AI scaling pathways: On grid, on edge, off grid, off planet

The incumbent path has three constraints, unequal in difficulty. Elder names physical grid capacity and deliverability, power quality, and social license to operate, pointing to blanket bans in some states and projects pulled years after announcement. Kann rates power quality the most manageable, since it is an engineering problem, and social license the most underappreciated. On capacity Kann separates two problems people conflate: generation equipment, transformers, switchgear and substation upgrades sit in a rough three-to-seven-year window, bad but finite, while new interregional transmission is, in his phrasing, essentially infinite years, because it largely is not happening. That matters because transmission is the constraint unique to this pathway. Going off grid still leaves you waiting for turbines and transformers; it only escapes the poles and wires. One clarification to carry: the behind-the-meter generation now being built at data centers is almost entirely hybrid, sites that are grid-connected or intend to be, and close to none of it is true off-grid operation.

Edge is where the episode sets up an intuition and knocks it down twice. The first-principles case is latency, and both conclude it is a red herring: the standard example cuts the other way, since a Waymo makes its driving decisions in the car and its calls to the cloud are not so latency-bound that a regional hyperscale site fails. The second case is cost, usually argued from cheap land at a substation or commercial property, but land is a very small share of a fully loaded data center, where the money is in GPUs, the building and labor, and a small site is subscale against a 300-megawatt one. Strip both away and what remains is speed: a site provisioned for five megawatts and drawing two can absorb three more without waiting for a system upgrade. But matching a single 300-megawatt development means a hundred successful site evaluations rather than one, and neither speaker has seen that demonstrated. Elder calls it frustrating because edge looks like it should be the right answer.

Off grid relaxes almost everything at once, which is why Kann finds it the most underrated option: land is not scarce, siting can avoid communities that do not want data centers, and a study Elder cites found over a terawatt of opportunity in the American Southwest alone. What stops it is that the grid absorbs shocks. Islanded, you build that absorber yourself, meaning inertia, fault response and black-start capability, and very few people know how to run a gigawatt-scale grid. Early anecdotal data, mostly from bridge-power projects, is that they are not holding even 90% uptime. Elder’s refinement is the useful part: 90% may be tolerable if you know when the missing 10% falls, and intolerable if it lands at random inside long training runs. He also notes that grid operators cannot yet manage the voltage swings and harmonic distortion from data centers while those are a small share of load, which is a bad sign for managing them when they are the only load. Kann’s counterweight is that five nines is partly a legacy of cloud service contracts, while engineering to it off grid means overbuilding generation and storage, at a cost that matters for a ten-billion-dollar asset.

Orbital gets taken seriously and still loses on cost. Both say plainly they do not believe Musk’s claim that space will be the cheapest compute within three or four years, and neither calls the idea insane. Heat rejection, the objection people reach for first, is real but not the killer, because heat radiates with the fourth power of temperature and chips keep getting denser and hotter. The harder problems are debris, since a gigawatt-scale orbiting asset is roughly four square kilometers of radiator and solar, and maintenance, since a failed GPU that an engineer swaps in near-real time on the ground stays broken in orbit. The economic case rests on free power, but energy is only 5% to 15% of an AI data center’s cost, chips cost the same either way and maintenance costs more. So the case for orbital is not that it is cheap; it is that Earth may not allow building fast enough. That sets up the comparison Kann says he rarely hears made. Off grid’s rate limiter is the supply chain for generation and electrical equipment, orbital’s is Starship launch cadence, and standing up a couple hundred gigawatts a year of turbine manufacturing seems to him more plausible than five Starship launches a day. Elder’s pushback is the episode’s strongest objection and it is not technical: would society tolerate 200-plus gigawatts a year of new gas infrastructure for twenty years, which is Musk’s own carbon argument? Kann answers that you could be maximalist on solar and storage, geothermal and nuclear instead. Elder half-concedes, then adds the constraint above both pathways, semiconductor fabrication, which he is confident cannot supply that build rate and so probably binds first. That points him to off grid moving faster, and leaves Kann surprised not that anyone would go to space but that they would skip the terrestrial waypoint on the way.