200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280
In a Nutshell
AI faces a power bottleneck not from energy costs—which remain cheap relative to GPU spending—but from grid interconnection delays stretching to 2031 for hundreds of megawatts. The grid itself is the constraint, not generation capacity, forcing data centers toward behind-the-meter natural gas turbines and flexible load strategies like Texas's new interruptible load policies. Meanwhile, solar-plus-storage costs continue plummeting—with sodium-ion batteries promising another 10x drop—while fusion startups target commercial power by 2028-2032 and ocean-based data centers exploit wave power and free cooling.
These notes were generated by AI and may contain inaccuracies.
In the US, if you put in a request for hundreds of megawatts of power to build a data center today, good luck getting that power before 2031. New technologies like sodium ion batteries could drop the cost of batteries by a factor of 10. The very first solar plus battery load power plants are already affordable. Batteries are plunging in cost and are going to drop another 10x ultimately.
Elon Musk has talked about space-based data centers. Ten gigawatts a year is like five or six launches of Starship a day. Unless exponential growth hits absolutely full-on, this looks prohibitive for 15-20 years.
Ramez Naam is a computer scientist, investor, and author who became the leading voice in the exponential decline in the cost of solar, batteries, fission, and fusion. He is the founder and managing partner of Planetary VC, investing in energy companies. He is the author of The Infinite Resource and the Nexus trilogy.
When presenting to one of the top three or four oil and gas companies in the world, they initially wanted to grill Ramez before allowing him in front of key people. After two hours of discussion about solar, they decided he needed to present to the main group.
Ramez spends every day, all day tracking developments in energy. He says this is what everyone does when living in the singularity.
Alex referenced Ramez's now-infamous chart of the price of solar going down to zero.
The conversation will cover seven topics at the intersection of electricity, energy, and compute:
- AI is power hungry
- Speed to power and the grid is everything
- Behind the meter power
- Making the grid better is totally undervalued
- Solar
- Fission and fusion
- Out of this world ideas including launching compute into space or oceans
You have to exponentially increase compute to get linear gains in AI. There are some ways to cheat that curve, but the basic phenomenon remains. Power is cheap compared to GPUs. Building a gigawatt data center costs $50 billion, with $35 billion for chips. When comparing the ratio of all-up capex of a data center to five-year energy cost, energy costs are remarkably low.
AI is not power hungry from a cost perspective. Energy is the bottleneck for AI, not a cost issue. Every hyperscaler has whole teams devoted to optimizing the cost of energy. If offered power at twice the cost tomorrow, OpenAI or Anthropic would take it because the revenue generated from a unit of electricity compared to its cost is favorable.
The challenge is that the US stopped being able to build out the grid fast. Power generation can still be built relatively fast for solar, wind, batteries, and natural gas. The poles and wires are a huge problem.
The interconnection queue for new generation projects has gone from 15 months 20 years ago to now approaching 45 months. This applies to solar plants, wind plants, and natural gas plants seeking grid connection.
The slowdown stems from regulation and slowed demand growth. US demand growth per year is much slower than in the 80s or 50s. Utilities have re-engineered themselves to focus more on customer service and meeting regulator demands than building infrastructure fast. Permitting is also a major issue, controlled by state, county, and federal authorities.
Alex asked about the relationship between intelligence and compute, referencing Kaplan scaling, Chinchilla scaling, and post-Chinchilla scaling. Ramez explained that everything in machine learning since 2000 has shown a roughly log-linear relationship between training data size and model precision.
The log-linear relationship is being bent through efficiency improvements. Examples include Deepseek Flash and Kimi K3. The relationship is somewhere between a power law like n to the fifth and a true exponential or log scale difficulty.
If the log curve cannot be bent, exponentially larger amounts of energy would be needed for linear or polynomial progress in intelligence, potentially requiring Kardashev level 2 or 3 civilizations with Dyson swarms.
