Verdict: The "AI bubble" framing is too blunt. Stock valuations have stretched and some projects will fail, but the demand is real: OpenAI is past a $20 billion run rate, Nvidia's data center revenue hit roughly $194 billion, and the largest tech companies are still expanding AI capex toward $700 billion this year. The real question is not whether AI is fake — it is which layer of the buildout gets paid back.
Last verified: 2026-06-17 · Core signal: paid enterprise AI revenue is growing, not shrinking · Risk: many companies will build the wrong capacity at the wrong time
Why "bubble" is the wrong first question
A stock correction tells you that investors think prices are stretched. It does not automatically mean the technology has no users. The two are related, but they are not the same.
The lazy version of the bubble argument compresses too much into one word:
- Inflated stock prices.
- Aggressive private valuations.
- Overbuilt data centers.
- Weak enterprise ROI in some pilots.
- Nvidia's revenue growth.
- OpenAI's revenue growth.
- The future of AI as a platform.
These are not the same question. A bubble detached from reality looks very different from a real buildout with speculative money piled on top.
History helps. Railroads were real infrastructure — and many railroad investors still got destroyed. Fiber-optic networks were transformative — and a lot of telecom investors lost money. Cloud computing was real — but not every cloud-adjacent stock captured value. The dot-com bubble did not prove the internet was fake; it proved markets can overprice the first version of a real platform shift.
AI looks more like those cases than like a pure South Sea Bubble of hype. There is real revenue, real chips, real power constraints, and real enterprise usage.
The demand signals are not theoretical
The most direct evidence of paid demand is the revenue at the companies closest to it.
| Company | Signal | Source | Date |
|---|---|---|---|
| OpenAI | Annualized revenue crossed $20B in 2025, up from ~$6B in 2024 | OpenAI / Sarah Friar, Reuters | Jan 2026 |
| OpenAI | Enterprise now ~40% of revenue; 1M+ business customers, 7M+ workplace seats | OpenAI company blog | Nov 2025 |
| Anthropic | Run-rate revenue reached ~$5B by Aug 2025; enterprise customers up sharply | Anthropic Series F announcement | Sep 2025 |
| Nvidia | Data Center revenue ~$193.7B in fiscal 2026 | NVIDIA Q1 FY2026 earnings, sector analysis | 2026 |
| Hyperscalers | Combined 2026 AI capex approaching ~$700B across Alphabet, Microsoft, Amazon, Meta | CNBC | Feb 2026 |
Nvidia's data center revenue is especially important. Those chips are not bought because a board member saw a cool demo. They are bought because training and inference workloads already exist and are expected to spike. Broadcom, another supplier, reported AI semiconductor revenue of $10.8 billion in Q2 FY2026, up 143% year over year — and the stock still fell because investors wanted even more forward momentum.
That reaction is the key clue. The debate is no longer "is there demand?" It is "does the growth justify the price?"
Why inference — not training — changed the math
For much of the early AI boom, the spending story was about training: build a giant cluster, run a training job, produce a model, repeat next generation. Training is episodic and capital intensive, but it is not continuous at scale in the same way.
Inference is different. Every prompt, every agent step, every tool call, every retry, every long-context document, every code review pass burns compute. A chatbot query was one interaction. An agent loop can be thousands or millions of tokens.
That is why the infrastructure buildout suddenly looks industrial. Microsoft, Google, Amazon, and Meta are not just shipping software features anymore — they are building inference factories. Factories need:
- Power contracts and land.
- Cooling and networking.
- Depreciation schedules and routing optimization.
- Batching, caching, and efficiency improvements.
AI is turning the world's most valuable software companies into capital-intensive infrastructure operators. That explains the capex. The open question is whether the work being done on those chips is valuable enough to justify the cost.
The two-sided picture: real buildout, real sorting
It is possible for both of these statements to be true at once:
- AI demand is real and growing fast.
- Many AI investments will not earn their cost of capital.
