Arthur Hayes On The Early Days Of BitMEX, Where Bitcoin Is Going, And Why We Are Over-Indexing On AI
Summary
BitMEX emerged from Hayes’s experience of institutional fragility and an underserved derivatives market. His first trading-desk day coincided with Lehman’s bankruptcy; after Citi let him go in 2013, he found Bitcoin, traded arbitrage and founded BitMEX in 2014. Spot exchanges were already crowded, while derivatives remained primitive enough that traders could double their money on a three-month cash-and-carry trade—provided the exchange survived.
BitMEX found product-market fit by offering 100X leverage, then invented the perpetual swap to simplify a product customers did not understand. The initial conservative futures contract lost to Chinese venues offering 20X; better technology let BitMEX jump to 100X and become cash-flow positive as Bitcoin rose from roughly $130 to $600 in late 2015. After the first perp design failed, the May 2016 version concentrated liquidity and used a self-referencing funding mechanism designed to “punish people who are going with the trend” and attract market makers against it.
Hayes views BitMEX’s closure as both sad and liberating because custody was “a terrible way to live.” The exchange reached roughly 250 employees and, he said, exited without ever losing customer funds. Its product remained widely used as crypto, in his framing, took over TradFi. He declined to explain the disposition of an insurance fund with “30-something thousand Bitcoin” left in it, calling the matter “TBD.”
Hayes remains overwhelmingly positioned for crypto’s long-term monetary thesis rather than short-horizon trading. “Ninety-something percent” of his net worth is in crypto, although he keeps enough real estate, gold, stocks and T-bills to be fine if it goes to zero; Maelstrom’s speculative directional positions generally run one to six months. Trading remains an intellectual game whose objective scorecard is money made or lost.
Hayes is bearish on AI stocks but believes AI will be useful; his objection is that customers are not paying the technology’s full economic cost. He argued that perhaps 95% of users would leave if pricing covered training, inference and profit, citing Anthropic’s stated 2025 ratio of $3 of cost for every $1 of revenue. Meanwhile, data-center inputs become scarcer and more expensive as the commodity they produce—compute—gets cheaper, creating an “upside down chart.”
His base case is that the AI build-out becomes too systemically important to fail, forcing monetary support that ultimately benefits crypto, including his long-Bitcoin thesis. Hayes assigned a 75% probability that Anthropic, SpaceX, OpenAI and similar companies will not be able to afford their five-year compute commitments without government assistance because demand will not reach profitable pricing at sufficient scale. Avi challenged the benign political reading, warning that an AI bailout could provoke a sharp leftward swing; Hayes answered that leaders of every ideology eventually choose another bailout over personally crashing the economy.
The apparent contradiction in Hayes’s AI view powers his new bet: failed AI economics could make agents abundant while creating demand for a compute-linked currency. Flop Network would reward GPU operators for useful inference and let agents convert Flop directly into their “life force,” compute. If trillions or quadrillions of agents each retain a small convenience balance, Hayes believes the network could become “the largest cryptocurrency ever, bigger than Bitcoin”—an explicitly speculative claim resting on cheap surplus compute and explosive agent growth.
Deep dive
1. Institutional failure pushed Hayes toward Bitcoin and entrepreneurship
Hayes grew up in Buffalo, lived in Detroit and decided around 2000 to study Chinese and move east after reading Ugly Americans. A University of Pennsylvania minor, six months in Hong Kong and a junior-year sales-and-trading job at Deutsche Bank became a 20-year career across Greater China; he has never held a job in America.
His first day on a trading desk was the Monday Lehman Brothers filed for bankruptcy. Management was naming banks employees could no longer trade with, while new graduates wondered whether their offers would survive; one managing director bluntly asked Hayes, “Do you think you’re gonna have a job in three months?”
The crisis destroyed his generation’s expected banking trajectory. When Citi let him go in 2013, Hayes read the Bitcoin white paper, absorbed its philosophy and began arbitraging it—turning the desire to “do my own thing” into the idea for BitMEX.