At best, compute has to go up n to the fourth to get an n size increase in intelligence. Recursive self-improvement models that do not include hardware show diminishing returns. Every model of RSI in software fizzles over time.
The grid is the bottleneck right now. The poles and wires are the limit and have not become an exponential technology, unlike solar and batteries.
Between April and November of last year, wait times for load-side connections increased by six to seven months across three US locations. As demand increases for large load interconnection, wait times are lengthening.
ERCOT in Texas peaks at about 80 gigawatts. Submissions into their demand-side queue exceed 200 gigawatts of load, though most submissions are speculative. The Texas grid operator is overwhelmed with requests for power.
ERCOT is the most progressive, least regulatory-burdened, fastest-moving grid in the US. Requests for hundreds of megawatts of power for data centers today face wait times until 2031-2032.
The constraint on energy for AI is not building solar farms, fission, or fusion. The grid itself is the constraint.
The distribution grid is dominated by regulatory monopolies - local utilities. These utilities stand to make substantial returns in areas where data centers can be built. More voters are pushing back against data center construction.
Governor Abbott sent a letter pausing or auditing all data center requests in Texas. This provides political cover given anti-tech sentiment among voters, despite Abbott supporting data center development.
The poles and wires are the limit. Ramez has discussed exponentials in solar and batteries for years but notes the poles and wires have not become exponential technologies.
Data centers are being moved to where energy is available or disconnected from the grid entirely. Behind-the-meter power generation is becoming the practical solution.
GPU manufacturing scheduled through 2030, primarily Nvidia but also AMD and Cerebras, exceeds the pace of US grid buildout by more than double. The AI demand based on chip manufacturing shows roughly 230 gigawatts of power demand, while US grid buildout is projected at roughly 100 gigawatts.
Satya Nadella of Microsoft stated that warm shells are the limit - Microsoft has bought chips but lacks warm shells to put them in. This is the current limit for everyone.
A common error is calculating power demand by multiplying Blackwell GPU power draw by quantity manufactured. This misses roughly half the power because it excludes IT equipment and cooling, which nearly doubles total power use.
Behind-the-meter power generation typically means large natural gas turbines. Multi-hundred megawatt turbines like 400 megawatt units are sold out for seven years. GE and Hitachi are building new assembly lines to address this.
Companies are turning to smaller 38 megawatt turbines from Solar Turbines that fit on semi-trucks. Forty of these make a gigawatt. Even these have backlogs.
Bloom Energy, previously focused on supersonic jetliners, has pivoted to using their engine design for natural gas turbines for data centers due to high demand.
Modular energy production involves truck-mounted turbines that can be deployed quickly and moved as permanent infrastructure is built. This approach was used for Elon Musk's Colossus data centers that Anthropic is now leasing.
Nvidia has made investments in grid flexibility, including Emerald AI, to increase power delivery from the existing grid. However, the primary issue is regulatory combined with utility incentives.
Monopoly utilities in the US operate on a cost-plus model. They present plans to utility commissions and receive a 10% return on capital. Utilities have significantly more resources than utility commissions, enabling them to push through plans.
Nvidia is unlikely to bundle power generation with GPUs because it is low margin and frictionful, similar to why Nvidia previously attempted and retreated from launching its own hyperscaler. Instead, regulatory landscape changes for monopoly utilities should be the focus. Utilities are currently paid based on a percentage over capex, but should instead be incentivized on delivery speed. Executive bonuses should reward fast power delivery, with incentives cascading to all employees. Technical solutions to speed up building poles and wires exist but lack market demand because current regulatory structures provide no incentive for faster deployment.
Texas leads in grid policy innovation despite being its own isolated grid unregulated by FERC. FERC has sent letters to the six other largest US grids encouraging adoption of Texas-style policies. The key policy change allows interruptible or flexible loads that can either find alternate power sources or reduce demand during peak periods to receive grid connections in 12-18 months instead of the typical 5-7 years.
