Volpe argues that AI bubbles follow the same pattern as dot-com and crypto: the technology works, but money runs out before revenue catches up. He draws on historical precedent to show that massive capital deployment without proven business models inevitably leads to a correction, regardless of the underlying technology's validity.
Volpe uses the dot-com crash as his central analogy: the internet was genuinely transformative and Amazon became a trillion-dollar company, but Pets.com and Webvan burned through capital without sustainable models. He argues AI is in the same position — the technology isn't fake, but the majority of current valuations assume revenue trajectories that haven't materialized.
The editorial frames the essay's timing as significant: three years into the generative AI era, OpenAI, Anthropic, Google, and dozens of startups have deployed billions in compute infrastructure. The central question is whether revenue growth can match the pace of capital expenditure, making this a pivotal moment for the industry's financial sustainability.
Martin Volpe published "How the AI Bubble Bursts" on March 30, 2026 — an essay that hit 257 points on Hacker News within hours, clearly striking a nerve with the developer community. The piece arrives at a moment when the AI industry sits at an inflection point: massive capital deployment on one side, growing questions about sustainable revenue on the other.
The essay's core thesis is deceptively simple: AI bubbles don't burst because the technology stops working — they burst because the money runs out before the revenue shows up. Volpe draws on historical precedent from the dot-com era and the crypto winter of 2022-2023 to argue that the pattern is always the same. The technology is real. The applications are real. But the business models haven't yet proven they can sustain the capital structures built around them.
This isn't a "hot take" from an outsider throwing rocks. It's a structured argument from someone who appears to understand both the technical and financial machinery of the current AI buildout.
The timing of this essay matters as much as its content. We're now roughly three years into the generative AI era, and the industry is entering what you might call the "show me the money" phase. OpenAI, Anthropic, Google, and dozens of well-funded startups have deployed billions in compute infrastructure. The question Volpe raises — and that the HN community clearly resonated with — is whether revenue growth can match the pace of capital expenditure.
The dot-com comparison is instructive not because AI is fake (it isn't), but because the dot-com crash didn't happen because the internet was fake either. The internet was transformative. Amazon survived and became a trillion-dollar company. But Pets.com, Webvan, and hundreds of others burned through capital faster than they could build sustainable businesses. The technology was real; the business models weren't — or at least, not at the valuations investors had assigned.
The parallel to AI infrastructure spending is hard to ignore. Major cloud providers have committed tens of billions to GPU clusters and data center buildouts through 2027. Nvidia's market cap reflects expectations of sustained hypergrowth in AI chip demand. Enterprise AI adoption is growing but — and this is Volpe's key point — the per-query economics of large language models remain brutal. Running inference at scale is expensive. Charging users enough to cover that cost, while competing with increasingly capable open-source alternatives, creates a margin squeeze that doesn't show up in top-line growth numbers.
The HN discussion around this piece splits roughly into three camps. The first agrees with Volpe's thesis and points to already-visible signs: AI startups offering unsustainably low prices to grab market share, enterprise pilots that don't convert to production deployments, and the growing gap between what investors expect and what the market actually pays for AI services. The second camp argues that AI is fundamentally different from prior bubbles because the productivity gains are immediate and measurable — code completion, document processing, customer support automation all show clear ROI. The third camp — perhaps the most interesting — holds that Volpe is directionally correct but too early: the bubble will burst, but not before another 18-24 months of irrational capital deployment inflates it further.
For developers and engineering leaders, the most actionable part of this analysis isn't the macro bubble question — it's the unit economics question at the product level.
Consider what it costs to run a production AI feature today. A mid-complexity LLM call (say, 2,000 input tokens, 500 output tokens on a frontier model) runs somewhere between $0.01 and $0.05 depending on provider and model. That sounds cheap until you multiply by millions of daily active users. A consumer app making 10 AI calls per user session at $0.03 each is spending $0.30 per session — numbers that would horrify any product manager who's worked in ads or SaaS where marginal serving costs approach zero.
The companies that survive a correction aren't the ones using AI — it's the ones who've figured out how to make AI features that users will pay a premium for, or that reduce costs elsewhere in the stack by more than the inference bill. This is the unsexy work of integration engineering: building retrieval pipelines that minimize token usage, caching aggressively, routing queries to the cheapest model that can handle them, and measuring actual business impact rather than demo impressions.
The open-source model ecosystem adds another pressure vector. When Llama, Mistral, and Qwen variants can run on $20K of hardware and handle 80% of enterprise use cases, the pricing power of API providers erodes. Volpe's framework suggests this is actually the mechanism through which the bubble deflates for some players: not a dramatic crash, but a steady compression of margins as commodity inference gets cheaper and differentiation gets harder.
If you're building on AI APIs today, the practical takeaway isn't to stop — it's to build with margin awareness. Specifically:
Architect for model portability. If your system is tightly coupled to a single provider's API, you're exposed to pricing changes and you can't take advantage of cheaper alternatives as they appear. Abstraction layers like LiteLLM or even a simple provider-switching interface pay dividends when the market shifts. The teams that will weather a correction best are those who treat their AI provider like a database — essential infrastructure, but swappable.
Measure cost per business outcome, not cost per API call. If your AI feature costs $50K/month in inference but generates $200K in retained revenue or saves $300K in manual processing, you're fine regardless of what happens to the broader market. If you can't tie your AI spend to a specific business metric, you're the kind of deployment that gets cut in a downturn.
Watch the enterprise procurement cycle. The leading indicator of a correction won't be a technical failure or a dramatic event. It will be enterprise renewal rates for AI products. When CFOs start asking "what did we actually get for that $2M AI platform contract?" and the answers are vague, budgets will tighten. If you're selling AI-powered tools, start building the ROI measurement story now.
Volpe's essay doesn't predict a date for the bubble to burst, and that intellectual honesty is part of why it resonated. The AI industry could sustain current spending levels for another year or three — bubbles always last longer than skeptics expect. What the essay does effectively is name the specific mechanism: not a Theranos-style fraud revelation, not a technical plateau, but the slow grind of unit economics forcing a repricing of what AI companies are actually worth. For developers, the safest position is the one that's always been safest — build things that generate more value than they cost, and don't assume today's pricing or funding environment is permanent.
It’s incredible how polarizing the AI rush is. I keep the perspective that the technology is an absolute step change but I have no idea where the cards will fall. I take a lot of issue with these style of articles. I get a sense that the authors are being overly defensive.The cost to serve tokens is
This article tries to build upon a lot of half-truths or incorrect facts, like this:> OpenAI is struggling to monetize. They turned to showing ads in ChatGPT,The ads aren’t going into your paid plans (except maybe a highly discounted tier, depending on the market). The ads are a play to offer a f
The thing I am struggling with is where is the impact of LLM tools, especially given the massive increase in token consumption from 2025 to now and the saturated presence of LLMs everywhere.Naively speaking, I have so many expectations for the impact of this tech.I'd expect a noticeable uptick
> nobody is sure if even their metered pricing is profitableThis is most likely wrong. Lab executives insist that serving tokens is profitable. It's the cost of training next-gen models that requires them to keep raising ever larger rounds. More importantly, many independent providers price
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
> RAM prices are crashing because new models won’t need as muchReality begs to differ [0] and following the link for that text goes to an article [1] where they talk about Google's TurboQuant which supposedly will lower the RAM requirements. Now if that means RAM prices come down (as specula