Core Scientific’s announcement to expand compute capacity from 500 MW to 2.5 GW, paired with AMD’s warrant acquisition, is being framed as a blueprint for post-halving mining survival. The narrative is seductive: a bankrupt miner reborn as an AI infrastructure partner. But beneath the surface, the deal’s true mechanics reveal a high-risk bet on repurposing stranded assets, with AMD holding the call option.
Context: The Migration from ASIC to GPU
Core Scientific emerged from Chapter 11 in January 2024, shedding debt and pivoting from pure Bitcoin mining to high-performance computing (HPC) hosting. The company’s existing sites—built for power-hungry ASICs—offer cheap electricity and existing cooling infrastructure. AMD, facing GPU supply constraints and rising demand for AI training, needs physical capacity to deploy its MI300X accelerators. The deal gives AMD access to over 500 MW of power immediately, with the option to scale to 2.5 GW as Core builds new facilities.
The transaction is structured as an equity-linked partnership: AMD receives warrants to purchase Core Scientific common stock at market price, effectively tying hardware procurement to a long-term capital commitment. On its face, this is a textbook win-win: Core secures a marquee customer and validates its hosting strategy; AMD locks in compute without building its own data centers.
Core: The Structural Vulnerability in Asset Repurposing
Let’s audit the economics. A standard Bitcoin mining container runs at roughly 25 kW per rack, consuming 600 kWh per day. An AI HPC rack—filled with GPUs, networking, and liquid cooling—can draw 40–60 kW, sometimes higher. The power infrastructure at Core’s Texas and Kentucky sites was designed for ASICs’ predictable load profile, not the bursty, thermal-sensitive demand of GPU clusters.
Based on my experience modeling power curves for arbitrage scripts during the 2017 ICO era, I can tell you that retrofitting an ASIC facility for GPUs costs between $500,000 and $1.5 million per megawatt in cooling and electrical upgrades. That’s a capital expenditure that Core must finance before seeing AI hosting revenue. The company’s Q2 2024 cash reserves were $180 million—enough for roughly 120–150 MW of conversion, not 2.5 GW. The rest will require debt or equity dilution beyond the AMD warrants.
The warrants themselves are the hidden tax. AMD can buy shares at the prevailing market price with no strike premium. That means if Core’s stock rises on AI optimism, AMD exercises and captures the upside. If the stock falls, AMD simply uses its warrants at a lower price, locking in a discount. The asymmetry is entirely in AMD’s favor. Core’s existing shareholders bear the dilution risk without the corresponding upside protection.
And here’s where the structural vulnerability bites: AI compute margins are compressing fast. CoreWeave, Lambda Labs, and even AWS are flooding the market with rented GPU time. The spot price for H100 compute has fallen from $4.50/hour to below $3.00/hour in six months. If Core signs long-term contracts at today’s rates, margin erosion will eat the cash flow needed to service the conversion debt.
Contrarian: The Short Thesis No One Is Discussing
Retail investors see AMD’s logo and assume a new revenue engine. The smart money sees an option structure that transfers value from Core’s shareholders to AMD. The warrants are essentially a free call on Core’s transformation—AMD pays no premium for the right to participate in upside. In 2020, during the DeFi rug-pull summer, I shorted protocols with similar token structures that gave insiders asymmetrical upside. The pattern is familiar.
Furthermore, the narrative ignores regulatory tail risk. Bitcoin mining is under increasing ESG pressure in the U.S., but AI compute—especially training large models—faces even steeper carbon scrutiny. Core’s power sources in Kentucky rely heavily on coal. A carbon tax or stricter EPA guidelines on data center emissions could render the entire 2.5 GW pipeline uneconomical. I’ve seen this before: in 2022, when Terra collapsed, I shorted LUNA derivatives after analyzing on-chain stablecoin flows. The early signal was always a mismatch between narrative and physical constraints.
There’s also the counterparty risk. AMD is not obligated to fill Core’s facilities. The warrants can be exercised at any time, but the commitment to purchase compute capacity is a master services agreement, not a take-or-pay contract. If AI demand softens, AMD can throttle its orders. Core would be left with 2.5 GW of mothballed capacity and a bloated balance sheet.
Takeaway: Watch the Revenue Split, Not the Headlines
Alpha isn’t leverage. It’s structural foresight. The only signal that matters for Core Scientific is the percentage of revenue derived from AI hosting in the next two quarters. If that ratio exceeds 35% and gross margins stay above 40%, the thesis holds. Below 25%, the conversion cost will destroy shareholder value.
We do not chase pumps; we engineer the squeeze. In this case, the squeeze is on Core’s management to execute a capital-intensive pivot with a partner who holds the cards. I’ll be watching the SEC filings for the warrant exercise terms. And remembering: yield is not free. Someone is always paying the risk.
Survival in this market isn’t about mining the next block. It’s about mining the next margin.