Arm's Neoverse V3 design (source: Azure blog on Cobalt 200).
Arm is a chip design company that licenses its processor designs to other firms, so Microsoft can build on a proven core and tune the rest. Cobalt 200 opened to customers as an early preview at Microsoft's Build conference in June 2026, in eight regions across the US, Europe and Asia (source: Azure blog on Cobalt 200).
CPUs matter more for AI than they used to. On Microsoft's July 29, 2026 earnings call, Satya Nadella said that for running agents, "CPUs are just as important as GPUs". He also said Microsoft expected to have Cobalt 200 racks in over 25 data centers around the world by the end of July (source: Microsoft FY26 Q4 earnings call). An agent is an AI tool that carries out a series of steps, and much of that work runs on ordinary processors.
Why Microsoft makes its own AI chips
Microsoft makes its own AI chips to run its AI services for less money and less power than bought chips allow. It also wants to depend less on any single supplier. Its 2026 shareholder letter, dated October 14, says it expects its own chips to power a greater share of its fleet over the coming years (source: Microsoft 2026 Annual Report).
Cost is the plainest reason, because a bought chip carries the seller's profit. Power is the second. A data center can draw only so much electricity. Microsoft now measures Maia in tokens per megawatt, which shows that it is trying to stretch power as well as money.
The third reason is fit. A company that designs the chip, the server, the rack and the software together can shape each part around the others. Microsoft described this in 2023 as building "from silicon to service" (source: Microsoft Source on the 2023 chips). The fourth is choice. On the July call, chief financial officer Amy Hood said Microsoft keeps looking "for the best price performance on silicon, including our investments in first party" (source: Microsoft FY26 Q4 earnings call). A buyer with its own chip has a second option every time it negotiates for the rest.
Crypto holders have seen a version of this before. Bitcoin mining moved from graphics cards to ASICs, chips built to compute bitcoin's hash and nothing else. One fixed task, repeated without end, rewards a chip made only for it. The parallel holds for inference, where Microsoft runs the same models all day. It weakens for training, where the newest models change too fast for a fixed design.
Cost per token, worked through
A token is a chunk of text, roughly a word or part of one, and AI services are measured by how many tokens they produce. The cheaper each token, the more an AI service earns from the same equipment. Two numbers show how much a bought chip adds to that cost: the seller's margin, and the gains Microsoft claims for its own chips.
Gross margin is the part of each sales dollar a seller keeps after paying to make what it sold. Nvidia's gross margin came to 75.0% in the three months to July 26, 2026 (source: Nvidia's Q2 fiscal 2027 results). Say you spend $20 billion with a seller earning that margin. Roughly $15 billion of it is the seller's gross profit, and about $5 billion pays for making the chips.
A custom chip wins back only part of that gap. Microsoft still pays TSMC to make it, pays for memory and design help, and carries years of engineering salaries before the first chip ships. Microsoft keeps whatever part of the seller's profit those costs leave over. That gain applies only to the work Maia can handle in place of a bought GPU.
Microsoft's own claims put a size on the gain. It says Maia 200 delivers 30% better performance per dollar than the newest hardware already in its fleet (source: Microsoft's Maia 200 announcement). It also expects Maia 300 to produce 40% more tokens per megawatt than leading GPU-based systems, without naming which ones (source: Microsoft's Maia 300 announcement). Both are Microsoft's own figures, not results from an independent test.
Cost of the same work = 1 ÷ (1 + the claimed gain)
Microsoft's claim | Same output needs | Saving |
|---|---|---|
30% better performance per dollar (Maia 200) | 1 ÷ 1.3 = 76.9% of the money | 23.1% less money |
40% more tokens per megawatt (Maia 300) | 1 ÷ 1.4 = 71.4% of the power | 28.6% less power |
A 30% gain sounds as if it should cut the bill by 30%. It trims it by 23.1%, because the gain is measured on the output side. Take work that would cost Microsoft $1 billion a year on bought hardware. At that rate, Maia would save about $231 million, and the saving grows with every rack it moves onto Maia.
Who designs and builds Microsoft's chips
Microsoft designs Maia and Cobalt itself, and TSMC builds them. Microsoft names TSMC's 3-nanometer process for Maia 200 (source: Microsoft's Maia 200 announcement) and for Cobalt 200 (source: Azure blog on Cobalt 200), and TSMC's N2P process for Maia 300's computing chiplets (source: Microsoft's Maia 300 announcement). Any other partner appears only in press reports.
