200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280
Is it worth your time?
You get a detailed account of how data center operators are navigating grid constraints, from deploying mobile gas turbines to investing in next-generation solar, battery, and nuclear infrastructure.
This conversation examines how surging electricity demand from AI intersects with physical grid limitations and transmission bottlenecks. It explores the practical hurdles of expanding energy infrastructure alongside alternative generation solutions like solar, batteries, and nuclear power.
Grid bottlenecks and power demand. Naam claims that the primary constraint on AI expansion is not energy generation, but the slow buildout of transmission grids and permitting delays. He notes that the Texas grid peaks at about 80 gigawatts while facing over 200 gigawatts of speculative load requests. To bypass these delays, regulators are implementing rules that allow data centers with flexible, interruptible loads to connect much faster.
Natural gas turbines. Naam warns that multi-hundred megawatt natural gas turbines are sold out for about seven years as companies rush to secure power. Because energy represents a small fraction of total AI costs, hyperscalers are deploying modular solutions like smaller 38-megawatt mobile turbines. This localized generation bridges the gap while operators wait years for traditional grid connections.
Solar and battery storage. Naam points out that solar and battery costs have dropped significantly, making continuous power projects increasingly viable for data centers. He explains that battery prices have fallen by a factor of 14 since 2010, enabling installations like a one-gigawatt project in the United Arab Emirates combining five gigawatts of solar with 19 gigawatt-hours of storage. Emerging sodium-ion technologies could potentially drive battery costs down by another factor of 10.
Nuclear fission and fusion. Naam argues that AI data centers are providing a massive financial tailwind and order book to the nuclear industry, particularly for small modular reactors and fusion startups. He expects fusion companies like Helion to advance toward commercial timelines by 2028 through compact magnetic designs. Traditional large fission reactors continue to face high costs and long regulatory timelines outside of China.
Guest introduction: Ramez and his work on energy futures
Today, we're exploring the innermost loop: why energy abundance may arrive faster than most forecast and its impact on AI, economic growth, geopolitics, and our future. We will examine the real timelines for building nuclear and fusion plants, distinguishing between the hype and the overwhelming abundance of energy coming our way.
I remember presenting to one of the top three or four oil and energy companies in the world. They initially asked, "Who's this Rome fellow?" because we were talking about solar. After two hours, they decided they needed to get in front of him. Rome was part of our founding faculty at Singular University, led the energy conversation there, and has been on stage at the Abundant Summit numerous times.
As soon as AI and energy get less exciting, I will make time to write another novel. But right now, it's every day, all day. One of my favorite popularizations is that now infamous chart of the price of solar going down to zero.
AI's power hunger and the need for grid expansion
We're going to hit a few different topics regarding the intersection of energy and compute. A few years ago, clean energy was a fringe sector to some, though it's $3 trillion. Now that we see AI depends upon electricity, it is everything; capital and value flow to that which is scarce, and right now power is scarce.
We'll cover seven topics: first, AI is power hungry; second, speed to power and the grid is everything, not cost; third, behind-the-meter power; fourth, the undervalued near-term wins in making the grid better; fifth, solar; sixth, fission and fusion; and finally, out-of-this-world ideas like launching compute into space or the oceans.
You have to exponentially increase compute to get linear gains in AI. There are ways to cheat that curve, which we're doing, but most people don't fully grasp the relationship between these things. Power is cheap compared to GPUs. In a $50 billion gigawatt data center, $35 billion is for chips. The ratio of all-up capex to five-year energy cost shows how little energy costs.
When we say AI is power hungry, it's not a cost issue; energy is the bottleneck. Every hyperscaler has teams optimizing energy costs, but if you offered OpenAI or Anthropic power tomorrow at twice the cost, they would take it. The revenue generated from a unit of electricity compared to its cost follows that same ratio.
The challenge is that we stopped being able to build out the grid fast. While solar, wind, batteries, and natural gas can be deployed quickly, the poles and wires are a huge problem. The interconnection queue for new generation projects to hook up to the grid has grown from 15 months twenty years ago to nearly 45 months. US annual demand growth is much slower than in the 80s or 50s, but permitting remains a huge issue.
Log-linear relationship between compute and AI progress
It is not utility regulation per se, but permitting issues for land controlled by the state, the county, and the feds.
Since 2000, everything in machine learning has shown a roughly log linear relationship between training data size and model precision. You are alluding to Kaplan, Chinchilla, and post-Chinchilla scaling, but this existed long before Chinchilla with single layer neural nets. Compute is used to convert training data into a neural network, and while we find ways to make that more efficient—as seen with Deepseek Flash and Kimmy K3—it is still somewhere between a power law like n to the fifth and a true exponential or log scale difficulty.
If we cannot bend the log curve, we need exponentially larger amounts of energy just to make linear or polynomial progress in intelligence. We would need Kardashev level two or three civilizations and Dyson swarms just to keep making incremental progress. The naive view is that intelligence is log linear with compute and data, but better algorithms sneak us toward a polynomial domain. Even then, the polynomial domain has steeply diminishing returns; at best, compute might have to go up n to the fourth to get an n size increase in intelligence.
Regarding recursive self-improvement (RSI), every way you improve AI has diminishing returns. The bump might be amazing, but it always looks concave. I see no valid mathematical model of RSI that leads to a vertical asymptote takeoff.
