The obvious reading of Monday’s Microsoft post is that a 40 percent efficiency gain just cut the Nvidia bill. It did not. Scott Guthrie, executive vice president of Cloud + AI, wrote on 12 October 2026 that Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems, with production deployments beginning in 2027, when Maia is supposed to become the primary inference system. Maia 200 is the chip already in production. Nasdaq’s official Monday close was $541.15, up $6.08, or 1.14 percent, from Friday’s $535.07. Satya Nadella posted at 13:03 UTC, 9:03 a.m. Eastern, and the shares still opened at $537.53. A token from an inference rack is not a training purchase order, and Microsoft published no dollar of chip spending moving off anyone else’s silicon.
Key facts
- Monday 12 October 2026 regular-session close $541.15, up $6.08, or 1.14 percent, from Friday’s $535.07. Open $537.53, high $542.74, low $535.7013, volume 18,314,940. Source: Nasdaq historical prices, downloaded 13 October 2026. The highest close in that file is $542.07 on 28 October 2025. Monday finished $0.92 under it.
- At 7:52 a.m. Eastern on Tuesday 13 October, Nasdaq still marked the session premarket, last sale $543.35, up $2.20, or 0.41 percent. The 52-week range on that page was $349.20 to $553.72. That print is not a close and is not on the chart.
- Guthrie, on the Official Microsoft Blog: “We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.” And: “we expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems.”
- Same post: a Maia 300 rack of 96 accelerators in a 167-kilowatt envelope, with compute chiplets on TSMC N2P and a 950-watt TDP. Maia 200, deploying in the United States, Europe and Australia, is at up to 1.7 times the tokens per watt of leading GPU infrastructure on Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry on Maia 200 are “to follow by end of year.”
- Nadella, on X at 13:03 UTC: “Today we are announcing our biggest bet yet on our own silicon.” He writes that Maia 300 “delivers 40% more tokens per watt compared to leading GPU-based systems.” Guthrie says “we expect” and “per megawatt.” The ratio matches. The verb does not. Maia 300’s production date is 2027.
- June-quarter release, 29 July 2026: revenue $90.0 billion, up 18 percent, and Azure and other cloud services revenue up 43 percent. Cash additions to property and equipment were $35.802 billion in the quarter and $115.948 billion in the year, versus $17.079 billion and $64.551 billion. Source: Microsoft’s FY26 Q4 release. Nadella said Azure revenue surpassed $100 billion for the first time, and Copilot reached over 30 million paid seats. Commercial remaining performance obligation was $678 billion, up 84 percent. The investor page lists FY27 Q1 earnings on 28 October 2026. The SEC 8-K feed, checked 13 October, shows nothing after 2 September 2026.
What just happened, and why the obvious reading is wrong
Monday’s document is an inference schedule. Maia 200 is in production and already serving Copilot. Maia 300 comes in 2027, and the primary-system sentence is attached to that year. GitHub Copilot and Microsoft Foundry on Maia 200 are still “to follow by end of year.” A 2027 role for the product stack is not a claim that the stack moved on Monday.
Guthrie does not say Microsoft will buy 40 percent fewer GPUs. He writes that “we expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems.” A token is served output. A megawatt is the power used to serve it. Training demand is another job, and an invoice is a third. Nadella’s “delivers” and “per watt” are the same ratio if the baseline matches, and a different tense. A chip whose “production deployments” begin in 2027 has not delivered a fleet result.
The live chip has the larger cross-GPU number, on a narrower job. Maia 200 is at up to 1.7 times the tokens per watt of leading GPU infrastructure on Teams, Copilot and Cowork. That is a ceiling of up to 70 percent, and the GPU clause is Guthrie’s. Nadella states the 1.7 times without it. The same day’s Azure note, by Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, corporate vice president of Systems Planning and Architecture, says Maia 300 does up to 7 times the rack-scale token throughput of Maia 200 and more than 2.5 times the performance per watt of Maia 200. Seven times versus the prior Maia rack is not 40 percent versus a GPU. Their sentence sets the scope: the aim is to “optimize the entire system for LLM inference so that every watt and every dollar delivers more intelligence.” That dollar is not a saving booked on Monday.
