Why Microsoft’s $190 billion AI bet may be the riskiest wager in big tech
Microsoft is guiding to roughly $190 billion of capital spending this fiscal year. Its AI business earns an annual run-rate of about $37 billion. That gap — spending on the order of five times what the business currently brings in — is the whole story. It’s also why I think Microsoft has made the most dangerous bet of the megacaps.
This is the long version of a call I’ve made publicly since January 2025 (the short version: Microsoft and Apple used to move together; I bet they’d diverge as the market saw through Microsoft’s AI spend — and they have). I scored that call against Apple in last week’s flagship piece; this is the standalone case against Microsoft. Here’s the full thesis.
The bet is bigger than the business
The tell isn’t Microsoft’s capex versus the whole company — it’s how much it’s spending on AI relative to what AI actually earns today. ~$190B of capital spending (the vast majority AI chips, servers, and datacenters) against a ~$37B AI revenue run-rate. Even granting that AI revenue is growing fast (up ~123% year over year), the company is betting well ahead of the revenue. That turns AI into an existential bet: win big, or destroy a lot of shareholder capital. There isn’t much middle ground at this scale.
It’s aimed at the wrong layer of the stack
Look at how the other giants positioned. Google built its own frontier model (Gemini) and its own chips (TPUs); Amazon built custom silicon (Trainium); Nvidia owns the picks-and-shovels. Each secured a differentiated layer.
Microsoft, leaning on OpenAI, effectively bet it wouldn’t need to own the model or the silicon — so it concentrated capital in the layer it could scale fastest: datacenter capacity. It has gestured at the other layers — in-house Maia chips, an MAI model team — but neither has reached the scale or quality to matter. Its next-gen Maia chip slipped from 2025 into 2026 and, by most reports, won’t rival Nvidia’s Blackwell (design changes — some requested by OpenAI — and staffing problems pushed it back). So in practice Microsoft still runs on Nvidia’s silicon (it remains one of Nvidia’s largest customers) and on OpenAI’s models.
That leaves it concentrated in the commoditizing layer. Compute is becoming a utility, and utilities compete on price — so as AI-hardware costs climb, Microsoft is defending margins in the part of the stack with the least pricing power, while the differentiated layers sit in others’ hands.
Copilot isn’t converting
Microsoft’s flagship AI product compounds the problem: heavy capex, weak pull-through. Only about 3.3% of its ~450 million commercial Office subscribers pay for the Copilot add-on, and of the seats that are provisioned, only ~36% are actively used (versus ~83% for ChatGPT). Asked to choose between Copilot, ChatGPT, and Gemini, one enterprise survey found 76% made ChatGPT their primary tool — and just 18% chose Copilot. A bet that leans on owning the customer relationship looks shaky when, given the choice, users keep picking someone else.
The keystone is becoming a competitor
Three or four years ago, OpenAI looked almost like a Microsoft subsidiary — and that closeness is exactly why Microsoft felt no urgency to build its own models or silicon. That alignment has unwound. In January 2025, Microsoft lost its status as OpenAI’s exclusive cloud provider (it now holds only a right of first refusal). OpenAI has since committed hundreds of billions to compute elsewhere — the ~$300B Oracle “Stargate” arrangement and a ~$38B AWS deal among them — restructured into a for-profit, and is reportedly preparing a 2026 IPO at a ~$850B valuation. It’s also pushing into enterprise products (ChatGPT Enterprise, agents) that compete head-on with Microsoft’s own. The partner Microsoft built its strategy around is becoming a rival.
The circular AI revenue problem
Here’s a question that goes underexamined when people talk about Microsoft’s AI growth: when Microsoft reports a $37 billion AI revenue run-rate, who is actually paying that bill?
The largest contributor by a long stretch is OpenAI itself. Microsoft holds a 27% equity stake** in OpenAI and booked a **$7.6 billion gain from that investment in Q2 FY26 alone. In return, OpenAI has committed to purchase roughly $250 billion in incremental Azure services. As of latest disclosure, OpenAI represents about 45% of Azure’s ~$625 billion in Remaining Performance Obligations. Forty-five percent of Azure’s contracted future cloud revenue is one customer.