Reuters also reported in January 2026 that TSMC makes Maia 200 on its 3-nanometer technology (source: Reuters via Yahoo Finance). The nanometer figure names a generation of TSMC's manufacturing. Newer generations pack transistors more tightly, so a chip can do more work on the same electricity. N2P belongs to TSMC's 2-nanometer family.
Design help is where the reporting gets thinner. In December 2025, The Information reported that Microsoft was in talks to design future chips with Broadcom. That would move its custom chip work away from Marvell, for which Microsoft is the second-largest cloud customer after Amazon, the report said (source: Sherwood News on the Broadcom talks). The coverage carried no confirmation from Microsoft, Broadcom or Marvell, and Microsoft's Maia 300 announcement names no design partner.
The partner question decides how much of the margin Microsoft keeps. A design partner charges for its work on each chip, so a custom chip trims the bill and leaves part of it with whoever helps build it.
OpenAI's chip designs as a third route
Microsoft also has rights to OpenAI's chip work, which gives it a third source of designs beside Maia and the chips it buys. OpenAI is designing AI chips with Broadcom. In a November 2025 interview with podcaster Dwarkesh Patel, Nadella said of OpenAI: "As they innovate even at the system level, we get access to all of it" (source: TechCrunch on Microsoft's chip plan).
The terms come from the companies' October 2025 agreement. It extended Microsoft's rights to OpenAI's models and products through 2032. Microsoft also kept its rights to intellectual property related to data center hardware and software, while OpenAI's consumer hardware was left out (source: Microsoft on the October 2025 OpenAI agreement).
The April 2026 revision kept the license through 2032 and made it non-exclusive, so OpenAI can grant the same rights to others. The two companies also named "collaborating on next-generation silicon" among their joint work (source: Microsoft on the April 2026 OpenAI update). The rest of that relationship, from the size of the stake to how it is counted, is covered in the guide to Microsoft's OpenAI stake.
Microsoft's Nvidia and AMD chips
Microsoft still buys AI chips from Nvidia and AMD alongside its own. On the July 29, 2026 call, Nadella said Microsoft keeps modernizing its fleet "with our own silicon innovation, alongside the latest from NVIDIA and AMD." That includes racks built on AMD's Helios and Nvidia's Vera Rubin designs (source: Microsoft FY26 Q4 earnings call).
Training the largest models is where Microsoft leans hardest on GPUs. 24/7 Wall St reported that Microsoft will keep leasing Nvidia chips for training and plans to be among the first clouds to roll out Nvidia's Vera Rubin systems (source: 24/7 Wall St on Maia 300). Hood described the infrastructure as "pretty fungible", meaning the same capacity can switch between tasks as demand shifts (source: Microsoft FY26 Q4 earnings call).
Chips are only half of Nvidia's hold on buyers. Most AI developers write their code for Nvidia's software platform, Compute Unified Device Architecture (CUDA). Reuters reported that Microsoft's Maia tools include Triton, an open-source tool that takes on the same tasks as CUDA (source: Reuters via Yahoo Finance). The more of Microsoft's own software runs through Triton, the easier it is to move work between Nvidia chips and Maia.
AMD's chips have been part of Azure since Microsoft announced MI300X virtual machines in 2023. Who buys AMD's AI chips sets out what Microsoft has committed to since.
For the seller's side of the same shift, Nvidia's moat and the custom-chip threat explains why cloud companies keep buying GPUs while they build their own. Microsoft's dual role as Nvidia's customer and rival also shapes the comparison of Microsoft and Nvidia as stocks.
Maia 300 and the October 2026 announcement
On October 12, 2026, Microsoft announced Maia 300 and said it will scale the Maia family to multi-gigawatt deployments starting in 2027. That would make Maia the primary inference system for Microsoft products and services. Maia 200, already serving Microsoft Copilot, is deploying across the United States, Europe and Australia (source: Microsoft's Maia 300 announcement).
