Bitcoin Miners Now Control the Scarce Input in AI: Power Capacity, per Bernstein

The AI buildout isn’t being throttled by capital or chips—it’s being choked by electrons. That’s why the planned Google–Blackstone AI cloud company, which would deploy Google’s custom chips with Blackstone investing $5 billion for a majority stake, puts an unexpected group in the driver’s seat: Bitcoin miners.

The constraint everyone underestimates is interconnection. Securing a single gigawatt of grid-connected power can take more than four years in many states. Meanwhile, miners have quietly assembled a pipeline exceeding 27 gigawatts of planned power capacity across the U.S. In an AI market starved for ready-to-energize sites, that inventory behaves like a call option on the future of compute.

Miners are already monetizing that option. According to Bernstein, the industry has unveiled more than $90 billion in AI-related contracts spanning 3.7 gigawatts of capacity. Roughly a third of these agreements are with hyperscalers; the rest sit with “neoclouds,” the independent AI compute providers that Big Tech is increasingly partnering with—or racing to replicate.

A few transactions show how the playbook is evolving: - IREN signed a $3.4 billion deal with Nvidia, including a $2.1 billion equity commitment tied to GPU deployment—an alignment mechanism that encourages both sides to scale coherently. - Riot Platforms struck an AI colocation agreement with AMD, signaling demand beyond a single silicon vendor. - Core Scientific and HUT 8 have landed major cloud customers of their own, reinforcing that miners can win across client types.

The core advantage isn’t just cheap megawatt-hours; it’s shovel-ready, grid-integrated megawatts plus land, permitting, substation capacity, and operational muscle. Converting ASIC-centric facilities into GPU-friendly data halls isn’t trivial—power density, cooling (air-to-immersion transitions), and network fabric all need upgrades—but miners are structurally closer to the finish line than greenfield entrants sitting in interconnection queues.

There’s a psychological shift at play too. Many investors still frame miners as levered bets on hash price. In reality, the best-positioned operators are morphing into “power landlords” with compute optionality: mine Bitcoin when margins are attractive; re-rate sites toward AI when long-term contracts surface; or blend both to smooth cash flows. Equity-linked partnerships (like Nvidia’s stake in IREN) further validate the pivot and may compress cost of capital as counterparties underwrite scale.

The business incentives here are strong, but not without friction. Communities and regulators are asking harder questions about grid stress, land use, and who should get priority access to scarce electrons. If policy tilts toward industrial load curtailment, miners need demand response programs that protect uptime SLAs for AI clients while maintaining grid-friendly profiles. Execution risk remains real: upgrading to data center-grade redundancy, fiber, and liquid cooling at scale is costly and time-bound, and power price volatility can erode contract economics if hedging is sloppy.

Still, the strategic positioning is hard to ignore. Whether hyperscalers internalize neocloud capacity or continue to outsource, the common denominator is energized, expandable power. For now, miners hold a disproportionate share of that asset.

On the equity side, Bernstein leans constructive on several names with this exposure: outperform ratings on IREN (price target $100), Riot Platforms ($25), CleanSpark ($24), and Core Scientific ($24), and a market-perform on MARA Holdings ($23). I’d expect dispersion to widen as the market distinguishes between miners with mature interconnection rights and retrofit-ready campuses versus those with paper pipelines and limited financing.

If you care about the AI cycle, track three tells: the maturity of miners’ interconnection queues, the quality and duration of their AI contracts, and their ability to balance Bitcoin mining with stable, high-IRR colocation. The next phase of AI won’t be won by whoever buys GPUs fastest; it will be won by whoever controls where the electrons land.