Public Bitcoin miner dumps Bitcoin for AI, historic mistake
Without a doubt, this is the era of AI. Companies are cutting their workforces in half to use cash flow to invest in hardware, while the stock market remains near all-time highs, thanks in large part to FAANG. OpenClaw, a self-hosted AI agent, has more stars on GitHub than Linux or React, but even Jack Dorsey is taking tough steps to rebuild Block in the face of digital artificial intelligence. But how much of this AI wave is hype, and how many of the companies building its infrastructure will actually benefit?
Public Bitcoin miners in the US are making their own choices, with some investing money in building AI data centers or completely rebranding themselves to distance themselves from Orangecoin. While there are various AI-related focuses and statements made by public Bitcoin miners on this issue, a few stand out as the most extreme.
Cypher Mining, estimated to be worth about $6 billion and one of the largest in the country, has announced a complete rebrand away from Bitcoin and towards AI hype. In its latest investment report, titled “Rebranding to Cipher Digital to Reflect Strategic Shift to HPC,” the company explained why it “sold its 49% stake in Alborz, Bear, and Chief Mining Sites.” Bitfarms Ltd, another large public miner valued at over $1 billion, has also pivoted completely to AI. Coindesk reported that CEO Ben Gagnon went as far as to say, “We are no longer a Bitcoin company,” but the company kept the “Bit” in its name.
Some of these publicly traded companies expect the returns from digital intelligence to be greater than those from Bitcoin, at least in the short to medium term, while others see it as an opportunity to diversify or be too big to pass up.
“The average cost to mine Bitcoin right now is about $87,000. The spot price of Bitcoin is about $70,000. So most of the industry is underwater, and public miners are using that as an excuse to pivot,” Kent Halliburton — co-founder and CEO of SazMining, explained in an exclusive interview with Bitcoin Magazine. Sazmining is a private Bitcoin miner that specializes in frontier energy sources and operates primarily outside the United States.
Halliburton also said, “$87,000 is an industry average and includes those operating older generation rigs on the Texas power grid. At our locations in Paraguay and Ethiopia, our customers receive 100% “We are producing Bitcoin for $50,000 to $64,000 based on energy costs, which is a spot discount of 10% to 30%. Profitability is right there.” That would require a longer investment horizon or cheaper energy, neither of which seem viable for U.S. public miners targeting quarterly reports in dollar terms.
But on the subject of cheaper energy, Halliburton suggests that public miners in the United States had the opportunity to be competitive, but were unable to capitalize on that resource. He didn’t say anything on the topic, saying that these publicly traded companies “had everything they needed to mine Bitcoin cheaply – power contracts, land, infrastructure – but they’re handing it over to Microsoft and Google in exchange for lease checks. They’re doing everything from securing the Bitcoin network to securing rack space for hyperscalers, and they call it a strategy. Meanwhile, they’re releasing over 15,000 Bitcoin from their balance sheets to fund the transition.”
Among the largest public Bitcoin miners, IREN Limited began pivoting to AI cloud services in April 2025, announcing a $9.7 billion, five-year agreement with Microsoft for 200 MW of critical IT workloads using NVIDIA GB300 GPUs. TeraWulf has completed multiple Google-backed HPC expansions through Fluidstack, securing a 10-year contract for over 200 MW.
Cipher Digital has completed a complete rebrand to HPC landlord with 600 MW of contracted capacity, including a 15-year 300 MW lease with AWS and a 10-year 300 MW lease with Google-backed Fluidstack. Hut 8 has signed a 15-year, 245 MW lease with Fluidstack, also backed by Google, with potential for future extensions and first offer rights of over 1,000 MW. Core Scientific has expanded its HPC focus to 270 MW through a partnership with Microsoft and CoreWeave to service OpenAI workloads.
Riot Platforms is strategically evaluating AI hosting expansion by partnering with AMD for a 10-year 25 MW operational lease and 600 MW AI/HPC evaluation at the Corsicana site, but no hyperscaler agreement has been announced.
MARA Holdings is diversifying into AI through a joint venture with Starwood Capital’s Starwood Digital Ventures, targeting 1 GW of IT capacity in the near term, scalable to more than 2.5 GW for hyperscale and AI workloads, with Starwood leading funding and tenant procurement, but no hyperscaler deal has been finalized yet.
CleanSpark pursues AI pivot by acquiring Texas land and power for AI/HPC. It includes 447 acres of land in Brazoria County with a potential of 300-600 MW and a site in Austin County contributing 890 MW of total power generation, and is being funded through Bitcoin sales, with tenant negotiations underway but no hyperscaler leases disclosed.
The AI gold rush is here, there is no doubt about it. Many of these public miners undoubtedly seem to see an opportunity to build infrastructure for profound technology trends. But in any case, in the long run, history has not been kind to those building the infrastructure of a new era. It tends to be a very high risk, medium reward kind of bet. For example, how many of the companies that built railroads are still around today? Or, without going that far back, can you name the companies that built fiber optic internet lines in the late 90s and 2000s?
There is a long list of railroad bankruptcies from the late 1800s that even led to a financial crisis called the Panic of 1873, many of which took on excessive debt to finance construction that was not yet in sufficient demand. After the panic, JPMorgan led the consolidation of bankrupt railroad companies, resolved debt disputes, and placed real estate assets under new ownership. They were the ones who finally took advantage of railroad construction.
And just as the century was coming to an end, the dot-com bubble of the 2000s left behind a graveyard of fiber-optic infrastructure companies that were eventually acquired by hyperscalers like Google and Meta for $1 during post-crash consolidation.
While the construction of railroads and fiber optic lines collectively helped expand global commerce in incredible ways, demonstrating the wisdom of the market as a whole, most individual companies involved did not survive the process, and venture capitalists focused on today’s AI boom are aware of this dynamic.
