The headlines are writing themselves: "AI bubble bursting." "Enterprise adoption stalling." "Cracks in the AI thesis." Polymarket gives a 24% chance the AI bubble bursts this year. The September 2026 Ramp AI Index — tracking spending across 70,000 companies — dropped a report titled "Cracks in the AI thesis, part 2." And AI Supremacy is asking point-blank: is AI adoption slowing down?
Read the full version with charts and embedded sources on ComputeLeap
View the full Ramp AI Index report
Here's what the bubble narrative gets wrong: it confuses a repricing event with a collapse signal. AI adoption isn't slowing — it's being reorganized. Value is migrating between layers of the stack, and the companies that understand which layer they're on will win. The ones chasing last year's playbook will confirm every bear thesis in the process.
The Numbers That Spooked Everyone
Let's start with what the bears are seeing, because they're not making things up.
The Ramp data is genuinely striking. Median AI expenditure among the top 1% of spenders fell nearly 10% in a single month — from $7,976 per employee in July to $7,205 in August. Overall, only 56% of Ramp customers are paying for AI products, and that number grew just 0.4% month-over-month. The adoption curve is visibly flattening.
Meanwhile, the LLM Token Expenditure Index dropped to $1.67, down 20% from its May peak. Effective token pricing on Ramp's platform fell 41% — from $1.15 per million tokens in March to $0.68 today. OpenAI and Anthropic have both announced aggressive price cuts over the past month.
At the enterprise level, the picture looks even bleaker for the "AI is transforming everything" crowd. Harvard Business Review reports that 88% of companies claim regular AI use, yet integration remains stubbornly shallow. Writer's 2026 survey found that 79% of organizations face AI adoption challenges — a double-digit jump from 2025 — and 54% of C-suite executives say AI adoption is "tearing their company apart." A Q1 2026 Morgan Stanley note found that only 21% of S&P 500 companies reported even one measurable AI benefit.
So yes, if you squint at these numbers through a "bubble bursting" lens, you can build a compelling case. But that case has a fatal flaw.
The Fatal Flaw: Confusing Price with Demand
Here's the number everyone is ignoring: token prices fell 41%, and spending fell 10%.
Do the math. If the price of your input drops 41% and your total bill only drops 10%, you are consuming dramatically more tokens. Companies aren't retreating from AI — they're buying more of it for less money. The Ramp report itself notes that volume growth is being driven by cheaper, standard-tier models like GPT-5.6 Terra and Claude's Sonnet series, while frontier model share dropped from 53% to 45% of total tokens.
View the full AI Supremacy analysis
This is exactly what AI Supremacy's Michael Spencer identified: the story isn't that companies are abandoning AI. It's that they're getting smarter about which models they use and when. Anthropic extended its lead to 43.8% of U.S. businesses (up 0.34 points), while OpenAI grew only 0.09 points to 39.8%. The market is consolidating around value, not retreating from the technology.
This pattern has a name in economics: price elasticity driving volume expansion. It's what happened with cloud computing in 2014-2016, with mobile data in 2012-2014, and with internet bandwidth in 2002-2005. The "bubble bursting" narrative in each case was actually a commoditization event that preceded the real adoption wave.
Where the Value Is Actually Going
The real story isn't about whether AI is growing or shrinking. It's about which layer of the stack captures the value. And Dylan Patel's SemiAnalysis just published the definitive analysis of this shift.
View the full SemiAnalysis report
The headline data point: Anthropic's ARR exploded from $9 billion to $44 billion-plus within a year. Gross margins on inference infrastructure climbed from 38% to over 70%. The model labs are capturing an increasing share of the total value created by AI, and they're doing it with improving unit economics.
But the most revealing data point isn't about a model lab — it's about SemiAnalysis itself. Patel's firm now spends $10.95 million annually on Anthropic Claude tokens. Their payroll is approximately $25 million. That means AI token spending is already at 25% of payroll — and Patel says it's on track to exceed 100% by year-end.
SemiAnalysis spends $10.95M/year on Claude tokens against a $25M payroll — 25% of human labor costs, heading toward 100%. That's not a company "slowing adoption." That's a company whose primary production input is shifting from humans to tokens.
This is the value-capture story the bubble narrative completely misses. At the infrastructure layer, Nvidia's Blackwell generates 30x more tokens per second than Hopper. Hardware costs per token are cratering. At the model layer, labs are repricing upward because their margins are expanding even as they cut sticker prices. And at the application layer, companies like SemiAnalysis are generating so much ROI from tokens that they're scaling spend faster than headcount.
The value isn't disappearing. It's moving.
The Three-Layer Framework
Think of the AI economy as three layers, each with a different value-capture dynamic:
Layer 1: Infrastructure (Chips, Data Centers, Cloud)
This is where the $700 billion is being spent. Hyperscalers guided over $700 billion in AI infrastructure capex for 2026. The bears' strongest argument lives here: that's a lot of concrete and silicon to amortize, and if demand doesn't materialize, depreciation schedules will crush margins.
But infrastructure is also where commoditization hits hardest. Token prices are falling because compute is getting cheaper — Blackwell's 30x throughput improvement means the same rack serves dramatically more inference. Infrastructure providers that can't differentiate will see margins compress. The value is passing through this layer, not accumulating in it.