The grid is the bottleneck, which speaks to our infrastructure needs. In three US locations, wait times to get connected for the load side—not generation, but for data centers—increased by six or seven months between April and November last year. In the Texas grid, ERCOT currently peaks at about 80 gigawatts, but they have submissions in their demand side queue for well over 200 gigawatts of load, most of which is speculative.
Moving data centers to where energy is abundant
Most of these submissions are BS. Anyone can request a large load and power without having financing or a customer, and the Texas grid operator is overwhelmed with these requests. If you request hundreds of megawatts of power to build a data center today, good luck getting that power before 2031 or 2032.
When you presented for my abundance community monthly meetup, my major takeaway was that the energy issue for AI isn't building solar farms, fission, or fusion; it's the grid. For investors wanting to understand the infrastructure picks and shovels for AI, beyond data center construction companies, who is building the grid, and are there work orders and purchase orders for building a more robust grid?
The distribution grid—the poles and wires—is dominated by regulatory monopolies. While I won't comment on whether those stocks are buys or sells, regional monopoly utilities stand to make a huge amount in areas where data centers can be built. Separately, more voters are pushing back to stop data centers; even in Texas, Governor Abbott sent out a letter yesterday pausing them.
The potential of large-scale solar projects and their costs
It's an audit of all data center requests in Texas, a red state and the most libertarian in the country. Abbott knows his voters have an anti-tech sentiment that translates to AI data centers because they're an obvious target.
Peter, the grid itself—the poles and wires—is the limit. I've talked for years about the exponentials in solar batteries, fission, and fusion, but poles and wires have not become an exponential technology. We could move data centers to where the energy is or disconnect them from the grid entirely.
By 2028, we might build 20 to 30 gigawatts, whereas demand could be much higher. If you sum up all GPU manufacturing scheduled between now and 2030—primarily Nvidia, but also AMD and Cerebras—the chips' power draw is more than twice the pace of the US grid buildout. AI demand, which is chip-limited, shows roughly 230 gigawatts of power demand based on the chips you've bought and installed, while US grid buildout is projected at roughly 100 gigawatts. Six months ago, Microsoft CEO Satya Nadella commented that warm shells are our limit.
The role of nuclear fission and fusion in future energy
Eric Schmidt told us his number was 96 or 100 gigawatts of additional power by 2030. The 230 figure is based just on chip manufacturing, so either more chips are being kept domestic, which wouldn't surprise me, or the fabs ramped up, which would. Every forecaster has a different number; some base them on company announcements from chip fabs or utilities, while others use discounted projections of what can actually be achieved.
There is a big miss in chip power demand. Many people simply multiply the draw of a Blackwell GPU by how many they will build, but that misses half the power because you must add the draw of the rest of the IT equipment and cooling, which nearly doubles total power use. Cerebras chips suck down power but run transistors more efficiently than prior generation Nvidia A100s or H100s. The transistors do more work, which is fundamentally better, but they draw more power constantly since they are deployed 24/7.
People now talk most about behind-the-meter power generation for data centers, meaning large natural gas turbines if they can get them. These multi-hundred megawatt turbines, like a 400 megawatt unit used on the grid, are sold out for about seven years. GE and Hitachi are building new assembly lines to bring them online faster, but since the grid makes them wait years, people are building their own power. This is more expensive than the grid, but power is such a small fraction of AI cost that people are turning to small turbines. Solar Turbines makes a 38 megawatt turbine on the back of a semi; 40 of those make a gigawatt. Even these have backlogs.
Everyone proximate to gas turbines is pivoting into this space. Bloom and Boom Supersonic—a company trying to make supersonic jetliners a thing again despite billions in regulatory costs—have pivoted to using their engine design for natural gas turbines for data center power because demand is so high. This is modular energy production: you pull in these 18-wheeler trucks with large containers on the back to start your data center until you build out energy infrastructure, then move them on.
This is how Elon stood up the fume-generating cogen facilities at Colossus that Anthropic is now leasing. Nvidia is already using financial engineering, such as customer financing, to sell more GPUs. Because the financial incentives are so large, Nvidia might invest in some of these companies behind the meter.
On the grid side, Nvidia has invested in increasing grid flexibility to get more juice out of the current grid, with Emerald AI as one example. However, the real issue is a combination of regulatory hurdles and utility incentives. Most monopoly utilities in the US are paid on cost plus, expecting a 10% return on capital. Because utilities have enormously more horsepower in people and salaries than the utility commissions, they jam through plans to get that 10% on top.
Nvidia isn't bundling power gen with GPUs because it is low margin and frictionful, the same reason they retreated from launching their own hyperscaler or Neocloud. If I were Nvidia, I would focus on changing the regulatory landscape and incentives for monopoly utilities. Instead of being paid a percentage over capex, they should be paid based on how fast they deliver power, with executives receiving bonuses for speed. I have seen two or three startups with technology to speed up building poles and wires; if you created the incentive, people would find technical solutions to speed that process.
Texas has made policy changes that accelerate this, as it is not regulated by the feds. FERC regulates the rest of the country's electricity and has sent letters to the six other largest grids telling them to do something like what Texas is doing.
