No AGI Without New Infrastructure
Cognition at scale demands substations, supply chains, and sovereign backends
1. Boca Chica, Meet Mountain View
On a wind‑scoured spit of scrubland outside Brownsville, Texas, Elon Musk built a launch factory he christened Starbase. Visitors expecting a NASA‑style campus, with hermetically-sealed clean rooms and slow bureaucratic rituals, are startled. The place feels less like Cape Canaveral and more like Pittsburgh circa 1905: heat plumes, clanging steel, half‑finished tanks rolling across the sand on crawler trucks. Yet within that rough‑edged carnival SpaceX solved a problem the rest of aerospace deemed impossible: iterate on super‑heavy rockets weekly, not yearly.
Musk took control of everything from rolled‑steel procurement to FAA negotiations, from cryogenic plumbing to livestream cameras. Only by owning the atoms, meaning the production lines, the launch pad, the airspace permits, and more, could he compress the test‑fly‑fail‑rebuild loop enough to make orbital rocketry feel like web‑app shipping.
Artificial General Intelligence (AGI) will require the same control over physical infrastructure.
We talk about AGI the way pre‑Internet dreamers once talked about “cyberspace”: ethereal, frictionless, unbound by geography. But intelligence at planetary scale is not weightless; it is a thermodynamic, jurisdictional, and supply‑chain project. To imagine the next leap as a clean YAML file riding a recursive function is to ignore the roar of the furnaces beneath.
Some argue that the AI frontier is fragmenting: open-source models grow more capable, edge inference spreads to phones, and fine-tuned 7B‑parameter variants outperform behemoths on narrow tasks. These are real trends, but they do not falsify the thesis. They mark the commodification of past breakthroughs, not the path to new ones. Decentralized cognition will flourish in downstream use cases. But upstream innovation, including agentic models, world simulators, and real-time multimodal inference, still demands petaflops, petabytes, and sovereign-scale infrastructure. The frontier remains centralized, because physics demands it.
If you want to forecast AGI, trade your LessWrong alignment posts for electrical‑interconnect diagrams, and your timeline analysis for zoning‑board minutes. Because the gating factor is not better software. It’s vast amounts of physical infrastructure.
2. The Metabolic Cost of Thought
A single NVIDIA H100 card, pushed to its 700‑watt limit, generates 38,000 BTU/hr. This is enough heat to heat a small house in a cold climate. And that’s one GPU. A GPT‑4‑scale training run used roughly 25,000 GPUs for sixty days. And each of these GPUs needs to have its heat dissipated. This isn’t software. This is complex infrastructure. Without the compute infrastructure and the cooling infrastructure, your software fails and returns no inference. You can just hand-wave away the infrastructure. You can’t assume that your software will work just because code is cognition and cognition is abstract.
And let’s talk about the electrical power required to run training and inference on frontier large language models. Do the math:
25,000 cards × 0.7 kW ≈ 17.5 MW
Add 10 % overhead for CPUs, NICs, and ancillaries → 19 MW
Real‑world PUE1 (efficient but honest) ~1.1 → 21 MW sustained draw
Build in N+1 redundancy and peak spikes → ≈25 MW site capacity.
That is one data‑center’s worth of power for one model, for two months. GPT‑5 is rumored to double the parameter count and triple the context window. Extrapolate naïvely and you’re looking at 50,000 GPUs and 40–50 MW of continuous power draw for most of a quarter. This isn’t cloud computing. This requires a facility akin to a mid-size steel mill. It demands a dedicated substation, immersion cooling, and grid interconnects carved out years in advance.
Critics object that better chips will slash the wattage. True: ASICs, photonic interposers, wafer‑scale integration, and cryo‑CMOS all promise order‑of‑magnitude jumps in joules per inference. But energy efficiency rarely lowers total consumption; it widens the aperture for demand. Britain’s coal imports rose, not fell, after James Watt perfected the steam engine. Jevons is not a footnote; he is the law.