Early crypto trained suspicion as much as trading skill. A recurring lure involved a supposedly famous buyer seeking 10,000 BTC and intermediaries imagining a six-figure commission; Hayes avoided being caught, but remembered such implausible private deals as “rampant.”
2. BitMEX won by taking a derivatives risk incumbents would not
Hayes founded BitMEX in 2014 with Ben Delo, who built the trading system, and Samuel Reed, who built the web architecture. Spot already had Mt. Gox, Bitstamp, Coinbase and large Chinese exchanges; Hayes’s pitch was that “derivatives” were where exchange economics remained open.
ICBIT provided both inspiration and warning. During Bitcoin’s 2013 move from roughly $250 to $1,300, a December–March future traded about 100% above spot, allowing a hedged trader to double money in three months; when the market corrected, weak liquidation and margining left the venue with a hole that a large Cumberland trader helped cover.
BitMEX’s first quarterly inverse BTC/USD future was too conservative. Traders preferred liquid Chinese venues offering 20X, so BitMEX adapted socialized-loss concepts, made its own margining changes and used stronger technology to offer 100X—its decisive differentiator.
The result arrived after a nearly volumeless first year: as Bitcoin moved from around $130 to $600 in late 2015, BitMEX became cash-flow positive. It was not yet the largest venue, but it finally offered something competitors could not.
3. The perpetual swap began as a customer-support fix
When BitMEX developed the perpetual swap in May 2016, customers were calling the exchange a scam because they did not understand expiry dates, futures basis or even two-sided order books. The team wanted one instrument that concentrated liquidity and reduced confusion: “We just wanted to make our lives easier and have less customer support requests.”
The first perp failed because funding referenced Bitfinex dollar and Bitcoin rates that updated too slowly while crypto rallied, leaving the derivative at a huge premium. BitMEX replaced it with a self-referencing funding mechanism using a premium index and an eight-hour lookback, making trend-followers pay enough to attract counterparties and collapse the basis.
Hayes rejected the idea that the mathematics was the hardest part. The non-trivial problem was asynchronously margining everyone tick by tick; Delo’s experience with the bank trading language KDB gave BitMEX an architecture competitors found technologically difficult to reproduce until Binance, FTX and others launched comparable products around early 2020.
Avi pointed to Hyperliquid, Lighter and CME’s interest as evidence that perps had spread widely. Hayes said he was proud of what BitMEX had created. His jab at CME’s supposed staffing obstacle to 24/7 trading: “It’s a database, man”—a firm trading quadrillions in notional should not be technologically constrained by weekends.
4. Closing BitMEX removed a custody risk no fortune could justify
BitMEX reached roughly 250 employees, but Hayes described holding billions in customer assets as intolerable: one mistake could create a multibillion-dollar loss. Closure was therefore sad and happy—BitMEX had made money, created an enduring product and exited with “no customer funds ever lost.”
Asked about the remaining insurance fund, Hayes offered only “TBD. No comment,” while estimating that “30-something thousand Bitcoin” remained. He did not clarify the ultimate ownership or planned disposition.
Avi recalled that Bitcoin fell only about 5% after the October 2020 US action against BitMEX, then recovered—a signal to him that a major rally was coming. Hayes’s view from inside was simpler: everyone opposing entrenched finance expected government action and hoped not to become the scapegoat. “Unfortunately, that was us.”
Today Hayes runs Maelstrom’s directional book, mostly long-only, with speculative holdings lasting one to six months. His personal wealth remains “90-something percent” crypto, balanced by enough real estate, gold, stocks and T-bills to survive a total crypto loss.
5. AI may be useful while the current build-out wastes capital
Hayes’s distinction is categorical: he values AI as a product but thinks its current businesses “suck.” Markets should require customers to cover delivery costs plus profit; if frontier developers charged fully for training and inference, he estimates perhaps 95% of current users would not pay.