July already blocked a single-supplier reading. On 20 July 2026 Guthrie wrote that Azure was still expanding its AI fleet with AMD’s most advanced AI and high-performance computing systems, beside Microsoft’s own silicon, on what he called an open and heterogenous platform. The page names ND MI455X v7 virtual machines for AI inference on AMD’s Helios rack. The 12 October post does not withdraw them. Primary is not a synonym for only.
Who is exposed
Holders are first. The June quarter’s weighted-average diluted share count was 7,443 million, so Monday’s $541.15 is about $4.028 trillion. That count ended 30 June, and the quarter returned $10.2 billion in dividends and repurchases, so the live count can differ. A $6.08 move does not reprice the equity, and it does not say which slice of the $115.948 billion in fiscal-year property and equipment additions was a GPU invoice.
The cloud print is why a real saving would matter later. Azure and other cloud services revenue rose 43 percent in the June quarter, inside Intelligent Cloud revenue of $39.3 billion. Nadella said that, in that fiscal year, Azure revenue surpassed $100 billion for the first time, and that Copilot had over 30 million paid seats. Amy Hood, chief financial officer, put Microsoft Cloud revenue at $59.3 billion, up 27 percent. Maia 200 is starting to serve those workloads. The October post gives no cost per token. Copilot is in production on it. Teams and Cowork are in the 1.7 times line. GitHub Copilot and Foundry are later in 2026.
AMD stays in the July inference lineup until Microsoft says otherwise. Nvidia is the name the tape reaches for, and this post gives that trade no unit cut and no named GPU. Nadella also wrote of “continuing to bring the latest innovations from across the ecosystem into our fleet.” FinanceFeeds has already covered how Nvidia traded the separate OpenAI stories. Hosting OpenAI models on Maia 200 is not the July non-GAAP bridge, which excludes the OpenAI investment and includes a $3.2 billion Anthropic gain. The stock is 53.4 percent above the 25 June close of $352.83 and only 4.1 percent above the 1 October 2025 close of $519.71. An earlier FinanceFeeds note used a different spot and a different bull and bear. Those prices are not reused here.
What the tape and the documents show
Change is versus the prior close in the Nasdaq file. Tuesday 13 October had not produced a cash close when the file was saved. The 7:52 a.m. Eastern premarket sale is not a row.
| Session | Close | Change | Volume | Range |
|---|---|---|---|---|
| Mon 12 Oct 2026 | $541.15 | +$6.08, +1.14% | 18,314,940 | $535.7013 to $542.74 |
| Fri 9 Oct 2026 | $535.07 | +$12.46, +2.38% | 18,500,610 | $526.2401 to $537.22 |
| Thu 8 Oct 2026 | $522.61 | -$7.15, -1.35% | 20,228,700 | $518.79 to $533.51 |
| Wed 7 Oct 2026 | $529.76 | +$0.46, +0.09% | 16,790,230 | $524.69 to $531.73 |
| Tue 6 Oct 2026 | $529.30 | +$4.12, +0.78% | 21,107,460 | $528.78 to $535.69 |
| Mon 5 Oct 2026 | $525.18 | +$7.65, +1.48% | 26,604,760 | $521.47 to $532.35 |
| Thu 25 Jun 2026 | $352.83 | low close in the file | 66,360,470 | $349.20 to $364.23 |
| Tue 28 Oct 2025 | $542.07 | high close in the file | 29,986,680 | $540.77 to $553.72 |
Monday continued Friday, which had already closed up 2.38 percent at $535.07. The open at $537.53 was only 0.46 percent above Friday, with Nadella’s post 27 minutes old, and the session added $3.62 to the close. The low held above Friday. The high, $542.74, got $0.67 above the old closing high, and the close still finished under $542.07. The post landed on the last step of a round trip back toward the October 2025 high. It does not explain the round trip. The same desk’s AMD note is about a different supplier’s supply comment, not this inference schedule.