Even granting heroic progress, every marginal watt saved would be devoured by new model variants, multimodal training, agentic swarms, and the simple compulsion to extend context windows until they embrace all of world history. The learning curve bends, the power curve climbs.
3. A Vignette From the Grid’s Front Line
Last autumn Amazon asked Dominion Energy for more power to feed its Ashburn data‑center cluster. Dominion, which serves the Pentagon and half of Washington’s suburbs, blinked. But Dominion has repeatedly reported that its power lines can’t handle more demand. The local county board, caught between tech‑lobby promises of “AI jobs” and homeowner rage over brownouts, sided with the grid operator. The most valuable compute corridor in the world slammed shut overnight. Microsoft, reading the tea leaves, quietly filed for permission to install small modular reactors (SMRs) on two of its own campuses in West Texas and Washington State.
Geopolitical concerns flow from these developments. As AGI capabilities harden into strategic assets, their underlying infrastructure becomes entangled with sovereignty itself. Advanced semiconductors are now governed by export controls more draconian than arms treaties. TSMC is no longer a supplier. It is a geopolitical tripwire. If SMRs become the power source of cognition, then whoever controls those reactors sets the tempo of thought. And if Western nations allow environmental review boards and zoning commissions to choke 400‑megawatt builds, compute will migrate to authoritarian regimes that greenlight substation buildouts by decree. This is not merely regulatory arbitrage. It is civilizational arbitrage. In the end, the substrate of intelligence will belong to the sovereigns who move fastest, build deepest, and control the grid.
That is the shape of the future: sovereign compute islands, welded to private generation, insulated from NIMBY politics: Starbases for thought.
4. Building an AGI Starbase
What does it take to build the substrate of artificial general intelligence? Start at the breaker panel.
AGI is a thermodynamic force, and it hungers for continuous power—100 to 500 megawatts of baseload capacity, depending on scale. These aren’t app servers in rented racks. These are multi-acre, sovereign-grade facilities, the kind that draw load profiles closer to aluminum smelters than to hyperscaler data centers. Ideal siting lies near hydroelectric dams, gas co-generation stations, or, increasingly, small modular reactors built not for the grid, but for cognition. Power independence is not an optimization. It’s the precondition for autonomy.
But electrons aren’t enough. Each rack, pushing 10 to 15 kilowatts, becomes a localized furnace. The thermal footprint demands immersion cooling or cold-plate liquid loops: no fans can keep pace. Waste heat must be cycled into closed-loop reclamation systems or vented via evaporative towers and coastal outflows. The AGI Starbase is a heat management system first, and a thinking machine second. The constraints are not virtual. They are hydraulic and thermodynamic.
Then come the chips. Not just GPUs, but wafer-scale slabs, ASICs, and whatever next-generation architectures survive the frontier. These components must be secured under multi-year take-or-pay contracts, and these contracts must be signed years before the build completes. Advanced packaging is the bottleneck. Facilities like ASE and Amkor, or whatever sovereign-backed alternatives emerge, must be locked in as upstream partners. High-bandwidth memory, measured in container-loads, not sticks, moves under guarded custody. In this regime, logistics is more than throughput. It’s custody of cognition.
Data follows. Text is no longer enough. The training sets now absorb petabytes of video, synthetic sensorium streams, and rendered self-play simulations. These bits cannot simply be scraped; they must be licensed, synthesized, or harvested from controlled environments. Fiber routes become geopolitical assets: satellite constellations introduce jitter; true latency determinism demands physical undersea cable rights, which must be negotiated with sovereigns. On-site pre-processing becomes mandatory; raw bits are too heavy to ship at scale.
