'What Not To Do'—A Bitcoin Miner Bets Against AI's Giant Data Centers

FORBES ·

Chinese auto-lender-turned-Bitcoin-miner Cango Inc. is making another bold pivot, launching its AI inference subsidiary, EcoHash. While many miners lease power to hyperscalers for AI training, Cango is taking a different path. It's betting on distributed AI inference, utilizing its network of smaller 10-50 megawatt sites, which are unsuitable for large training clusters but perfect for low-latency inference. Cango aims to partner with independent miners, providing AI tech and clients, creating a symbiotic relationship. The company sold most of its Bitcoin and reduced debt to fund this shift, despite market skepticism about the AI bubble and the financial demands of the transition. Cango believes the "long tail" of smaller sites represents a significant, untapped market for AI compute. "From outside, people will definitely be thinking that the company is crazy," said Juliet Ye. "Who are they? They're making this bold move, and they do not know anyone in this industry." She was talking about the day a Chinese auto-lending firm spent hundreds of millions of dollars to become a Bitcoin miner. That was a year and a half ago. Now it is doing the same thing in reverse. Almost every public Bitcoin miner is rushing to lease its power to the hyperscalers building AI's giant training clusters. Cango Inc. is walking the other way. Cango (NYSE: CANG) is on its third life. It listed in New York in 2018 as China's only US-listed auto-financing platform. In November 2024 it agreed to acquire about 50 exahash of Bitmain rigs and became a pure-play Bitcoin miner. Then, on April 13 this year, it launched an AI-inference subsidiary called EcoHash, with its own software layer, EcoLink. No AI training. No giant new data centers. Just a bet that the small, scattered miners the hyperscalers can't use are where a lot of AI compute will end up. "What not to do is as important as what to do," said Ye, Cango's senior director of communications. She comes back to that line again and again. It is the whole strategy in nine words. Ye says the company never set out to mine Bitcoin at all. It set out to own energy. She would know the history. She has spent eight years at Cango, after the Wall Street Journal and the consulting firm FTI. The story she tells starts with cars. Cango took an early stake in Li Auto, the Chinese electric-vehicle maker, before it went public. When Li Auto listed in 2020, Cango booked a fair-value gain of about 3.3 billion yuan, roughly $508 million, and an appetite for the power business underneath the cars. By 2023 it was scouting energy projects in Australia and the Middle East. "During a trip in the Middle East, to look for solar projects, the management bumped into Bitmain," Ye said. That is how an auto lender met Bitcoin mining. What clicked was not the coin. It was the wiring. "All these mining sites are basically, literally, just energy infrastructure," Ye said. "The only reason there are mining farms is because they use the energy and turn the energy to coin. We can still turn energy into other things." Mining was a way in. "We are not thinking of doing Bitcoin mining from day one. We're thinking of running energy infrastructure from day one." Getting in was expensive. Cango paid $256 million in cash for 32 exahash of Bitmain machines in November 2024, then took another 18 exahash in stock that closed the following summer, the shares going to a company run by a former Bitmain finance chief. To escape its "China concept stock" label, it sold its entire domestic auto business for about $352 million. It brought in crypto-native leadership, including a new chief executive and a chairman who founded Antalpha, a financing firm tied into the Bitmain world. By mid-2025 the lender was gone. A miner stood in its place. Cango is not the only miner running for the AI exit. The math of mining has met the math of AI, and both businesses fight over the same thing: electricity. "AI HPC's future might be Bitcoin mining's past," Leo Wang, a Canaan executive, said on the On The Margin podcast. In 2021 miners were the villains, blamed for burning power. Now that same power is the prize. "It is all energy play," Wang said. "We think in the future, energy will be a scarcer asset for everybody." What miners hold that AI labs crave is not chips . It is a plug. A new substation and a long-term grid contract can take years to land. "When hyperscalers are looking for someone who can supply them short-term guaranteed power, they turn to Bitcoin miners, because the Bitcoin mining companies have already put money and secured power," Wang said. The miners, he added, "got lucky" that AI showed up just as block rewards got thinner. The timing tracks the cycle. "We have been following the four-year cycle unbelievably well," crypto investor Michael Terpin said on the On The Margin podcast. After each halving, mining margins tighten, and operators go looking