Microsoft’s AI capex is not the surprise; the surprise is that demand is still ahead of it

The print matters because it separates what investors already expected from what they did not. What was priced in was a large Microsoft quarter: the Street was already at $80,308.7 million of revenue and $3.90 of EPS, leaving little room for a conventional “AI is working” reaction. What actually surprised was the quality of the beat: revenue came in at $81,273.0 million, a +1.2% surprise, while EPS came in at $4.14, a +6.2% surprise, meaning the model did not just clear on cloud demand, it cleared on below-the-line and operating leverage even as the company is spending into the cycle. The mispricing is that investors are likely to focus on the gross margin line falling to 68.0% in Q2 FY2026 from 69.0% in Q1 FY2026, while the more important signal is that revenue accelerated to $81.27 billion after $77.67 billion and still carried enough earnings power to beat EPS by +6.2%. In a semiconductor portfolio, that combination says hyperscale AI demand has not yet hit the budget wall that would normally follow a spending surge.

That revenue trajectory is the first defense of the thesis, because Microsoft is no longer merely growing off a pandemic or seat-expansion base. Revenue was $69.63 billion in Q2 FY2025, $76.44 billion in Q4 FY2025, $77.67 billion in Q1 FY2026, and $81.27 billion in Q2 FY2026, with revenue YoY at +16.7% in the just-reported quarter. The sequence matters more than the single beat: QoQ growth of +4.6% in Q2 FY2026 followed a smaller +1.6% in Q1 FY2026, so the quarter did not show the deceleration investors fear when AI capacity is constrained and enterprise optimization resurfaces. Amy Hood’s company-accounting framing was deliberately simple, saying, “This quarter, revenue was $81.3 billion, up 17% in constant currency.” That wording matters because management chose to frame the quarter around constant-currency growth rather than one-time items, which supports the idea that the business base is compounding through AI demand instead of leaning on accounting noise.

The capacity story explains the margin guide, because the cost of supporting that demand is now visible in the P&L and the balance sheet at the same time. Gross margin was 68.0% in Q2 FY2026, down from 69.0% in Q1 FY2026 and 68.7% in Q2 FY2025, and the Q3 FY2026 history line shows 67.6%, so the model is absorbing real dilution. The bear case will argue that this is the beginning of an AI infrastructure margin reset, and that argument has evidence: Hood said capital expenditures were $37.5 billion in the quarter, with “roughly two-thirds” on short-lived assets, primarily GPUs and CPUs. But the variant read is that this is not capex without utilization; cash flow from operations was $35.8 billion and up 60%, and Microsoft Cloud revenue was $51.5 billion and grew 26% in constant currency. If short-lived asset spend were running ahead of demand, the first place it would show up is weaker cloud growth or weak collections, yet the company printed both $51.5 billion of Microsoft Cloud revenue and $35.8 billion of operating cash flow.

That distinction between capex burden and demand absorption is why the Q3 guide is more important than the Q2 beat. Hood guided revenue of $80.65 to $81.75 billion, or growth of 15 to 17%, which at first glance sits below Q2’s $81.27 billion midpoint optics and invites a “beat but guide not enough” reaction. The sharper read is segment mix: productivity and business processes is guided to $34.25 to $34.55 billion, intelligent cloud to $34.1 to $34.4 billion, and more personal computing to $12.3 to $12.8 billion. The semiconductor relevance is in Azure, where management expects Q3 revenue growth between 37-38% in constant currency. That figure is the crux: if Azure can sustain 37-38% while COGS are guided to $26.65 to $26.85 billion and operating expense to $17.8 to $17.9 billion, the AI infrastructure cycle is still supply-constrained, not demand-disappointed. If Azure slips materially below that 37-38% range, the same capex becomes the wrong kind of operating leverage.

The product-level evidence supports the supply-constrained interpretation because the AI usage metrics are unusually broad, not just one workload or one customer class. Nadella said the Microsoft Cloud surpassed $50 billion in revenue for the first time, up 26% year over year, and the narrower AI platform signals were also concrete: Fabric’s annual revenue run rate is now over $2 billion, with over 31,000 customers and revenue up 60% year over year; the number of customers spending $1 million plus per quarter on Foundry grew nearly 80%; and over 250 customers are on track to process over 1 trillion tokens on Foundry this year. Those are not interchangeable with generic enterprise software metrics. For semiconductor investors, “over 250 customers” on a path to “over 1 trillion tokens” is a workload-intensity signal, while “nearly 80%” growth in $1 million-plus Foundry customers is a monetization signal. The two together argue that Microsoft is filling expensive AI clusters with paid enterprise demand rather than subsidizing empty inference capacity to protect strategic positioning.

The enterprise software layer also matters because it creates an internal funding source for the hardware cycle. Productivity and business processes revenue was $34.1 billion and grew 16% in constant currency, while paid Microsoft 365 commercial seats grew 6% year over year to over 450 million. Copilot Pro Plus subscriptions for individual developers increased 77% quarter over quarter, and paid Copilot subscribers reached 4.7 million, up 75% year over year. Security is becoming another monetization surface around AI, with 24 billion Copilot interactions audited by Purview in the quarter, up 9x year over year. These numbers do not prove that every Copilot SKU is margin-accretive today, and they do not need to. They show that Microsoft has multiple places to attach AI consumption to existing commercial relationships, which helps explain why the company can spend $37.5 billion on capex in one quarter while still returning $12.7 billion to shareholders through dividends and share repurchases.

