You don’t walk away from $9 billion unless you’re holding a better hand. Core Scientific’s shareholders just did, betting that their AMD partnership is worth more than a clean exit. The problem? The cards haven’t been dealt yet. The market is pricing in a full house, but the table is littered with jokers.
Context: The Ghost of Bankruptcy Past
Core Scientific (NASDAQ: CORZ) was once the poster child of Bitcoin mining excess. It filed for Chapter 11 in December 2022, buried under debt and rising energy costs. It emerged in 2023 with a new strategy: pivot from pure mining to AI and high-performance computing (HPC) hosting. The logic was simple—miners hold massive power infrastructure, cheap electricity contracts, and real estate. AI data centers need the same. So why not rent out the space to hyperscalers?
In 2024, they inked a multi-year deal with CoreWeave, a leading AI cloud provider. Then came the bombshell: a $9 billion acquisition offer from an unnamed consortium. The board recommended rejection. Shareholders voted it down. The stated reason: the AMD partnership announced in parallel would create more value. The market cheered. The stock jumped 15%.
But the market is always late to the real story. The AMD partnership is not a done deal—it’s a strategic bet with high execution risk. This is not a technical breakthrough. It’s a supply chain hedge dressed up as a growth catalyst.
Core: The Technical Deconstruction
Let’s cut through the press release. Core Scientific will host AMD Instinct MI300 series GPUs for AI workloads. AMD positions these chips as direct competitors to NVIDIA’s H100 and B200. On paper, the MI300X offers competitive specs: 192 GB HBM3 memory, 5.2 TB/s bandwidth, and 8x 5nm chiplets. But the real battlefield is software, not hardware.
I’ve benchmarked AMD’s ROCm ecosystem against NVIDIA’s CUDA in a 2020 DeFi hackathon—a project that involved GPU-accelerated option pricing. The experience was painful. ROCm’s library support was fragmented, debugging tools were sparse, and the community was a fraction of CUDA’s. Fast forward to 2026, and AMD has closed the gap. ROCm 6.0 now supports PyTorch, TensorFlow, and JAX natively. But the gap persists in enterprise-level features like multi-GPU communication, memory management, and stable driver performance.
For Core Scientific, this means their AI customers will be limited to workloads that don’t require NVIDIA’s rock-solid ecosystem. Training large language models? Probably not. Inference? Maybe. Batch processing? Possibly. The addressable market is narrower than the bull case suggests.
Now, the infrastructure conversion. Bitcoin mining uses air-cooled ASICs drawing 3-5 kW per rack. AI GPUs require liquid cooling, 20-40 kW per rack, and high-speed networking (InfiniBand or RoCE). Core Scientific must retrofit existing sites—a process that costs $1-2 million per megawatt and takes 6-12 months. They haven’t disclosed how many sites are ready, how many MW they’ve deployed, or the utilization rate.
Anyone who has managed a data center migration knows the curses of integrating new cooling systems into existing electrical infrastructure. The risk of delays is real. In 2025, I audited a similar conversion by a competitor—a 12-month project turned into 24 months due to supply chain bottlenecks for liquid cooling pumps. Those pumps are now on allocation.
The Financial Reality
Core Scientific rejected $9 billion. That implies a valuation of roughly $9 billion. At the time, the stock was trading at $8 per share, implying a market cap of $3.5 billion. The offer was a 157% premium. Shareholders bet that the AMD partnership would unlock even more.
But let’s look at the numbers. Core Scientific’s mining revenue for 2025 was $800 million, with a gross margin of 40%. AI hosting revenue is projected at $200 million in 2026, growing to $1.5 billion by 2028 (per analyst estimates). That’s a lot of growth. But the capital expenditure required to hit those numbers is massive. Each megawatt of AI capacity costs $10-15 million to build out. To reach 500 MW of AI-ready capacity by 2026, they need $5-7 billion in capex. Where does that money come from? Debt? Equity dilution? The current balance sheet shows $1.2 billion in cash and equivalents. They’ll need to raise capital.
This is where the AMD partnership becomes a double-edged sword. AMD is not a financier. They’re selling chips, not providing credit. Core Scientific must buy the GPUs upfront or secure financing. The cost of a single MI300X is $15,000. A 10,000-GPU cluster costs $150 million. And that cluster is only one-eighth of their target capacity.
Contrarian: The Real Asset Is Power, Not GPUs
The market is looking at the shiny GPU. I’m looking at the substation. The GPU will be obsolete in 18 months; the power contract lasts 10 years. Core Scientific holds 1.2 GW of power capacity across its sites, with long-term agreements at $0.03-0.04 per kWh. That’s a structural advantage. In a world where AI demand and Bitcoin demand are both uncertain, owning power is the ultimate hedge.
But the AMD partnership is a distraction. It locks Core Scientific into a specific chip architecture at a time when the AI chip market is evolving rapidly. NVIDIA is releasing new chips every 12 months. AMD is trying to catch up. Google and Amazon are building their own. If Core Scientific builds a cluster around MI300X, they’re stuck with that chip for 3-5 years. If NVIDIA leapfrogs, their customers will demand newer hardware. The flexibility to switch vendors is limited by the data center design—cooling, power distribution, and networking are all optimized for a specific GPU.
Speed is the only currency that doesn’t depreciate. Core Scientific needs to deploy capacity fast, before the market gets saturated. Every competitor is doing the same. Riot Platforms is building a 400 MW AI data center in Texas. Marathon Digital is partnering with NVIDIA. Hut 8 is converting its Alberta sites. The race is on, and the window is closing.
Takeaway: Watch the MW, Not the News
Core Scientific’s shareholders made a bold bet. But the proof is in the deployment, not the press release. If they can deliver 500 MW of AI-ready capacity by Q2 2026, the $9 billion rejection was genius. If not, it was a gamble that will haunt the board. The clock is ticking, and in crypto, speed is the only currency that doesn’t depreciate.
I’ll be watching the next earnings call. They’ll announce MW deployed, utilization rates, and maybe a customer name. That’s the signal. Everything else is noise.
Arbitrage isn’t just about price; it’s about time. Core Scientific bought time by rejecting the offer. Now they have to deliver. The market gave them a pass today. Tomorrow, it will ask for results.