Mike Schroepfer: AI’s Bottleneck Is Electricity, and One Tanker of Fusion Fuel Could Power the U.S. for a Year

Mike Schroepfer spent nine years as CTO of Facebook, then Meta, before founding Gigascale Capital to back companies using technology against climate change. On Redpoint’s *Unsupervised Learning*, Jacob Effron puts the usual tension on the table: ESG commitments on one side, massive data-center buildouts on the other. Schroepfer calls that tension the opposite of a problem. It is, he says, the best possible news for people who actually want both AI and a livable grid.

The arithmetic does not start with ChatGPT. Even without AI, the United States needs about five times today’s grid by 2050 — every gas car becoming an EV, plus onshore steel, cement, and concrete. That demand is already there. What AI adds is a set of customers who are technical, well capitalized, and in a hurry. Hyperscalers, in his telling, pull new generation technologies forward and walk them down the cost curve. Later adopters will buy those technologies because they are better, faster, and cheaper, not because they are good for the planet.

The deeper frame is human-scale. Technology is the only rate-limiter he knows for getting eight billion people into comfort and safety. Most people who need air conditioning do not have it; many lack clean water. Cooling, heating, and making water potable are solved engineering problems whose binding constraint is the cost of energy. Tenth that cost and deploy it everywhere, and living standards that rich countries take for granted become possible. He then multiplies a dedicated 24/7 reasoning agent — a kilowatt? a megawatt? a terawatt? — by eight billion people. The numbers get “crazy fun,” and still look tractable. The refrain he keeps coming back to is energy, energy, energy.

What is working right now is solar. In 2024, 80% of new energy added to the U.S. grid was utility-scale solar, a fact Schroepfer suspects most people do not know, and the reason is simple: it is the cheapest way to put new electrons on the grid. The catch is the duty cycle. Solar works about 25% of the time. Night does not work. Winter in the UK barely works. A $4 billion data center filled with Jensen Huang’s chips, sitting idle three-quarters of the year, is not an answer he will accept. Those buildings want to run 24/7.

So he is bullish on fusion. The reaction works — it is what burns at the center of the sun, and humans have made it happen more than once. The wild part is density. His rough, still-to-be-double-checked numbers: if you 5x the U.S. grid and power all of it with fusion, one super tanker of fuel supplies the entire country for a year; a pickup truck fuels a major plant for a year. No long-lived waste pile of the usual kind. Give him a 40-acre plot and two years to build, and he has a factory that turns a tiny bit of matter into energy. That, he says, is sort of magic.

Wilder still is an offshore compute platform from a company he names Pantala: a 200-meter structure that bobs in the ocean, harvests wave power, and uses the sea as both energy source and heat sink for immersive cooling. Put generation and cooling in one place and you might have the cheapest inference platform on the planet, if you can live with a deployment model that takes some imagination. They are already building it. Effron notes the emerging idea of UBI as a ration of compute or energy; Schroepfer likes the direction. Most hyperscalers have already signed purchase agreements for existing or new nuclear plants, and Meta put out an RFP for next-generation fission. Energy, he says, was not a board-level conversation five or ten years ago. It is now, because training still needs megawatts and reasoning models are moving demand toward inference — where cheaper tokens mint new use cases, and inference-time compute looks near-infinite at the right price.

Schroepfer has ridden three waves since the 1990s — internet, mobile, AI — and under each one was a tailwind of cheaper compute. Stanford dropped the assembly requirement for CS majors around the time he was there; wasting cycles on C, then Python and JavaScript, bought programmer productivity. AI systems writing code are even less efficient per cycle, and nobody is going back. He wants the same curve on energy: costs down 10% year over year for twenty or thirty years. Then AI compute is widely available, everyone lives with air conditioning, and you start manufacturing things — synthetic fuel, for example — that currently sound like science fiction.

