The shareholders of Core Scientific just rejected a $9 billion exit. They didn't just say no to a check—they lit a fuse on a bomb that could either launch the company into the AI infrastructure stratosphere or blow up the remaining value. The trigger? A partnership with AMD that reads more like a strategic bet than a delivered product.
I've spent nineteen years watching capital flows in crypto, from the 2017 ICO fog to the 2022 Terra collapse. Every time I see a company reject a premium buyout in favor of a technology pivot, my first instinct is to trace the liquidity ghosts. Where does the money come from? Where does it go? And who is left holding the bag when the music stops?
Let me be clear: I am not a miner. I am not a data center operator. I am a researcher who models cross-border payment systems—systems that rely on latency, throughput, and settlement finality. And what I see in Core Scientific's AMD partnership is a microcosm of a macro trend: the desperate attempt to repurpose Bitcoin mining infrastructure for AI workloads, without fully accounting for the engineering and economic friction.
Context: The Phoenix and the Chimera
Core Scientific is a Nasdaq-listed company (CORZ) that emerged from Chapter 11 bankruptcy in early 2024. Its core business is Bitcoin mining—operating vast warehouses of ASICs that solve SHA-256 hashes for block rewards. But the halving in April 2024 slashed the block subsidy from 6.25 BTC to 3.125 BTC, squeezing margins. The company needed a new narrative.
Enter AI. The boom in large language models and generative inference created an insatiable demand for GPU compute. Hyperscalers like AWS, Azure, and Google Cloud are building new data centers, but they can't keep up. The bottleneck is not chips—it's power and real estate. Bitcoin miners sit on massive power contracts, often locked in at low rates for years. They also have existing facilities with cooling, security, and network infrastructure.
The logic is seductive: take a Bitcoin mine, rip out the ASICs, install Nvidia H100s or AMD Instinct GPUs, and sell compute to AI startups. It's the same playbook that CoreWeave and Hut 8 are executing. But the devil is in the plumbing.
Core Scientific already has a hosting deal with CoreWeave, signed in 2024. That deal provides a revenue floor. The AMD partnership, announced in early 2025, is supposed to be the next step: diversify the chip supply chain and potentially secure better pricing or terms. But the original article—the source of this analysis—contains no contract value, no minimum purchase commitments, no delivery timeline. Just a handshake and a press release.
Core: The Technical Trap
Let me tell you a story from my days modeling arbitrage in DeFi Summer. I was analyzing Uniswap V2 pools against traditional FX forward markets. I found a 15% risk-adjusted yield advantage in cross-border settlement times. I built a bot, ran it for three weeks, and then abandoned it. Why? Because the operational complexity—the constant monitoring of gas prices, the need to rebalance liquidity, the risk of a single sandwich attack—consumed all the theoretical edge. The model was right, but the execution was a nightmare.
Core Scientific's AMD partnership faces a similar chasm between theory and practice. Converting a Bitcoin mine to an AI data center is not a simple hardware swap. Here are the technical barriers that the press release obscures:
1. Cooling. Bitcoin ASICs are air-cooled. They run at 30-50°C with fans. AI GPUs, especially high-density clusters like the AMD Instinct MI300X, require liquid cooling—direct-to-chip or immersion. Retrofitting an existing facility with liquid cooling loops, chillers, and coolant distribution units is a multi-million dollar capital expenditure per megawatt. The article does not mention any CapEx budget.
2. Networking. Bitcoin miners communicate via a simple pool protocol. The network is a star topology: each miner talks to the pool server. AI training, on the other hand, requires high-bandwidth, low-latency interconnects like InfiniBand or RoCE (RDMA over Converged Ethernet). The nodes must synchronize gradients every few milliseconds. A standard Bitcoin mining facility has a 1Gbps or 10Gbps switch. An AI cluster needs 400Gbps per node. The networking upgrade is a separate infrastructure rebuild.
3. Software Stack. Nvidia's CUDA is the de facto standard for AI compute. AMD's ROCm is catching up, but the ecosystem is still fragmented. Many popular libraries (e.g., PyTorch with FlashAttention) are optimized for CUDA. Porting to ROCm requires engineering effort, and the performance gap, while narrowing, remains. For a company like Core Scientific to sell compute, they must offer a seamless experience. If the customer's code breaks on AMD hardware, the contract is worthless.