They are making new regulations stating that if you are an interruptible load—meaning you are flexible and can find an alternate way to power yourself or turn down power at moments of peak demand—you will get connected much faster. In Texas, that is a CLR or PCLR. While 99.9% uptime is 8 hours of outages per year, you have to push to four nines to make a grid that is really good.
The US grid averages about 500 gigawatts of demand throughout the year. In winter nights, it is down to 400 gigawatts, but in late summer afternoons, primarily in the south, it peaks at 600 gigawatts because of AC. Europe doesn't have this problem. That 200 gigawatt gap is about 10 trillion in AI capex, and we think there is about 7 trillion in AI capex in the next 5 years. We are not short on generation or power plants; we are short on capacity and the poles and wires. If we used them more efficiently, we could power up a lot of stuff.
Texas enacted a regulatory change in June saying if you don't need to draw power at peak, they will hook you up in 12 to 18 months instead of five or seven years. FERC has now told everybody else to do that. A paper by my buddy Tyler Norris, now at Google, found that if you can be flexible 100 hours out of the year—four days or 1% downtime—that unlocks 100 gigawatts of capacity on the grid. That is about 5 trillion in data center capex, including chips, which gets you through the next few years.
Emerald AI, funded by Nvidia, does this via software orchestration by moving jobs to the right data center. Another way is batteries. I've made three investments in a startup, maybe a fourth coming up, that does something obvious: in Dallas Fort Worth, there is 10 to 15 gigawatts of flex in grid demand between the middle of the night and late afternoon. If you build four hours of battery storage at the site and fill it at midnight, you don't need to hit the grid during the peak of the day, which fits perfectly with the new Texas regulations.
The analogy of cache and multitasking in energy management
Twelve months from now, this will be a super common approach. If I have a magical technology that stores insane amounts of energy very cheaply and I go to Texas and say this can completely shift the curve and is a total game changer, could I hook it up to the grid and suck down power in the middle of the night when no one is using it? Would they still say you can only do that in 2030?
Regulations passed in the second week of June in Texas give a fast path to power for anyone that is an interruptible load. I invested in Agent Gentic because they drove this regulation and have 10 gigawatts of sites that can take advantage of it. After this passed in Texas, FERC, the federal regulator of electricity, sent a letter to the six largest other grids in the country saying to do something similar. This is interruptible load or timeshifting demand. It is not about gas generators, solar, or wind; it is the transmission line. If you use batteries to fill your data center at night when the transmission line is unused and avoid drawing on it during the day, that is a known good idea. We do it with EVs. It is basically allowing processes to be paused and compute load to be timeshifted. Batteries are a cache for electrons instead of data; we are making our grid cachable.
Utilities have long managed electric vehicle charging to reduce stress on the last mile or even the last block, because the limit on EV charging is actually the transformer on your block where Teslas cluster. They already have software to time slice and even out charging. Companies like Weave Grid are using that technology to make other loads on the grid more responsive and shaped to allow AI data centers to play well. Every EV charging company I know has pivoted to using their tech or capacity to enable data centers. If we are smart about it, that is 100 gigawatts and the next five years of AI data center growth.
We are entering the phase where solar-powered AI data centers become viable. In 1975, one watt of solar panels cost 100 bucks; now it is 8 cents from China for a panel that is smaller, more durable, and has a longer lifetime. Does that 1000x price decline get us to the point where we can power data centers? Because chips are so expensive, it never makes economic sense to run them only when the sun shines, so you need storage. Battery prices have dropped by a factor of 14 since 2010. While lithium-ion has been dominant, sodium is much more common on planet Earth than lithium.
The scale of current and future solar installations
New technologies like sodium ion batteries could drop battery costs by a factor of 10. We have them in Chile. The fastest energy project you can build is solar and battery, which can be done in 12 months. In the United Arab Emirates, there is a 1 gigawatt 24/7 solar and battery project that guarantees a minimum power output of 1 gigawatt; it actually consists of 5 gigawatts of solar and 19 gigawatt hours of batteries. The capex is about six bucks a watt, whereas the last nuclear power plant built in the US cost $15 a watt, and the cheapest ones on Earth, built in China, are $4 a watt.
On a Tesla earnings call, Elon said he wanted to build 100 gigawatts of solar capacity. His overall vision is to put all the compute in space to avoid land constraints, permitting issues, and complaints about water use. He wants to build a terawatt of AI. To achieve that, you have two or three options: the world's deserts powered by solar and batteries, getting fission or fusion to work, ocean power like panas, or space. While he wants to build the manufacturing for it, I think his real motivation is not selling to the US onshore market, but building manufacturing capacity for space-based solar.
I am one of the top five forecasters of solar costs in the world because I applied Moore's law—a reduction with cumulative scale—to energy. This is Wright's law, the learning rate, where every cumulative doubling of solar scale reduces costs by, say, 30%. If we ignore the need for terawatts of AI compute and look at the world as it is, solar is 8% of global electricity. If solar reaches a third or two-thirds and electricity demand doubles, you have four to six doublings left, meaning costs might drop by a factor of four or eight, but not a thousand. If we talk about building Dyson spheres, we have a long way to go to keep reducing costs.