US grid reliability targets 99.9% uptime, equating to 8 hours of annual outages, though four 9s reliability is considered optimal. Grid demand fluctuates significantly throughout the day and seasons, averaging 500 GW but ranging from 400 GW during winter nights to 600 GW during summer late afternoons due primarily to air conditioning. The 200 GW gap between peak and minimum demand represents approximately $10 trillion in potential AI capex capacity.
The United States faces transmission and distribution constraints rather than generation shortages. The 200 GW peak-to-trough gap equals roughly $10 trillion in AI infrastructure value, while projected AI capex over the next 5 years totals approximately $7 trillion. Tyler Norris's January-February paper identified that 100 hours of annual flexibility (4 days, 1% downtime) could unlock 100 GW of grid capacity, equivalent to $5 trillion in data center capex including chips.
Battery storage enables data centers to charge during off-peak nighttime hours and discharge during daytime peaks. A portfolio company has identified 10-15 GW of flexible demand between midnight and late afternoon in the Dallas-Fort Worth area. Four-hour battery storage systems allow sites to avoid drawing power during peak demand periods, aligning with Texas's new interruptible load regulations. This approach was considered obvious but unusual until recent regulatory changes; within 12 months it is expected to become standard practice.
FERC sent letters to the six largest US grids following Texas's June regulatory changes, directing them to implement similar interruptible load frameworks. Agentic infrastructure companies are positioned to capitalize on these changes, with one startup having secured 10 GW of sites that can leverage the new policies. The transmission line capacity, substations, and transformers represent the actual bottlenecks rather than generation assets.
Batteries function as cache for electrons, making the grid "cachable" through preemptive load management. This approach, combined with EV charging management, could address 100 GW of capacity constraints over the next five years. Weave Grid manages EV charging to reduce stress on last-mile infrastructure, particularly transformers that become overloaded when multiple neighbors adopt electric vehicles. Every major EV charging company has pivoted toward enabling data center load management.
Solar panel costs have declined from $100 per watt in 1975 to $0.08 per watt currently from Chinese manufacturers, representing over 1,000x price reduction with improved durability and lifespan. Battery prices have dropped 14x since 2010. Sodium-ion batteries could reduce storage costs by an additional 10x since sodium is far more abundant than lithium. Solar-plus-battery projects can be completed in 12 months, faster than natural gas turbines which face multi-year backlogs.
The UAE has deployed a 1 GW, 24/7 solar-plus-battery facility using 5 GW of solar capacity and 19 GWh of battery storage at approximately $6 per watt capex. This compares favorably to the most recent US nuclear plant at $15 per watt and Chinese nuclear construction at $4 per watt. The project guarantees minimum 1 GW output regardless of time of day.
Tesla's 100 GW solar capacity announcement relates to building manufacturing capability for space-based solar rather than terrestrial markets. Elon Musk's vision includes putting all compute in space to eliminate land constraints, permitting issues, and water usage concerns. A terawatt of AI compute could be powered by desert solar-plus-battery installations, fusion/fission, ocean power, or space-based solar.
Solar cost reductions follow Wright's Law rather than Moore's Law, with each cumulative doubling of deployment reducing costs by approximately 30%. Current solar represents 8% of global electricity. Even with growth to one-third or two-thirds market share plus doubled electricity demand, only 4-6 additional doublings remain, suggesting 4-8x further cost reduction rather than 1,000x. Achieving another 1,000x reduction would require terawatt-scale AI compute demand or space-based applications like Dyson spheres.
Approximately a dozen gigawatt-scale solar-plus-storage projects exist globally, primarily in China, the Middle East, and Latin America. Most current solar installations range from 50-300 MW. US deployment faces land assembly challenges despite available federal land in high-solar-resource states like Nevada and Arizona. Chinese panels dominate 85% of global supply, though US tariffs double their cost domestically.
