200GW Hiding in Grid, Sodium Batteries 10x Cheaper, Wave-Powered Datacenters w/ Ramez Naam | EP #280

Ramez Naam argues that power availability, not cost, is the binding constraint for AI infrastructure. Building a one-gigawatt data center costs roughly 50 billion dollars with 35 billion going to chips alone, making electricity a small fraction of total capex over five years. Hyperscalers would willingly pay twice the current electricity price if power were available immediately because revenue per unit of electricity far exceeds what they pay for it.

The grid interconnection queue is the central bottleneck. Wait times have grown from 15 months two decades ago to nearly 45 months today, and a request for hundreds of megawatts in Texas right now would likely not receive power before 2031 or 2032. ERCOT peaks at about 80 gigawatts but has over 200 gigawatts of load submissions in its demand-side queue, though Naam acknowledges most are speculative and will evaporate. The structural cause is that slow US demand growth over prior decades caused utilities to re-engineer themselves around customer service and regulatory compliance rather than fast construction, and utility commissions are far less staffed than the utilities they regulate.

Naam estimates that 200 gigawatts of unused grid capacity sitting between winter-night lows of roughly 400 gigawatts and summer-afternoon peaks of roughly 600 gigawatts is equivalent to approximately 10 trillion dollars in AI capital expenditure, exceeding the projected 7 trillion dollars in AI capex over the next five years. Texas passed a regulatory change in June allowing interruptible loads to connect in 12 to 18 months instead of five to seven years, and FERC subsequently directed the six largest US grids outside Texas to implement something similar. Naam estimates that being flexible for just 100 hours per year could unlock 100 gigawatts of capacity worth roughly 5 trillion dollars in data center capex including chips. The startup Agentech drove the Texas regulation and has 10 gigawatts of sites positioned to benefit. Emerald AI, funded by NVIDIA, routes jobs to whichever data center has available capacity to achieve similar flexibility through software.

Large natural gas turbines of around 400 megawatts from GE and Hitachi are sold out for approximately seven years. The fastest energy project currently buildable is a solar plus battery installation achievable in roughly 12 months. A one-gigawatt 24-7 solar and battery project outside Dubai requires five gigawatts of solar capacity and 19 gigawatt-hours of batteries and costs approximately six dollars per watt in capex. GPU manufacturing power demand scheduled through 2030 is estimated at roughly 230 gigawatts, more than twice the projected US grid buildout of approximately 100 gigawatts through that period. Power demand calculations frequently miss roughly half of total draw by excluding IT equipment beyond GPUs and cooling.

Solar panel costs have fallen from 100 dollars per watt in 1975 to 8 cents per watt from China, a more than 1000-fold decline, and battery prices have dropped by a factor of 14 since 2010. Naam argues sodium ion batteries could drop battery costs by a factor of 10 relative to lithium ion because sodium is far more abundant, and predicts battery costs will ultimately fall another 10 times from current levels. Approximately 85 percent of US solar panels come from China, and US tariffs effectively double their price, slowing the domestic buildout. The core problem for solar at higher latitudes is seasonal variation, not cost or nighttime storage. London receives one sixth to one seventh as much solar insolation in January as in June, and seasonal storage batteries cycling only twice per year carry far higher unit energy costs than daily-cycling batteries.

On nuclear, Naam says the simplest near-term actions are to stop shutting down existing plants, extend lifetimes, and restart safely closed ones. Outside China and possibly South Korea, nuclear fission is ruinously expensive because it is built infrequently. The last US nuclear plant cost 15 dollars per watt while the cheapest Chinese-built plants cost 4 dollars per watt. Naam does not expect any company to have a commercial small modular reactor operating in 2030 or 2031 despite optimistic projections, and cautions that first units will not be cheap. He identifies Altoe as his favorite SMR company and describes factory-built SMRs as getting cheaper faster than field-assembled plants because manufacturing costs fall with repetition while construction costs do not.

Commonwealth Fusion Systems uses thin-film superconducting magnets that dramatically shrink magnet size compared to ITER and plans to achieve breakeven at around 600 megawatts. Helion captures fusion energy directly as electricity rather than through a steam turbine, avoiding roughly 60 percent energy loss from steam conversion, and has a power purchase agreement with Microsoft for 50 megawatts underwritten in 2028, though Naam considers that timeline plausible but unlikely. A regulatory decision made during COVID treated fusion reactors more like radiological imaging machines than fission plants, which Naam calls a huge deal for the industry, resting on the fact that fusion reactors fail safe because the reaction simply stops if something goes wrong.

Naam argues that CUDA, not chip performance, is NVIDIA's primary competitive moat, and that moat is beginning to break due to AI-assisted code recompilation allowing workloads to run on competing hardware, with roughly 5 trillion dollars of market cap at stake. He also notes that 90 to 95 percent of AI compute load has shifted to inference rather than training, and inference requires only single-rack coherence rather than global supercluster coherence, making interconnect far less critical for the dominant workload.