The thesis: sell the shovels
Green's central analogy is the California Gold Rush. In his telling, the speculators often went broke while the suppliers, Levi Strauss and Wells Fargo among them, "made bank." Applied to AI, the Big Four infrastructure builders are Microsoft, Amazon, Alphabet and Meta. They rent compute and lock customers into ecosystems, and they do so while the model makers burn cash.
He argues that OpenAI is the most exposed of the American players because it "rents everything but the horse." He says Microsoft Azure, chiefly as the GPT compute supplier, will profit before OpenAI does. He also notes that Microsoft's Copilot competes with OpenAI's models.
That's a sharp framing, and it fits how the cloud business works. Still, it is one columnist's view. Green says openly that his best friend's complaints about OpenAI may bias him. He also hedges on China versus the U.S. with a plain "Dunno."
Checking the numbers
The essay mixes several kinds of spending figures: corporate capex, AI-specific capex, and forecasts. They aren't interchangeable, so here is what each one supports.
- Past versus projected spending. The essay says the Big Four spent roughly $1.6 trillion since 2023, about half of it in 2026. I couldn't verify that aggregate against the sources I checked, so treat it as the author's estimate.
- Goldman's forecasts. Coverage of Goldman Sachs' research puts roughly $7.6 trillion of AI infrastructure investment from 2026 through 2031, covering compute, data centers, and power. Goldman's own "Tracking Trillions" piece says its baseline model implies $765 billion in annual AI capex in 2026, growing to $1.6 trillion in 2031.
- Hyperscaler spending. Reporting on Goldman's note says hyperscaler capex will reach $1.2 trillion in 2027 and $1.4 trillion in 2028. The essay's "$140 billion in 2023 to around $800 billion this year" is close to that coverage, which says the five largest hyperscalers went from roughly $150 billion of annual spending in 2023 to an estimated $800 billion in 2026. The essay's 2023 base is slightly lower, and it describes the group as the "Big Four" where the coverage counts five.
- A forecast, not a ledger. The $7.6 trillion is a projection with stated assumptions. Goldman itself says the figure is highly sensitive to a small set of assumptions about how the infrastructure itself is built and renewed.
- Defense comparison. The essay compares the AI figure with roughly $7 trillion of defense spending over the same five years. That is the author's comparison, and I did not verify it.
Microsoft's side of the ledger
Microsoft's own filings give the cleanest view of the "sell the compute" strategy.
- Capex. On the fiscal 2026 third-quarter call, CFO Amy Hood said Microsoft expected to invest roughly $190 billion in capital expenditures in calendar 2026. The figure includes about $25 billion from higher component pricing. These are management expectations, not realized returns.
- Capacity. On that same call Microsoft said customer demand continues to exceed available capacity, and it expected to stay constrained at least through 2026.
- Latest quarter. The July 29, 2026 release for the fiscal fourth quarter reported revenue of $90.0 billion, up 18%, and operating income of $40.6 billion, up 18%. Azure and other cloud services revenue increased 43%. CEO Satya Nadella said Azure revenue surpassed $100 billion for the first time and Microsoft 365 Copilot reached over 30 million paid seats.
- Margin pressure. Earlier in the year, Microsoft said the Intelligent Cloud gross margin percentage decreased, driven by continued investments in AI infrastructure and sales mix shift to Azure, offset in part by Azure efficiency gains.
- Backlog. Microsoft's commercial remaining performance obligation rose 84% to $678 billion in the fourth quarter. The company's third-quarter commentary noted that a portion of the backlog includes OpenAI.
The numbers show Microsoft growing quickly while spending heavily and carrying margin pressure. They don't show how much of the revenue or profit comes from AI specifically. Microsoft's own disclosures also show it holds investments in OpenAI. The release notes a $3.2 billion gain from our investment in Anthropic as well. Microsoft is a supplier, an investor and a competitor to the model makers at once. That makes the simple "supplier wins" story less tidy than the essay suggests.
Amazon and the other builders
Green points to Amazon's most recent results, with AWS revenue up 37% and operating profit up 64%, as proof that renting compute works. I couldn't retrieve Amazon's release to confirm those numbers. They do fit the essay's broader point that cloud providers can sell infrastructure to many AI customers. They don't isolate how much of that growth comes from AI.
His sketches of Alphabet and Meta are interpretive. He argues that Gemini mainly serves to strengthen Google's search and advertising business, and that Meta's AI mainly improves ad targeting. Neither claim is independently verified here.
The SpaceX wildcard
The essay's boldest claim is that SpaceX's planned orbital compute, which it calls Starmind, could turn the Big Four into the Big Five. The plan, as Green describes it:
- Fabricate AI chips at the Terafab facility in Grimes County, Texas.
- Build solar panels at a SpaceX plant in Bastrop, Texas.
- Assemble satellites at the Gigasat facility, also in Bastrop.
- Launch them in batches on Starship.
My research packet pointed to a SpaceX update on Terafab dated August 6, 2026, but I couldn't retrieve the page. Treat the plan as an announced project, not proof that the plant is producing chips or that orbital compute is cheaper.
The "free solar power, no water, no local authorities" pitch is Green's summary of a business case. It ignores the costs the essay itself mentions: solar cells, radiators and laser links. It also skips launch costs, maintenance, hardware replacement and regulatory exposure. Musk's quoted comparison says an AI satellite is easier to design than a Starlink satellite. That is a claim about design simplicity, not a total-cost comparison.
What this means for Windows and IT readers
- Capacity is the constraint. Microsoft says demand outruns supply through 2026. Expect continued pricing pressure and possible regional or service availability limits for GPU-heavy Azure workloads.
- Costs are moving to consumption. Microsoft's third-quarter commentary noted weaker Dynamics 365 bookings as customers balance per-seat pricing against an emerging seats-plus-consumption model. IT budgeters should model AI costs as variable usage, not only license counts.
- Vendor concentration matters. The essay's lock-in argument applies to enterprise buyers too. Tying workflows to a single cloud's AI stack is easier on day one and harder to unwind later.
- Copilot adoption is real but only partly measured. Over 30 million paid Microsoft 365 Copilot seats is a meaningful count. It doesn't tell you how much those seats are used or what return they deliver.
- The bubble question is open. Green leans toward "shakeout, not crash." Goldman's own framing is more cautious: the debate is whether adoption will justify the spend. Reports on Goldman's work also flag power as a major constraint.
Bottom line
Nobody is declared the winner here, and the evidence doesn't support naming one. Microsoft's reported numbers back the essay's claim that infrastructure sellers are well placed right now. The proof of a payoff from the $7.6 trillion forecast build-out is still missing. Green's most speculative claims, about OpenAI's economics, China's strategy and orbital data centers, remain opinion.