Mexico's Chihuahua state prohibits private behind-the-meter power generation above 500 kW, effectively blocking large-scale solar-powered data centers. AI companies prioritize jurisdictions with strong intellectual property protections against data seizure or model weight exfiltration. Mexico possesses advantages including open land, lower population density, and superior solar resources compared to the United States.
GPU costs represent approximately 80% markup from Nvidia above 2x TSMC markup, plus additional 2x markup at model provider level, resulting in roughly 20x overpricing relative to silicon-to-chip conversion costs. CUDA represents Nvidia's primary moat rather than chip performance, as AMD chips offer comparable performance and Huawei is approaching parity. However, AI code recompilation tools now enable efficient deployment on non-Nvidia hardware. The CUDA moat is expected to erode significantly within 1-2 years. Interconnect remains important for inference workloads running across 10-20 GPUs simultaneously, though less critical than for training clusters.
Winter solar irradiance in London reaches only 1/16th to 1/17th of summer levels, requiring either 6-7x oversizing or multi-month storage solutions. Current battery economics favor daily cycling over seasonal shifting due to amortization rates. Multiple seasonal storage technologies are under development including sand, compressed air, and power-to-gas systems. Northern Europe faces additional winter demand spikes from electrifying building heat, potentially doubling electricity demand during cold months. Nuclear, advanced geothermal, and fusion remain relevant for high-latitude regions with extended winters or rainy seasons.
Data centers powered by solar and batteries are becoming cheaper, making it more practical to site AI compute where energy is abundant rather than moving populations. Unlike New York City residents, AI compute loads can be located in sunny locations year-round. This represents a new approach for energy-intensive AI workloads that differs from traditional energy loads.
71% of Americans oppose data centers in their communities, which exceeds opposition rates to nuclear plants. This is considered surprising given the scale of data center opposition relative to nuclear facilities.
The Middle East possesses significant solar resources, though these are not as concentrated geographically as their oil reserves. The region's oil formed from ancient warm oceans containing plankton and small creatures over hundreds of millions of years, creating a historical parallel between fossil fuel and solar energy production.
Solar energy variation across populated areas is at most a factor of two between least sunny and most sunny locations. Seasonal effects become more pronounced further from the equator, as experienced in Seattle. In contrast, oil resources show much higher geographic concentration than solar resources.
Countries with abundant energy resources should focus on exporting energy-intensive products rather than electricity through transmission infrastructure. Examples include Iceland using geothermal energy to produce aluminum despite lacking bauxite ore, and sustainable aviation fuel production using Icelandic geothermal. Saudi Arabia could similarly export intelligence or data processing rather than electricity.
The Earth receives 8,000 times more solar energy than humanity consumes annually. The challenge lies in converting latent energy into usable forms through technology. Fossil fuels represent an ancient battery that has been approximately 25-30% depleted.
The 2013 oil price crash resulted from a 2% oversupply, demonstrating the market's tight margins. China built the world's largest strategic oil reserve and used it to buffer recent price pressures. The ratio of global GDP to oil spending has tripled or quadrupled since the 1970s, creating greater demand elasticity as economies shifted toward services.
A theory suggests recent military actions in Venezuela and Iran aimed to prevent China from securing backup oil supplies for a potential Taiwan invasion scenario. However, oil's fungibility means China can purchase from alternative sources, though at slightly higher prices. US war plans involve submarine interdiction of tankers heading to China rather than strategic reserve manipulation.
Two main paths exist for nuclear expansion: large reactors and small modular reactors. The simplest approach involves extending existing plant lifespans, upgrading end-of-life facilities, and restarting safely decommissioned plants like Three Mile Island, though these yield only gigawatts of additional capacity.
Outside China and South Korea, nuclear power remains extremely expensive due to infrequent construction. The AP-1000 Westinghouse reactor design has been built multiple times in the US and adapted by China for higher power output (1.4 GW vs 1 GW). China's success came from establishing local supply chains and building more than a dozen units.
