Last month, a client in Vienna — a conservative portfolio manager who had never touched a crypto asset in his life — asked me a question I couldn't answer cleanly. He'd just read that Bloomberg Intelligence projects AI investment to reach $750 billion by 2026. "If the money is that certain," he said, "why does nobody talk about where the return comes from?" He wasn't being rhetorical. He was genuinely confused, and his confusion was the most honest thing I'd heard in weeks. That number — $750 billion — has been circulating through crypto media and traditional finance channels alike, repeated as though repetition equals verification. But when I sat down to audit it the way I'd audit a smart contract, I found something unsettling: the figure is enormous, the confidence behind it is high, and the actual mechanism connecting investment to return is almost entirely absent from the conversation.
To understand why this matters, we need to remember how we got here. The AI capital cycle didn't emerge from a vacuum — it followed a recognizable pattern I've watched play out in crypto three times now. First comes the technical narrative: a genuine breakthrough that nobody can fully explain but everyone can feel. Then comes the capital narrative: institutions translate that excitement into spending commitments, and the commitments themselves become the story. In 2017 we called it ICO fever. In 2021 we called it the NFT boom and DeFi summer. Each cycle had the same skeleton — technical promise, capital flood, and a long, painful gap between what was invested and what was earned. The AI buildout is running that same skeleton, but at a scale that makes the crypto cycles look like rehearsal. When Bloomberg Intelligence puts a number like $750 billion on the table, it is not describing revenue. It is describing spending. And spending, unlike revenue, requires no customer to say yes.
The core mechanism here is what I've started calling the investment-revenue vacuum — the measurable distance between what capital commits and what the market can actually pay back. Sequoia's David Cahn framed a version of this in 2024 with his "$600 billion question," arguing the industry needed roughly $600 billion in annual revenue just to justify its infrastructure spending. The $750 billion projection doesn't close that gap. It widens it. If global AI revenue that is actually monetizable — API calls, SaaS subscriptions, enterprise licenses — still sits in the $100 to $200 billion range, then we are looking at a three-to-five-fold suspension between what is being poured in and what is being pulled out. I have spent the last several years auditing tokenomics models, and I can tell you the first thing that kills a protocol is not a bad idea. It is a spending curve that outruns its earning curve for one quarter too long.
What makes this cycle harder to read than a crypto cycle is the circular financing structure underneath it. Chipmakers invest in cloud providers and model companies. Those companies then purchase chips from the chipmakers. On the surface, demand looks explosive. Beneath the surface, a meaningful slice of that demand is being manufactured by the same parties who profit from it. I've seen this before, and not in AI. In 2021, several DeFi protocols were propped up by their own treasuries buying their own tokens through intermediary wallets. The volume was real. The demand was not. When sentiment reversed, the circularity unwound in days. The AI version of this loop is slower and more institutional, but the accounting logic is identical — apparent demand that includes vendor financing is not the same as end demand, and the gap between the two is where valuations go to die.
The physical layer adds a second constraint that capital narratives almost always skip past. You cannot spend $750 billion on compute without confronting electricity, advanced packaging, and memory bandwidth. Roughly 40 to 50 percent of that figure, if it lands, becomes chips, servers, and data centers. That translates into hundreds of thousands of GPUs, and every one of them needs power, cooling, and HBM memory that the supply chain may not be able to deliver on the same timeline as the money arrives. I watched a similar mismatch in 2021, when capital flooded into protocols faster than their infrastructure could scale, and the result was not growth — it was congestion. Money arriving faster than physical capacity can absorb it does not create value. It creates idle assets and depreciation. And GPU depreciation cycles run two to three years. If utilization lags, the write-downs come quickly.
Here is the part that I think most coverage gets backwards. The dominant narrative frames $750 billion as evidence of confidence — proof that the smartest institutions in the world believe in AI's future. I want to gently push against that. A capital expenditure forecast is not a bet on outcomes; it is a bet on not being left behind. The hyperscalers cannot afford to underinvest, because underinvesting guarantees irrelevance while overinvesting only risks losses. This is a prisoner's dilemma dressed as optimism. Each player is forced to escalate regardless of whether the collective spend is rational, and the industry-wide return on capital falls even as individual commitment rises. That dynamic does not produce a supercycle. It produces a subsidy war, and subsidy wars end with consolidation, not coronation.
The second thing we tend to forget is who wrote the forecast. Bloomberg Intelligence is sell-side research. Sell-side research has a structural incentive toward optimism, because optimism moves capital, and moving capital is the business model. I don't say this to accuse anyone of bad faith. I say it because I spent years translating institutional narratives for traditional finance clients, and I learned that the most dangerous number in any report is the one presented without its definition. Is $750 billion annual or cumulative? Does it include operating expenditure or only capital expenditure? Which companies does it actually cover? The crypto brief that carried this figure answered none of these questions, and that silence is more informative than the number itself. A prediction without a denominator is not a forecast. It is a mood.
So where does this leave us, and what should we actually watch? I am not here to tell you AI is a bubble, because that word has become a substitute for thinking. What I am telling you is that the honest question is not whether AI will matter — it clearly will. The honest question is whether the revenue curve can bend upward fast enough to meet a spending curve that is already committed. Watch three things. First, the ratio of AI capital expenditure to actual AI revenue, quarter by quarter. Second, the share of that spending financed by debt rather than equity, because leverage turns a soft landing into a hard one. Third, the grid connection schedules for new data centers, because physics does not negotiate with enthusiasm. If those three signals stay healthy, the cycle holds. If they diverge, the vacuum reveals itself — and it will not announce its arrival in advance. The story was never in the size of the number. It was always in whether anyone could explain where it comes back from.