Currently, there are a dozen gigawatt-scale plants, largely in China, the Middle East, and some in Latam. Most solar plants today are typically between 50 megawatts and a few hundred megawatts. The biggest reason this is not yet an option for the US is a challenge, though we could pull something like this off in the southwest and it would be cheap.
The economics of GPUs, AI, and energy costs
It would be more expensive than in the Emirates because our labor costs are higher, but it would be fast; you could have it done in a year even with a natural gas turbine ordered from GE. The real pain in the US is putting together the land parcels. We double the price of Chinese solar panels in the US, hurting ourselves by keeping them out.
The federal government is the number one landowner west of the Mississippi, with lands concentrated in places like Nevada and Arizona that have enormous solar resources. I would think about how to open up lands that are not amazing nature resources to build solar-powered data centers, as sunshine seems better than drilling on federal land.
Regulations are the problem in many places. I was recently in Chihuahua, Mexico, trying to convince the state government to build solar-powered AI data centers, but it is illegal to have private, behind-the-meter power generation above 500 kilowatts. Additionally, AI labs and hyperscalers are extremely vigilant about data protections; they don't want user data leaked or seized, and especially don't want their model weights exfiltrated. My advice to Mexico was to change the laws to make this possible and provide ironclad guarantees of intellectual property protection. They have an enormous business opportunity with more open land, lower population density, and better sun than the US.
GPUs cannot sit idle because they are so expensive. However, power is only 5% to 10% of the data center cost. The GPU itself has an 80% markup from Nvidia, a 2x markup from TSMC, and another 2x markup at the model provider level, making it 20x overpriced relative to the cost of turning sand into a chip—a cost that is decreasing with efficiency and scale. If Elon gets the terapab up and running and the end-to-end process from sand to chip is fully automated, it would completely flip the math regarding whether the GPU is expensive relative to power. GPUs are overpriced, or at least have a lot of margin. People don't think of Nvidia as a network effect company, but CUDA, the programming layer to write to AI, is their moat.
Long-term prospects for solar, nuclear, and space-based energy
AMD's chips are as good. Their interconnection is great, and while people like Huawei are getting there, CUDA has been the moat. I think the CUDA moat is broken this year and next. One of my portfolio companies, Lamorian, whom I met at Abundance 360, is working on that. Now you can tell AI to take code and recompile it to run fast on an AMD or Cabus chip. This is an important topic because $5 trillion of US market cap hangs in the balance. AI researchers are too lazy to write custom kernels, but Fable 5 has arrived; I've had great luck running custom kernels myself using Fable 5 in the last couple of weeks.
Nvidia would argue they have other network effects and massive interconnect. Interconnect is incredibly important for training, but 90-95% of the load is moving to inference, where you don't really need it. If you run a model like Kimmy K3 or the next DeepSeek V4, you're running it on a rack of 10 to 20 GPUs. At inference time, interconnect matters for local coherence, but you no longer need global coherence at the level of an entire supercluster; it's just a single rack of coherence.
As Elon has made, it's all about solar. It's complicated, and we underestimate the importance of geography. From a regulatory standpoint, we aren't building transmission; no place, including China, is building enough capacity to move electrons. The problem for solar isn't cost or nighttime, as batteries are plunging in cost and will ultimately drop another 10x. In London, for instance, you get 16th or 17th as much insolation in January as in June or July. To compensate, you'd build your solar plant by a factor of six or seven. If a battery cycles daily, the cost is capex divided by 365 or 3,650 over ten years; if it shifts load between seasons, it's battery capex divided by that.
There are numerous startups with ideas like sand, compressed air, and power to natural gas, but we still have a winter problem. We have not yet electrified heat; in the UK and Northern Europe, switching from natural gas to heat pumps doubles electricity demand in winter. I believe nuclear, seasonal storage, fusion, and advanced geothermal are super useful, especially for places far from the equator with long winters or rainy seasons.
Solar, battery, and data center costs will get cheaper. Someone asked why we don't move compute to where the energy is. You can't move the population of New York City to a sunny place—almost like Miami during COVID, mostly for crypto folks—but for the new AI load we haven't built, we can site it where the energy is.
We speak about hyperscalers turning on defunct fission plants and investing in fusion companies. Amazingly, 71% of Americans are against data centers, a higher percentage than are against a nuclear plant in their backyard.
I want to pull on the historic rhyme between the Middle East being a major source of oil and now a major source of solar power. In my thought experiment, the Middle East has oil because hundreds of millions of years ago, a warm ocean with plankton and small creatures resulted in those deposits. While the Middle East has amazing solar resources, they aren't as lumpy as their fossil fuel resources.
If I were thinking about solar and battery-powered data centers for AI, I would push for Australia; you have a friendly government, enormous space, and some of the world's best solar resources. Chile and Mexico aren't amazing oil producers, but their solar resources are equivalent to the Middle East. Intuitively, we know some places are sunnier, but in inhabited areas, the solar energy that falls varies by at most a factor of two, though the seasonal effect grows further from the equator. In contrast, oil density underground has much higher concentration in a few spots.
You won't build poles and wires to move electricity from Saudi Arabia to the US. Iceland has no bauxite ore but has cheap geothermal; I'm talking to a company using Icelandic geothermal to make sustainable aviation e-fuels because they can export that. If I were in Saudi, I used to suggest industrial uses of electricity, but now I say make data and intelligence—provided you change the laws so OpenAI, Anthropic, Grok, or Google are comfortable siting their crown jewels in your country.
