Crypto underworld in the AI circle

Author: Xiao Bing, Deep Tide TechFlow

In Abilene, Texas, an about 1,000-acre construction site is gradually bringing eight H-shaped data centers into operation.

This is the first large campus of OpenAI’s “Stargate” project. The overall project is planned with 1.2 GW capacity; the first two buildings have already gone live, and the remaining buildings are still under construction.

The park is operated by Oracle, and the developer behind it is called Crusoe, but what it did first was Bitcoin mining.

Its founder, Chase Lochmiller, used to be a partner at the crypto fund Polychain Capital.

In 2018, he and his childhood friend, Cully Cavness, found that U.S. oil fields were burning large amounts of associated natural gas every day that couldn’t be transported. The two moved power generation equipment and mining rigs to the wellhead, using the natural gas that would otherwise be wasted to mine Bitcoin.

The business logic is simple: find power in remote areas, and turn electricity into computing power within a very short time.

Seven years later, the customer switched from the Bitcoin network to OpenAI. In 2025, Crusoe will sell its Bitcoin business, which will exceed 425 modular data centers, to NYDIG, focusing on AI.

Crusoe’s transformation looks like a big leap, but its core capability hasn’t changed: finding power, building data centers, and operating and maintaining them.

Like Crusoe, there are many companies and people that have made a dazzling shift from Crypto to AI, but they are not like the PayPal Mafia, who all came from the same company, and there is no single organization connecting them. What links them are three types of assets left behind by the last Crypto cycle:

• Electricity, land, and grid-connection permits held by mining firms;

• Engineers and founders trained by Crypto companies;

• Capital accumulated during the last bull market.

After 2022, these three types of assets began flowing into AI at the same time.

Mining firms sell AI not mining rigs, but electricity

“Humanity’s biggest contradiction in the future is the growing demand for data processing versus limited computing power.”

In 2019, Wu Jihan, founder of Bitmain, wrote this in an article titled “The Beauty of Computing Power,” and stated directly that this is why Bitmain invested in AI chips.

At the time, those words looked like public-relations language; six years later, looking back, it feels more like a pre-written prophecy.

In February 2026, Wu Jihan’s mining company Bitdeer announced it would clear all its Bitcoin inventory to provide liquidity for building AI data centers, making the shift from Crypto to AI—decisively.

When Bitcoin mining firms pivot to AI, people often understand it as “converting mining rigs into AI servers,” but that’s not the actual situation.

Most Bitcoin mining rigs rely on ASIC chips that can execute only specific hashing algorithms and can’t be used to train large models. Even GPUs left behind at Ethereum mining farms can’t directly meet the requirements that today’s large AI clusters have for networking, video memory, liquid cooling, and reliability.

The mining firms’ truly valuable assets are data centers already connected to the power grid.

When building an AI data center, the hardest part is often not buying GPUs, but securing hundreds of megawatts of stable power, obtaining land, substations, transmission lines, and construction permits. This process may take years. In the past, mining firms had already completed a large amount of upfront work in North America, Northern Europe, and the Middle East—partly to reduce mining costs and for compliance.

As Bitcoin mining profitability fell and AI companies were willing to sign long-term high-price contracts, mining firms naturally started switching customers.

CoreWeave is one of the earliest to complete the transformation.

In 2016, three commodities traders placed a GPU on a pool table in their Manhattan office and started mining Ethereum. After the Crypto winter arrived, they used falling prices to acquire large quantities of used graphics cards, then expanded the business into film rendering and machine learning.

The company originally had a different name, Atlantic Crypto, and later became CoreWeave. In its listing documents, it shows that before 2022, most of the company’s revenue still came from crypto mining; afterward, the crypto business was completely stopped.

Today, CoreWeave is a leading AI cloud company backed by NVIDIA, and its path is being replicated across the entire crypto mining industry.

In 2026, TeraWulf signed a roughly 401 MW, 20-year data center lease with Anthropic, with an initial contract value of about $19 billion;

Cipher Mining signed a 15-year deal with AWS for 300 MW, about $5.5 billion;

Core Scientific will provide large data center capacity to CoreWeave on a long-term basis.