The human infrastructure is no less critical. Power systems must be monitored by engineers who understand substation arc faults and reactive loads. Cryogenics techs oversee cooling infrastructure. Reliability engineers probe silicon at the edge of electromigration failure. Red-team adversarial squads stress-test agentic systems that may rewrite their own subroutines mid-inference. Most of all, the site must run in shifts, 24/7, without burning out the few hundred humans on Earth capable of rebooting a 10-trillion-parameter model that’s crashed at 3 a.m.
Security is non-negotiable. These facilities are hardened against ISR overflight, drone incursions, and plain-old sabotage. Compute weights will be treated as munitions under ITAR equivalents. Legal architectures will span Delaware LLCs, Singapore holding companies, and Cayman SPVs. This is not just for tax efficiency, but to navigate the CFIUS maze and avoid strategic asset seizure. The legal shell becomes part of the infrastructure stack.
Remove any pillar—power, cooling, chips, data, doctrine, labor, or legal envelope—and recursive self-improvement halts. The mind is not above the substrate. It is the substrate. AGI is not a clever abstraction. It is a supply chain running at the edge of failure.
5. The Steel‑Man Counterargument
Optimists reply that the entire premise is a Kodak‑moment fallacy. As models compress, they argue, we will finetune offline on cheap edge devices; training will blur into inference; the “GPU famine” will fade as photonics drops watts per token by 100×.
All partially true. Yet the ceiling simply rises.
A ratchet hides inside scaling laws: every time tokens per joule improve, inventive researchers spend the dividend not on thrift but on fresh parameter space: longer sequences, richer modalities, self‑play universes. Efficiency liberates ambition faster than it erases cost. The net load goes up.
If a doubter still objects, ask her to name one computational epoch, such as mainframes, PCs, or smartphones, that ended with less aggregate electricity consumed than before it began. Information wants to eat power.
6. From Polemic to Prescription
So what? Polemics are cheap; substations are not. If AGI truly rhymes with Starbase, three audiences must act:
Investors. Shift DCF models from ARR fantasies to cap‑ex velocity. The moat will be who can finance 300 MW campuses at single‑digit WACC, not who can prompt‑engineer best.
Policymakers. Fast‑track SMR licensing, overhaul export‑control regimes that treat GPU clusters like kidney dialysis machines, and pre‑empt local zoning choke points, or watch the industry migrate to whichever jurisdiction writes the shortest environmental‑impact statement.
Strategic Operators. Secure back‑end packaging capacity and HBM supply now. Futures contracts or outright equity stakes in ASE, TSMC’s advanced‐packaging lines, or upcoming India fabs will matter more than yet‑another‑road‑tested‑LLM slogan generator.
7. Terraforming the Substrate
Starbase was less a rocket pad than a terraforming project. It reshaped regulation, labor, and coastal geography to turn orbital logistics into a routine service. AGI demands an analogous terraforming of the grid, the chip stack, the labor market, and the legal code. This is not optional overhead; it is the work.
If you are not pouring concrete, negotiating easements, and soldering cold plates, you are not building AGI. You are writing fan‑fic. Intelligence will not float above the planet like a benevolent ghost. It will thrum inside caverns of steel and coolant, sucking electrons the way a refinery gulps crude, guarded by men with biometric badges and anti‑drone jammers.
Brains in boxes make tidy thought experiments. But the mind that changes history will arrive in a roar of electrical and algorithmic transformers which bear more resemblance to a blast furnace than to a whitepaper. When that day comes, we will recognize the shape. We have seen it once already, rising out of the mangroves at Boca Chica.
The next Starbase will not hurl metal skyward. It will mint cognition on the anvil of infrastructure. It will be powered by sovereign reactors, cooled by militarized pipelines, and shielded by treaty-proof legal armor. The future will not be led by those who align models. It will be led by those who align substations.
Power Usage Effectiveness. It’s a data center efficiency metric defined as: PUE = Total Facility Power / IT Equipment Power
IT Equipment Power is the power consumed directly by servers, GPUs, storage, networking, etc.
Total Facility Power includes everything: cooling, power distribution losses, lighting, etc.