Energy Grid Optimization AI

AI buyers are locking up physical power supply directly

Power is now treated as the AI supply-chain bottleneck. Anthropic agreed to buy commitments for 1.7 GW of heavy-duty generators from OnePWR Solutions in a $3.4B deal, which moves the buyer of last resort from utilities to the model labs. Parallax (3D-printed gas turbines), Bloom Energy (fuel cells) and ERock (onsite microgrids, NYSE IPO filing) are all pitching onsite generation to data centers. Vensum Power frames its near-lossless power conversion as a way to optimize existing grids for AI data center demand. Grid optimization software is being pulled toward the load that needs power fastest.

Weekly totals peaked at $14.5B across 30 deals in the week of Oct 12 and $11.9B the following week. They then fell to $441M in the week of Oct 26 and $1.3B across 11 deals in the week of Nov 9. The theme's velocity reads -0.35 even though 28-day capital is still $18.7B. Series D+ and unknown-stage rounds account for $39B of the 90-day total, so a handful of large checks dominate the numbers.

Why it matters · Headline totals overstate breadth, so investors should look at deal counts and early-stage activity to judge whether momentum is real.

Blackstone and Goldman Sachs are among the top investors by deal count, alongside Nvidia and Sequoia. LS Power agreed to buy six NRG gas plants for $4B in cash. Magnetar led or took part in repeated $500M rounds. Blackstone is expanding in the Gulf and is described as a return-stream provider across sectors. Infrastructure and credit capital is financing generation assets, and optimization software is the add-on layer.

Why it matters · Software that boosts returns on acquired assets gets a built-in buyer in infrastructure funds, but valuations will follow asset-finance logic rather than SaaS multiples.

Series A is the most common stage, with 45 deals and $3.9B over 90 days. An $80M Series A led by B Capital was reported on Nov 10. Heron Power (solid-state transformers, $140M Series B, 50 GW of orders) and Utilidata (power platform for AI data centers) show investors paying up for hardware-software hybrids at the grid edge.

Why it matters · Winning early-stage companies pair software with proprietary power hardware, so pure-software entrants face a higher bar.

General Matter's enrichment push (a $900M DOE contract) and filings such as Deep Fission's $156M IPO target keep nuclear on the agenda. Grid-optimization-specific signals now outnumber nuclear ones in the latest data.

Why it matters · Nuclear capital is increasingly separate from the software-optimization thesis, so exposure to one does not imply exposure to the other.