for a second way to make money . Markets have run with it. Core Scientific was an early mover, leasing capacity to AI cloud CoreWeave, and miners from IREN to the firm once called Bitfarms have followed. "Crypto mining warehouses are quietly switching to AI inference, and pulling in around four times the revenue," the analyst behind the @0xCristal account wrote on X. "A GPU warehouse makes more serving LLM inference than mining blocks." This is where Cango breaks from the herd. The popular move is to turn a few huge sites into AI training campuses and sign one long lease with a hyperscaler. Cango said no to that. "We are definitely not doing AI training," Ye said. "That sector is already crowded with hyperscalers. It's not realistic for us to compete." The decision came out of the company's own shape. Cango works with more than 30 sites around the world, most of them 10 to 50 megawatts. Too small for a hyperscaler chasing 100-megawatt campuses. But, Ye argues, just right for the other half of AI. "For AI inference, you have to be distributed. You have to be close to your clients to lower the latency," she said. "Ten to 50 megawatts is too small for hyperscalers, but it's perfect for AI inference." Then she gets to her favorite number. "Over 70% of the power in the mining sector is actually owned by individual players, smaller sites," Ye said. "Only 30% is controlled by those public miners." Those small operators own land and power. They don't own the AI technology, the customers, or the financing. Cango wants to bring all of it. "We are offering them a symbiotic relationship. We come to the sites, we bring the AI playbook, and they own the land, they own the power," she said. "If one thing can make Cango stand up in the next three to five years on the AI front, it's the symbiotic relationship between these smaller sites." EcoLink is the glue. One small site can't match a hyperscaler's always-on uptime, so Cango spreads the reliability around instead. "If one side is down, we can direct the workload to another site, in milliseconds," Ye said. The buyers, so far, are what she calls the long tail . GPU marketplaces like Runpod and Vast.ai. Distributed inference clouds like Zenlayer. AI startups too small to sign a hyperscaler's terms. Price is the hook: a top provider might charge several dollars per GPU per hour, and a marketplace rents the same chip for under a dollar. None of the early test clients took an exclusive deal, Ye said, and most renewed anyway. "The customer demand is definitely real." Cango has not quit Bitcoin. It still runs about 31.7 exahash, which brought in $98.4 million of mining revenue in the first quarter. That is the cash that keeps the lights on while the company raises money for AI. "Most miners just drop Bitcoin mining for good," Ye said. "For us, it's more a hybrid approach." The cleanup was brutal. "We're basically clearing the decks," Ye said. "Investors might want to invest in our AI pivot, but they do not want their money used to pay the old debt." So Cango sold 6,451 Bitcoin for around $442 million and cut long-term debt from $557.6 million to $30.6 million in a single quarter, a 94.5% drop. Its coin hoard shrank to about a thousand. Then it raised $75 million tied to the EcoHash launch. The first AI node is going into a 50-megawatt site Cango owns in Georgia, bought last August for $19.5 million. Ye calls it a "living showroom." Two or three more are due by year-end. Not everyone is sold. "People are a little bit cautious about it," Wang said of the AI rush, "because people are worried about a bubble." The story is running years ahead of the revenue. Retrofitting a warehouse full of fans into a liquid-cooled AI data center costs a fortune. Plenty of miners have spiked on a press release and nothing more. The one once called Bitfarms jumped hundreds of percent on its AI rebrand before it booked a dollar of AI revenue, and analysts who track the pivots keep warning that the money needed to finish them runs into the billions. Bitcoiners have a different worry. As miners switch off rigs, the network's hashrate has slid, and some say the security cost is being waved away. "Bitcoin miners are abandoning the network for AI money," one widely shared X post warned. Cango's own cushion is thin. It had just $7.2 million in cash at quarter's end after the debt purge, and at least one outlet has questioned its standing on the NYSE. Even the marquee deals wobble: CoreWeave's $9 billion bid for Core Scientific fell apart earlier this year. Ye's answer is the discipline that runs through everything she says. The mega-sites and the marquee training leases will go to the giants. Cango is betting on the rest: the thousands of megawatts spread across small, independent miners, the power the giants can't easily touch. That, she thinks, is where a lot of AI inference will quietly run.

DYAX Investor Sentiment

Bullish (Long) 53% · Bearish (Short) 47%

287 participants

Related News

원문 보기 — FORBES