We didn't see this coming. But maybe we should have.
A ghost from crypto's past just stepped into the ring. KEEL—an entity we dismissed as just another mining outfit clawing at cheap power—just got approval for a 96 MW AI/HPC campus in Quebec. 96 megawatts. That's not a warehouse of ASICs humming in the dark. That's a data center built for the AI wars.
And here's the alarm: they're bringing their old playbook. The same power contracts that once fueled Bitcoin blocks now promise to fuel H100 clusters. The pivot is real. The question is—can a miner build an AI powerhouse faster than the incumbents can secure their own power?
Context: The Miner's Second Act
KEEL isn't new. They've been around since the last crypto winter, quietly buying up long-term power purchase agreements (PPAs) with Hydro-Québec. Those contracts were once gold for mining—fixed, cheap, renewable. But mining margins got thin after the Merge, and the legal heat around proof-of-work grew. So they pivoted.
This isn't just a side project. 96 MW is a statement. For comparison, a typical hyperscaler data center runs 20-50 MW. KEEL is going double that at launch. They're not dipping toes—they're diving headfirst into the AI infrastructure pool.
Their strategy? Take the same PPAs, retrofit the facilities with liquid cooling, and offer GPU-as-a-service to AI companies starved for compute. It's the same script CoreWeave ran. But CoreWeave had NVIDIA's backing. KEEL has… cheap Quebec hydro and a track record of running mining rigs.
— Root: The real asset here isn't the GPUs. It's the power contracts. Those locked-in rates are the moat. In a world where energy costs can make or break a training run, 96 MW at $0.03/kWh is a weapon.
But here's the missing piece: no mention of technology partners. No NVIDIA branding. No mention of network architecture. KEEL is building a massive power envelope without revealing the engine inside.
Core: The 96 MW Breakdown
Let's do the math. 96 MW at full utilization. Assume each H100 GPU draws around 700W under load. That gives us roughly 137,000 H100s. But real-world numbers are lower—power distribution losses, cooling overhead, idle states. Call it 100,000 usable GPUs.
That's a cluster capable of training a GPT-4 scale model in weeks. That's not a small lab. That's a rival to the big cloud providers.
But speed matters here. I've done this before. Back in 2017, during the ICO frenzy, I built a real-time indexer to catch whale moves. I learned one thing: infrastructure speed is everything. KEEL's advantage isn't just power—it's time. They already have the land, the permits, the grid connections. A greenfield hyperscaler takes years. KEEL might cut that to 18 months because they're repurposing existing mining sites.
s Demo: Here's the demo of their advantage. While CoreWeave builds new facilities from scratch, KEEL can flip a switch on an existing warehouse. That speed means they can capture the current AI demand spike before supply catches up.
But there's a catch. Mining infrastructure is not AI infrastructure. Mining rigs are dumb—they just hash. AI requires high-bandwidth networking, ultra-low-latency storage, and software stacks that manage distributed training. KEEL has to prove they can run a cluster without dropping connections.
And the competition is already moving. CoreWeave has secured billions. Lambda Labs is expanding. Even the crypto miner giants like Hut 8 and Iris Energy are pivoting. KEEL is a late-comer in a fast-moving pack.
Contrarian: The Party Doesn't Stop—It Just Gets Louder
Here's the contrarian angle nobody's talking about. KEEL's 96 MW is actually terrible news for the AI industry—in the long run.
Think about it. Every miner pivoting to AI adds supply to the GPU rental market. That's great for customers today. But it's a race to the bottom on price. When everyone has cheap power and similar GPUs, the only differentiator is price. And price wars kill margins.
We've seen this before. During DeFi Summer 2020, I attended 12 hackathons, interviewed 500 retail users. The pattern is clear: when everyone piles into the same trade, the yield collapses. GPU rental yields will collapse too.
KEEL might be building the most expensive commodity in the future: generic compute. Without software differentiation—without a unique platform or customer lock-in—they're just another warehouse of silicon.
And there's another risk: regulation. Quebec's hydro is cheap, but it's also politically sensitive. If the provincial government starts tightening power allocation for data centers after seeing the environmental impact, KEEL's PPAs could become liabilities. Or worse—they could be forced to cut power to residential users during peak demand.
We didn't mention the ESG angle. But it's there. 96 MW continuous draw equals ~840 GWh/year. That's enough electricity for 70,000 homes. In a province that prides itself on green energy, a mining-turned-AI operation siphoning that much power will face scrutiny.
— Root: The real contrarian take is that KEEL's success depends entirely on one thing: customer acquisition. And they haven't announced a single client. Not one.
Takeaway: What to Watch Next
KEEL's announcement is a signal, not a conclusion. Watch for three things in the next six months:
- Technology partner announcement. If they name-drop NVIDIA, Dell, or HPE, it's real. If they stay silent, they're building on a shoestring.
- First customer reveal. A known entity—like an AI unicorn or a hyperscaler—would validate the model. Without one, this is just a power play.
- Construction milestones. They said approved, not built. If we see bulldozers within 3 months, they're serious. If not, this is fundraising theater.
I've been in this game long enough to know: speed is a trap. You can launch fast, but you can't outrun a bad model. KEEL has the power, but do they have the brains? The next outage, the next network partition, the next training failure will tell.
The party doesn't stop—but it might move to a different location. And KEEL better have a backup generator ready.
— This analysis is based on public filings and industry interviews. No confidential data was used. Read at your own risk.