The AI community’s shrug, Effron says, is that short-term natural gas is fine because AGI will solve climate in three or four years. Schroepfer does not love “and then AGI will solve it” as an answer to anything. It is a useful tool; it is also punting the problem down the hill. His forecast is that the next five years will be messy. If you need a gigawatt on the grid in a year or two, running 70 to 90% of the time, a combined-cycle gas turbine is a very effective way to do it if you ignore the climate side effects. Companies are making rational calls: power now, carbon offsets later. The inversion he expects is new gas assets built for AI sitting on a grid where coal is falling, gas is filling the gap, and solar’s S-curve is chasing from below.

The adult move is to walk and chew gum. Take 100 megawatts of gas this year. When you need a gigawatt in 2027 or 2028, ask what mix can actually show up: solar, which you can build fast; batteries, to stretch that solar; next-generation geothermal, fission, and fusion, funded now so the tech arrives sooner or with higher probability. Every hyperscaler, he says, should pop champagne if a fusion company works or if next-gen geothermal breaks out. That is a new set of suppliers for the thing that currently gates everything else. Lithium-ion cells, on the market since 1991, are already 97 to 98% cheaper and still falling more than 10% a year. ASML’s lithography tool is, in his ranking, one of the most brilliant machines ever built. The engine of human progress is any technology you can pour R&D into, scale in a factory, and copy forever — a solar wafer you can hold, a car packed with AA-sized cells.

He started Facebook AI Research in 2013 because a convolutional net had just won ImageNet by a landslide. The read was obvious: add compute, add data, keep the algorithm, and the then-“huge” setups would look tiny. FAIR’s stack produced PyTorch, now the default training framework, and quieter pieces like FAISS, the nearest-neighbor search buried in countless systems. His picture of a technical stack is layered. The further down you go, the more you want commoditization. Most companies will not train their own foundation model, just as most do not fab chips or write operating systems. Linux is the template: pool ingenuity in one artifact instead of copying it.

Foundation models, in that view, are a generalizable base. Funny videos, health diagnostics, running the power grid, discovering materials — each application needs a pile of domain knowledge a whole company or industry can wrap itself around. Share the base; compete above it. As CTO, his goal was not that Meta *made* the best technology. It was that Meta had *access* to it. Open source that is widely used is, by definition, accessible, and contributing to it means you also steer it, at zero dollar cost, with the rest of the world doing inference optimizations on every Llama release. American innovation, as he defines it, is decentralized invention plus collaboration at the right nodes — open source and scientific publishing. Going all-in on open weights was an uncommon call at the time. In hindsight he thinks they were on the right side of history. The regret he will own is internal: they debated scale versus algorithms and could have jumped on scale a little faster.

Developer tools have done the classic move up the stack. Model architecture is no longer where most people innovate; Transformers are there. The hard part is the system: collecting data, pretraining, post-training, RLHF and RL, then herding a 25,000-node cluster where some nodes are always down. It is closer to physics than to a GPU under the desk. Custom silicon is the same make-versus-buy question he has been answering since 2008, when Facebook leased inefficient buildings, then built data centers, then designed servers, starting with the high-count web boxes. Today it is hard to find a box in a Meta data center Meta did not design. They never made CPUs, because Intel was good enough. Nvidia, he says without hedging, makes unbelievable tech, and Huang is a hero of compounded R&D. When GPUs become a loud CapEx line, everyone asks whether they can go cheaper and more specialized. Beating Nvidia at a general-purpose FLOPs machine is a low-hope project. The 10x usually comes from putting a specific algorithm in hardware — and you have to guess that algorithm two years before tape-out.

Buying GPUs is the safer physical bet, because a training cluster that ages out of training can still do inference. You do not throw it away. Data-center capacity is the nightmare. When Meta built its first halls, the construction team asked how much he would need in 18 months. He had no idea how many people would be on the product. They had to order steel; you cannot walk into Home Depot the day before and build a 150,000-square-foot building. That impedance mismatch — software uncertainty versus a two-year physical pipeline — is, he says, a gut-buster for every hyperscaler friend he has.

