The news hit my terminal this morning: Galaxy Digital advancing Phase II of its Helios AI data center in West Texas. The crypto press is calling it a “Crypto-to-AI pivot.” I call it a real estate arbitrage dressed in server racks.
Let’s cut through the PR. Galaxy Digital is a publicly traded financial services firm that happens to own a Bitcoin mine. Helios was originally a 200-megawatt (estimated) mining facility. Now they’re retrofitting it for AI compute. The market is cheering. I’m not buying the narrative without a ledger.
Context
Galaxy Digital’s Helios site in West Texas sits on a massive power grid (ERCOT) with access to cheap wind and solar. The original mining operation was a classic ASIC farm: power-hungry, heat-intensive, and single-purpose. Converting such a facility to AI/HPC (High-Performance Computing) is not a simple swap of GPUs for ASICs. It’s a full-scale infrastructure rebuild.
AI data centers require dense computing clusters, typically NVIDIA H100 or A100 GPUs, with liquid cooling, high-bandwidth networking (800Gbps+), and redundant power distribution. The cooling alone is a different beast. Mining rigs use air cooling; AI clusters need direct-to-chip or immersion cooling. Retrofitting existing power transformers and electrical substations is possible, but the thermal and noise constraints shift entirely.
Based on my experience auditing mining operations during the 2022 bear market, I’ve seen the engineering gaps firsthand. One miner I consulted claimed to “pivot to AI” but after a site visit, I found the power distribution units (PDUs) were rated for 20kW per rack, not the 60-100kW required for GPU clusters. The retrofit cost was 40% of building new. Galaxy likely faces similar realities.
Core Analysis
Phase II construction is underway. That’s a positive signal: it means the dirt is moving and the permits are pulled. But the devil is in the kilowatts per square foot. The original Helios mine was built for ASICs, which are power-dense but not compute-dense in the same way as GPUs. An ASIC miner consumes about 3kW per unit and has a flat layout. A GPU server rack can consume 15-30kW and requires much higher network throughput. The cooling infrastructure must be upgraded to handle the heat density.
The real bottleneck isn’t power; it’s cooling and networking.
Galaxy’s Phase II announcement lacks technical specifics: no mention of GPU count, no cooling technology (air vs. liquid), no network backbone provider. This is a red flag. If they are building a world-class AI data center, they would boast about the cluster size. The silence suggests the project is still in the design phase, not operational.
Let’s compare to peers. Core Scientific signed a 12-year deal with CoreWeave for AI compute. Hut 8 secured a $150M financing for AI data center build-out. Both have disclosed specific GPU counts and power contracts. Galaxy’s Phase II is a construction update, not a revenue commitment.
I didn’t see a single AI client name in the release. That’s the story.
Financial reality: Galaxy Digital is a diversified crypto finance company with trading, lending, and asset management. Their mining segment is a small part of the balance sheet. The Helios conversion is a capital-intensive project with uncertain returns. The bull case is that AI compute demand will outstrip supply for the next three years, and Galaxy can capture a slice of that market by leveraging existing land and power. The bear case is that the operational complexity of running an AI data center is far beyond mining, and Galaxy will end up as a landlord to a hyperscaler, earning thin margins.
Contrarian View
The market is pricing this as a transformational event. I see it as a defensive move. Mining margins have compressed post-halving. Bitcoin’s price is up, but mining difficulty has followed. The “Crypto-to-AI” narrative is a liquidity lifeline for miners who can’t survive on Bitcoin alone. Galaxy is no exception.
But here’s the blind spot: the AI data center market is already crowded. CoreWeave, AWS, Azure, and Google are building at scale. They have long-term relationships with AI companies. Galaxy’s Helios will compete for scraps unless they secure a hyperscaler tenant. The most likely path is a wholesale lease to a cloud provider, which would cap returns at 8-12% ROIC, not the 30%+ that crypto traders expect.
This isn’t scaling; it’s slicing already-scarce AI compute demand into fragments.
Another layer: the Texas power grid. ERCOT is a deregulated market with volatile pricing. During winter storms, prices spike. Mining operations can curtail to sell power back. AI data centers cannot simply shut down without losing customer trust. That means Galaxy will need to secure power purchase agreements (PPAs) or backup generation, adding cost.
From my 2022 Celsius short, I learned that when companies pivot away from their core business, it’s often a sign of distress. Celsius pivoted to institutional lending after retail yields dried up. It didn’t end well. Galaxy is not Celsius, but the pattern is familiar: a company with a weakening core business (crypto trading/mining) chasing a hot narrative (AI). The execution risk is high.
Takeaway
Galaxy Digital’s Helios Phase II is a real infrastructure project, but it’s not a crypto moonshot. It’s a real estate and energy play. The value creation will come from leasing the space to AI tenants, not from mining. The market will eventually realize that Galaxy is becoming a data center REIT, not a crypto native.
Watch for the AI client signings, not the construction updates.
I’ll be tracking the power capacity disclosures and the cooling technology choices. If they announce a liquid cooling installation and a tenant like CoreWeave, I’ll reconsider. Until then, this is a narrative trade, not a fundamentals trade.
The question is not whether Galaxy can build a data center. It’s whether they can run it profitably in a market dominated by hyperscalers. The ledger doesn’t lie. Neither does the P&L.