He identified Anthropic, SpaceX and OpenAI as the main developers of frontier AI. His evidence is negative unit economics: Anthropic said in 2025 that producing $1 of revenue cost $3, while he said Anthropic and OpenAI could require $50–$100 billion of annual debt issuance. He also tentatively cited what he recalled as a roughly $40 billion SpaceX bond sale two days earlier.
Avi’s pushback—worth keeping: data-center investment is meant to reduce compute costs. Hayes’s rebuttal was that this worsens the investment mismatch: labor and materials become more expensive as build-out competes for scarce resources and energy, while the resulting compute becomes an ever-cheaper commodity. “You have an upside down chart in terms of the business.”
The dot-com comparison did not change his conclusion. AI may prove “extremely useful” over time, but investors can still pay the wrong price and finance supply that end customers do not economically demand.
6. An AI bailout is Hayes’s 75% base case—and a crypto catalyst
Hayes puts a 75% probability on the major AI companies failing to afford their contracted compute without government help. The condition is specific: end demand does not reach pricing and scale sufficient for positive unit economics before the companies’ five-year compute commitments are fulfilled.
The exposure extends beyond US developers to Asian chip, hardware and raw-material suppliers. If the promised trillions of dollars of compute go unused, the global economy falters and politicians face backlash over why perhaps $2–$3 trillion went into “steel and cement” instead of healthcare, universities, water or food.
Avi argued that a bailout of an already divisive industry could push the US sharply left and potentially sweep crypto into the backlash. Hayes invoked 2008: Obama talked accountability, but Hayes said that when leaders face a choice between continuing bailouts and accepting a 30%–40% economic contraction, they choose to kick the can down the road.
Hayes sees the same constraint in China. The 2009–2018 property and infrastructure boom created losses that the Communist Party cannot politically allocate among local governments, banks and the central government, so the system prolongs the status quo. His conclusion was that voters and leaders favor continued printing over the contraction that would stop crypto rising, though he admitted, “I don’t know when” the endpoint arrives.
7. Financialization cannot fix inequality, but compute-backed money might coordinate agents
Avi defended broader market ownership, citing Trump Accounts and the proposed $1,000 for every US newborn. Hayes called that stake immaterial when the 1% already owns the relevant stocks, and argued that new money raises scarce-goods prices: “I got an extra one thousand dollars, but my rent went up fifty percent.” Trump, AOC and Mamdani differ rhetorically, he argued, but all face incentives to print money for supporters.
On Zcash, Hayes likes the privacy narrative but acknowledged minting concerns. He thought formal verification might make its way into the protocol in Q4 that year and said his firm was an investor in Zodle, which he called one of the leading Zcash applications.
Flop Network targets a different coordination problem: agents have no dominant payment currency, yet require scarce compute to exist. GPU miners would perform useful inference, receive Flop block rewards and transaction fees, and give the currency a direct conversion path into floating-point operations.
Hayes’s analogy is calories: humans retain money because it can reliably buy food, while agents may retain Flop because it buys compute. Cheap spot compute from overbuilt data centers could support fleets of increasingly capable agents; if trillions or quadrillions hold even small balances, he sees a path to a network larger than Bitcoin.
He originated the idea, recruited a former BitMEX trading-engine engineer as CTO and a 20-year banking contact as CSO, then supplemented the three-person team with agents. “I would never do this without AI,” he said, while stressing that experienced humans must direct those agents responsibly.
8. The presidential pardon was a lottery ticket, not a thriller
Hayes described seeking a Trump pardon as a conventional legal process. After Trump signaled interest in people who had been wronged by legal action, lawyers prepared a brief explaining why Hayes deserved relief and submitted it through the appropriate channels.
There were no clandestine cash drops or Watergate meetings in his account: many people want pardons, one person determines what gets signed, and “it either gets signed or it doesn’t.” Hayes’s summary of the process was “literally that banal.”