4. Power Density. A Bitcoin mining rack consumes about 30-40 kW. An AI GPU rack can consume 50-100 kW per rack, depending on the density. The facility's electrical infrastructure—transformers, switchgear, busbars—must be upgraded to handle the higher per-rack load. This is not trivial. It often requires renegotiating power agreements with utilities, which can take months.
I have seen this movie before. In 2021, I modeled the correlation between Ethereum gas fees and US CPI data. I called NFTs "digital land grabs"—speculative stores of value against fiat depreciation. The same dynamic is at play here: Core Scientific is selling a story of AI infrastructure scarcity, but the underlying asset is a converted mining barn with aging power contracts.
Tracing the liquidity ghosts through the ICO fog. The capital flowing into Core Scientific's stock is chasing the AI narrative, not the mining reality. The $9 billion rejected offer was a liquidity event that would have crystallized value. By rejecting it, shareholders are betting on a future that is far from guaranteed.
Contrarian: The Bear Case That No One Is Talking About
Let me play the structural skeptic. The consensus view is that Core Scientific's pivot to AI is a no-brainer: power is scarce, AI demand is infinite, and the company has a captive asset. But I see three blind spots.
1. The AMD Partnership Is a Distraction. AMD is not the dominant player in AI compute. Nvidia holds over 80% market share. AMD's Instinct GPUs are competitive on paper, but the software ecosystem is a lagging indicator. If Core Scientific builds its AI hosting business around AMD, it will be serving a niche market. The hyperscalers (AWS, Azure, GCP) are buying Nvidia because their customers demand it. Core Scientific's AMD partnership limits its addressable market to cost-sensitive startups or organizations that want to avoid vendor lock-in. That's a smaller pool.
2. The Capital Expenditure Spiral. The article notes that Core Scientific has no disclosed CapEx plan for the AI conversion. But the company retains debt from its bankruptcy restructuring. To fund the conversion, it will likely need to issue new equity or take on more debt. Both options dilute existing shareholders or increase financial risk. The rejected $9 billion offer was a premium precisely because an acquirer could have funded the conversion with cheaper capital. By going it alone, Core Scientific is betting on its own ability to raise funds at favorable terms. In a rising interest rate environment, that's a dangerous bet.
3. The Macro Liquidity Squeeze. I am a macro watcher. The global liquidity cycle is turning. The Federal Reserve is holding rates higher for longer. The M2 money supply growth is slowing. In such an environment, capital-intensive infrastructure projects face higher hurdle rates. The returns on AI hosting—which are essentially a spread between power cost and compute revenue—are sensitive to electricity prices and utilization rates. If the economy slows, AI startups tighten budgets, and compute demand softens. Core Scientific's fixed costs (leased facilities, power contracts) remain, but variable revenue may drop. This is a classic operating leverage trap.
Digital land prices don't reflect the underlying plumbing. The market is pricing Core Scientific as if the conversion is a fait accompli. But the plumbing—the cooling, networking, and software stack—is still being designed. The stock price may be a phantom mirror of the AI hype, not the operational reality.
Takeaway: The True Test Is Delivery, Not Announcements
I have studied enough macro cycles to know that the most dangerous moment in a bull market is when everyone agrees on a narrative. The Core Scientific story is compelling: a phoenix rising from bankruptcy, leveraging its power assets to capture the AI wave. But the narrative is not the business.
What matters is not the AMD partnership announcement. What matters is the number of megawatts of AI-ready compute delivered by Q4 2025. What matters is the utilization rate of those GPUs. What matters is the unit economics: after paying for power, cooling, and capital depreciation, does the company generate positive free cash flow?
Until those numbers are published, the stock is a bet on the management team's ability to execute a complex technical transformation. And I have learned from the 2022 Terra collapse that execution risk is the most underestimated variable in crypto-adjacent narratives.
The bubble breathes. Don't mistake the exhale for a new breath. Core Scientific may succeed, but the path is narrower than the press release suggests. The $9 billion ghost will haunt every quarterly earnings call until the company proves it can deliver real AI infrastructure, not just a press release.
Watch the delivery. Ignore the hype. The liquidity ghosts are always hiding in the details.