The semiconductor supply-chain read-through is therefore direct: Microsoft’s print supports continued demand for the AI compute stack, but it also points to mix moving beyond merchant GPU boards into custom silicon, racks, and liquid-cooled systems. TSMC is tied to 5nm custom AI chip fabrication for Maia and Cobalt, so Microsoft’s $37.5 billion capex and “roughly two-thirds” allocation to short-lived GPUs and CPUs imply sustained advanced-node and advanced-packaging pressure where custom accelerators can reduce dependence on external GPU supply. Marvell is named as the silicon design partner for Maia ASICs, making Foundry customer growth of nearly 80% in $1 million-plus quarterly spend relevant to custom silicon attach rather than only cloud software revenue. Quanta Computer, Wiwynn Corporation, and Auras Technology are the rack and liquid-cooling read-throughs, because the data pack ties them to integrated rack-scale AI/cloud server hardware, finished hyperscale data-center racks, cold plates, CDUs, and quick-connectors for Microsoft self-developed and GB200/GB300 hyperscale liquid-cooling rack programs. Nuvoton Technology has a narrower but still concrete exposure through server SCM and BMC management chips for Microsoft self-ASIC AI server platforms. The magnitude is not a vague “AI demand” call; it is Microsoft buying $37.5 billion of capex in a quarter while guiding intelligent cloud to $34.1 to $34.4 billion and Azure to 37-38% constant-currency growth.

The peer comparison reinforces why Microsoft’s spend should not be treated like any other hyperscaler’s spend. In the latest peer set, MSFT shows $82.89 billion of revenue, 67.6% gross margin, and +18.3% revenue YoY. That growth is close to AMZN’s +16.6% and AAPL’s +16.6%, below GOOGL’s +21.8%, and far below NVDA’s +85.2%, but Microsoft’s 67.6% gross margin is materially above AMZN’s 51.8%, AAPL’s 49.3%, and GOOGL’s 62.4%, while below NVDA’s 74.9% and META’s 81.9%. The investable implication is that Microsoft is one of the few AI capex spenders whose software gross margin base can partly cushion hardware depreciation, while NVDA remains the purer monetizer of the same cycle. For semiconductor PMs, that means Microsoft’s gross margin pressure is a demand signal for suppliers, not necessarily a warning that AI economics are broken across the chain.

The call delivery slightly complicates the bull case, and that complication is worth taking seriously because the numbers moved against an unqualified enthusiasm read. In the tone history, Q2 FY2026 sentiment was 0.24 versus 0.26 in Q1 FY2026 and 0.28 in Q4 FY2025, while guidance_tone rose to 0.51 from 0.42 in Q1 FY2026. The positive part is that ai_optimism increased to 0.36 from 0.29, consistent with the product data around Foundry, Fabric, Copilot, and Purview. The negative part is that tone_confidence was 0.33, below 0.42 in Q1 FY2026 and 0.44 in Q4 FY2025, while uncertainty was 54.8. That is not a reason to fade the print, but it is a reason to define the thesis around utilization evidence rather than management enthusiasm. The call sounded more positive on guidance than the prior quarter, but not more confident.

The RPO language is the bridge between management tone and revenue durability, because it gives a backlog anchor without pretending all of it converts immediately. Hood highlighted the breadth of future demand with an unusually pointed comment: “55% or roughly $350 billion is related to the breadth of our portfolio, a breadth of customers, across solutions, across Azure, across industries, across geographies.” The phrase matters because it commits management to breadth, not just one mega-customer or one AI training cycle. Still, this is where the hedge belongs: a large RPO balance supports durability, but the same quarter also guides COGS growth of 22 to 23% against revenue growth of 15 to 17%, so the next debate is not whether Microsoft can sell AI capacity, it is whether the cost curve improves fast enough. The conflicting numbers are explicit: Azure is guided to 37-38% in constant currency, but gross margin has moved to 68.0% in Q2 FY2026 and 67.6% in Q3 FY2026 history, while COGS is guided to $26.65 to $26.85 billion. Demand is winning, but the margin bill is still arriving.

What to watch next is therefore narrow and actionable. The thesis is confirmed if Q3 revenue lands within or above the $80.65 to $81.75 billion guide, intelligent cloud lands within or above $34.1 to $34.4 billion, and Azure holds the 37-38% constant-currency growth range while gross margin does not deteriorate beyond the 67.6% Q3 FY2026 history line. It is further confirmed if Microsoft Cloud stays near or above the $51.5 billion base and capex remains interpretable as utilization-driven rather than anticipatory, with operating cash flow not breaking from the $35.8 billion level that was up 60%. The thesis breaks if Azure misses the 37-38% range, if COGS comes in worse than the $26.65 to $26.85 billion guide without offsetting cloud growth, or if tone deteriorates from Q2 FY2026’s 0.51 guidance_tone and 0.36 ai_optimism into the Q3 FY2026 tone pattern of 0.34 guidance_tone and 80.8 uncertainty. For semiconductor exposure, the next quarter’s read is simple: if those levels hold, Microsoft remains a buyer that can keep pulling through TSMC, Marvell, Quanta Computer, Auras Technology, Nuvoton Technology, and Wiwynn Corporation; if they fail, the market will start treating $37.5 billion of quarterly capex as a digestion problem rather than a capacity shortage.