Imagine a family where nobody ever meets a stranger. The parents fund the kids’ businesses, the kids hire their cousins, the cousins marry each other, and every problem is solved by shuffling money and favors around the same dinner table. On paper, the household looks rich. In reality, nobody knows if anyone outside that house would pay for anything they do.
That, roughly, is the current AI economy.
You have three main branches of the family. At one end, chip vendors like Nvidia and AMD, printing silicon and profit. In the middle, the landlords of compute: Microsoft, Amazon, Google, Oracle and friends, building data centers and renting them out. At the far end, the AI labs and startups: OpenAI, Anthropic, xAI, plus hundreds of logo-heavy, margin-light “platforms” sitting on top.
The headline deals already read like family bookkeeping. OpenAI signs a commitment to spend around 300 billion dollars on Oracle data centers. Oracle uses that to order more Nvidia chips. Nvidia, sitting on its “dragon’s hoard”, writes checks into OpenAI and other labs, which then have more money to buy Nvidia hardware. This gets presented as three independent businesses responding to booming demand. It looks a lot less independent if you zoom out: the same few actors wiring money to each other, while the rest of us are told to admire the “ecosystem”.
If this were an isolated curiosity, it would be fine. But the same pattern repeats. Nvidia invests in AI companies so they can afford its chips. Labs borrow billions to buy those chips, sometimes putting their existing hardware up as collateral, and call that “building the future”. A Bank of America chart that now circulates in videos and explainer threads shows this mess of arrows, with the polite label “vendor financing”. Translation: suppliers are subsidising their own demand and hiding the discount under “investments”.
Meanwhile, the story told to investors is that this is the safest way to play a revolution. Do not worry about whether OpenAI, Anthropic or the 487th “AI copilot for your vertical” ever make money. Own the chips. Own the cloud. Be “picks and shovels” in a gold rush. That line sounds wise until you realise the shovel maker is also handing out gold so that miners can afford shovels. At that point you are not above the frenzy; you are underwriting it.
The numbers people latch onto are loud enough. OpenAI valued at around 500 billion dollars on a bit over 10 billion in revenue and losses larger than that revenue. Nearly 500 AI unicorns with billion-dollar valuations, inside a field where the honest facts are “we do not know yet” and “we almost certainly do not need 500 of these”. A handful of firms account for 99 percent of AI token spend, two companies buy close to 40 percent of Nvidia’s chips, and ten firms, most of them leaning hard on the AI story, now make up about 40 percent of the S&P 500’s weight. If even one of these pillars cracks, you do not just have “volatility in tech”. You have a market-level migraine.
It is tempting, from here, to just shout “Ponzi” and move on: Circular deals, no clear end-customer, lavish spending justified by marketing claims about “AGI” that current models cannot technically deliver. The technology itself has limits we politely underplay. Today’s large language models and diffusion systems are powerful pattern machines, but they are still pattern machines. They interpolate between examples; they do not conjure abstract structure out of nowhere. They still fall for prompt injection, still confuse handles and data, still treat any text that looks like an instruction as an instruction.
The mismatch is obvious. You have marketing that hints at “human-like cognitive abilities”, and underneath it, a system that breaks in mundane ways when someone sticks a malicious string in the input box. So yes, there is a gap between promise and reality. Yes, some of the revenue projections look like science fiction written in Excel. Bain talking about needing two trillion dollars in annual AI revenue by 2030, more than Microsoft, Meta, Alphabet, Amazon, Apple and Nvidia’s 2024 revenue combined.
Still, declaring the whole thing doomed is just another lazy comfort loop. There is real demand for automation, summarisation, content generation, boring workflow glue. Companies are already laying off people on the assumption that chatbots and internal copilots can take over slices of work. In parallel, a lot of this capex is not pure delusion, it is defensive. Meta, Microsoft, Amazon, Google are not stupid; they are terrified of underinvesting when the next general platform shift might eat their margins. They view AI spend partly as an option: maybe it pays off directly, maybe it just stops someone else from owning the layer between users and their services.
History also refuses to play the same movie twice. The dot-com era gave us thousands of vapor companies and an eighty percent crash in the NASDAQ, but it also laid foundations for cloud, smartphones, modern logistics and advertising that later did produce enormous profits for a small subset of survivors. Research on past technological booms is pretty consistent: revolutions are usually bad investments in aggregate, but very good for a few firms and, eventually, for users who enjoy cheaper and better services while someone else eats the loss.
So no, this is not a simple question of “bubble or no bubble”. Some parts of this thing are clearly overcooked: startup valuations, certain chip names, the idea that every SaaS feature with “AI” stapled on deserves a ten-times multiple. Other parts might be grimly rational: hyperscalers locking in supply, real estate and utilities quietly riding the data center wave, boring infrastructure that will outlive this particular marketing cycle.
The deeper problem sits somewhere else. It is not just that the same firms keep trading money and hardware with each other. It is that they keep trading the same assumptions. Everybody inside the circle has converged on a shared story: that enterprise adoption will be fast, that users will eventually pay real prices, that scaling current architectures will fix their own economics, that the grid and regulation will somehow stretch to meet whatever number of GPUs they feel like ordering. On the other side, the skeptical chorus reuses its own script: dot-com, Greenspan, Nortel, metaverse 2.0, “we have seen this movie before”.
Both stories are comforting. Both flatten the details. It is much easier to tweet “AI is a Ponzi” than to sit with the fact that vendor financing is a fraction of total spend, not the whole thing. It is easier to chant “this changes everything” than to admit that most current AI companies are money furnaces with questionable business models.
In a healthy market, outside capital tests these stories. Investors ask what happens when the easy users have signed up, when free tiers are cut, when regulations clamp down, when power gets more expensive, when the first big AI darling misses its revenue promises. In the world we have, a lot of that capital is coming from inside the same small circle of winners, and it is being used not just to build capacity, but to keep the story running.
That is where the incest metaphor actually fits. When a family keeps reproducing only with itself, the damage is subtle: weaker immune systems, hidden defects, less resilience when something unexpected hits. The AI economy today looks similar. You have concentrated gains, recycled money, shared narratives, and a shrinking pool of genuinely independent checks on reality. The damage will not arrive as a single, cinematic pop. It will show up as disappointing returns, quiet consolidations, ugly write-offs, and eventually a much smaller set of firms quietly using AI as plumbing while the original grand promises are edited down.
What we are watching is not divine destiny or pure scam. It is a family system that refuses to bring in new blood, convinced that if they just keep coupling capital with capital, something transcendent will emerge. Maybe it will. But if you are not sitting at their table, be careful which branch of this family you marry your money into.