Hut 8 in Texas’ Beacon Point campus continuously signed two 15-year leases, each with a base contract value of about $9.8 billion.

IREN, after reaching a $9.7 billion cloud services agreement with Microsoft, continued to disclose a new contract totaling $2.8 billion in July 2026.

According to CoinShares, as of Q1 2026, listed mining companies have already announced more than $70 billion in AI and high-performance computing contracts. At the same time, per-unit computing revenue from Bitcoin mining once fell to about $30 to $35 per day per PH/s. A batch of mining operations using older equipment or located where electricity prices are higher is already nearing losses.

Mining firms have flipped from computing-power infrastructure in the Crypto era to computing-power infrastructure in the AI era—and they remain on the front edge.

From OpenSea to OpenRouter

Besides mining farms, people in the Crypto industry are also moving to AI.

Alex Atallah is a co-founder and former CTO of OpenSea. During the peak of NFTs, OpenSea’s monthly transaction volume once exceeded $4 billion. In July 2022, Atallah left the company to prepare for a new venture.

In 2023, he founded OpenRouter.

OpenRouter solves a very direct problem: there are more and more large models, and their prices, speeds, and capabilities differ. Developers don’t want to integrate APIs with every model provider. With OpenRouter, they only need to connect to a single interface to call hundreds of models and have requests automatically allocated according to price, performance, and availability.

In 2025, OpenRouter completed a total of $40 million in financing, with a valuation of about $500 million.

In May 2026, it completed a $113 million Series B round led by CapitalG, and its valuation rose to $1.3 billion. In the past six months, the Token volume processed per week by the platform grew from 50 trillion to 250 trillion.

OpenRouter and OpenSea don’t do the exact same things, but the business structure is quite similar.

OpenSea aggregates NFT buyers and sellers; OpenRouter aggregates models, computing supply providers, and developers. The former matches digital asset trading, while the latter matches inference requests. The product changes, but the ability to build a marketplace and integrate fragmented supply remains.

Some Crypto traces even remain directly in the product. Open the registration page for OpenRouter—next to the Google and GitHub login options, MetaMask is still there, and the platform also accepts USDC payments.

Fal.ai is another example.

Founded by Burkay Gur, who previously worked on building a machine learning platform at Coinbase, Fal.ai started in 2021 and initially developed machine learning data pipelines and deployment tools.

After Stable Diffusion went open source, they found that although there are increasingly many image and video models, inference is slow, deployment is troublesome, and GPU utilization is low. So Fal.ai shifted its focus to generative media inference.

This choice paid off quickly.

By mid-2025, Fal.ai’s annualized revenue was already approaching $95 million. In the same year in December, the company completed a $140 million D round led by Sequoia Capital, reaching a valuation of $4.5 billion. Companies such as Adobe, Canva, and Perplexity are all using its generative media infrastructure.

Use Crypto money to support AI

What mining firms provide to AI is electricity and data centers; the capital accumulated during Crypto cycles enters AI in another way.

The most direct example is Jed McCaleb.

He created the crypto exchange Mt.Gox, and later co-founded Ripple and Stellar—among the earliest groups of crypto billionaires.

In 2023, the Navigation Fund backed by McCaleb put up about $500 million to buy 24k NVIDIA H100s in one go, and then established Voltage Park to rent GPUs to AI companies and research institutions.

He didn’t create a new public chain. Instead, he converted the money he earned from Crypto into the scarcest assets in the AI industry.

In 2026, Voltage Park merged with Lightning AI, an AI development platform. The valuation that the deal counterpart provided for the merger entity was about $2.5 billion. The wealth accumulated from the previous crypto cycle thus became an asset and liability on the balance sheet of an AI cloud company.

The investment portfolio left behind by FTX founder SBF—now already blamed for failure—provides an even more dramatic case.