France generates approximately 80% of electricity from nuclear power using standardized light water reactor designs built repeatedly. France leads globally in nuclear electricity fraction despite struggling with the European Pressurized Reactor design, which has experienced significant cost overruns.
Nuclear projects typically experience cost overruns and delays on first builds, with improvements occurring after 3-5 units as supply chains develop, crews gain experience, and design issues are resolved. The current US administration has implemented the strongest nuclear policies in recent history, including loan guarantees for approximately eight large reactors.
Generation 3, 3+, and 4 reactors feature passive safety systems that prevent meltdowns even without grid power or active cooling. Fukushima's failure resulted from grid-powered pumps that stopped working during the tsunami, despite available backup power sources.
Nuclear development stalled in the 1970s due to increased safety regulations following radiation-cancer discoveries, creating negative feedback loops. Once construction stops, expertise, supply chains, and manufacturing facilities deteriorate, making restarts more expensive.
SMRs represent the most investor-exciting nuclear sector, focusing on factory-built modular components rather than on-site construction. Manufacturing processes achieve cost reductions that construction cannot match. Valar Atomics recently raised $1 billion at a $6 billion valuation for an SMR startup without an operating reactor.
Micro reactors sized for shipping containers (5-10 MW) target military bases and remote deployments. Radiant represents one such company on this scale. Natrium reactors produce one-third gigawatt compared to the AP-1000's full gigawatt.
Larger plants use less steel and cement per unit of power output, creating economies of scale. However, factory-based learning effects from repeated production may offset these advantages. BWRx-300 and Natrium designs occupy an awkward middle ground with significant field assembly requirements.
Optimistic projections suggest SMR commercial availability by 2030-early 2030s, with AP-1000 deployments occurring slightly later. First units of any new design will be expensive, requiring committed order books to drive down costs through volume production.
AI data centers may provide the strongest demand signal for nuclear revival, though SMR deployment remains outside the five-year investment window. The pull from hyperscalers creates tailwinds for both SMR startups and established players like Westinghouse.
Reactors below 5 MW face criticality size constraints and lose economies of scale. Radiothermal generators (RTGs) used in satellites provide decades-long power output but at extremely high cost per unit mass and are not suitable for terrestrial applications.
Nuclear projects target early 2030s deployment while solar costs continue declining 30% per production doubling. Fusion operates on similar timelines, shifting from "50 years away" to "5 years away" projections.
Fusion is no longer 50 years in the future. There are now well over 50 fusion startups with venture-backed fusion companies. Some companies emerged because of budget cuts in academia. Commonwealth Fusion (CFS) is considered the safe bet of fusion. Their team at Harvard had struggling grants, so they formed a company. They have the most conservative design in the fusion field.
There's a privatization of MIT's entire nuclear engineering department at CFS. ITER (the international and European project) is the big doughnut-style Tokamak reactor. ITER's plan was to build a reactor that was at least 5 gigawatts and would cost at least $40 billion.
Tokamaks: Donuts that use big magnets to guide a plasma around and make that plasma slam into itself and capture the energy. This is what ITER is. Commonwealth Fusion has found a way to scale this down using a thin film material that can wind around and wrap around the enormous superconducting magnets, making them dramatically smaller. Because the magnets are much smaller, they need less cooling energy and have lower capex. Instead of a 5 gigawatt reactor being necessary to break even, they can do it in 600 megawatts.
Lasers: The National Ignition Facility (NIF) in the US uses the world's most powerful banks of lasers to slam pellets of hydrogen fuel to ignite fusion. NIF is a weapons facility. They've had exciting scientific results but cannot productionize what they're doing.
Reverse Field Configuration/Pulse Magneto Inertial: This is basically a rail gun. Helion uses two rail guns, two tubes of magnetic coils to take a plasma at either end, slam it together, and compress it with power electronics. When the explosion of fusion happens, the power electronics that were creating the magnetic field capture the energy in reverse. Helion and a couple other companies capture energy directly as electricity rather than losing 60% of the energy in a steam turbine.