That customer’s economics tell their own story. OpenAI’s annual cloud spend has grown past $60 billion**. Its reported revenue is closer to **$25 billion. The gap — well over $30 billion a year — is filled by recycled investment capital flowing in from Microsoft, Nvidia, and others, then flowing back out as compute purchases. The same loop runs across the rest of the AI economy: Nvidia announced a strategic investment of up to $100 billion in OpenAI tied to OpenAI deploying 10 gigawatts of Nvidia-powered datacenter capacity. Amazon’s $5 billion investment in Anthropic came with a $100 billion Anthropic commitment to spend on AWS Trainium compute. Bloomberg, Built In, and a growing chorus of independent analysts now call this the “circular economy of AI” — the same capital cycling between investors and customers that are largely the same entities.
This isn’t fraud, and I want to be careful with the framing. Real GPUs are being deployed. Real services are being delivered. The financial reporting is technically accurate, and regulators specifically distinguish circular deals from sham “round-trip” transactions. But it does materially weaken the interpretation of “AI revenue growth.” Microsoft’s $37B run-rate isn’t $37B of external customer demand competing for compute. A meaningful portion of it is Microsoft’s own investment capital, recycled through OpenAI, showing back up on the income statement.
For the bull thesis on Microsoft, this matters in two specific ways:
First — concentration risk is enormous. When 45% of Azure’s contracted future revenue is one customer, and that customer has already started diversifying to Oracle and AWS at staggering scale, the future-revenue picture isn’t as locked-in as the headline run-rate implies. “OpenAI was Azure’s biggest customer” is a great line in a bull deck. It looks very different described as “OpenAI was 45% of Azure’s RPO and is now moving most of its incremental compute elsewhere.”
Second — the growth rate is partly endogenous. As long as Microsoft and others keep investing in OpenAI, OpenAI keeps buying Azure capacity. Stop the investment flow — or have OpenAI IPO into a future where it’s funded by external investors rather than its compute providers — and the dynamics change materially. “AI revenue is up 123% year-over-year” is impressive. But how much of that growth is being driven by capital recycling versus independent customer adoption?
I don’t think this is a fatal flaw in the Microsoft story. But it does mean the $37B run-rate is a softer floor than it looks, and it makes the $190B capex bet harder to defend on cash returns alone. It also means the bull-case calculation — “AI revenue compounds at high rates and eventually justifies the capex” — needs to assume that external customer demand grows faster than the capital-recycling flows can sustain. That’s a much harder bet than the headline numbers suggest.
So the “single existential bet” isn’t just large — it’s concentrated in the weakest part of the value chain, dependent on a partner-turned-rival, supported by revenue that’s partly self-funded, and exposed on multiple fronts. The market has noticed: the stock is down ~24% from its late-2025 high, its worst quarter since 2008.
What would make me believe again
I’m not a permabear — and I own the stock. Here’s the concrete scorecard. Real progress on these would move me from skeptic toward believer:
- Proof the capex is earning its keep — not just AI revenue growth, but AI profitability; Azure AI margins holding or expanding as it scales. The $190B has to visibly earn its cost of capital.
- The free-cash-flow bleed stops — capex moderating, or revenue catching up enough that free cash flow turns back up instead of shrinking.
- Copilot inflects — rising paid-attach, climbing active usage, and signs it’s winning engagement rather than losing it to ChatGPT.
- A differentiated rung of the stack — Maia silicon deployed competitively at scale, or an in-house model that genuinely rivals the frontier, cutting the dependence on Nvidia and OpenAI.
- The OpenAI overhang resolved — a durable, clearly favorable arrangement, or Microsoft demonstrably standing on its own.
- Capital discipline returns — guidance that spending is matched to returns, a signal Microsoft will spend because the ROI is there, not to keep up.
None of this is out of reach. Microsoft is a great company with extraordinary assets, and if it strings several of these together, I’ll change my mind in public — the same way I made the original call in public. Until then, the risk/reward points the other way.
Four audacious paths — speculative, not predictions
None of these is happening based on public signals. I include them because each maps directly to a structural problem identified above, and thinking through them clarifies what would actually fix the bear case — and the boldest of them would change my view fast.
