The bubble paradox

There's a phrase that gets trotted out during every mania: “pessimists sound smart, optimists make money.” Paul Kedrosky, playing the part of the pessimist, made a detailed and persuasive case on Derek Thompson's podcast last week that AI is a bubble. And I wanted to agree with him. It is intellectually satisfying to point out that the consensus trade is crowded, that valuations are untethered, that we've seen this movie before. But the analogies to past bubbles didn't quite hold.

This is a bubble. Of course it's a bubble. Every general purpose technology produces one. The question isn't whether there's a bubble. The question is whether this bubble has structural features that change it in meaningful ways.

I think there are three.

Tokens aren't electrons

From the beginning, there's been this assumption that AI models would converge toward becoming utilities. Intelligence in, intelligence out. Tokens as electrons. And therefore, the reasoning went, the model layer would commoditize, open source would win, and value would accrue to the application layer on top.

That hasn't happened yet. And I think part of the reason why is that the models are genuinely different. You can feel the difference when you use ChatGPT or Claude or Gemini or Grok. The training process, the post-training alignment, the harness, all of it imbues each model with a character that users notice and that builders depend on.

This matters for two reasons. First, users develop personality preferences, as we’ve already seen, which creates a level of consumer stickiness that you wouldn’t normally expect from pure infrastructure. Second, developers who integrate these models into their products build a significant amount of model-specific scaffolding, which means it’s actually quite difficult to simply swap models for each other. There’s a switching cost. And it makes the model layer stickier than a model-as-utility thesis would have predicted.

Physics as a governor

Previous infrastructure bubbles overbuilt into a demand vacuum. During the dot com era, companies laid enormous amounts of fiber optic cable before anyone needed the bandwidth. That supply was speculative as the demand hadn't yet materialized.

The AI buildout has the opposite problem. Despite demand, there are real bottlenecks limiting supply: power generation, chip fabrication, the fact that ASML can only produce so many EUV lithography machines per year. Just listen to Dylan Patel from SemiAnalysis on Dwarkesh Podcast last week and you’ll quickly realize that these are physical chokepoints that capital alone cannot solve.

What's interesting about this is how these constraints may change the shape of the bubble itself. Past bubbles overbuilt ahead of demand. This one is building into a supply wall. As a result, these constraints may act as a governor on overbuilding. I doubt they’ll prevent a bust, but perhaps they will smooth out the cycle, containing our worst, speculative excesses simply because the infrastructure can't be built fast enough to match the capital available.

A collapsed stack

The most consequential difference, in my opinion, is that the infrastructure providers are also the builders of end-consumer products.

The railroads and cloud were infrastructure that others built on. With AI, the frontier labs are building the platform and the application.

OpenAI's models are both the API that developers build on and ChatGPT, the product that consumers pay for directly. This collapses the stack in a way that makes historical analogies misleading. Monetization happens faster because the infrastructure provider is in direct conversation with the end user and can build what they want. That's a feedback loop between the model's capabilities and the consumer's willingness to pay that didn't exist when Union Pacific was laying track. This means the infrastructure layer is positioned to capture far more value than in previous buildouts.

Bubbles don’t create too much risk. They destroy it.

So yes, this bubble has features that make it structurally different from past ones. The models are stickier than a pure utility. The buildout is running into physical constraints. The stack has collapsed. These are real differences and they create real opportunities.

They're also the most dangerous thing about this moment. Because every one of those arguments gives investors and founders permission to believe the “long AI” thesis is forever safe.

And when everyone agrees the question shifts from what you're building to how you're positioned. And then we all, founders and investors alike, start optimizing for legibility and fitting cleanly into the “long AI” narrative, because the narrative is so strong that proximity to it feels like enough.

This is how bubble thinking kills risk-taking. Not by encouraging recklessness, but by making the safe, legible, consensus play seem like vision. I've referred to this in the past, inspired by Adam Mastroianni and Derek Thompson, as the "cover band economy." In music, investors are buying up rights to old hits because they're safe assets. In startups, founders are building the legible, obvious thing (e.g. the coding agent, the orchestration layer, the vertical AI targeting a market that half the latest YC class is also chasing) and hoping to ride the wave and fundraise on a theme we've all decided is "smart." 