A gigawatt is a billion watts, the unit data center builders use for the electricity their chips draw. Each Maia 300 rack holds 96 chips inside a 167-kilowatt power limit. Microsoft says a full rack produces up to seven times as many tokens as a Maia 200 rack (source: Microsoft's Maia 300 announcement). One gigawatt of those racks would be about 6,000 racks and roughly 575,000 chips. That counts only the racks' own power and leaves out cooling and the rest of the building.
Nadella wrote on X that day: "Today we are announcing our biggest bet yet on our own silicon." Microsoft's shares closed 1.14% higher, while Nvidia's fell 1.29% (source: 24/7 Wall St on Maia 300). Two days later, his shareholder letter called Maia 300 and Cobalt 200 the year's progress in "our own first-party silicon" (source: Microsoft 2026 Annual Report).
A move of about 1% tells you little on its own. The chips change Microsoft's costs over years, and the shares that day also traded on every other piece of news in the market.
Maia delays and the 2027 schedule
Maia 300's 2027 date rests on a schedule that has slipped before. In June 2025, The Information reported that the next Maia chip, code-named Braga, was at least six months late, pushing mass production from 2025 into 2026. The report blamed design changes, staff limits and high turnover, and said the chip was expected to trail Nvidia's Blackwell chips (source: Business Today on the Maia delay).
Microsoft announced Maia 200 seven months later, in January 2026. A delay costs money in a plain way. While a chip is late, the work it was meant to take on runs on bought GPUs, at Nvidia's or AMD's prices. TechCrunch described Microsoft's chip effort in November 2025 as struggling and behind Google's and Amazon's (source: TechCrunch on Microsoft's chip plan), which is part of why the OpenAI route matters.
Launch dates, delays and stock reactions all reach the share price. You can watch how chip news lands on the MSFTUSDT Perpetual page, where the price moves with the shares during the US session.
Majorana 1, Microsoft's quantum chip
Majorana 1 is Microsoft's quantum computing chip, unveiled on February 19, 2025, and it is a research project that sits apart from Maia and Cobalt. Microsoft says the chip holds eight topological qubits on a design meant to scale to one million, and that the approach could tackle industrial-scale problems within years (source: Microsoft Source on Majorana 1).
A qubit is the basic unit of a quantum computer, the counterpart of a bit in an ordinary one. Eight qubits on a chip built to hold a million shows how early the work still is. Microsoft is one of two companies invited into the final phase of a US defense research program on large-scale quantum computers, DARPA's US2QC (source: Microsoft Source on Majorana 1).
For the stock, quantum is a long-dated research bet. Microsoft does not report quantum revenue separately, and the cost of the work sits inside its research and development spending. So Majorana news can change how investors see Microsoft's future without showing up in its quarterly results.
What Microsoft's own chips mean for the stock
Microsoft's own chips matter to the stock through cost. If Maia does the same AI work for less money and power, each new data center earns more from the same spending. That supports cloud margins and free cash flow. If a generation slips or falls short, Microsoft buys more GPUs at outside prices, and the bill rises.
The chips are one part of a much larger spending plan that also pays for land, buildings, power and bought GPUs. Microsoft's AI capex and cash flow breaks that total down. A cheaper chip also lowers depreciation, the yearly charge for wear on equipment, which comes off profit in each year the servers run.
Maia and Cobalt run inside Azure, Microsoft's cloud, so their effect shows up in the cloud business. How Microsoft makes money sizes that business against the rest of the company. Each report gives three signs of whether the chip plan is working. Watch how fast capital spending grows against cloud revenue, what Microsoft says about its own chips' share of the fleet, and whether its cloud gross margin holds up.
Microsoft set Wednesday, October 28, 2026, after the Nasdaq close, as the release date for its July to September 2026 results. The Microsoft earnings preview lists what the company guided for that quarter. The call is where Nadella and Hood usually update the Maia and Cobalt rollout.
Trading Microsoft's chip news on BloFin
On BloFin, you trade Microsoft's chip news through the MSFTUSDT Perpetual, which follows Microsoft's share price, pays out in USDT and allows leverage of up to 20x. It never expires, and its funding normally settles at 8AM, 4PM and 12AM UTC each day (source: BloFin MSFTUSDT contract information).
