The retail analyst writes "Microsoft: Buy or Sell?" and then fills 2,000 words with P/E multiples and consensus estimates. That analysis misses the single mechanism that matters.
Microsoft stock has fallen from its $555 all-time high in July 2025 to around $460 as of early June 2026 - a 17% drawdown that most investors dismiss as "AI sentiment cooling." It isn't. The market is pricing in something harder: the capex/revenue math is broken, and the custom silicon escape hatch may not exist.
Here's why.
The $130B Question No One Is Asking
Microsoft reported Q3 FY2026 revenue of $82.9 billion, up 18% year-over-year. Azure grew 40%. The AI revenue run rate hit $37 billion annually. Wall Street cheered. Consensus has 32 analysts still on Buy.
Then look at capital expenditures.
Q3 capex was $31.9 billion, up 49% year-over-year. Q4 guidance puts capex above $40 billion. Consensus projections put FY2027 total capex at approximately $130 billion - nearly double the $65 billion spent in FY2025. That number is accelerating, not plateauing.
Here's the arithmetic nobody wants to do: Microsoft is projecting $130 billion in annual capex to support an AI revenue run rate of $37 billion. Every dollar of AI revenue requires roughly $3.50 of capex investment. Not $1.50, not $2.00 - $3.50. That is not a business model. That is a venture-style bet dressed as cloud infrastructure.
Microsoft does not break out AI-specific capex from total infrastructure capex, but given that Azure's AI services grew 123% year-over-year in Q3, the marginal capex is overwhelmingly AI-directed. General cloud infrastructure doesn't require this kind of spending acceleration. AI does.
The implication: unless AI revenue growth consistently outpaces capex growth - and it hasn't yet - each additional dollar of spending generates less return than the last. Diminishing capital efficiency on a $130 billion base is how you destroy shareholder value while posting headline growth.
The Maia Problem
Microsoft has been selling the custom silicon narrative since November 2023, when it unveiled Maia 100 alongside Athena and Cobalt chips. The pitch was clear: build inference ASICs (application-specific integrated circuits - chips designed for one specific workload) to reduce reliance on NVIDIA, improve inference economics, and capture the margin that merchant GPU vendors otherwise take.
Maia 200 was the promised evolution. Microsoft announced it on January 26, 2026, claiming it is "three times more performant" for inference, built on TSMC's 3nm process, and uses a redesigned on-chip SRAM (static random-access memory) architecture to reduce dependence on expensive HBM (high-bandwidth memory).
The problem: SemiAnalysis reported on January 7, 2026 - three weeks before the announcement - that Maia 200 has been internally assessed as a failure, forcing Microsoft back to the drawing board. The public launch was a PR recovery play, not a product milestone. Independent semiconductor watchers including Dylan Patel have since noted that Microsoft's ambition to move Maia onto advanced process nodes (18A) has also been abandoned.
This matters because custom silicon was supposed to be the margin lever that makes the capex math work. Without it, Microsoft is stuck buying NVIDIA GPUs at full merchant pricing and absorbing the entire infrastructure cost on its own balance sheet. Amazon faces the same problem with Trainium. Google has managed the TPUs better, but even Google's custom silicon doesn't solve the aggregate capex question.
When the custom chip thesis collapses, the capex thesis collapses with it.
What The Azure 40% Number Actually Means
Azure's 40% growth is real. No one is disputing the top-line. But growth is not the same as capital efficiency - and in infrastructure businesses, capital efficiency is what separates compounders from value traps.
The framework is simple: for every dollar of capex, how many dollars of incremental revenue does it generate? In Microsoft's best years as a software company, the answer was multiples - software margins approach 70-80%, and incremental capex was minimal. In an AI infrastructure business where capex is growing at 49% and AI revenue at 123%, the ratio improves - but only until the spending curve outpaces adoption, and that's where Microsoft is heading.
Q4 guidance of $86.7–87.8 billion in revenue implies 13–15% growth. Azure is guided at 39–40% constant currency. Meanwhile capex is guided above $40 billion for the quarter alone. Revenue growth is decelerating. Capex growth is accelerating. When those two curves cross, the market re-rates.
The Copilot Pivot Doesn't Fix The Math
Microsoft Copilot generated an estimated $4.8 billion in 2026 revenue through enterprise subscriptions, Azure AI usage, and GitHub Copilot. That's respectable - but it's also the number that changes how it's sold. GitHub Copilot is shifting to usage-based billing on June 1, 2026, replacing flat subscriptions with pay-per-token pricing.
The market reads this as a confidence play. It may be the opposite: when you move to usage-based billing, you're admitting that fixed pricing was masking the real unit economics. If customers actually measured how much inference compute each Copilot interaction consumed, they'd find the margin is razor-thin at current NVIDIA GPU prices. Usage-based billing lets Microsoft pass the infrastructure cost to the customer - which is what you do when your own TCO (total cost of ownership) doesn't support the old pricing model.
The Cross-Currents
The cross-currents are:
- Azure 40% growth is real - Microsoft's enterprise moat and cloud positioning remain the strongest of any hyperscaler. This is not a fake number.
- Capex trajectory is unsustainable - $130 billion in annual spending requires AI revenue to scale to $100 billion+ to justify. That's nearly 3x the current run rate. No one has a credible path there in the next 18 months.
- Custom silicon delay destroys the margin thesis - If Maia 200 is genuinely in trouble, Microsoft's inference economics improve only as fast as NVIDIA's pricing generosity, which is to say: not at all.
- Stock has already fallen ~17% - The decline from $555 to $460 represents the first pricing of this risk. The question is whether it's enough.
Directionally, the capex/revenue gap is the dominant force. It hasn't been resolved. It's widening. The Maia setback means Microsoft's internal margin lever is broken. Azure growth is too real to ignore but too slow to justify the spending.
What It Means For The Stock
Microsoft is not a buy at $460 unless you believe AI infrastructure economics will improve faster than the company's own execution suggests. The Azure growth story is genuine. The capital efficiency story is not.
The break condition is simple: if Microsoft can show Azure AI revenue growing at or above capex growth for two consecutive quarters, the thesis repairs itself. If capex continues to outpace revenue for the rest of FY2026 and into FY2027, the stock has further to fall.
Right now, the evidence points in only one direction. You decide whether the market has priced enough in, or whether the $555-to-$460 move is just the beginning of what happens when a $130 billion annual spend meets a $37 billion revenue reality.
It is not as good as it looks.