The Anchor Tenant: AI's Power Bill Just Rebuilt Climate Tech

The wave right now

For a decade, climate tech's problem was the buyer. Utilities move in regulatory time, corporates buy sustainability in goodwill time, and consumers don't buy it at all. The 2023–2026 cohort solved this not by finding a better pitch but by getting a new customer: the AI data center, the only energy buyer on Earth that is desperate, rich, and in a hurry. Voxel Energy exists because grid interconnection now takes two to three years — so they ship solar plus repurposed EV batteries and route around the utility entirely. Madrone exists because chillers are now the constraint on compute density. Even the nuclear startups pitch datacenter demand before they pitch decarbonization. If you're deciding what to build this quarter, internalize this: the climate thesis and the compute thesis have merged, and the compute side is writing the checks.

The landscape today

Powering and cooling the build-out

The most direct play is selling watts and thermal headroom to compute. Madrone claims 30% less power and water than traditional chillers by supplying 20°C water from outside the building; Inviscid AI attacks the same HVAC bill with simulation — 1000× faster than traditional CFD, wired into the BMS. NetworkOcean takes the thermodynamics argument to its logical extreme and puts the data center in the ocean. The pattern: nobody here is selling "green" — they're selling more compute per megawatt, and the carbon win is a side effect. That's the correct order of operations, and it's why this sub-group will outlive any ESG budget cycle.

P26

Madrone

Maximizing compute via hyper-efficient cooling

ActiveData Center Cooling Infrastructure2 ppl

W26

Voxel Energy

Energy independent data centers with solar and repurposed batteries.

ActiveData Center Energy Independence3 ppl

W26

Inviscid AI

Real-time Physics Simulations for Industrial Facilities & Data Centers

ActiveBuilding Physics Simulation AI2 ppl

S24

NetworkOcean

We build and operate underwater data centers.

ActiveIndustrials4 ppl

Software is finally eating the grid

The grid itself is the second beachhead, and the wedge is no longer dashboards — it's decisions. Squid builds AI agents for grid planning against SCADA/EMS systems; Condor Energy puts trading-grade models into enterprise procurement; Dartboard Energy turns raw grid telemetry into audit-ready incident logs. What's changed since the last cycle is that these are workflow-replacement products, not analytics overlays — they sit where money or compliance moves. The earlier entrants here, Rewbi and Atmeto, validated battery-revenue optimization in 2023–24; the 2026 batches moved up-stack to planning and procurement, which is where the harder moats are.

W26

Squid

AI agents for power grid planning 🦑

ActiveAI Grid Planning3 ppl

W26

Condor Energy

Software for enterprise energy procurement.

ActiveEnergy Procurement Optimization3 ppl

W25

Dartboard Energy

AI Analyst for Electricity Markets

ActiveIndustrials2 ppl

Nuclear, top to bottom of the stack

The boldest bet in the cohort is that the nuclear thaw is real, and it's being made at every layer at once. Apollo Atomics is building radically compact reactors targeting a 10 MW commercial pilot by 2028; Terranox AI goes upstream to AI-driven uranium discovery, trained on 70+ years of exploration data; Maritime Fusion targets ships, where diesel's incumbency is weakest and grid alternatives don't exist. Notice the structure: reactor, fuel, and niche deployment are all being founded simultaneously — that's what an ecosystem forming looks like, not a single moonshot.

P26

Apollo Atomics, Inc.

We make the most compact nuclear machines.

ActiveCompact Nuclear Reactors

W26

Terranox AI

The first AI-powered uranium discovery company

ActiveAI Uranium Exploration2 ppl

W25

Maritime Fusion

Fusion reactors for ships

ActiveIndustrials6 ppl

Simulation as the wedge into atoms

The quietest pattern is the most reusable: sell physics simulation as software, and let someone else carry the capex. Silurian builds foundation models for Earth systems starting with weather; Ionworks sells battery simulation to automotive incumbents; Matforge compresses semiconductor materials discovery from years to months with AI agents. These are climate-and-energy companies with SaaS economics and hardware-shaped TAMs — the best risk-adjusted entry point in the whole category for a software-native founder.

S24

Silurian

Foundation models to simulate Earth, starting with weather.

ActiveIndustrials6 ppl

S24

Ionworks

Simulation software used by leading battery and automotive companies

ActiveIndustrials5 ppl

P26

matforge

AI scientists to discover materials for the semiconductor industry.

ActiveAI Materials Discovery2 ppl

The cohort signal

This is accelerating, and it looks deliberate. The Summer 2023–Winter 2024 entries were scattered — utility-bill plumbing (Nectar), battery ML, zero-upfront C&I solar. The 2024–2025 batches added simulation and grid software. Then Winter and Spring 2026 went all-in: of the ten current-cohort companies in those two batches, at least seven — from Matforge and Madrone through Squid and Voxel Energy — are aimed squarely at the AI-energy collision. When a program concentrates that hard in consecutive batches, it's a thesis, not noise. Expect the next batch to push further into nuclear and behind-the-meter power.

Lessons from the last cycle

The prior cohort's outcomes draw the map of what kills companies here. Oklo (Summer 2014) is the existence proof that nuclear patience pays — it's public — but it took a decade-plus; price your runway accordingly. Moxion Power and Sunfolding both died as capital-intensive hardware businesses selling into slow-moving buyers — the cautionary inverse of today's "sell to data centers" discipline. Heimdal shut down selling direct air capture into voluntary carbon markets, and the acquired-not-iconic exits of Pachama and Pina Earth say the same thing more politely: markets built on goodwill produce modest outcomes; markets built on mandates or megawatts produce real ones.

If you're building here

Three concrete openings:

• The interconnection layer itself. Voxel Energy routes around the queue and Squid plans the grid, but nobody in this cohort owns the interconnection process — the studies, filings, and utility negotiation that create the 2–3 year delay. That's a workflow product with a desperate, well-funded buyer on both sides.
• The electrical chain inside the data center, beyond cooling. Cooling is getting crowded; PowerMatrix is nearly alone on power conversion for AI hardware. Everything between the substation and the rack — switchgear, dynamic load orchestration, power quality — is under-attacked.
• Nuclear's missing middle. Apollo Atomics builds reactors and Terranox AI finds fuel, but licensing operations, regulatory tooling, and fuel logistics are unclaimed. Boring, defensible, and every SMR company will need it.

Tarpits to avoid by name: carbon accounting and utility-bill dashboards (Nectar already owns the plumbing, and the last cycle showed the exit ceiling); generic battery revenue optimization (Rewbi and Atmeto got there in 2023–24); anything monetizing voluntary carbon credits — that's the Heimdal lesson.

What you'd have to believe: that AI power demand is structural rather than one capex cycle; that behind-the-meter generation is a permanent architecture, not a stopgap until utilities catch up; and that the nuclear regulatory thaw survives a change of political weather. If you believe all three, this is the rare climate moment where the customer is pulling harder than the founders are pushing.