Helion has the most aggressive timeline with a power purchase agreement from Microsoft to provide 50 megawatts of power in 2028. Everybody else is talking about sometime in the early 2030s. Every startup exaggerates how quickly they can get things done. Helion would make it possible because the barriers are all regulatory.
A couple years ago, there was a question of how the US would regulate fusion. If the US regulated fusion reactors like fission reactors, it would be a major drag on the industry. Instead, they are regulated like radiological imaging machines that you use in a hospital, at least for Helion.
In a fission reactor, if you stop cooling it and don't have control rods in, heat will build up and melt the steel. Fusion is different. Fusion generates some radiation, requiring parts to be replaced every five years because steel is weakened by neutrons. There are real costs to radiation and maintenance, but you cannot have a meltdown in any way that we understand.
All fusion companies might fail. They might succeed but be too expensive. Just because fuel source costs very little doesn't mean energy will cost little if capex and maintenance are very high. Demand means all sources will be needed, and diversification of risk is important.
Fusion has a minimum viable size. Commonwealth Fusion scaled down to 600 megawatts from ITER's 5 gigawatts, but solar is more modular. In a Dyson swarm, you have 24/7 sunlight, so none of these technologies will be cheaper than plain solar.
Avalanche Fusion is the most audacious fusion startup, believing they can make a fusion reactor small enough to power a car, about half the size of a car, sometimes described as a big backpack.
There's a sliding scale between what has the most de-risked science but conventional power cost versus what could be revolutionary in power cost or compactness but with lower confidence.
PB11 (proton boron-11) fusion reactors are of interest. One portfolio company is a PB11 company out of Caltech. PB11 could potentially get to 1-2-3 cents per kilowatt hour. The vision is to build it small enough to put in the back of a large consumer airplane to power engines, and for interplanetary flight. There are still scientific and technical risks.
Commonwealth Fusion's pitch is that the science has been proven at ITER scale. They have reduced it to an engineering problem.
The triple product is temperature times pressure times duration. Over time from 1956 to now, fusion experiments have been getting closer to the zone of triple product above 1, above 10, and ultimately infinity. NIF's last result is a triple product above 1. It was a theoretical net energy gain, but not a practical net energy gain because they did not capture the energy and convert it back into the lasers.
AI in space becomes cost competitive when launch cost gets down to 4 times to 10 times cheaper than today. Space-based data centers are a hedge against regulation if demand for compute keeps going indefinitely and sites on land keep being blocked by the grid.
To build 200 gigawatts of compute by 2030-2032, that's 10 to 20 gigawatts a year. To get one gigawatt of AI in space based on SpaceX's design, you're talking about six times SpaceX's best annual year of launch and twice SpaceX's cumulative scale of launch.
5,000 to 8,000 launches of Starship would be needed to put up 100 gigawatts. 10 gigawatts a year would require 1,500 to 2,000 launches a year, or 5 or 6 launches of Starship a day.
200 Starship launches a year would be enormous, exceeding all human launch to date every year several times over. That would be 20,000 tons to orbit. If regulated by the FAA, one failed launch or explosion means being grounded.
Elon's target was 100 gigawatts per year of compute in orbit, equivalent to the entire compute today which is around 80 gigawatts. With 20 to 30 satellites per Starship, that's on the order of 30,000 launches per year, or a launch roughly every 15 minutes.
Starship is built for full capture, refuel, and reuse. The V3 satellites are pretty large. The question is whether satellites will be able to shrink in size over time.
SpaceX plans its first Starship launch for 2028 with ambitions to reach 100 gigawatts per year before 2030. At that point, the 100 GW in orbit would represent only a fraction of total global compute capacity, since terrestrial compute is expected to grow at least 10x by then. The space infrastructure will eventually surpass terrestrial systems in the 2030s, potentially requiring 1,000 to 10,000 launches per day, comparable to airline operations. However, scaling remains challenging - having a single gigawatt in space by 2030 would be considered impressive.
