There is a particular kind of vertigo that comes from watching numbers this large move this fast. OpenAI carries a valuation north of $852 billion. China has reportedly lined up a national compute plan measured in the hundreds of billions of dollars. And the money flowing between the chipmakers, the clouds, and the labs is starting to circle back on itself in ways that make even bullish investors uneasy. The question every founder, marketer, and operator is quietly asking is the only one that matters: is this a durable buildout of the next computing platform, or a bubble that leaves behind some very expensive infrastructure as a consolation prize? The honest answer is that it can be both at once — and understanding why is more useful than picking a team.
The case for ‘bubble’
Start with the valuations, because that is where the discomfort begins. By the figures circulating in trade coverage (per dentro.de/ai, which we’d flag against company disclosures), OpenAI is generating roughly $2.6 billion in monthly revenue and serving around 900 million weekly ChatGPT users. Those are not vanity metrics — they are genuinely strong fundamentals for a company barely a decade old. And yet an ~$852 billion valuation still implies years of compounding, near-flawless growth before the math closes. The same data point feeds both sides of this argument: bulls see momentum, bears see a price that has run far ahead of present cash flow. When a single company’s worth assumes it captures an enormous slice of a market that does not fully exist yet, you are pricing narrative as much as revenue.
The more structural worry is circular financing. The clearest tell is when capital, compute, and customers start to be the same parties wearing different hats. Consider the trading firm Jane Street’s move into AI investing — a reported $1 billion CoreWeave investment in April plus a stake in Anthropic (per LLM Stats citing the WSJ’s Gregory Zuckerman, which we’d verify against the primary report). Layer that onto the broader pattern: chipmakers investing in the labs that buy their chips, clouds taking stakes in the model companies that rent their capacity, and labs committing to spend that capacity back. Money loops. Each loop inflates the apparent demand the next loop is built to satisfy. None of this is necessarily fraudulent or even irrational, but it does make demand look more proven than it is.
And that is the third leg of the bubble case: capex is outrunning verified demand. The buildout is being justified by projected usage curves, not banked ones. If those curves flatten — if enterprise adoption stalls, if margins on inference compress, if the killer applications stay niche — a great deal of capital will have been committed against a future that arrived smaller and later than the spreadsheets promised.
The case for ‘buildout’
Now the other side, which is stronger than skeptics like to admit. The usage is real. Hundreds of millions of weekly users and billions in monthly revenue are not phantom demand conjured by a financing loop; they are people and businesses paying for something they find useful, repeatedly. The top labs are not pre-revenue stories selling a dream — they are operating businesses with genuine, growing, recurring income. That distinguishes this moment sharply from the purely speculative manias of the past, where the underlying product barely existed.
Then there is the infrastructure itself, which has a habit of outlasting the hype cycle that funds it. Data centers, power agreements, fiber, and fabrication capacity are durable assets. Even if the equities correct violently, the silicon and the substations do not evaporate. They get repriced, change hands, and keep computing. The capacity built in a frenzy becomes the cheap, abundant base layer of whatever comes next.
Finally, sovereign compute is emerging as a category of demand that is largely indifferent to quarterly sentiment. A reported $295 billion Chinese national compute plan — and the parallel ambitions of other governments — reflects a view that AI capacity is strategic infrastructure, like ports or grids. States do not buy compute to flip it; they buy it because they have decided they cannot afford not to have it. That is structurally durable demand, insulated from the mood swings of public markets, and it puts a floor under the buildout that private speculation alone would not.
Why it can be both
The resolution to this debate is that ‘bubble’ and ‘buildout’ are not mutually exclusive — they describe different layers of the same phenomenon. The financial structure can be a bubble while the physical structure is a buildout. Equities can be mispriced by a wide margin and still leave behind rails that genuinely matter.
The railroad and dotcom analogies are the standard reference here, and they are useful only if handled with care. Britain’s railway mania of the 1840s and the late-1990s telecom and dotcom boom both featured wild overinvestment, ruinous losses for many investors — and, crucially, the survival of the infrastructure. The track stayed in the ground; the fiber stayed lit. A later generation built profitable businesses on top of assets bought for cents on the dollar after the crash. The lesson is not ‘it all works out.’ The lesson is that the rails surviving and the investors getting wiped out are perfectly compatible outcomes. The analogy breaks, though, if you assume today’s leaders are guaranteed to be tomorrow’s winners — most of the railway companies died, and the dominant web companies of 2010 were mostly not the darlings of 1999.
So the question that actually matters is distributive: if it pops, who eats the losses? The circular financing makes this harder to trace, which is itself a warning sign. When chipmakers, clouds, labs, and trading firms are all cross-invested, a correction in one node propagates through the others. The taxpayers funding sovereign compute, the institutional investors holding late-stage stakes, the employees compensated in paper valuations — they sit at different points on the risk curve, and not all of them know it.
What founders should do regardless
Here is the practical part, because most readers cannot move markets but can decide how to build. The strategic posture that survives either outcome is the same one that has always survived hype cycles.
- Build on durable demand, not narrative. If your business only makes sense in a world where AI valuations keep climbing, you are not a company — you are a leveraged bet on someone else’s stock. Anchor to a problem people would still pay to solve if the word ‘AI’ disappeared from your pitch deck tomorrow.
- Keep burn defensible. Cheap capital and cheap compute are both features of the boom, and both can reverse. Assume your next raise is harder and your inference costs are higher. A burn rate that only works at peak-bubble pricing is a liability disguised as ambition.
- Own a workflow, not a wrapper. The thin layer between a user and a frontier model is the most exposed position in this entire stack — squeezed by model providers above and commoditised by competitors beside you. Embed yourself in how customers actually get work done: their data, their processes, their decisions. Workflows are sticky; bets on a valuation are not.
The buildout is real and the bubble is plausible, and a clear-eyed operator can hold both ideas at once. The infrastructure being poured today will likely still be computing in a decade, regardless of what the equities do in between. The winners will be the ones who built something the rails were always going to need — not the ones who confused the price of the network for the value of being on it.