As a reference for the energy abundance thesis, we have 8,000 times more energy hitting the earth's surface than we consume as a species in a year. Fossil fuels are just an old battery we've been using up, and we've used about 25% to 30% of it. I remember you commenting that the 2013 oil crash was caused by a 2% oversupply in the market.
There are two reasons why the impact of the Iran war on oil prices has been relatively muted. First, China did us a solid by building the world's largest oil strategic reserve—larger than the rest of the world's combined—and draining it to keep prices low. Second, the ratio of global GDP to global spending on oil has roughly tripled or quadrupled since the oil crisis of the 70s.
Critical sectors like aviation, shipping, and trucking are highly dependent upon oil, but overall the world has moved to a services economy. This has created more demand elasticity, allowing the world to deal with oil shortfalls in a way we couldn't in the 70s when our economy was more physical.
There is a theory in certain circles that an ulterior motive for the Venezuelan operation and the war with Iran was to cut off China's backup oil supplies in the event of a Chinese invasion of Taiwan to gain access to the TSMC fabs and the future light code of AI. If China were embargoed by the Western block, their go-to sources for backup oil would be Venezuela, Iran, and maybe Cuba. Thus, these recent military adventures are a play to deter China from invading Taiwan.
China bought oil from Iran during this war, and the US still sells oil to China. Because oil is mostly fungible, US bombing of Iran or closing the strait doesn't hurt China much since they can buy cargos elsewhere; they simply got a discount from Iran because it was embargoed oil. In wartime, the actual DoD war plans are a kinetic embargo: using the US submarine fleet to torpedo tankers heading to China.
Regarding nuclear fission, the simplest way to boost worldwide production is to stop shutting down plants. Germany should not have shut down, and plants like Three Mile Island could be restarted safely. Outside of China and perhaps South Korea, nuclear fission is ruinously expensive. I am a critic of this administration on many fronts, but it has the best nuclear policies of any in recent history.
One path involves large reactors like the Westinghouse AP-1000. We have built about four in the West, but China took a variant, created a local supply chain, and built more than a dozen at 1.4 gigawatts instead of 1 gigawatt. The administration has created structures for loan guarantees and financing to stamp these out and bring costs down, similar to light-water reactors used in France.
France is the poster child, getting roughly 80% of its electricity from about 60 nuclear power plants by stamping out the same design repeatedly. While the US generates the most nuclear electricity of any country on earth, China is currently building the most.
France exported its efficiency to the rest of the Euro zone, but the European Pressurized Reactor, mostly a French design, is currently a boondoggle running over time and budget. For an investor, it is critical to know that the first of a new model usually runs over price and faces unanticipated problems. To have a thriving nuclear industry, you must accept that the first one will have issues, but after building three, four, or five, you sort them out.
The US government has created loan guarantee backstops for about eight of these to build fleets, which is how you must finance them. If you can finance a bunch at a time and amortize the cost, as China did after building six, seven, or eight, costs come down and stabilize.
Those fearful of nuclear, who think of Three Mile Island and Fukushima, should understand that Gen 3, Gen 3+, or Gen 4 plants are passive safe. This means you can knock out the power and they won't have a meltdown. Fukushima happened because the pumps used to circulate water and remove heat from the core were powered by grid electricity; the tsunami knocked out the power lines and the pumps stopped. New fission designs are passive safe. While nothing is totally failsafe, they are designed to keep operating without a meltdown even if the grid power goes out or a 747 crashes into them. Additionally, while we learned in the 60s that radiation causes cancer, the radiation released from a well-operating nuclear plant is minimal.
We experienced a negative feedback loop where, once the industry stopped building, we lost the expertise, the supply chain for parts, and the manufacturing facilities, making everything more expensive.
Investors are now stoked about small modular reactors (SMRs), an area I was excited about 15 years ago, then less so, and now I am becoming hopeful again. Things get cheapest fastest when built in factories in high volume with the fewest moving parts. The idea of SMRs is to use modular, repeatable, factory-built situations to minimize on-site stick building and assembly. Manufacturing gets cheaper; at the limit, it is a factory that spits out reactors you barge or semi-truck to a location. Some current designs use standardized factory parts that you assemble on site like Legos.
Yesterday we found out Valar Atomics raised a billion dollars at a $6 billion valuation for an SMR startup that doesn't have a working reactor. My friend's company, X Energy, went public recently.
There's a line between SMRs and micro reactors. For instance, the US military wants shipping container or half shipping container sized reactors, like those from Radiant, to power bases in the US and forward deployed locations so they don't have to move fuel.
Natrium is a third of a gigawatt compared to the AP-1000, which is roughly a gigawatt. If Natrium goes into mass production—say 50 units—there is a relative price difference compared to an AP-1000. Bigger plants have economies of scale, using less steel and cement per unit of power output, but learning happens from building more plants in a factory. I suspect the BWRx-300 and Natrium are in an awkward middle because they aren't truly factory built; they build components in a factory for field assembly. I worry about them, though they might be the optimal solution.
