Power Players

Mike Schroepfer on AI's energy crunch as an economics problem

Mike Schroepfer· Founding Partner at Gigascale Capital
·~31 min·English·Yahoo Finance
AI InfrastructureInferenceBusiness Strategy
TL;DR

Former Meta CTO Mike Schroepfer argues AI's energy crunch is an economics problem, not a physical limit -- and lays out the untapped headroom, from wave-powered inference buoys to fusion, that could keep supply ahead of demand.

01Core Idea

The Ocean Runs Your Next AI Query

Schroepfer's boldest bet puts AI inference on autonomous, wave-powered buoys in the open ocean, where bandwidth is plentiful and cold water makes cooling straightforward.

this turns out to be the perfect platform to run AI inference nodes because Starlink is plenty of bandwidth and you also are sitting in very cold ocean water.

— Mike Schroepfer, Power Players
Key Insight
The quiet unlock is that inference distributes in a way training cannot. A frontier training run wants one giant, tightly-networked campus; serving is many independent nodes, so it can chase cheap power and simple cooling to places a training cluster never could -- even a buoy at sea.
02Ocean vs Space

Not a Physics Problem, an Economics One

When people debate whether a moonshot can work, Schroepfer reframes it as a cost sheet: the ocean beats space not on physics but on dollars per watt delivered.

order of magnitude or more cheaper for me to put a ton of material in the ocean than it is to lift it all the way up into space.

— Mike Schroepfer, Power Players
Key Insight
His Yeti-cup point is sharper than it sounds: the same vacuum that keeps coffee hot makes space a terrible radiator, so orbital data centers pay a permanent tax to dump heat. The lesson generalizes -- most 'is it possible' arguments are really 'does it pencil out' arguments in disguise.
03The Worldview

More Degrees of Freedom Than People Realize

Schroepfer's optimism rests on a stack of independent levers -- efficiency, solar, new generation -- where any single one buys enough time to matter.

you can make AI models 10x, 100x more efficient by making some trade-offs and quantization or using a light model.

— Mike Schroepfer, Power Players
Key Insight
The pessimist's mistake is treating today's model size and today's grid as fixed inputs rather than variables. Because the levers are independent -- you can win on efficiency or on generation -- the system has far more slack than a single-variable panic assumes.
04Energy Bet

Fusion Is the Endgame

Fusion's remaining risk has moved from science to economics: humans can already make it happen; the open question is whether it pencils as a power plant.

to power Austin for a year, a pickup truck's worth of fuel is what we're talking about.

— Mike Schroepfer, Power Players
Key Insight
Notice how he prices the risk: the physics is treated as done, and what's left is purely commercial. That reframing is exactly why a venture investor, not a national lab, is now a natural funder of fusion -- the bet is on cost curves, not on whether a mini-sun can be lit.
05Investing Lens

Find the Angle In

The micro reactor's edge is not cheap power -- it is picking a market where the incumbent option, shipping diesel, is painful enough that price barely matters.

would you like a box that shows up on site and makes power for five years without any fuel? How's that sound?

— Mike Schroepfer, Power Players
Key Insight
A startup's default state, in his words, is dead, so the first job is not the biggest market but the sharpest wedge: the customer for whom the status quo genuinely hurts. Win there, earn the right to compete on the grid later -- the same play that took solar and EVs from niche to mainstream.
06Innovation Engine

Two Questions: Why, and Can We Now?

Two questions drive Schroepfer's whole portfolio: why is this expensive, and can new tools build today what was impossible five years ago?

Why does that part cost $1,500? Is it because the materials are worth $1,500 or because it's 15 bucks in parts and someone's charging me a thousand% margin

— Mike Schroepfer, Power Players
Key Insight
New tools do more than speed up old work -- they make previously-uneconomic designs testable. Metal 3D printing lets a team try twenty nozzles instead of modeling one, which quietly reopens engineering problems everyone had filed away as settled.
07Industry Structure

The $2 Trillion Question

The frontier labs' valuations assume the gap to the next-best model stays wide and stays worth a premium -- and both halves can fail on their own.

their economic model is premised on leadership on models meaning that they have the latest and greatest models and others don't.

— Mike Schroepfer, Power Players
Key Insight
The louder risk is a fast-follower closing the capability gap. The quieter one is customers deciding 'good enough and cheap' beats 'best and expensive' -- at which point the premium collapses even while the frontier lab still technically leads.
08Leadership

The Default State Is to Slowly Die

Left alone, a profitable company coasts on its success and fades; founder mode is the forcing function that overrides those incentives.

the default state for every big company is to slowly die

— Mike Schroepfer, Power Players
Key Insight
Founder mode is less a personality than a fight against the physics of success: every local incentive in a profitable firm points toward coasting. Overriding them takes someone with the standing to spend the company's comfort -- which a founder has and a caretaker CEO usually does not.