Yes, AI Is a Bubble. The Bubble Just Isn’t Where You Think.
We sell AI for a living. So take this as a confession, not a hot take: there is a bubble. It’s enormous, it’s obvious, and pretending otherwise to protect a sales pitch is how you end up holding the bag.
But “AI is a bubble” and “AI is worthless” are two completely different claims, and almost everyone collapses them into one. The bubble is in the financing and the hype. The value is in a small, boring, unglamorous corner that keeps working no matter what the stock market does.
Here’s the line between the two.
The bubble is real — follow the money
The capital going into AI infrastructure has detached from the revenue coming out of it. This isn’t a vibe; it’s on the balance sheets.
| The number | What it says |
|---|---|
| $600–690B hyperscaler capex planned for 2026 — nearly 2× 2025 | Microsoft, Alphabet, Amazon, Meta and Oracle are spending like there’s no tomorrow[1][2] |
| 45–57% of revenue now goes to capex at the hyperscalers | A ratio that was previously unthinkable for tech companies[3] |
| ~$60B/yr OpenAI compute spend vs ~$13B revenue | A ~$47B annual gap, and a projected ~$14B loss in 2026[1][4] |
| $600B in revenue needed to justify spend so far (Sequoia) | The industry is nowhere near it[6] |
Then there’s the part that should make anyone nervous: the money is increasingly circular. Nvidia agreed to invest up to $100B in OpenAI; OpenAI committed ~$300B to Oracle; Oracle buys Nvidia chips to fulfil it. Dollars spent by one player return as revenue for another, which props up everyone’s valuation at once.[4][5]
And the marquee deal in that loop is already wobbling: by March 2026, Nvidia’s Jensen Huang told investors the $100B OpenAI investment was “not in the cards.”[4]
When demand has to grow into financing this aggressive — and the financing is partly the companies paying each other — you have the textbook setup for a correction. Critics comparing it to 1999 aren’t being dramatic.[5]
The bubble is real — follow the results, too
If the spending were quietly paying off inside companies, the financing wouldn’t matter. It isn’t.
| Finding | Source |
|---|---|
| 95% of enterprise generative-AI pilots delivered zero measurable return | MIT, The GenAI Divide, 2025[7] |
| 88% of orgs use AI somewhere, but only 39% see any EBIT impact | McKinsey Global AI Survey, Nov 2025[8] |
| 60% of deployments generate no material value; only 5% create value at scale | BCG, The Widening AI Value Gap, 2025[8] |
| Only 4 of every 33 proofs-of-concept reach production | IDC[8] |
| Enterprises abandoned 2.3 AI initiatives each in 2025, $7.2M sunk per initiative | S&P Global[8] |
This is the real bubble: not the technology, but the belief that buying it is the same as benefiting from it. Most of the spend is going into pilots that die before production.
So why isn’t this the end of AI?
Because we’ve run this exact experiment before, and we know how it ends.
Railway Mania, 1840s Britain. Parliament authorised thousands of miles of track, much of it redundant. Investors got wiped out. The rails stayed — and became the backbone of the industrial economy.[9]
The dot-com fiber glut, late 1990s. Telcos laid 80+ million miles of fiber on a delusion that traffic doubled every 100 days. The crash turned most of it into “dark fiber.” That same fiber, bought for pennies on the dollar, later carried YouTube, Netflix, and the entire cloud.[10]
So when people ask “will the AI bubble pop?”, they’re asking the wrong question. A better one: what survives the pop?
Where the value actually is
Even the skeptics leave a door open. Goldman Sachs ran a report literally titled “Gen AI: Too Much Spend, Too Little Benefit?” — and MIT’s Daron Acemoglu, about as bearish as economists get, estimates AI will meaningfully touch maybe 5% of tasks in the next decade.[11][12]
Five percent of all economic tasks is a catastrophe for a $7-trillion build-out. It’s a goldmine for anyone targeting the right 5%.