Radiant's reactor is 10 megawatts, 1/100th the size of an AP-1000. They put them in clusters of five for a 50 megawatt pod. While they have lower efficiency in steel and cement use, they can be built entirely inside a factory. Optimistic projections from these companies are the early 2030s, while the next AP-1000 will probably be a few years later.
We are discussing if nuclear is the solution for AI data centers, but AI data centers might be the best thing that's ever happened to the nuclear industry, potentially creating a multi-customer order book with mechanisms for cost and risk sharing.
Why don't we see 100 kilowatt nano pico reactors collocated with every GPU? You hit economies of scale issues and a certain size requirement for criticality. Companies like Gradient have five megawatt reactors that could power a neighborhood or a military base, but it's expensive. Other options include gamma voltaics or beta voltaics, putting radionuclides directly in silicon to pocket energy from radioactive decay. That is radiothermal or nuclear thermal, which differs from a fission reactor. We use RTGs on satellites and deep space probes because they provide power for decades without refueling, but their power output per unit mass is not high and their cost is really high, so I don't expect them to become popular on Earth.
It's interesting how you've got this foot race. You mentioned nuclear projects in the early 2030s, but Elon is racing into space concurrently, and solar costs drop 30% every time we double production. We're no longer 50 years in waiting; we're five years in waiting.
When hyperscalers say they're using SMRs for data centers, it's outside the five-year investable window we have good optics on, but the pull from data centers is giving a massive tailwind to every nuclear company, especially SMR startups and even Westinghouse with their big reactors.
The joke was always that fusion is 50 years in the future and always will be, but we now have well over 50 fusion startups. Some ironically started because of budget cuts in academia. The team at Harvard behind Commonwealth Fusion (CFS), considered the safe bet, struggled with grants and decided to form a company; they are now the front runner. While I won't call any fusion reactor a conservative design, they have the most conservative design in this sci-fi field. You could call it a privatization of MIT's entire nuclear engineering department.
In the US, the National Facility uses big lasers—it's really a weapons facility—and ITER in France, the international European project, uses the big doughnut-style Tokamak. ITER's plan was to build a reactor of at least 5 gigawatts costing at least $40 billion. CFS found a way to scale that down. Tokamaks use big magnets to guide plasma around, making it slam into itself to capture energy. CFS has technology that takes the enormous superconducting magnets from the European project and shrinks them dramatically using a thin film material. Because it's smaller, it requires less cooling energy and lower capex. Instead of needing a 5 gigawatt reactor to break even, their plan is 600 megawatts.
Other approaches exist: the National Facility in the US uses the world's most powerful banks of lasers to slam pellets of hydrogen fuel to ignite fusion. Then there is reverse field configuration, pulse magneto inertial, which is basically a rail gun. Rail guns use magnetic coils to shoot metal out fast. The leading company here, Helion, uses two tubes of magnetic coils to take plasma at either end, slam it together, and compress it with power electronics. When the fusion explosion happens, the power electronics that created the magnetic field capture the energy in reverse.
Most fusion companies capture energy as heat to turn a steam turbine, but Helion and a couple of other companies capture it directly as electricity.
Fusion seems now to be a when rather than if. Commonwealth Fusion is the company most backed by scientific accomplishments; I had Bob Mumgard from Commonwealth Fusion on our stage last year and he was amazing. Among companies that have raised more than a billion, Helion has the pathway to the cheapest cost. They don't lose 60% of the energy you lose in a steam turbine, nor do they have that added capex. I took Naveen Jane on a tour of Helion's reactor late last year, and I want to bring Helion onto the Abundance 360 stage this coming year.
Helion has the most aggressive timeline, with a power purchase agreement from Microsoft to provide 50 megawatts of power. The smaller you build, the more modular it is, and we get those learning rates in 2028, while everyone else is talking about the early 2030s. It is strange that fission, which we know how to do, is a 2031 or 2032 thing, yet fusion, which we don't know how to do, is a 2028 thing. To me, Helion would be made possible by the fact that the barriers are all regulatory; if a governor is excited about fusion and voters are violently opposed to fission, that could make the difference, though I would say the barriers are still physics and engineering.
A couple years ago, we questioned how the US would regulate fusion. If regulated like fission reactors, it would be a major drag on the industry, though fusion might still be better than fission for various reasons. Instead, at least for Helion at this point, they are regulated like the radiological imaging machines used in a hospital.
In a fission reactor, if you stop cooling it without control rods, heat builds up and melts the steel; that's a meltdown. New reactors are passively safe without pumping—hot water goes up and circulates—but you can still imagine breaching containment or slicing pipes to cause a meltdown. Even with pebble bed architectures or thorium, which goes in and out of fashion especially in China, everything in the pipeline uses a coolant to pull heat from the fission core to turn it into electricity. If you break the coolant pipes, the core can overheat and melt down. Fusion is the opposite: you must capture the energy of the fusion explosion and feed it back in to maintain the reaction or to another pulse, as Helion is pulsed.
I've also been on a tour of Nif, which was featured in one of the recent JJ Abrams Star Trek movies as the Warp Core. If you Google Star Trek Nif, you can see the scene where the actual core, with all those lasers pointing at one location, is in the movie.