This way of thinking destroys the risk-taking required to build great startups. We’re building companies designed to get funded, not companies designed to exist. We’re optimizing for the probability of success rather than the magnitude of success.

If you find yourself convinced that you are not taking risk, or that your risk has been materially reduced simply because the macro trend is so powerful, you should stop and ask yourself a question: are you building a company, or are you making a trade?

Bubble psychology makes the consensus trade feel costless and the contrarian bet feel reckless. But during a bubble, consensus is where the risk, strange as it sounds, is hiding.

David Peterson

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

Bribing PE - Reuters reports on OpenAI and Anthropic are both courting private equity firms to form joint ventures, aiming to deploy AI tools across their portfolios of companies. An aggressive enterprise land-grab ahead of potential IPOs for both companies. OpenAI is offering significantly sweeter deal terms (17.5% guaranteed minimum returns on preferred equity), while Anthropic's version offers no guaranteed returns. It’s not for everyone, and firms like Thoma Bravo have walked away citing concerns about long-term profitability.

Smuggling Compute - In a dramatic turn of events the CEO of NVIDIA’s third largest customer Super Micro Computer was found to be smuggling Nvidia AI chip-equipped servers to China. The operation was elaborate: defendants used fabricated documents, staged bogus equipment to pass audit inventories, and used a pass-through company to conceal their misconduct — even using hair dryers to remove labels and serial numbers from real machines and placing them on dummy machines left behind after the real machines had been shipped to China. Noah Smith recently wrote about the US crackdown on grey market chips "With no AI chip exports to China and no smuggling, we estimate the US would hold a 21–49x advantage in 2026-produced AI compute, depending on whether FP4 or FP8 performance is used for Blackwell chips."

HOW TO STARTUP

EU Inc? - The EU shared the first draft of their response to EU-Inc’s lobbying for a European Delaware style system. This would be a single pan-European company structure designed to make it easier for startups to scale across all 27 member states without navigating 27 different legal systems. The Financial Times outlines why the proposal is underwhelming. On flexible funding, harmonised stock options, fast digital registration it seems relatively founder friendly. However, on legal certainty for investors it falls short, because national courts will still interpret the rules, meaning investors will still face 27 different legal environments. The FT’s view is that the Commission aimed low pre-emptively to get something approved, and the result is a proposal that looks European on paper but feels national in practice. EU-Inc themselves commented that it runs the risk of becoming ‘another GDPR’.

HOW TO VENTURE

Fragmentation & Consolidation - Keith Teare makes a case for the continued existence of early stage funds, as VC starts to fragment into distinct asset classes. Venture is now several distinct businesses (seed, venture, growth, and increasingly retail-facing secondary vehicles) with different economics and market structures. The trend towards concentration where a tiny number of firms are capturing a growing share of early-stage dollars is a structural issue according to Teare. At seed, the top 5 firms went from 3% of all seed dollars in 2020 to over 16% in 2026 so far. At Series A it was already concentrated and is getting more so.

“For founders, that means there is no single fundraising market anymore. There is a hierarchy. If you can access the top firms, the value of that access is rising. If you cannot, you need to be realistic about which market you are in and what kind of syndicate can actually help you.

For LPs, the implication is even sharper. If the market keeps concentrating into a handful of firms, manager selection matters more than ever. But if the same firms winning that concentration battle are doing so in a business whose economics are deteriorating with scale, then the obvious allocation path may not be the right one.

For emerging managers, the message is brutal. There is less and less room for the generic venture fund. You need either genuine structural differentiation or the ability to drive outcomes through access, ownership, and involvement. Probably all three."

Reco has released Reco AI Agent Security to fill the visibility gap for AI agents across the enterprise.

Hornet
Senior Research Engineer (Trondheim, Norway)

Motorica
Customer-facing Senior Technical Animator in North America (Remote)

Sourcix
Frontend Developer (Tel Aviv)