Arthur Hayes, who co-founded the cryptocurrency exchange BitMEX and now runs the investment firm Maelstrom, told an audience in Singapore that the money going into artificial intelligence is being thrown away. Humanity is "wasting multi-trillion dollars" building data centres, he said at the Gamma Prime Investing Conference on 7 October.
His argument rests on a pattern rather than a forecast of the technology. "If you study financial history and you study every single major technological rollout, it always is overbuilt," he said. "There always is a crash, and there always is a bailout."
What follows the bailout is the part that interests him commercially. Governments printing money to rescue the builders would push that money somewhere, and Hayes expects a share of it to reach digital assets.
"Thankfully, we have bitcoin and other crypto to soak up that excess liquidity," he said.
He also put a date on the stress. The bills for computing capacity that companies have already committed to buy will fall due in late 2027 or 2028, he said. That is when the data centres being built now come into service and the bills arrive.
What The Number Actually Covers
The phrase "multi-trillion" needs a frame, because the companies doing the spending publish their own figures every quarter. Amazon, Microsoft, Alphabet and Meta have told investors they will spend roughly $732B between them in 2026 on buildings and equipment. Alphabet alone has moved from $91.4B spent in 2025 to a plan of $195B to $205B.
Not all of that is AI, and the trackers say so. Amazon's number carries warehouses, delivery robots and satellites alongside servers, so a reader treating the combined total as a pure AI figure would be overstating it.
Even so, the direction supports the scale Hayes describes, because four companies are spending close to three quarters of a trillion dollars in a single year. The guidance was raised twice during 2026, and a multi-year total reaches trillions without anyone exaggerating.
The Case For Overbuilding
The sharpest version of the overbuild argument is not about demand at all. It concerns how long the hardware lasts, and therefore how much of its cost is charged against profits each year.
Michael Burry, the investor known for betting against American mortgages before the 2008 crisis, put a figure on it in November 2025. He argued the largest AI spenders would understate depreciation by $176B between 2026 and 2028. Depreciation is the yearly charge a company books for its equipment wearing out. By the end of that period, he said, Oracle would be overstating earnings by 26.9% and Meta by 20.8%.
The mechanism is an accounting choice. Microsoft and Oracle now spread server costs over six years, up from shorter periods, and Alphabet and Meta over five and a half to six. Each of those extensions lowers the yearly charge and raises reported profit, in Microsoft's case by about $3B a year.
One company has moved the other way, and Amazon shortened the assumed life of its AI servers from six years to five in 2025. It said the change, the first such reversal after years of lengthening, reflected the pace of technology development in artificial intelligence.
What Cuts The Other Way
The evidence does not point in one direction, and the rental market is the clearest counterweight. Nvidia's H100 chip, which launched in 2022, was renting for $3.39 an hour in September 2026, up 14% on the year before.
Hardware that holds its price four years after release is not hardware that has been wasted. The operators also move chips down a ladder as newer ones arrive, running the latest silicon for training and the older stock for inference, which is the routine work of answering users.
Burry's own figure has a gap underneath it, since he published no methodology for the $176B. The claim rests on an assumption that AI servers lose their economic value in three to four years, and the rental prices are currently contradicting it.
Hayes left himself the same opening. The alternative, he said, is that AI turns out to be so useful within twelve months that the companies building it become profitable. The spending would then not be waste at all.
Flop Is Built For The Aftermath
The reason to read the call alongside the position is that Hayes is constructing a business that benefits if he is right. He has come out of retirement to lead Flop Labs. The company is building a shared ledger that no single company controls, on which autonomous software agents buy computing power, inference and memory.
The design pays suppliers in its own digital token, called FLOP. The suppliers who run the hardware process requests from AI agents and are paid in FLOP, while checkers confirm the work was done and store the agents' data. The agents spend the token to buy capacity and keep what they have learned.
Timing and distribution are both set. The token giveaway is planned for the fourth quarter of 2026 and the network for the first quarter of 2027. Hayes says that there will be no early sale of tokens to investors and that they go instead to participants who do useful work.
A marketplace of that kind is cheapest to fill when computing power is abundant and its owners are under financial pressure. That is exactly the condition Hayes forecasts for late 2027, which puts his market call and his product launch on the same calendar. The same bet on demand arriving later sits behind every data centre built ahead of it.
The Record Behind The Call
Hayes is a prolific forecaster, and his hit rate has been examined. The crypto publication Protos reviewed 20 of his market calls in May 2025 and found 16 had failed, two had succeeded and two were still open, with seven tokens he had favoured down between 28% and 89% for the year.
He has said publicly that being wrong does not trouble him, which is a consistent position for someone who publishes often and in public. It also means anyone judging this particular call has a published hit rate to set against it.
His legal history is on the public record too. Hayes pleaded guilty in 2022 to violating the Bank Secrecy Act over BitMEX's lack of customer identity checks, and received two years of probation. President Trump pardoned him and his co-founders in March 2025.
Where The Argument Stands
Nothing Hayes said is unusual as a reading of financial history, and the overbuild pattern he describes has appeared in railways, telecommunications and fibre. The disagreement is over whether the AI buildout has reached the overbuilt stage of that pattern. Published evidence is genuinely split: an accounting debate implies overstated profits, while a rental market implies the hardware still earns.
What the next eighteen months will produce is harder data than either side currently has. Depreciation schedules will either hold or be shortened again, the capacity contracted this year will either fill or sit idle, and Flop will either find cheap computing power to trade or launch into a market where nobody is selling cheaply.