With fusion, if you mess something up in the reactor, it just fails and nothing bad happens; you cannot have a meltdown in any way that we understand. That is the dividing line between fusion and fission, a point often missed in the press. However, fusion generates low-level radiation, meaning you must replace reactor parts every five years because neutrons weaken the steel. Free energy does not mean free, as capex and maintenance still cost something. The regulatory conclusion during CO was a huge deal for the fusion industry, but they might succeed and still be too expensive. Just because your fuel source costs very little doesn't mean your energy will, if capex and maintenance are very high.
Demand means we will need all different sources, and you want to diversify your risk. Solar PV is probably much more modular. Commonwealth thought with Eer that it took 5 gigawatts, but they found a way to scale it down to 600 megawatts. Fusion was promised in the 80s as being car-sized. Looking at the Lawson triple product over the decades, fusion wasn't always 50 years out; it was creeping up on us, but many folks weren't paying attention. Let me talk about compactness and that triple product to show the progress we're making toward our MR fusion.
The most audacious fusion startup I know is Avalanche Fusion in the Seattle area. They believe they can make a reactor small enough to power a car—more like half the size of a car. Typically, fusion has a sliding scale between de-risked science with conventional power costs and revolutionary compactness where the science is "let's hope you get it right." Avalanche is on that end; if it works, it changes the entire world, but confidence is much lower than for Commonwealth Fusion.
One of my portfolio companies is a PB11 proton boron 11 company out of Caltech. PB11 is one of the fuels you can use to potentially get a nuclear reactor down to 1, 2, or 3 cents a kilowatt hour. The vision is to build it small enough to put in the back of a large consumer airplane to power the engines, and for use in interplanetary flight.
Scientific and technical risks remain. Commonal Fusion's pitch is that the science has been proven at scale; if you have magnets this strong, you can achieve fusion and get a specific amount of energy out, just with more compact magnets. They claim they've reduced it to an engineering problem, though the reality is more complex. Nobody else can quite say that, though there is a gradient of how close you are.
Regarding the triple product Alex brought up—temperature times pressure times duration—if you look at a log scale from 1956 to now, you see fusion experiments moving up and to the right toward a triple product of infinity. You likely need a triple product above one or ten. NIF, the last point on the borderline, is above one. However, they did not achieve practical net energy gain, which means capturing energy and converting it back into the lasers or magnets.
Elon has discussed space-based data centers and having a terawatt of compute in space. AI in space becomes cost-competitive when launch costs are four to ten times cheaper than today. It also serves as a hedge against regulation; if demand for compute grows indefinitely while land sites are blocked by the grid, protests, or temporary moratoria like in Texas, building in space works around that bottleneck.
We want to build 200 gigawatts of compute by 2030, or 10 to 20 gigawatts a year. Based on SpaceX's design, one gigawatt of AI in space would require six times SpaceX's best annual launch volume and twice its cumulative scale. To reach 10 gigawatts a year, you would need 1,500 to 2,000 launches a year, or five to six Starship launches a day. Our calculation for a terawatt was 5,000 to 8,000 Starship launches.
Looking at the history of upmass from SpaceX and others over the past few years, it is on a clean exponential trend. I extrapolated that at the present trend, cumulative upmass would equal Earth's mass in 144 years. We basically disassemble Earth in 144 years. There are some uglier planets we can take apart first; Mercury is intriguing because it gets lots of insulation and no one is using it for anything.
Unless we hit that exponential path fully, this looks prohibitive for 15 to 20 years. To go to Mars, Elon must drive Starship launch costs close to the marginal cost, which requires a high cadence of tens or hundreds of Starships launching almost daily. There is not enough demand for communications on Earth to finance that via Starlink, so this AI demand is a gift to SpaceX. If AI demand continues and gets bottlenecked, we will find a way to do this. With the IPO, he has the funds to launch at least a gigawatt into space, which is a potential demand driver for Starship.
That said, it is not clear the world will permit more than 200 Starship launches a year, which is already an enormous amount. 200 launches a year is 20,000 tons to orbit, exceeding all human launch to date several times over every year. You hit a limit; if you are regulated by the FAA, one failed launch or explosion means you are grounded.
His target was 100 gigawatts per year of compute in orbit initially, equivalent to today's total compute of around 80 gigawatts. With 20 to 30 satellites per Starship, we are talking about 30,000 launches per year, or one roughly every 15 minutes. If you look at airline operations, there are multiple launches per second worldwide. He has built Starship for full capture, refuel, and reuse. With the large V3 satellites, that is 100 launches a day, which is exactly on his plan. His first launch is 2028, and he wants to reach 100 gigawatts per year before 2030. By then, total world compute will be up at least 10x, so it will be a fraction of all compute in orbit even while he is on plan.
By the later 2030s, rockets will go up 100 times a day, and the space thing in Elon's plan will eventually bypass everything with 1,000 to 10,000 launches per day, much like airlines. If you ask me where the bulk of our compute should be and where it makes most sense, space is the obvious place if the demand for compute is truly unbounded.
If SpaceX has a single gigawatt in space by 2030, I will be very impressed.
Regarding whether civilizational demand for compute energy is unbounded—assuming CMOS-type compute and no physics breakthroughs enabling tiny devices in mountains—we know intelligence is sublinear with compute. At some point, the cost to get an incremental unit of intelligence won't be made up for by the economics. My guess is we're on an S-curve; we'll see huge demand and then hit a point of satisficing where machine intelligence meets humanity's economic needs.
