Points of Focus

• Bernstein says power access, not chips, is now the biggest AI infrastructure bottleneck.
• Bitcoin miners control over 27 gigawatts of planned power capacity and have already signed $90+ billion in AI-related deals.
• Converting mining sites into AI-grade data centers can cost up to $1.6 million per megawatt and take six months per module.

A fresh wave of AI infrastructure deals has thrust Bitcoin miners into Wall Street’s spotlight. In a client note reportedly titled “Bitcoin Miners: Google-Blackstone Neocloud News – Follow the Gigawatts,” Bernstein analysts argued that listed miners (publicly traded Bitcoin mining companies) have become critical suppliers of the resource now constraining AI expansion the most: grid-connected electricity. 

The note followed Google and Blackstone’s May 18, 2026 announcement of a joint AI cloud venture targeting CoreWeave’s market. CoreWeave is one of the largest independent AI cloud providers, supplying Nvidia-powered compute infrastructure to major AI firms, including OpenAI, Meta, and Anthropic. Backed by $5 billion in Blackstone equity, the venture plans to deploy Google TPUs through a compute-as-a-service model, targeting 500 megawatts of capacity by 2027 and up to $25 billion in total compute investment, including leverage. For Bernstein analyst Gautam Chhugani’s team, the deal reinforced a thesis already gaining momentum. 

Bernstein just went bullish on Bitcoin miners amid $90B+ in AI data center deals.

Outperform ratings for IREN, RIOT, CLSK & CORZ.

Miners' power capacity is now a massive AI advantage.

Power is the new oil. pic.twitter.com/nLkP2i7neK

— Crypto Banter (@crypto_banter) May 19, 2026

Electricity, not silicon, becomes the AI bottleneck

Bernstein’s core argument is reportedly that power, not chips or capital, has become the main bottleneck in AI data center expansion. Securing a single gigawatt of grid capacity in the US can take more than 40 months, even in development-friendly markets like Texas. Against that backdrop, Bitcoin miners collectively control over 27 gigawatts of planned power capacity, giving them access to a rare, pre-energized infrastructure that can’t be replicated quickly.

The firm estimates miners have already inked $90+ billion in AI-related deals covering 3.7 gigawatts of capacity, with approximately one-third involving hyperscalers directly and the rest tied to neocloud providers. Bernstein assigned Outperform ratings to IREN (Iris Energy), Riot Platforms (RIOT), CleanSpark (CLSK), and Core Scientific (CORZ), while keeping MARA Holdings (MARA) at Market Perform.

I like these insights from an exec at Schneider Electric (key player in data center buildouts).

"
AI hardware is rewriting the data‑center energy equation. One rack of NVIDIA’s Blackwell 2 now draws roughly 180‑200 kW, driving whole‑site demand from ~300 MW to 1.2‑1.5 GW—a…

— Ben Bajarin (@BenBajarin) August 4, 2025

Deal flow cements miners’ strategic position

Recent deals highlight how quickly miners are moving into the AI supply chain. 

On May 7,2026, IREN committed to a 5-gigawatt AI expansion on Nvidia’s AI factory architecture, backed by a $3.4 billion, five-year GPU cloud services agreement under which IREN will provide Nvidia with managed GPU cloud services. The partnership also gives Nvidia a five-year option to buy up to $2.1 billion in IREN shares. 

Meanwhile, in January 2026, Riot Platforms signed an AMD co-location agreement under which it’ll lease 25 MW of power and data center space for AMD’s AI and HPC workloads, with an option to scale to 200 megawatts at its Rockdale, Texas, facility.

The valuation gap remains striking, with miners holding AI contracts trading at roughly $6 million per planned megawatt, double the $3 million valuation implied for pure-play Bitcoin miners. Yet the sector still trades at an estimated 90% discount to established AI data center operators on enterprise-value-per-megawatt metrics. 

This reflects investor skepticism over whether mining companies can successfully execute large-scale AI infrastructure conversions and secure long-term enterprise demand. That said, Bernstein argues that demand for shovel-ready, energized land will remain strong, and Bitcoin miners control an outsized share of it.

$IREN $247.8M net loss was heavily impacted by a $140.4M asset impairment, which relates to decommissioning bitcoin mining equipment to clear space for new AI servers.

Iren is retrofitting its bitcoin mining operations because the company can generate higher revenues and… pic.twitter.com/YjHrq7zC56

— Ray Myers (@TheRayMyers) May 9, 2026

Power is the moat, but retrofitting is the hidden cost

While Bernstein views grid access as miners’ key advantage, the transition requires far more than power alone. Most mining sites use air-cooled ASIC racks at lower power densities with flexible uptime. AI and HPC workloads require comparatively more sophisticated infrastructure, including liquid cooling, low-latency networking for GPU clusters, Tier-3/4 redundancy, and near-continuous uptime backed by SLA commitments. 

Retrofitting existing mining campuses into AI-grade facilities can cost roughly $1.2 million to $1.6 million per megawatt, and take four to six months per module. Those engineering costs and deployment timelines may start slowing some early 2026 conversion projects, tempering the sector’s plug-and-play narrative even as investor expectations for AI revenue continue rising.