Layer 2: Model Labs (Anthropic, OpenAI, Google DeepMind)
This is where value is concentrating right now. Anthropic's margin expansion (38% to 70%+) tells the story: model labs are cutting prices while improving profitability because their compute efficiency gains outpace their price cuts. They're the tollbooth — every application-layer company pays them per token.
The SemiAnalysis analysis argues Nvidia and TSMC are actually underpricing relative to the value they deliver — strategic restraint to maintain ecosystem stability. If that's true, the model labs' margins have room to expand further as they negotiate better hardware deals while maintaining premium application-layer pricing.
Layer 3: Applications (Where the ROI lives)
This is the layer the "adoption is slowing" narrative ignores entirely. Elena Verna told Lenny's Newsletter that "60-70% of traditional growth tactics no longer apply" in AI-native companies. Lovable hit $200M ARR in under a year with 100 employees. That's not a bubble metric — that's a structural efficiency gain.
The application layer is where tokens convert into business value. SemiAnalysis isn't spending $10.95M on tokens because they're caught up in hype. They're spending it because the ROI justifies it — and the ROI improves as token prices compress. When the ratio of token cost to human cost keeps falling (thanks to Layer 1 commoditization), the application layer's ROI only improves.
Contrarian Corner: The Bears' Math Is Real
Warning: The gap between AI capex and AI revenue is real, and "repricing" doesn't make it disappear. $700 billion in infrastructure spend against ~$100 billion in AI revenue is a 7:1 ratio. History suggests capex cycles this aggressive don't end gracefully. The optimistic read assumes demand will compound faster than depreciation — but what if enterprise AI adoption is an S-curve nearing its plateau, not an exponential in its early innings?
The bears aren't wrong about the numbers. RAND reports that roughly 80% of enterprise AI projects fail to deliver business value. MIT's Project NANDA found 95% of generative AI deployments produced no measurable P&L impact. Only 21% of the S&P 500 report any AI benefit at all.
The Hacker News community has been debating this extensively, with one widely-upvoted thread arguing that "a technological revolution and its adoption curve and a financial bubble are two completely different phenomena with different physics." That's exactly right — and it cuts both ways. The technology can be genuinely transformative while the financial instruments built on top of it are simultaneously overpriced.
Another HN thread raised the specter of hyperscalers "left holding debt from infrastructure that looks increasingly unneeded." If the application layer captures most of the value (as the data suggests), then the infrastructure layer is building capacity that generates returns for its customers, not for itself. That's the railroad-builder's dilemma all over again: the railroads went bankrupt, but the economy they enabled boomed.
The honest answer is that both things can be true. AI can be genuinely transformative at the application layer while simultaneously being a bad investment at the infrastructure layer. The bubble narrative and the growth narrative aren't contradictions — they're describing different layers of the same stack.
What the Prediction Markets Say
Polymarket's AI bubble contract is instructive. It started the year at 17% probability and climbed to 24% by April. The market is pricing in meaningful uncertainty — but notice what "bubble burst" means in the contract: a 40%+ decline in AI-related equities. That's a statement about stock prices, not about technology adoption.
This distinction matters. The dot-com bubble "burst" destroyed trillions in market value while leaving behind Amazon, Google, and the entire modern internet economy. The technology adoption curve never reversed. What reversed was the financial premium placed on unrealized potential. If the AI bubble follows the same pattern, we'll see equity corrections in infrastructure-heavy names while application-layer companies continue scaling.
What This Means for You
Tip: The actionable question isn't "is AI a bubble?" It's "where in the stack am I building, and who captures the value from my work?"
If you're a developer, founder, or engineering leader, here's the framework:
Stop watching the macro and start watching the margins. The top-line "AI spending is slowing" narrative is noise. What matters is the margin structure at your layer. If you're building applications that convert tokens into business outcomes, your economics improve every time token prices drop. If you're building infrastructure, you're in a commodity race.
Build at the application layer. SemiAnalysis isn't spending $10.95M on tokens because they're caught up in hype. They're spending it because the ROI justifies it — and the ROI improves as token prices compress. The application layer is where technology value converts to business value. That's where you want to be.
Don't confuse price compression with demand destruction. Token prices falling 41% while spending falls only 10% means usage is up roughly 50%. The AI market is growing in volume even as it contracts in price. This is healthy maturation, not collapse.
Watch the Anthropic vs. OpenAI share shift. Anthropic gaining share while OpenAI stalls isn't a random fluctuation — it reflects enterprises voting with their wallets on model quality, reliability, and developer experience. The Ramp data makes this concrete: 43.8% Anthropic vs. 39.8% OpenAI among U.S. businesses.
Revisit your vendor mix quarterly. The market is consolidating fast. Open-source adoption is still only at 6.4% among AI-spending businesses per Ramp, but Chinese open-weight models like DeepSeek are advancing rapidly on cost-per-quality metrics. Your vendor strategy should be a portfolio, not a monogamous relationship.
Expect a bumpy 12 months for AI equities, not for AI utility. If the bubble-burst scenario plays out, it will look like 2001: stock prices correct, weak startups die, and the survivors emerge with better unit economics and less competition. The technology doesn't un-invent itself. The value-capture analysis from SemiAnalysis suggests the model labs and application-layer winners will be those survivors.
The bubble debate is a distraction. The real game is value capture by layer. Know which layer you're on, and act accordingly.
Related reading: The Subsidy Clock Is Ticking on AI Tokens | When AI Scaling Laws Break the Capex Math | AI Unit Economics vs. Human Labor
Originally published at ComputeLeap
Top comments (0)