The error is treating the technology as one uniform bet. It is not. AI is a general-purpose technology being applied across thousands of workflows with very different economics.
| Workload type | Economics | Example |
|---|---|---|
| Coding agent saves weeks of engineering time | Strong ROI | Claude Code, GitHub Copilot-style workflows |
| Legal review agent processes thousands of contracts | Strong ROI | contract analysis, compliance review |
| Customer support resolves real tickets | Moderate-to-strong ROI | automated triage, agent handoff |
| Shallow chatbot on stale knowledge base | Weak ROI | basic FAQ bots with no workflow integration |
| Generic enterprise dashboard "pilot" | Often negative ROI | demo-driven projects with no production follow-through |
That is why enterprise AI ROI looks like a mess right now. Some companies are capturing value. Others are spending money on tools without changing the process underneath. Most companies were already bad at software implementation, data projects, and cloud migration — so it should not be shocking that many are also bad at AI transformation.
The technology can be real while adoption is uneven.
What to watch instead of "bubble or not"
Whether you are an investor, a builder, or a business operator, these questions are more useful than the binary debate:
- Is the usage paid or just engaged? A pilot is not a product. A press release is not a production workload.
- Is it improving a workflow with clear economics, or creating more human review work? The latter is a cost, not a saving.
- Is capacity being built because customers are waiting, or because the board wants an AI strategy? One is demand pull; the other is supply push.
- Is the model being used where expensive reasoning matters, or is premium compute being burned on cheap tasks? That is the operating question of 2026.
- Who controls the bottleneck? Power, memory, advanced packaging, and efficient inference routing are scarce. Commoditized inputs are not.
- Can the company finance its own buildout, or does it need someone else to? Balance sheet matters when capex is this large.
These questions separate what is real from what is narrative. They also explain why a stock correction does not settle the issue. Markets can reprice for valuation, crowded trades, interest rates, or capital rotation while the underlying business keeps growing.
What this means for you
If you run a small business or team, do not ask "is AI a bubble?" Ask "where is the paid demand in my own workflows?"
The places to look first:
- Repetitive cognitive work that currently eats staff time: email triage, document review, coding, support responses.
- Work that fails or stalls because of capacity limits: analysis that takes days, code reviews that sit in queues, customer inquiries that miss SLAs.
- Tasks where a small team needs expert-level output without hiring for it: research synthesis, contract review, competitive analysis.
Start with one workflow that has clear before-and-after metrics. Measure cost, time, and error rate. If the AI tool does not change the workflow enough to move one of those, it is probably not worth expanding. If it does, the "expensive" inference is likely cheaper than the human time it replaces.
The winners over the next few years will not be the companies that talk about AI the loudest. They will be the ones that turn intelligence into reliable, measurable, high-utilization workflow value.
Related reading
FAQ
Is AI a bubble?
Parts of the market around AI show bubble-like dynamics: stretched valuations, hype-driven stock pops, and some overbuilt capacity. But the underlying demand — paid revenue, chip purchases, power constraints, and real enterprise usage — is not imaginary. It is better described as a real buildout with speculative layers on top.
Why are AI stocks falling if demand is real?
Stocks price in expectations of future growth and profit. When capex rises faster than visible payback, investors reprice the stocks downward even while the business grows. A correction in price is not proof that the demand is fake.
What is the difference between AI training and AI inference?
Training builds a model, usually in large episodic runs. Inference runs the model every time someone uses it. Agents, code tools, and long-context workflows have made inference the dominant ongoing cost — and the reason infrastructure spending has exploded.
Why is enterprise AI ROI so mixed?
AI is a general-purpose technology. Some workflows have clear economics; others are badly implemented or chosen for novelty. Companies that change the underlying process usually see returns. Companies that layer a chatbot on top of a broken process usually do not.
What should a small business actually do?
Pick one workflow with a clear cost or time metric. Run a measured pilot. Expand only if the metric moves. Avoid buying AI because competitors are buying AI.
Discussion
0 comments