1. The Anthropic merger of equals
The single move that would most cleanly invert this entire bear thesis. Microsoft needs the differentiated model layer it doesn’t own. Anthropic needs unlimited compute that only a hyperscaler can provide. The numbers fit unnervingly well. Anthropic is reportedly heading toward a ~$1 trillion IPO. Microsoft has lost roughly that much in market value from its $538 ATH. A merger of equals — Microsoft trading a meaningful stake in the combined entity for Anthropic’s model franchise, its talent, and the Claude product line — instantly fixes the central critique of this whole piece: the lack of differentiated ownership at the model layer of the stack.
The combined entity would own the Claude family alongside Azure’s infrastructure, with a unified roadmap that Microsoft has been trying to construct through partnerships and failing. It would create a credible frontier-model competitor to OpenAI without depending on OpenAI — replacing a slow-motion divorce with a deeply aligned partner. Microsoft and Anthropic already have a meaningful working relationship (Anthropic models on Azure, joint enterprise pursuits); a deeper combination is a smaller cultural and operational leap than it sounds. Decades-defining capital allocation moves look obvious in retrospect — Disney–Pixar, Google–Android, Facebook–Instagram. An MSFT–Anthropic combination at this scale and this moment could be the AI-era equivalent. It would also remove most of the reasons to stay skeptical in my scorecard above.
2. The friendly Salesforce takeover
There is a piece of Microsoft folklore — possibly apocryphal but widely told — that Microsoft considered acquiring Salesforce around $40 billion in the mid-2010s and walked away because the price felt expensive. Today Salesforce trades at roughly $155 billion, owns Slack, Tableau, MuleSoft, and Informatica, and dominates the enterprise CRM and data-platform layer where Microsoft Dynamics never caught up. A friendly takeover today would consolidate two of the three most important enterprise software franchises into one company, with massive cross-sell into Microsoft’s existing 450M+ commercial seat base.
The honest constraint: antitrust reality in 2026 makes this nearly impossible to land. Microsoft barely got Activision through. The FTC and DOJ would treat a $200B+ MSFT–Salesforce combination as an existential market-structure event. The window for this was 2014. Whether or not Microsoft truly walked at $40B back then, that decision now costs the company well north of $100B in lost optionality. The lesson generalizes: capital discipline applied at the wrong moment is the most expensive kind of discipline.
3. Spin off LinkedIn
LinkedIn was acquired in 2016 for $26B; it now generates $15B+ in annual revenue and would credibly trade as a standalone at $150-200B. It has almost nothing to do with the AI thesis, the Azure thesis, or the enterprise productivity thesis at the company’s core. The conglomerate-discount argument is straightforward: spin it off, return capital to shareholders, let LinkedIn operate at its own velocity without Microsoft bureaucracy, and refocus the parent company on what actually matters in the AI era.
4. Sell or spin off Xbox
The Activision acquisition closed in 2023 at ~$69B and was, by any honest read, late-stage diversification rather than strategic focus. Gaming is a great business; it isn’t Microsoft’s core business anymore. Selling or spinning Xbox + Activision would surface meaningful capital, signal disciplined refocusing, and let a buyer (or an independent entity) operate gaming with the dedicated attention it deserves. It would also send a clear message to the market: “We’re not trying to be a conglomerate. We’re trying to win the AI era.”
I am not predicting any of these. But the existence of this list is itself an argument: when a company of Microsoft’s caliber has this many available moves while the market is pricing in continued underperformance, the question becomes whether management has the imagination and the conviction to act. History suggests the audacious move tends to come from a CEO who knows the narrative is shifting — not from one who’s certain it isn’t.
Full disclosure: I own Microsoft — a large, legacy position from my years working there (2013–2023), held mostly because the tax bill to sell would be brutal. Nothing here draws on inside knowledge; it’s entirely from public filings and reporting. If anything, I’d rather be proven wrong. I compare Microsoft directly against the company I think wins this era — Apple — in the companion piece, Head to Head: Apple vs. Microsoft.
Seek Value Now is published for educational and informational purposes only and is not investment advice. I may hold positions in the securities discussed. Do your own research before investing. — HG
Social distribution drafts (Notes series, Reddit outlines, X/Bluesky threads) live in this folder’s social.md.