Arthur Hayes: Humanity is Wasting Trillions on AI Data Center Construction
Former BitMEX CEO and co-founder of crypto investment firm Maelstrom, Arthur Hayes, stated at the Gamma Prime Investing Conference in Singapore that humanity is wasting trillions on AI data center construction. He believes this round of construction will ultimately make computing power extremely cheap and abundant, leading to oversupply.
Hayes identified end buyers as SpaceX, OpenAI, and Anthropic, stating that all three are currently unprofitable. He predicts that after the data centers are completed, infrastructure providers will collect payments from customers who have committed to computing power, with a key testing window around the end of 2027 or 2028. He also presented a bullish scenario: if AI proves sufficiently useful in the next 12 months and demand keeps pace, these companies could potentially turn profitable.
Hayes expressed skepticism about shorting AI companies, considering it a poor investment opportunity. He noted that profitable companies include storage chip manufacturers and Nvidia, and investors need to assess whether future profit multiples are reasonable. He drew historical parallels, stating that major technological rollouts often lead to overbuilding and crashes, followed by bailouts, citing the aftermath of the 2008 financial crisis and similar situations over the past two decades.
He bets that the excess liquidity from bailouts will be absorbed by Bitcoin and other crypto assets, emphasizing the need for patience. Under the same logic, he is advancing the AI agent payment project Flop, which is set to launch in Q1 2027, aiming to establish a spot market for computing power where participants provide GPUs and earn Flop tokens after completing inference.
Hayes mentioned that there is currently no payment network for AI agents. His bet is that if agents can directly exchange a currency for the computing power they consume, they will use that currency. The related public statements also noted that Flop is designed without venture capital allocations or presales, with an airdrop planned to last 90 days starting in late October, distributing about 25% of the planned supply over 10 years, with the mainnet also targeted for Q1 next year.
This is an event-driven liquidity trade, not a bullish stance on AI revenues. Buyers are betting on bailouts after overbuilding and positioning themselves for cheap computing power, while sellers rely on long-term commitments of unprofitable labs' data center debt and forward computing contracts. Money flows from AI capital expenditures and related credit towards potential bailout liquidity, which Hayes directs towards Bitcoin and the computing settlement network. Profitable chip and storage companies will still benefit in the short term, while unprofitable cutting-edge labs and leveraged data centers will face pressure at the payment points in 2027 to 2028.
Source: Public Information
ABAB AI Insight
Hayes's path has always been about the cash flow of exchanges and macro liquidity. In 2014, he co-founded BitMEX with Ben Delo and Samuel Reed, turning Bitcoin volatility into a fee market with perpetual contracts up to 100x, reaching daily trading volumes in the billions during the peak of 2020 to 2021. In 2022, he admitted to violating the Bank Secrecy Act and failing to establish an effective anti-money laundering program, receiving a two-year probation and a $10 million fine, subsequently shifting his public identity to the family office Maelstrom. His articles from 2022 to 2023 primarily focused on Treasury bond issuance, Federal Reserve operations, and dollar liquidity, rather than project fundamentals.
This capital movement operates on two levels. One level is Maelstrom's allocation logic: AI-related debt absorbs new money, leading to a lack of liquidity buying for Bitcoin, and the money printing triggered by the bubble burst is what he aims to catch. The other level is Flop, which turns GPU inference and agent payments into on-chain spot transactions, profiting from the settlement fees after the revaluation of excess computing power, rather than buying another unprofitable model lab. The motivation is to avoid the payment timing of data center debts and position at the downstream of bailouts and cheap computing power.
Structurally, this is similar to the overbuilding of fiber optics in 2000, rather than just an internet stock bubble. Global Crossing and Level 3 laid far more fiber than demand in the late 1990s, and after their stock prices went to zero, bandwidth prices collapsed, making it cheaper for later companies like Google, Amazon, and video distribution. Hayes places the delivery of data centers in 2027 to 2028 in the same context: OpenAI, Anthropic, and SpaceX are the current demand committers, while Nvidia and storage manufacturers are the shovel sellers, with Flop aiming to create a bandwidth exchange post-crash. The industry phase has shifted from a model competition to a credit commitment period, where the constraint on expansion is no longer chip delivery, but who can pay after delivery.
The essence is a transfer of pricing power. Currently, computing power prices are supported by capital expenditures and long-term leasing contracts, rather than by the real payments of agents. Once commitments face accounting in 2027 to 2028, pricing power will shift from "builders who can borrow money" to "payers who can continuously consume inference." If the government chooses to print money rather than allow credit defaults to clear, pricing power will again shift to assets that can absorb excess fiat currency. Hayes's mechanism judgment is that overbuilding itself creates cheap computing power, and bailouts create excess currency, with neither side requiring AI labs to be profitable first.