Mapping the AI Stack: Where Demand Is Real vs. Where Money Is Speculative 

The AI sector is under correction, and the bubble question is back. But asking 'is AI a bubble' compresses too many distinct layers into one verdict. The real work is mapping where demand is already locked in, and where capital is still speculative. That distinction matters because the companies closest to actual usage aren't pulling back. According to public earnings reports and filings, OpenAI's annualized revenue grew from roughly $2 billion in 2023 to over $20 billion in 2025. Anthropic reported even faster growth. Nvidia disclosed that its data center business reached nearly $194 billion in fiscal 2026. The hyperscalers (Google, Microsoft, Amazon, Meta) are collectively on pace to spend around $700 billion this year on AI infrastructure, and they're still talking about capacity constraints, not demand problems. A stock correction tells you investors think prices are stretched. It doesn't tell you the demand is fake.

What 'capacity constrained' actually means

When Microsoft CEO Satya Nadella told investors on the company's Q3 2025 earnings call that Microsoft will spend $190 billion on capex in calendar 2025 and still expects to be capacity constrained through year end, that wasn't a GPU shortage story. The constraint is a layer below: it's whether manufacturers can package enough chips with the high-bandwidth memory (HBM) that modern AI workloads require. The supply problem isn't logic chips. It's the part of the stack almost nobody in procurement is fluent in yet.

Capacity constraints at this scale change the shape of AI vendor contracts. Six months ago, an AI vendor contract looked like a software agreement. Now it's effectively a supply contract tied to hyperscaler allocation. It should have capacity terms, fallback provisions, and line items that didn't exist a few quarters ago.

Which parts of the build-out are speculative versus locked in?

The lazy version of the bubble argument treats inflated stock prices, aggressive private valuations, overbuilt data centers, weak enterprise pilots, Nvidia's revenue, and OpenAI's growth as if they're all the same question. They're not. You can have a correction in AI stocks and still have tremendous unmet demand for production inference workloads (running trained models at scale). You can have some companies overbuild capacity and still have the world be dramatically underbuilt for the inference layer that supports real applications.

The segments with locked-in demand share a pattern: they're automating tasks with clear, measurable labor substitution. Coding assistants that reduce pull-request cycle time. Support chatbots that deflect tier-one tickets. Research agents that compress literature review from days to hours. These aren't aspirational use cases. They're production workloads with revenue attached.

The speculative layer is generic enterprise pilots with no clear ROI path. Companies running 'AI innovation sprints' with no defined success metric. Proof-of-concept projects that never graduate to production because the per-inference cost (what it costs each time the model processes a query) isn't justified by the value created.

What this means for positioning and procurement

The companies winning enterprise deals right now are the ones that can articulate a narrow, defensible use case with transparent cost structure and measurable output. They're not selling 'AI transformation.' They're selling a specific workflow improvement with a clear before-and-after. When GitHub Copilot launched, it didn't promise to 'revolutionize development', it promised to reduce the time developers spend writing boilerplate code, with metrics tied to pull request velocity and code completion acceptance rates. That specificity converts.

If you're evaluating an AI vendor or use case, the filtering questions have shifted. Does this use case have measurable labor substitution, or is it vibes-based innovation theater? Does the vendor have guaranteed capacity allocation from their hyperscaler partner, or are they reselling spot access that could get rationed? What happens to service levels if the infrastructure provider they depend on restricts capacity further? Can they articulate per-inference cost, and does that cost structure hold at scale, or does it blow out under production load?

These aren't typical software procurement questions. They're supply chain questions, because AI infrastructure is now industrial in a way SaaS never was. The companies that understand that, and structure their vendor conversations accordingly, are the ones that will scale without hitting invisible capacity walls six months into deployment.

Sources

• OpenAI (2025) Annualized revenue grew from roughly $2 billion (2023) to over $20 billion (2025). Referenced in source material; original disclosure not independently reviewed by Markedeen.
• Nvidia (2026) Data center business revenue reached nearly $194 billion in fiscal 2026. Referenced in source material; original filing not independently reviewed by Markedeen.
• Microsoft (2026) Q3 FY2026 earnings call: CEO Satya Nadella stated the company would spend $190 billion on capital expenditures and still expects to be capacity constrained through year end. Referenced in source material; original earnings call transcript not independently reviewed by Markedeen.
• Source material (2026) Estimate of combined 2026 AI infrastructure spend by Google, Microsoft, Amazon, and Meta at roughly $700 billion; not attributed to a specific named report in the source.