Look 001 · October 12, 2026 · Reid Petrie · Long term
Reid’s Fund Look 001: The Power Wall
The thesis
The next decade of AI rewards whoever makes intelligence cheaper per watt.
Every AI company and every person making money off AI depends on data centers. Data centers are warehouses of GPUs that pull huge amounts of electricity and turn most of it into heat. Today the industry answers that by building bigger and buying more power. I think that works for a while, then hits a wall, and the money moves to whoever gets more useful work out of each watt.
I see three layers where that happens: making power, moving data more efficiently, and new kinds of compute that need far less energy to begin with. This look maps the companies I think matter in each.
Why now
Power, not chips, is becoming the bottleneck. You can order GPUs; you can't plug them in if the grid can't deliver the electricity. That's why big tech is signing direct nuclear deals: Microsoft paying to restart a reactor at Three Mile Island, Google contracting small reactors from Kairos, Meta prepaying Oklo for a 1.2 GW campus.
The other signal is where smart money is going. Nvidia and AMD are both investing in optical interconnect startups, and the U.S. Commerce Department signed a letter of intent for up to $75M toward thermodynamic chips. The incumbents are funding the efficiency fix to their own problem.
Layer 1: Energy
The most direct bet: whoever supplies reliable power to data centers gets paid no matter which chips or models win.
- Oklo builds small sodium-cooled fast reactors (15 to 75 MW) and sells the power as a service. Its order book is around 14 GW across Switch and Meta, its test reactor hit first criticality in August 2026, and the first commercial unit targets 2028. Public (NYSE: OKLO), so it trades on hype and regulatory news as much as progress.
- Kairos Power uses molten-salt cooling and has NRC approval for a 35 MW demonstration reactor. Google signed to buy up to 500 MW from seven reactors, the first by 2030. Private.
The catch: no commercial small modular reactor has run yet, so the economics are still unproven. This layer is about timing as much as technology.
Layer 2: Moving data with light
A big share of data center power goes to shuttling data between chips over copper, not doing math. Replacing copper with light cuts that waste, and this layer is already selling.
- Ayar Labs makes optical I/O chiplets that go directly inside GPU and accelerator packages. It raised a $500M Series E in March 2026 at a $3.75B valuation with Nvidia and AMD participating, plus a $150M extension in September, taking it past $1B raised.
- Lightmatter builds photonic interconnects and processors. It's raised around $850M and was last valued at $4.4B, with GV and Fidelity among its backers.
This is the most proven layer of the thesis, but it's also priced in. These are late-stage companies; my angle is the next generation of startups building around them.
Layer 3: New kinds of compute
This is the highest-risk, highest-upside layer: hardware that doesn't just do the same math more efficiently, but does it a different way.
- Extropic builds thermodynamic chips that use the natural noise in transistors to compute probabilities instead of fighting it. It has prototype silicon (X0) and a production chip (Z1) in development, plus a CHIPS letter of intent for up to $75M. Its headline 10,000x efficiency claim comes from a simple benchmark plus modeled hardware, not real workloads yet.
- Normal Computing is Extropic's most direct competitor in thermodynamic hardware.
- Cortical Labs grows human neurons on silicon chips. Its CL1 is the first commercially sold biological computer, backed by Horizons Ventures, Blackbird, and In-Q-Tel. Near-term use is drug research.
- FinalSpark rents cloud access to brain organoids, starting around $1,000 a month. Organoids currently live about 100 days, and training them is still unpredictable.
Biocomputing is the purest version of my thesis and also the furthest out. I'm tracking it as a 10-year bet, not a near-term one.
Early-stage picks
The companies above prove the thesis is real. These are the early ones I'm actually tracking, where the outcome is still open.
They are the cards at the bottom of this page, each with its own page.
Energy has no true early-stage pick yet. The nuclear startups are all raising $100M+ rounds, so that layer is late by the time it's fundable.
Where Bittensor fits
Bittensor isn't part of the efficiency thesis. It decentralizes who owns and coordinates AI, not how much energy it uses. Its miners still run on GPUs, mostly in data centers.
I treat it as a separate, small bet: that AI ends up more open and distributed than the frontier labs expect. The case for it is a 21M hard cap, 128 subnets competing for rewards, and real work shipping, like Templar's 72B-parameter decentralized training run in March 2026.
The case against: one analysis estimated its top subnet's token rewards run 22 to 40 times what customers actually pay, and the next emissions halving will test that. TAO also trades with crypto and AI hype, so it isn't a hedge against the rest of this thesis.
How I'd size it: small, held or staked rather than mined, with one signal to watch, whether subnets start earning real revenue.
What would prove me wrong
- Silicon keeps winning: if Nvidia's per-watt gains plus cheap gas and solar keep up with demand, there's no wall and no reason to switch.
- Reactors slip: if small reactors stall in regulation past 2030, Layer 1 becomes a story stock trade, not a power supply.
- New compute stays in the lab: if Z1 and organoids can't run real workloads reliably, Layer 3 stays a research curiosity.
- AI demand cools: if spending on AI slows, the power crunch eases on its own.
Changes
- October 14, 2026: Moved the early-stage picks table to the company cards at the bottom of the page, and removed the separate watchlist. Its companies are in the layers above, and Cortical Labs is a card.
- October 14, 2026: Set the horizon: long term.
Sources
Companies under this thesis
Published October 12, 2026. An opinion, not investment advice and not an offer.