And that 5% is identifiable. The same MIT study that found 95% failure found the winners had one thing in common: they bought narrow tools from specialised vendors instead of building generic ones in-house. Buying succeeds ~67% of the time; internal builds succeed about a third as often.[7]
Look at where AI clears an honest, measurable return today:
| Use case | What it costs / returns |
|---|---|
| AI customer support | $0.99–2.00 per resolved ticket vs $6–12 for a human; reported ~$3.50 back per $1 spent[13] |
| Coding assistants | Reported ~376% 3-year ROI, payback under 6 months, 25–39% productivity gains[14] |
(Read those vendor-sourced returns with a grain of salt — they’re the optimistic end. But the direction squares with the MIT data: narrow, integrated, outcome-first AI is the 5% that works.)
Notice what every one of these has in common. It’s narrow. It’s wired into a real workflow. It replaces a cost you can measure. It doesn’t need AGI, a $100B GPU cluster, or a circular financing deal to pay off. It just has to do one job a human was already paying to have done.
The dividing line
| The bubble (will pop) | The value (will compound) |
|---|---|
| Valuations priced for AGI | Tools priced against a real cost line |
| Demos that wow in a meeting | Software wired into a daily workflow |
| Circular vendor financing | Revenue from customers who’d miss it if it vanished |
| “AI strategy” with no defined outcome | One job, one measurable return |
| Generic in-house “let’s build a chatbot” | Narrow tool from a vendor who lives in that niche |
When the froth burns off, the first column disappears and the second column gets cheaper to run — because the overbuilt compute becomes dark-fiber cheap.
That’s not a reason to be scared of the bubble. If you’re actually using AI to do a specific job for a specific return, it’s the best thing that could happen to you.
We built WisWes on the right-hand column on purpose. One job — recovering the sales a store loses when a shopper can’t get an answer — measured against a number a merchant already tracks. That’s the bet: not that AI is magic, but that the boring, narrow, ROI-positive slice survives whatever the market does to the rest.
References
- Goldman Sachs — “Why AI Companies May Invest More than $500 Billion in 2026.” https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026 ↩
- Futurum Group — “AI Capex 2026: The $690B Infrastructure Sprint.” https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/ ↩
- Introl — “Hyperscaler CapEx Hits $600B in 2026.” https://introl.com/blog/hyperscaler-capex-600b-2026-ai-infrastructure-debt-january-2026 ↩
- Bloomberg — “AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other.” https://www.bloomberg.com/graphics/2026-ai-circular-deals/ ↩
- The Register — “AI’s trillion dollar deal wheel bubbling around Nvidia, OpenAI.” https://www.theregister.com/2025/11/04/the_circular_economy_of_ai/ ↩
- Sequoia’s “$600B question,” via Calcalist. https://www.calcalistech.com/ctechnews/article/ekwf7tdrj ↩
- MIT — “The GenAI Divide: State of AI in Business 2025,” via Fortune. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/ ↩
- Master of Code — “AI ROI: Why Only 5% of Enterprises See Real Returns in 2026” (compiling McKinsey, BCG, IDC, S&P Global). https://masterofcode.com/blog/ai-roi ↩
- A Wealth of Common Sense — “The Railway Bubble vs. the AI Bubble.” https://awealthofcommonsense.com/2025/11/the-railway-bubble-vs-the-ai-bubble/ ↩
- IEEE ComSoc — “Big tech AI data-center spending vs the dot-com fiber buildout.” https://techblog.comsoc.org/2025/09/27/big-tech-spending-on-ai-data-centers-and-infrastructure-vs-the-fiber-optic-buildout-during-the-dot-com-boom-bust/ ↩
- Goldman Sachs — “Gen AI: Too Much Spend, Too Little Benefit?” https://www.goldmansachs.com/images/migrated/insights/pages/gs-research/gen-ai--too-much-spend,-too-little-benefit-/TOM_AI%202.0_ForRedaction.pdf ↩
- 404 Media — “Goldman Sachs: AI Is Overhyped, Wildly Expensive, and Unreliable” (Acemoglu interview). https://www.404media.co/goldman-sachs-ai-is-overhyped-wildly-expensive-and-unreliable/ ↩
- Fin.ai — “ROI of AI Customer Service: 2026 Benchmarks & Data.” https://fin.ai/learn/roi-ai-customer-service-agents-benchmarks ↩
- TeamDay — “AI Agent Use Cases with Proven ROI (2026 Data).” https://www.teamday.ai/blog/ai-agent-use-cases-2026 ↩