Funding is a small fee that long and short holders of the contract swap at set times, nudging its price back toward the stock's. How crypto funding rates work explains who pays whom. Trading hours matter too. Outside US market hours, on weekends and on holidays, BloFin can switch stock futures to reduce-only trading. In that mode you may cut or exit what you already hold, while new positions and top-ups wait for the market to reopen (source: BloFin stock futures notice). Microsoft's results come out after the US close, so a position you want to hold through a report has to be open before the 4PM ET close, which is 8PM UTC in summer and 9PM UTC in winter.
BloFin also lists perpetuals on Microsoft's chip suppliers and reported partners, tracking the shares of Nvidia, AMD, Broadcom, Marvell and TSMC. One chip headline can move several of them at once, as Microsoft and Nvidia moved in opposite directions on October 12. Pair trades on BloFin covers trading one contract against another.
Leverage multiplies every move, in both directions, against the margin you post. A 5% fall in Microsoft's shares takes 50% off the margin of a 10x long position before fees and funding. A gap after the close can also carry the price past your liquidation level before you can act.
Looking to trade Microsoft around its chip news? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the MSFTUSDT Perpetual trading page.
Frequently asked questions
What software runs on Microsoft's Maia chips?
Maia runs models through the Maia SDK, a set of tools Microsoft offers developers. It includes PyTorch support, a Triton compiler and a library of tuned code for common AI tasks. Together they let teams move models built for other chips onto Maia with less rework. PyTorch is the widely used framework many AI models are written in. Microsoft previewed the SDK alongside Maia 200 in January 2026 (source: Microsoft's Maia 200 announcement).
Where are Microsoft's Maia chips running?
Maia 200 went live first in Microsoft's US Central region near Des Moines, Iowa. In January 2026, Microsoft named its US West 3 region near Phoenix, Arizona as next (source: Microsoft's Maia 200 announcement). By October 2026, Microsoft said it was deploying Maia 200 across the United States, Europe and Australia. Maia 300 starts production deployments in 2027 (source: Microsoft's Maia 300 announcement).
Which AI models run on Maia 200?
Maia 200 runs OpenAI models and Microsoft's own MAI models, and it already serves Microsoft Copilot. Microsoft said in October 2026 that GitHub Copilot and Microsoft Foundry, its service for building AI apps, would follow by the end of the year (source: Microsoft's Maia 300 announcement). At launch in January, it named OpenAI's GPT-5.2 models and Microsoft 365 Copilot among the first users (source: Microsoft's Maia 200 announcement).
How does Maia 200 compare with Amazon's and Google's chips?
By Microsoft's own figures, Maia 200 has three times the 4-bit performance of Amazon's third-generation Trainium chip, and 8-bit performance above Google's seventh-generation TPU. Microsoft gives over 10 petaflops at 4-bit precision and over 5 petaflops at 8-bit (source: Microsoft's Maia 200 announcement). A petaflop is a thousand trillion calculations a second. On the rivals' side, Amazon's own AI chips and Google's TPU chips follow the same cost logic.
Which companies use Microsoft's Cobalt chips?
Microsoft's own services run on Cobalt, and Azure offers Cobalt 100 virtual machines to outside customers in 32 regions. Microsoft names Databricks and Snowflake among the companies that adopted Cobalt 100, with Amadeus, Siemens, Sprinklr and Temenos also reporting gains (source: Azure blog on Cobalt 200). On the July 2026 call, Nadella added Adobe, Arm, Elastic and OpenAI to the list of customers running work on Cobalt servers (source: Microsoft FY26 Q4 earnings call).
Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include Microsoft's announcements of Maia 100 and Cobalt 100 (2023), Maia 200 (January 2026), Cobalt 200 (June 2026) and Maia 300 (October 2026), its fiscal 2026 fourth-quarter earnings call, its 2026 Annual Report, its October 2025 and April 2026 posts on the OpenAI partnership, its Majorana 1 announcement and Nvidia's second-quarter fiscal 2027 results. News sources include Reuters, The Information, TechCrunch, Sherwood News, Business Today and 24/7 Wall St. BloFin's sources are its stock futures notice and its MSFTUSDT contract information, current as of October 28, 2026.
Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold Microsoft in any form. Microsoft's chip savings are its own estimates, and a late or weaker Maia generation would leave it paying outside prices for more GPUs. The MSFTUSDT Perpetual gives you no shares or dividends, funding payments can cost you while you hold it, and a leveraged position can be liquidated if Microsoft's price gaps against you after the close. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.