Hook: A Data Anomaly in the Boardroom
Shareholders of Core Scientific rejected a $9 billion acquisition offer. Simultaneously, they announced a partnership with AMD. These two events should not be read as separate. They are the same signal: a vote against selling at a fixed price, and a bet on a future that is not yet built. The $9 billion figure is now a psychological anchor. Every future decision will be measured against it. But the AMD partnership, presented as a strategic leap, carries a different kind of weight. It is a promise of transformation—from a Bitcoin mining giant to a high-performance computing (HPC) colossus. I have spent years inside the guts of DeFi protocols, but this is not a protocol. This is a physical plant. And physical plants obey different laws.
Context: From Mining to Mission Control
Core Scientific emerged from bankruptcy in early 2024 with a plan. The company had spent years building massive Bitcoin mining facilities—sites with hundreds of megawatts of power, robust cooling, and a culture of operational reliability. After the crypto winter, the narrative shifted. Bitcoin mining alone was not enough. The infrastructure—the land, the power purchase agreements, the substations—could be repurposed for AI data centers. The market loved this pivot. Core Scientific signed multi-year hosting contracts with CoreWeave, a pure-play AI cloud provider. Then came the acquisition offer. The details of who offered $9 billion are not public, but it was likely a major player seeking to acquire the entire infrastructure portfolio. Shareholders said no. The board said: we can do better. Then they announced AMD.
Core: The Technical Anatomy of a Pivot
Let me be clear: converting a Bitcoin mining facility into an AI data center is not a trivial software upgrade. It is a massive engineering challenge. From my years auditing DeFi bridges, I learned that the most dangerous assumptions are the ones that look like progress but are actually shortcuts. The same applies here.
Power Density and Thermal Dynamics
A typical Bitcoin mining site uses air-cooled ASICs. Each unit draws about 3.5 kW. The power density is low, and the heat is manageable with fans and evaporative cooling. An AI data center running NVIDIA H100 or AMD Instinct MI300X GPUs draws 30–40 kW per rack. Some racks hit 100 kW. That is a 10x increase in power density. The existing cooling infrastructure cannot handle it. Core Scientific will need to retrofit with liquid cooling—direct-to-chip or immersion. That requires tearing out the old racks, installing new piping, and dealing with coolant distribution units. This is not a plug-and-play upgrade. It is a capital-intensive rebuild.
Networking: The Silent Killer
Bitcoin mining nodes communicate over a simple network. They do not need low latency. AI training, particularly for large language models, requires InfiniBand or RoCE (RDMA over Converged Ethernet) with microsecond latency. The switch fabric must handle east-west traffic at hundreds of gigabits per second. Core Scientific’s existing network infrastructure was designed for a different workload. To upgrade, they need to run new fiber, install new switches, and reconfigure the entire topology. This is a civil engineering problem disguised as a networking problem.
Software Stack: The AMD Gambit
AMD’s Instinct GPUs are powerful. The MI300X has 192 GB of HBM3 memory—more than NVIDIA’s H100. But the software ecosystem is the real battlefield. NVIDIA’s CUDA is the entrenched incumbent. Every major AI framework—PyTorch, TensorFlow, JAX—is optimized for CUDA. AMD’s ROCm is catching up, but it is not a drop-in replacement. The gap is real, and it eats into total cost of ownership. If Core Scientific hosts AMD-based clusters, their customers will need to port their workloads. That is friction. The best hardware in the world is useless if the software stack is brittle. From my experience integrating AI oracles with blockchain consensus, I know that small software incompatibilities can cascade into 30% overhead. Trust is not a variable you can optimize away. Neither is software compatibility.
Power Contracts: The Real Asset
Core Scientific’s true moat is its long-term power purchase agreements (PPAs). These are contracts signed years ago, locking in electricity rates at $0.03–$0.05 per kWh. In a world where spot power prices are $0.08–$0.12, that is a 2x advantage. AI data centers are power-hungry. The margin is in the watts. If Core Scientific can deliver AMD GPUs at a 30% lower power cost, they win. But the PPAs are not transferable to the new load easily. The utility may need to approve a new capacity request. The existing contract may have a minimum take-or-pay clause that locks the customer into a fixed power draw. Renegotiating these contracts is a legal and regulatory minefield. I have seen similar issues in the DeFi space where smart contract parameters could not be changed without community governance. Here, the governance is the local utility commission.
Contrarian: The AMD Partnership Is a Distraction
Everyone is focusing on the AMD partnership as the main event. I think the real story is the failed sale. Shareholders rejected $9 billion because they believe the company is worth more. That is a bold statement. But it also sets an expectation. The company must now prove it can generate value beyond that number. The AMD partnership is a tool, not a solution. The real value driver is the existing infrastructure—the power, the land, the operational expertise. AMD is just a vendor. The partnership could be as simple as a purchase agreement with no exclusivity, no revenue sharing, and no technical integration. The press release did not disclose any of those details. That is a red flag. In my experience auditing DeFi protocols, the projects that announce partnerships without concrete milestones are the ones that later fail to deliver. The absence of detail is a data point.
Furthermore, the AI hosting market is becoming crowded. CoreWeave, the original partner, is now a competitor. The same infrastructure that Core Scientific built for CoreWeave could be used to host Core Scientific’s own AMD clusters. But that creates a conflict of interest. The shareholders who rejected the sale are betting on management execution. But management has not yet proven they can run a profitable AI hosting business. The history of Bitcoin mining companies pivoting to AI is mixed. Riot Platforms announced a similar pivot and then struggled to attract customers. Hive Blockchain has had more success, but they focused on a specific niche. Core Scientific is entering a market where the incumbents have years of networking experience, established customer relationships, and optimized software stacks. Promising to deliver AI compute is not the same as delivering it.
Takeaway: Watch the Metrics, Ignore the Press
Over the next six months, the only numbers that matter are: (1) megawatts of AI-ready capacity deployed, (2) utilization rate of that capacity, and (3) the average revenue per megawatt. If Core Scientific can show 100 MW of AI capacity with 80% utilization, the stock will soar. If they deliver 20 MW with 50% utilization, the $9 billion rejection will look like a mistake. The AMD partnership is a bet on the future of compute. But the physics of power and cooling are unforgiving. Code executes. Intent diverges. Infrastructure trusts no one. I will be watching the quarterly filings, not the news releases. The real story is written in the kilowatt-hours.