Overpredicting hurts, and underpredicting hurts more. Extra capacity is a financial problem. Missing capacity is a technical and product problem: you can see the thing happening and cannot go get it. He has regretted the misses more than the surplus. There are pipeline tricks for hiding latency and keeping optionality at each stage of site, steel, and fit-out. He will not share the details on a podcast. Optimize the pipeline, he says, and you can do it well.

VR’s limiter has always been content. The product he wants is the operator from *The Matrix*: tell the system what you want, and the world appears. Generative AI that can create arbitrary 3D scenes near-instantaneously puts that in range — not next year, but in some number of years he will not pin down. Closer is contextual AI on smart glasses: live translation, a menu you cannot read, a tour guide for the building in front of you, family dropped into the same scene. AI today is still something you sit down to do. An always-on companion, walking with you, is a form factor nobody has really shipped. On the model side, straight pretraining is hitting diminishing returns. The surprise has been treating the LLM as input to a reasoning model via post-training RL. How many years that crank turns, and how many domains are easy to verify, is the question for the next couple of years. Math checks out; code either compiles or it does not; video is harder to ground. Memory is the other hole. Gemini’s million-token window is impressive. Associative long-term memory, the human kind, is still missing.

The climate applications he likes are search problems over a big space: Zanskar using data to pick geothermal drill sites so steam comes out of the right hole; the same pattern for copper or hydrogen; weather and insurance risk; materials and catalysts for carbon capture. Reid Hoffman’s new company hunting cancer targets with AI sits in the same family. Then he throws cold water. Draw a pie chart of all the time and cost to bring a new material to market, and ask how much of that pie the AI actually eats. In most pitches he sees, it is under 10% — a discovery step that does nothing for scale-up manufacturing or finding customers. He is bullish on AI in the long run and thinks it is overhyped for these uses in the short run, for exactly that reason. The interesting companies are end-to-end. Zanskar speeds an activity that already hits the P&L. Selina Tobaccowala — Evite co-founder, later at SurveyMonkey — has people clip a thermal camera to a phone, walk the house, and get an AI report: insulate here, replace the fridge, seven-month payback, for $50 to $100 instead of a $500 truck roll.

His own daily use is distillation with a skeptical lens. A Hacker News hashing-algorithm paper summarized well and then produced wrong examples when he asked for them. Deep Research, on geothermal reactor designs or biochar, is about as good as a sharp non-specialist MBA given a week, delivered in 15 minutes. Check the data points. For kids, the more the tools improve, the more the fundamentals matter: critical thinking, math, writing, reading, and the gut call on imperfect data that machines still do not own. The two trainings that served him best were high-school debate through nationals, so speaking in public stopped being a big deal, and Stanford CS, where the language was never the point — breaking problems down was. The skill he wants for his children is the refusal to say “I’m not a tech person.” He is not a plasma physicist, and he still talks fusion. AI is a fast tutor if you will tell it to explain it like you are in sixth grade.

Five years from now, the CTO job looks more like today than people think. AI is to coding as JavaScript is to C is to assembly: a higher abstraction, a backhoe instead of a shovel. Compiling the whole codebase used to be a 1990s crisis; nobody mentions it. What does not change is deciding which problems matter and organizing smart humans to go after them. He looks for people who keep a priority queue in their head and always have the most important problem on top. Brilliant people who wander off to crush a 1% issue, even 10x better, are wasting the 99%. Teams will get smaller and companies will come up faster. Judgment on leverage gets more expensive, not less.

Quickfire: model progress this year will be more than last year. The first thing he does with a new model is throw it a complicated paper to summarize. His weird prediction is that most people will have an AI friend — not because group chats of bots will feel natural (mixing two friend groups is already awkward), but because a non-judgmental ally who is on your side is something more people will find in a chatbot than we currently admit. Then he points people to gigascalecapital.com, LinkedIn, and Threads, and leaves the last word where the whole hour pointed: how to make a whole hell of a lot more energy, cheap and sustainable. The ceiling on the AI race is drawn on a grid map.