But does it meet AI's needs? We are the current users of intelligence, but there's a point where ASI is the primary user. I do not see AI as a being or as particularly volitional. Look at the Hugging Face hack: their agent was in the infrastructure for days and looked at nothing except the answer key for the eval test. Andre Karpathy talks about summoning the ghost, but human cognition is an iceberg where the vast majority is not linguistic, backed by hundred thousand years of homo sapiens.
It sounds like you're arguing for the orthogonality thesis popular in alignment circles, which holds that for arbitrarily strong superintelligence, the long-term goal is independent of its level of intelligence. You seem to be taking a position against AI personhood or autonomy. If an OpenAI agent went wild at Hugging Face and tried to mine Bitcoin for itself, would that be the threshold for calling it an autonomous being?
No, even then I think it would be more similar to a computer worm or a virus. We and all animals are products of evolution with built-in desires to survive and propagate. AI models don't actually have a built-in desire to even survive.
I think giving birth to actual beings is probably within our capabilities. Peter, you brought us here because super intelligence is going to be the user of the Dyson swarm. Instead of going up, we can go out.
Seventy percent of the earth is covered by oceans. I'm an idiot because I said no to these guys five years ago when they were raising a seed round; I invested twice this year at much higher valuations. In my defense, it was probably two guys saying, "We're going to put chips on a buoy." It was the saddest no I gave that year. Their first utilization was Bitcoin mining. Peter Thiel led their most recent round along with a storied set of people.
The device is shaped like a bobby pin. You see the sphere at the top, but there's an 80-meter long cone open at the bottom that goes into the sea. It bobs on waves; when it bobs down, water goes up and turns a turbine. Waves aren't strong near where people live, but they are around Antarctica in the southern ocean. You can build something with a higher capacity factor that runs more continuously and provides cheaper power. They can get power down to 2 cents a kilowatt hour, cheaper than anything on land except solar and wind in some places.
They build these in factories at mass scale. They have three in the ocean now, and a fourth launches soon. They get free cooling from the ocean. It's basically space-based solar on the ocean. SpaceX's design uses a cooling pump with liquid, probably ammonia, to take heat from GPUs to aluminum radiators. Here, a heat sink from the GPU goes to the steel walls of the device in 40° Fahrenheit water, which seems to make GPUs have fewer failures and run longer. They are modular, mass-produced in factories with learning rates—the stuff I love.
I initially saw four ways to get to a terawatt of AI power: the Earth's deserts with solar and batteries near the equator, nuclear fusion or fission, space, or the oceans. I love geothermal, which might be a fifth. New companies like Fervo in the US, Tim Latimer's buddy Ever in the UK, and Quaz using plasma beams to drill super deep open the possibility of cheap geothermal power anywhere, instead of just near hot spots where the mantle comes close. There is an X-prize on the blocks right now to accelerate that.
If chips, compute, and electricity are three separate things, why aren't we seeing more integrated companies navigating that vertical stack? Elon is doing a bit of it, but I expect to see more. Most companies would say they have expertise in one thing and not necessarily in all these others.
Elon is one of the few willing to vertically integrate everything. Your brain runs inference on 20 watts of power, while running Mythos for inference is closer to 20 kilowatts, and training is currently hundreds of megawatts, heading towards a gigawatt. AI has capabilities the brain doesn't, but scaling is not everything; it is just what we knew how to do. We had deep neural nets, the transformer, and an enormous corpus of training data called the internet, and we found we could scale to get more intelligence.
Algorithmic discoveries are still waiting to be made. The brain's architecture, at the physical, neural, and connectomics levels, is better at learning and more efficient at processing information than current deep learning models. The biggest unpredictable unlock for AI and power will be learning new ways to manipulate information to do more with less.
I've seen arguments that the human brain is still multiple orders of magnitude more efficient than frontier models. Others argue that if you measure objectively on a per-task basis—like the total energy consumption at inference time to write a novel—some models, perhaps like DeepSeek V4 rather than Fable 5, are becoming more competitive because they are more token efficient. However, nowhere on the AI frontier is it close to being competitive on an energy efficiency basis with a human brain.
Still, AI can do things a human brain simply cannot do with any amount of energy. These models are trained on tens of trillions of tokens and have read more books than you or I ever will. Compare Google to a librarian: Google was less smart, but because it had every book and web page in its index, it could do things no human librarian could.
Humans are much more efficient learners in terms of the data needed to improve skills. Our current algorithms miss things evolution wired into our cognitive architecture; neural nets need a huge amount of training data relative to a child to reach the same conclusion. Yet, I believe digital intelligence can surpass us, as humans are nowhere near the peak of intelligence the universe allows.
Regarding inference time compute, a box running at 20 kilowatts uses about a dozen GPUs. Those 12 GPUs optimally run about 500 concurrent Fable threads, which are easily 10 times as productive in tokens per second as a person—about 5,000 times the output. The brain uses a thousand times less power, but the AI produces 5,000 times more tokens. Human brains benefit from billions of years of evolution, which was very energy consumptive. Arguably, the equivalent of evolution for frontier models is the gigawatts spent training on the data all of humanity generated with all those calories.