In the early days of the Ethereum whitepaper translation, I learned a lesson that has never left me: infrastructure carries ideology. The wiring, the cooling, the chip architecture — these are not neutral. They are the physical manifestation of a value system. When I heard that Core Scientific, a Bitcoin miner that filed for Chapter 11 in 2022, had struck a 2.5 gigawatt partnership with AMD, I didn't see a merger of equals. I saw a marriage of convenience between a debt-starved energy broker and a chipmaker desperate to break NVIDIA’s monopoly. But beneath the press release, there is a deeper tremor: the soul of mining infrastructure is being auctioned to the highest bidder, and that bidder is now AI.
2.5 gigawatts. Let that number settle. It is roughly the peak power draw of the entire country of Norway at any given moment. It is enough electricity to run 300 million laptops, or, more relevantly, to operate a cluster of AMD MI300X accelerators capable of training a frontier model every two weeks. Core Scientific is not selling Bitcoin mining capacity anymore; it is selling location, wire, and the illusion of sovereignty. And AMD, after years of chasing NVIDIA's shadow, has finally found a partner willing to bet the farm on its non-CUDA ecosystem.
Context: The Mining Inflection Point
Bitcoin mining has always been a game of energy arbitrage. You find stranded power — hydro in Quebec, wind in Texas, flare gas in the Bakken — and you convert that cheap electricity into the world’s most immutable digital asset. For a decade, the narrative was simple: Bitcoin is digital gold; miners are the refiners. But as the hashprice collapsed in 2022, and as the FTX hangover exposed the fragility of crypto-native balance sheets, a new whisper began circulating in the boardrooms of Riot, Marathon, and Core Scientific: what if the infrastructure we built for SHA-256 could be repurposed for matrix multiplication?
The pivot to HPC was inevitable. A Bitcoin ASIC burns ~30 Joules per terahash, and it can only do one thing: process double-SHA-256 hashes. An AMD GPU, on the other hand, can train Llama 3, render a Pixar film, or simulate protein folding. The same 1 MW of power that yields 130 PH/s of Bitcoin mining yields roughly 500 TFLOPS of FP16 compute. In a bull market for AI, that compute can be rented out at $2-$4 per GPU-hour. In a bull market for Bitcoin, that same 1 MW might generate $100-$200 of daily revenue. The math is stark.
But the math is not the only variable. In 2020, I spent 600 hours auditing the initial Aave V2 interest rate models. I found three critical logic errors that could have drained $4 million. What I learned was simple: code is law, but ethics is soul. The miners are not just selling compute; they are selling trust in a system that has never been tested for AI workloads at scale.
Core: The Technical Moral Hazard
Let’s peel back the 2.5 GW announcement. First, the technical challenge: deploying 2.5 GW of HPC infrastructure requires not only hardware but also a complete rethinking of cooling, networking, and redundancy. Bitcoin mining farms are typically air-cooled, noisy, and tolerant of partial failure. If one ASIC overheats, the pool just reallocates the work. AI clusters, however, are hyper-connected with NVLink or Infinity Fabric. A single GPU failure can stall a training run for hours. Core Scientific will need to retrofit dozens of facilities with liquid cooling, high-speed interconnects (likely 400Gbps InfiniBand), and redundant power feeds that can sustain PUE below 1.1. Based on my audit experience, the gap between ‘stranded power’ and ‘AI-ready data center’ is not a gap — it’s a chasm.
Second, the AMD factor. AMD’s MI300X is a beast on paper: 192 GB of HBM3 memory, 1.3x the theoretical FLOPs of NVIDIA H100. But in practice, the ROCm software stack remains immature. The CUDA moat is not a technical wall; it is an ecosystem of libraries, debuggers, and mindshare. When I helped translate the Ethereum whitepaper into Portuguese, I learned that language is power. CUDA is the language of AI. ROCm is a promising dialect, but dialects don't run production pipelines at scale. If Core Scientific deploys 2.5 GW of AMD accelerators, they are betting that the entire AI industry will eventually learn a new language. That is a bet on human inertia, not on hardware.
Third, the capital problem. Core Scientific emerged from bankruptcy with a restructured balance sheet, but its net debt still exceeds $400 million. Building 2.5 GW of HPC capacity will cost somewhere between $10 billion and $20 billion, depending on location and retrofitting costs. No mining company has ever raised that amount for non-mining purposes. The partnership with AMD itself does not provide capital; AMD sells chips, not equity. Core Scientific will need to tap debt markets, issue new stock, or — and this is where my ears perk up — tokenize the future compute power. I have seen DePIN projects promise to bridge this gap, but the reality is that retail token holders are not institutional lenders. Transparency isn't the oxygen of trust.
Contrarian: The Danger of the Infrastructure Catch-22
The mainstream narrative is optimistic: miners have power, AI needs power, so let the marriage begin. But there is a deeper, more unsettling truth: when a miner becomes an AI landlord, it becomes a landlord of a system that is fundamentally centralizing.
Consider the political economy. If Core Scientific controls 2.5 GW of compute, who decides which models get trained? The same committee that once voted on Bitcoin improvement proposals? Or a small board of private equity investors? The Ethereum whitepaper taught us that decentralization is not just about node count; it is about the distribution of power. A mining farm that is capable of training GPT-6 is a single point of failure for the entire ecosystem of decentralized AI. If that farm is shut down by regulators, or if it goes bankrupt, the AI models that depend on it vanish. This is not resilience; it is fragility disguised as growth.
Moreover, the AMD partnership could create a vendor lock-in paradox. Mining companies have historically championed open hardware — anyone can build an ASIC. But by committing to AMD, Core Scientific is placing a two-year bet on a closed ecosystem (AMD’s RDNA/CDNA architecture). If NVIDIA later offers a better price or performance, Core Scientific cannot switch without ripping out all the power infrastructure. This is the opposite of the modular, open-source ethos that the crypto community claims to cherish.
Takeaway: The Soul of the Machine
I am not a Luddite. I believe that the integration of Bitcoin mining infrastructure with AI compute is inevitable and, in many ways, efficient. Stranded power should not go to waste. But I also believe that the moment we stop asking "who controls the compute" is the moment we surrender our digital sovereignty.
Core Scientific and AMD are not villains. They are rational actors in a market that rewards speed over reflection. But as someone who spent years translating the philosophical underpinnings of decentralization, I can only watch with quiet alarm. The herd is stampeding toward a future where the same companies that once mined the hardest money now rent out the most powerful machines. And unless we build true decentralized compute marketplaces — where anyone can buy and sell GPU time without rent-seeking intermediaries — the 2.5 GW agreement will be remembered not as a breakthrough, but as the moment the elephant sat down in the village.
We need more than a press release. We need an open-source benchmark of ROCm against CUDA, conducted by independent auditors. We need a public commitment from Core Scientific that the compute will be available to small developers, not just hyperscalers. And we need a regulatory framework that treats large compute clusters as critical infrastructure, not just power loads.
Until then, I will keep my miner’s hands clean. Code is law, but ethics is soul. And the soul of this deal is still up for grabs.