Gap between capital investment and revenue AI
Various investor groups are beginning to question where the revenue for this massive infrastructure spending will come from. Goldman Sachs argued in an October 2025 report titled “AI: Is We in a Bubble?” that while previous investments may have been supported by strong technology-related returns, valuations for some companies are starting to “bubbly.”
Sequoia’s David Chan points out that the gap between AI revenue and capital expenditures (Capex) has widened since 2023, with a widely reported difference of $600 billion. The hyperscaler’s 2026 capex commitment is over $700, but where is the return?
While OpenAI’s $20 billion in annual recurring revenue (ARR) is impressive for a startup, FuturumGroup reports that it represents “approximately 3% of total hyperscaler capex in 2026,” and “Anthropic’s $9 billion run rate puts it in a similar position, with 9x year-over-year growth. Cohere ($150 million) So is the entire cohort of pure-play AI vendors, including . ARR), Mistral (approximately $400 million), and Perplexity ($148 million annually), among others, will probably account for less than $35 billion of total projected revenue in 2026. ”
Skepticism about where the value of AI can actually be captured is also expressed by venture capitalists like Chamath Palihapitiya. He is a prominent investor in Groq, a company developing custom silicon for the AI era that was licensed by NVIDIA in a $20 billion deal last year, and has been a Facebook insider throughout the company’s growth into a hyperscaler. If he has doubts about the profitability of artificial intelligence railway construction, then perhaps there is something worth looking into very closely.
Palihapitiya also argued on a recent All In Podcast that companies will soon realize that they are exposing their trade secrets to cloud AI and may choose to self-host instead. Building an in-house GPU farm may seem like a bit of a side quest, but are you really willing to risk your trade secrets by working with an AI provider that trains on user data? After all, new versions of models trained on that data will have it incorporated into their knowledge base and exposed to the world. And even if corporate agreements to not use corporate data for training become the norm, it creates a very high level of trust, creating a systemic risk for a given company: the risk that the data could be leaked or viewed by the wrong insiders within the cloud AI provider company.
For the same reason, there is also the question of whether the market fundamentally wants cloud AI. Would you hire a personal assistant if you knew the data you shared with it would end up on the internet? Probably not, but that’s what’s happening with AI. In fact, the Southern District of New York recently ruled that users are not entitled to client-attorney privilege when receiving legal assistance from an AI chatbot, meaning that confidential discussions with the AI can be legally subpoenaed and used against the client in court. This illustrates the risks associated with blindly trusting AI. Some speculate that new types of terms and contracts will need to be created to support this use case. But this case illustrates the fundamental elements of the demand for AI. In other words, people want reliable humanoid intelligence, digital or otherwise.
AI loyalty and trust
Ah, “trust,” the ubiquitous, almost supernatural word that does so much to carry the weight of the world. But what is trust? Fundamentally, it is predictability, a person’s confidence that another person, system, or AI will behave in a certain way in a reliable, predictable, and positive manner toward his or her benefit. However, if the AI is hosted in the cloud, no such guarantees can be made. Data basically leaves the user’s machine and is processed in the “cloud”, and what happens there is beyond our human comprehension. In fact, “the cloud” comes with legal risks that can impede your loyalty as a user in certain scenarios. So perhaps the public is fascinated by OpenClaw.
In recent weeks, a new open source project in the world of AI has taken the tech industry by storm. With 289,000 stars on GitHub, it has more stars than Linux has, despite supporting software infrastructure around the world, and more than React, one of the world’s most popular web development languages. And it’s only been live for how many weeks? How can this happen? Why do people like it so much?
Well, there are probably two reasons. It feels more like a human assistant than a chatbot. It updates automatically, remembers what you are interested in, writes a diary and develops based on your preferences. But most importantly, you can host it on your own machine. People were buying Mac minis en masse to run OpenClaw, combined with Claude Max API token plans for about $200 per month. Some argue that this is a revolution in self-hosting, even though the above setup still relies on the cloud. But what’s really happening here is that OpenClaw appears to be loyal, remembers you, and is “inside” your PC. This is not a chat interface where the context window eventually grows unmanageably large and dies a small death, replaced by a new chat tab. OpenClaw is not a chatbot. This is a type of AI entity that users build relationships with. And good relationships are built on trust.
So what does all this have to do with public Bitcoin miners? Perhaps self-hosted AI is the future, and China’s AI model is increasingly lean and capable of running on machines that are far from cutting edge, perhaps with pressure to innovate due to sanctions on specialized AI hardware such as high-end Nvidia chips. All kinds of open source tools for managing and hosting models locally are being released and improved regularly. If history is any guide, the mass production of AI hardware will lead to the commoditization of powerful computers that can reach end users’ homes and process AI.
In fact, Apple, the worst FAANG of all AI products ever introduced, could end up being one of the biggest winners of the AI race. why? Because the user hardware is better. Modern Macs do not distinguish between RAM and VRAM. This is a problem seen with all other computers that rely on GPUs, such as Nvidia. This limits the size and speed of models that can be self-hosted. Instead, modern Mac machines have all their RAM integrated, allowing users to run powerful models locally that cannot easily be run on non-Apple hardware. Self-hosted AI is the future.
Therefore, public Bitcoin miners who were pursuing medium-term fiat profits may have just fallen into a trap. It’s the same trap that the giants of the dot-com bubble fell into. It’s the same trap that the giants of the industrial age who built the railroads fell into. The infrastructure that powers the future doesn’t necessarily make a profit.
The post Public Bitcoin Miners Are Dumping Bitcoin for AI, a Historical Mistake originally appeared in Bitcoin Magazine and is written by Juan Galt.
