Cinderwood

The Cinderwood Thesis

On the power constraint underneath AI compute, and the capital that clears it.

Nicholas Roche · Managing Partner · 2026

§ 1 — Silicon gives way to electrons. Exhibit 1 maps where the pressure sits.

§1 The Bottleneck Moved

For twenty years, software scaled against a grid that barely noticed. A data center was a rounding error on a utility's load forecast, and compute planning meant negotiating with chip vendors rather than transmission planners.

AI training loads ended that. Campuses now arrive asking for hundreds of megawatts at a time, in markets where interconnection queues run for years and large transformers are quoted in lead times that would have sounded absurd a cycle ago. The binding constraint on compute has moved from silicon to electrons.

Exhibit 1 · Where the queue pressure sits
Interconnection queue pressure
Active capacity seeking interconnection, by state
Projects in queue ~10,300
Generation queued 1,400 GW
Storage queued 890 GW
Queued GW <30 → 400+
Active generation and storage capacity seeking grid interconnection, by state, year-end 2024. Glow intensity scales with queue depth. Compiled from Lawrence Berkeley National Laboratory queue data. Hover a state for detail.

Total installed US capacity is roughly 1,200 GW. The queue holds more than that, waiting on a study process measured in years. That claim would have needed defending in 2023. Now the market defends it for us.

***
§ 2 — The tape is already pricing the shortage.

§2 The Market Is Already Clearing

Watch where the largest buyers of compute are putting capital and the thesis stops being a prediction.

YearTransaction
2025 Google acquires Intersect Power for $4.75B. A hyperscaler buying a power developer outright; the offtake agreement stopped being enough.
2025 Form Energy signs 30 GWh of iron-air storage with Google. Multi-day storage contracted at a scale that only makes sense if firm power is the product being bought.
2026 Bell anchors a Saskatchewan AI campus with CoreWeave and Cerebras. Compute siting to stranded Canadian generation; the load is moving to the electrons, not the other way around.

These are the transactions we write about every week. Each one prices the same shortage.

***
§ 3 — Quarters against decades. Exhibit 2 is the spread over time.

§3 Two Clocks

The market wants gigawatts on a software timeline. The grid delivers them on an infrastructure timeline. Hyperscalers plan in quarters, utilities plan in decades, and the spread between those two clocks is where this entire cycle gets decided.

Exhibit 2 · Demand against firm additions
Firm power
Grid metrics · weekly
Queue depth 2,600 GW
DC contracted 48 GW
Avg queue wait 5.0 yr
Hyperscale demand Firm additions BESS dispatch Curtailed
Q1'24Q2Q3Q4Q1'25Q2Q3Q4Q1'26Q2Q3Q4Q1'27e
Hyperscale power demand against firm additions, storage dispatch, and curtailment, by quarter, 2024–2027e. Bars left of the now line are printed data and hold still; bars right of it are projections and breathe with forecast uncertainty. Illustrative series pending sourced data.

Every serious answer to that spread (behind the meter generation, modular gas, nuclear restarts, storage stacked against data center load) is a company somebody is building right now. Nearly all of them need two things at once, engineering that works and a capital stack that closes.

***
§ 4 — The scarcest product in the transition, and the name of our Sunday letter.

§4 Firm Power

A training cluster buys every hour of the year, including the winter peak and the still summer night. Average electricity is worthless to it. The product it actually needs is called firm power, and firm power is the scarcest asset in the entire energy transition.

Whoever can deliver reliable electrons at scale, on schedule, decides who gets to train.

Firm Power is also the name of our weekly letter, because the whole thesis compresses into those two words. Every Sunday we cover the deals, the projects, and the capital moving through the power and compute stack.

***
§ 5 — Deal flow that starts before the deck exists.

§5 The Edge

Our deal flow starts earlier than a deck. Alongside the fund, Cinderwood runs a capital advisory practice (cinderwood.capital) that sits inside live raises across the power and compute stack, while terms are still forming and before rounds are rounds. The fund underwrites with information a cold pitch never carries.

Both sides of the table get something real from that. Founders get an investor who has structured and closed this kind of capital before, who knows what a credit committee will actually sign. Allocators get a manager whose view of the corridor comes from inside the transactions, not from reading about them afterward.

Cinderwood invests at seed in the companies shortening the path from electron to token. Generation and storage built for compute load, grid software, the picks and shovels of interconnection, and the financial infrastructure this buildout still lacks.

***
§ 6 — The advisory record to date.

§6 The Record

Capital advised across mandates $140M+
Advised exits 5
Blended MOIC on advised exits 9.5×

Advisory track record of Cinderwood's capital advisory practice. Fund I materials are available to qualified investors on request.

***
§ 7 — Two doors.

§7 The Path Forward

Power demand from compute is a new baseline, and it compounds. The capital markets infrastructure needed to fund it, from development equity through equipment financing to offtake credit, is still being invented deal by deal.

The next great compute companies will be energy companies that learned software's clock speed. We plan to have met them first.

Building in the corridor

nicholas@cinderwood.vc

And every Sunday, Firm Power.