| Item | Maia 200 | Maia 300 | Not stated |
|---|---|---|---|
| Status | In production, serving Copilot. GitHub Copilot and Foundry later in 2026 | Production deployments begin in 2027 | Any share of chip purchases moving on a date |
| Versus GPUs | Up to 1.7 times tokens per watt on Teams, Copilot and Cowork | Expected 40 percent more tokens per megawatt | The GPU generation in the baseline |
| Versus prior Maia | The baseline | Up to 7 times the rack-scale tokens, and more than 2.5 times the performance per watt | A combined figure versus GPUs |
| Scale | United States, Europe, Australia | Multi-gigawatt from 2027. 96 accelerators, 167 kilowatts | How many gigawatts |
| Work | OpenAI, MAI and open-source models, plus Copilot | Primary inference system from 2027 | Any move of training demand off merchant GPUs |
What Microsoft said, and what it left off the page
“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services,” Guthrie wrote. Multi-gigawatt is not a count. A rack at 167 kilowatts is 0.167 megawatts, so one gigawatt is about 5,988 racks, or about 575,000 accelerators. Microsoft did not say how many gigawatts. Chip power and rack power also differ. Ninety-six times the 950-watt TDP is 91.2 kilowatts, 54.6 percent of the rack. Hosts, memory, storage, cooling and switches are the other 75.8 kilowatts. The 40 percent claim is across the system, per megawatt, not 40 percent off the chip alone.
Memory checks out and still does not name a replacement. 288 GB times 96 rounds to Guthrie’s 27.6 TB, and 18 TB/s times 96 is 1.728 PB/s against his 1.7 PB/s. Nadella’s fuller sentence is less triumphant than the first line: “Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services.” Missing from both pages: a percent of AI-chip spending, a unit cut, a token price against a named GPU, and any 8-K. The SEC feed stops at 2 September 2026.
Power, capex and a quarter that already closed
Fiscal 2026 cash additions to property and equipment were $115.948 billion, against $64.551 billion, up 79.6 percent on those two lines. The June quarter was $35.802 billion, against $17.079 billion. That cash is spent. Maia 300 does not begin production deployments until 2027, so a more efficient rack does not refund buildings already paid for, and the cash-flow line does not split accelerators from power gear. Tokens per megawatt become profit only if the power is on. The architecture note says liquid cooling removes more than 80 percent of the rack’s thermal load, and that the rack can take facility water or dump heat through an exchanger in an air-cooled hall. The blog does not say how many halls are ready.
Commercial remaining performance obligation was $678 billion at 30 June, up 84 percent. The 43 percent increase in Azure and other cloud services was earned on the mixed fleet Guthrie described in July. FY27 Q1 earnings are listed for 28 October 2026, covering a quarter that ended 30 September, twelve days before this post. That call cannot contain Maia 300 shipments. It can be asked whether the purchase mix changes. Until there is a dollar, the mix is July’s description plus October’s intention.
Bull, base and bear
These are not bank targets. The last 90 daily log-returns in the Nasdaq file run from the 3 June 2026 close of $427.34 through Monday. The mean is 0.00262352 and the sample standard deviation is 0.02337094. The window includes the drop to $352.83 on 25 June, which is why the mean is high. Fifty-six sessions remain through 31 December 2026 after removing Thanksgiving, 26 November, and Christmas, 25 December. Base is drift alone: $541.15 times e to the power of 0.00262352 times 56, or $626.79, 15.8 percent above Monday. Bull adds 1.5 standard deviations, $814.81, up 50.6 percent. Bear subtracts one standard deviation, $526.22, down 2.8 percent.