In 2022, SBF invested $500 million in then-not-yet-well-known Anthropic, taking about 13.5% equity. After FTX filed for bankruptcy, the liquidation team sold these shares in batches in 2024, recovering about $1.3 billion. Today, Anthropic’s post-money valuation is $50k. If FTX hadn’t sold the shares, its ownership would have been about 6.7%, corresponding to a value of about $65 billion—roughly 130 times the original $500 million investment.

Cursor’s story is even more extreme.

In April 2022, Alameda—funded by SBF—participated in Anysphere’s early financing with $200k. The company later launched the AI programming tool Cursor. After FTX entered bankruptcy proceedings, the liquidation team sold that stake for $200k in April 2023—essentially buy at cost, sell at cost.

In June 2026, SpaceX announced it would acquire Anysphere for $60 billion in an all-stock transaction. According to published reports, the equity Alameda initially received was about 5%. If you completely ignore dilution from Anysphere’s later financing, the paper value of this stake could reach $250k—15k times the original $200k investment.

Of course, this can’t be explained solely by the idea that SBF is an investment genius. More accurately, before the release of ChatGPT, a subset of the most aggressive, highest-risk-tolerance capital in the Crypto bull market had already started looking for AI projects.

When the Crypto market was booming, much of the capital believed two judgments: computing power would become more and more valuable, and software networks could expand globally in a very short time. AI satisfied both conditions at once.

So after Crypto capital entered AI, what they bought wasn’t just graphics cards—they were also funding new technical and organizational experiments.

Nous Research is a typical case.

Hermes Agent developed by Nous is an open-source AI agent that can accumulate long-term memory and automatically generate skills. According to openrouter’s statistics, Hermes Agent Token usage ranks #1 globally, surpassing Claude Code.

In 2025, Paradigm led Nous Research’s $50 million A round financing.

According to reports at the time, this financing corresponded to a valuation of about $1 billion for its yet-to-be-issued Token. Its prior investors also included crypto VC Distributed Global and former Coinbase CTO Balaji Srinivasan.

Besides Hermes, Nous is also developing Psyche, a distributed model training network built on Solana.

Traditional AI labs need to concentrate massive GPU resources in a single data center. Psyche wants to test another route: connect GPUs scattered across different regions, belonging to different participants, to jointly train models. Then, smart contracts coordinate training progress, verify participants, and allocate rewards.

At this stage, Psyche is still an experiment, and the testnet Token has been explicitly labeled by the official as having no economic value. But it represents another impact after Crypto capital entered AI.

OpenAI’s early team also seriously considered a similar direction.

OpenAI was founded in 2015 as a nonprofit organization, but the funding required for cutting-edge models quickly exceeded what a donation model could support. By the end of 2017, Sam Altman and Greg Brockman had already begun discussing new fundraising structures, one of which was an ICO.

Later disclosed internal emails show the team seriously studied issuing Tokens in early 2018. Musk explicitly opposed it, believing that an ICO would seriously damage OpenAI’s credibility. OpenAI later added that by the end of January of that year, the team itself was gradually losing interest in that plan.

In the end, OpenAI chose to form a for-profit entity and received a large investment from Microsoft, but Sam Altman didn’t leave Crypto.

In 2019, he co-founded Worldcoin together with Alex Blania and Max Novendstern. The project uses an iris recognition device called Orb to verify that a user is a real and unique human, and builds an identity and payments network through World ID and the WLD token.

From Crusoe and CoreWeave to OpenRouter, Fal.ai, and Nous Research—these individual stories filled with survivor bias don’t necessarily mean that Crypto companies’ AI pivots are more likely to succeed.

Mining firms leave behind electricity, land, and grid-connection permits; exchanges and Web3 companies cultivate engineers familiar with distributed systems, GPU scheduling, and global product building; and the wealth created by token price increases becomes capital used to buy graphics cards, invest in model companies, and fund technical experiments.

Crypto didn’t turn into AI out of thin air—it simply channeled resources left behind by the previous cycle into the next industry that needs them more.