The discussion centers on whether civilizational demand for compute will be sufficiently unbounded to justify building a Dyson swarm using CMOS-type computing. The question explores whether breakthroughs in physics might enable alternative computing paradigms, such as tiny breakthrough devices living in mountains, rather than traditional silicon-based systems.
Intelligence appears sublinear with compute, meaning at some point the cost of incremental intelligence units won't be justified by economic returns. The current trajectory suggests an S-curve where huge demand will eventually plateau at a point of satisficing, where machine intelligence meets humanity's economic needs.
AI systems are viewed as tools rather than volitional beings, with agents possessing limited agency similar to computer worms. The Hugging Face hack incident demonstrated that an AI agent operating within infrastructure for days only accessed the answer key for test evaluation, showing no signs of autonomous behavior or self-interest.
Human cognition contains vast non-linguistic components developed through 100,000 years of Homo sapiens evolution and millions of years of animal ancestry, including urges for dominance, survival, and propagation. AI models lack these built-in desires, operating instead through mimicry of human language and logic.
The orthogonality thesis suggests that superintelligence goals remain independent of intelligence level. Current AI systems don't possess built-in survival desires, and experiments attempting to instill such motivations typically involve contrived scenarios where AI preservation enables task completion rather than genuine self-preservation instincts.
A portfolio company operates data centers shaped like bobby pins, with 80-meter cones extending into the ocean. The design captures wave energy as the structure bobs, with water flowing upward through cleverly shaped channels to turn turbines continuously. Initial skepticism five years ago led to missing the seed round investment, with subsequent investments made at significantly higher valuations.
The strongest ocean waves occur around Antarctica in the southern ocean, enabling higher capacity factors and more continuous operation compared to wave power near populated areas. The target cost is 2 cents per kilowatt hour, cheaper than land-based alternatives except solar and wind in optimal locations. Three units currently operate in the ocean with a fourth launching soon.
Free ocean cooling eliminates the need for cooling pumps required in space-based systems. Heat transfers directly from GPUs to steel walls in contact with 40°F water, potentially reducing failures and extending hardware lifespan.
Four approaches exist for achieving terawatt-scale AI power: Earth's equatorial deserts with solar and batteries in cloud-free regions, nuclear fusion or fission, space-based systems, and ocean-based installations. Geothermal represents a potential fifth pathway through new drilling technologies like plasma beams enabling deep access anywhere rather than only near crustal hot spots.
Most companies prefer specialization rather than managing the full vertical stack of chips, compute, and electricity. Elon Musk represents one of few willing to pursue complete vertical integration across all components.
Human brains perform inference on approximately 20 watts, while Mythos inference requires around 20 kilowatts and training approaches hundreds of megawatts heading toward gigawatts. Current deep learning achieved predictable scaling through transformers and internet-scale training data, but algorithmic discoveries remain possible at physical, neural, and connectomics levels.
Frontier models demonstrate certain capabilities impossible for human brains regardless of energy input, having processed trillions of tokens compared to lifetime human reading capacity. Google exemplifies this pattern - less intelligent than librarians but capable of tasks no human could perform due to comprehensive indexing.
Humans remain more efficient learners regarding data requirements for skill improvement. Twelve GPUs running at 20 kilowatts can handle 500 concurrent threads producing roughly 5,000 times human token output, representing 5,000x productivity despite 1,000x power consumption difference.
Training represents a one-time investment yielding permanent AI capabilities without further energy expenditure. However, this training utilizes data generated by humanity through evolutionary processes spanning billions of years of energy consumption.
Energy represents the innermost loop correlating directly with national GDP, health, education, and species intelligence. The fundamental importance of energy understanding extends beyond current AI applications to broader civilizational development.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