The bear is shallow because one standard deviation over 56 sessions is about 17.5 percent and the drift is about 15.8 percent. They nearly cancel. With the drift removed, one standard deviation is $454.32, and 1.5 standard deviations is $416.28 down or $703.48 up.
| Case | Price | Versus $541.15 | What has to be true |
|---|---|---|---|
| Bull | $814.81 | +50.6% | The rebound’s daily drift continues through 31 December and the close finishes 1.5 standard deviations above it. GitHub Copilot and Foundry on Maia 200 have to look like a cost item by then. This is a volatility path, not company guidance. |
| Base | $626.79 | +15.8% | Only the mean continues. The 40 percent claim stays an expectation, no gigawatt count appears, and the 28 October call does not price the purchase mix. |
| Bear | $526.22 | -2.8% | The drift holds and the stock finishes one standard deviation under it. The 2027 rack stays a slide and Monday’s failure to clear $542.07 stands. If the drift breaks, the stress is near $454, not a second official case. |
This is not financial advice. The band cannot know if the 40 percent survives a named GPU, because Microsoft has not named one.
What happens next
On 28 October 2026 Microsoft reports a quarter that ended 30 September. Maia 300 is not in those shipments, there was no 8-K on 12 October, and the blog gave no purchase mix. The call is the first scheduled time to ask whether the primary-inference sentence changes capex or the Nvidia and AMD mix. A dollar on that call would be the first figure this story does not yet have.
Ignite, which the architecture banner puts on 17 to 20 November 2026 in San Francisco and online, is the product date Guthrie pointed to. The rack spec is already public. Ignite can still add the gigawatt count and the GPU behind the 40 percent. If it does not, the hole is the same. The check inside 2026 is Maia 200, not Maia 300. Guthrie scheduled GitHub Copilot and Foundry on Maia 200 by year-end. If they are on it by 31 December, the in-house footprint is wider than Copilot, Teams and Cowork. If they are still “to follow,” the 2027 sentence is ahead of the chip that is supposed to be in production now.
Frequently asked questions
Does making Maia the primary inference system from 2027 change the Nvidia bill?
It is still a 2027 slide. Guthrie dated the primary-inference role and the multi-gigawatt scale-up to 2027, and he attached the 40 percent token claim to Maia 300 as an expectation, not as a purchase cut. Monday’s official close was $541.15, up 1.14 percent. The post names no dollar of spending moving off Nvidia, and it says Microsoft will keep bringing other companies’ silicon into the fleet.
What is the 40 percent claim, and whose words are they?
Guthrie wrote on 12 October that Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. Nadella wrote the same day that Maia 300 delivers 40 percent more tokens per watt than leading GPU-based systems. The ratio matches. The verb does not. Neither sentence names the GPU, and neither is a measured result from a rack that is in production.
Is Maia 200 already in production, and how is that different from Maia 300?
Yes. The 12 October blog says Maia 200 is in production, serving Copilot, and deploying in the United States, Europe and Australia, at up to 1.7 times the tokens per watt of leading GPU infrastructure on Teams, Copilot and Cowork. Maia 300 begins production deployments in 2027. GitHub Copilot and Foundry on Maia 200 are still scheduled for later this year, so even the current chip is not fully rolled out.
What did Microsoft stock do on 12 October 2026?
It closed at $541.15, up $6.08, or 1.14 percent, from $535.07, on 18,314,940 shares. The open was $537.53, about 27 minutes after Nadella’s post. The high was $542.74 and the low was $535.7013. The close finished $0.92 under the highest close in the Nasdaq file saved here, $542.07 on 28 October 2025. Volume was slightly below Friday.
What are the bull, base and bear prices?
Through the 31 December 2026 close, using the last 90 daily log-returns and 56 sessions, the bull case is $814.81, the base case is $626.79 and the bear case is $526.22, against Monday’s $541.15. Bull is 1.5 standard deviations above the drift. Base is the drift alone. Bear is one standard deviation below it. They are not broker targets, and they are not financial advice.
When can this acquire a dollar figure?
The next scheduled chance is the FY27 first-quarter call on 28 October 2026, which reports a quarter that ended before the Maia 300 post. Ignite, on 17 to 20 November, is the product forum Guthrie pointed to. Neither date creates a saving by itself. A figure appears only if management states a purchase mix, a gigawatt count, or a cost per token against a named GPU.