The Physical Limits of Artificial Intelligence: Inside the Compute Rush

The artificial intelligence boom is usually described in terms of models, parameters, and breakthrough capabilities. But beneath the chatbots and image generators lies a frantic, physical scramble for the raw materials of computation: specialized silicon, massive data centers, and gigawatts of reliable electricity. Christian Hamilton's Compute Rush steps away from the algorithmic hype to document the supply chains, grid constraints, and geopolitical maneuvers that will determine whether AI's promise can actually be built.

What the book is about

Compute Rush is organized into 25 chapters that trace the AI infrastructure stack from chip fabrication to community permitting. The book's structure follows a "silicon, steel, and steam" framework: early chapters examine what's driving compute demand (model scaling, training vs. inference economics, latency requirements, data bottlenecks); middle sections dissect the semiconductor supply chain (GPU dominance, challengers, foundry concentration, export controls, networking); later chapters confront the physical realities of data center siting, cooling revolutions, power generation (gas, renewables, nuclear), and the social license to operate. Hamilton draws on interviews with executives, engineers, grid operators, and policymakers, plus case studies like Oracle's 131,000-GPU supercluster and Microsoft's 24/7 carbon-free energy push. The intended reader is anyone who needs to understand AI's physical constraints — investors, policymakers, infrastructure professionals, and technically curious observers — rather than specialists in any single domain.

The compute curve is driven by economics, not just model size

Hamilton argues that AI's insatiable compute demand isn't simply about researchers building bigger models for prestige. Chapter 1 traces how "killer use cases" across healthcare, finance, and manufacturing create tangible business value that justifies exponential spending. Chapter 2 introduces a crucial distinction: training is a one-time capital expense, but inference — serving models to users — becomes a relentless operational cost that scales with adoption. A case study of a major hyperscaler ("CloudCo") shows how inference costs threatened to exceed the entire R&D budget, triggering a $500 million annual savings program through hardware specialization, quantization, and dynamic scaling. The "cost-per-token" and latency metrics that drive product decisions are, in Hamilton's framing, the economic engines pulling the physical infrastructure behind them.

Chips are a geopolitical choke point, not just a supply chain issue

Chapters 6 through 9 reveal how a handful of companies and facilities control the world's most advanced AI silicon. Nvidia's CUDA ecosystem creates what Hamilton calls "GPU Nation" — a de facto monopoly reinforced by software lock-in and allocation power. But the deeper bottleneck is fabrication: TSMC produces roughly 92% of leading-edge chips, and ASML holds a near-monopoly on the EUV lithography machines required to make them. Chapter 8's "Foundry Funnel" explains how yield rates, advanced packaging (especially CoWoS for stacking high-bandwidth memory), and multi-year capacity planning create rigid constraints no amount of money can quickly solve. Chapter 9 then layers on export controls, industrial policy (CHIPS Act, EU Chips Act), and the emergence of "friend-shoring" as governments treat compute access as national security. The 2024 Taiwan earthquake case study makes the geographic concentration visceral: even a brief disruption at TSMC rippled through global AI roadmaps.

Power delivery, not generation, is the immediate crisis

The "Great Reforecast" (Chapter 16) documents how utility load forecasts have been upended: U.S. data center demand may double from 35 GW to 78 GW by 2035, and global data center electricity could reach 945 TWh by 2030 — roughly Japan's current total consumption. But Chapter 17, "Wires, Not Just Watts," identifies the real bottleneck: transmission and interconnection queues. Over 2,600 GW of projects sit in U.S. interconnection queues — more than double the entire installed grid capacity. Lead times for large transformers and substation upgrades stretch to years, while data centers can be built in months. Hamilton details how utilities like Dominion Energy and PJM are pursuing dedicated high-voltage corridors and AI-assisted grid management, but the mismatch between digital speed and physical infrastructure pace remains a central tension.

Cooling and water have become design constraints, not afterthoughts

Chapter 12's "Cooling Revolutions" and Chapter 14's "Water, Community, and Permitting" show how thermal management has migrated from the mechanical room to the boardroom. Air cooling tops out around 15–20 kW per rack; AI workloads now demand 40–100 kW, forcing adoption of direct-to-chip liquid cooling, rear-door heat exchangers, and immersion cooling. A retrofit case study (Synapse Corp) demonstrates how legacy facilities can adapt via hybrid approaches. But liquid cooling shifts the resource burden to water: evaporative cooling towers consume millions of gallons, sparking community opposition in water-stressed regions. The Loudoun County case study shows a startup forced to redesign for 60% water reduction and invest in community benefits to secure permits. Noise, land use, and the disparity between construction jobs and minimal operational employment further complicate the "social license" Hamilton explores in Chapter 24.

Three divergent futures for 2025–2030

The final chapter presents three scenarios rather than a single forecast. "Constrained Grid" sees transmission bottlenecks and community resistance throttling AI growth, favoring incumbents with secured power and accelerating edge/on-device AI. "Nuclear Renaissance" envisions SMRs and microreactors providing dedicated, carbon-free baseload for AI campuses, with tech giants signing long-term PPAs to de-risk nuclear deployment. "Efficient AI" bets on software breakthroughs — quantization to 4-bit, sparse architectures, compiler optimization — decoupling capability growth from compute demand. Hamilton treats these as overlapping forces: "The reality will likely be a hybrid, a constant dance between accelerating demand, infrastructural constraints, technological ingenuity, and evolving policy." The scenarios give readers a framework for tracking which leading indicators (interconnection queue clearance rates, SMR regulatory approvals, open-weight model efficiency gains) signal which future is materializing.

Who should read this

Compute Rush is essential for infrastructure investors, utility planners, policymakers, and technology strategists who need a systems-level view of AI's physical dependencies. Engineers working on chips, cooling, or power will appreciate the cross-disciplinary connections Hamilton draws. Readers looking for a treatise on AI algorithms, model architectures, or societal implications of AI capabilities should look elsewhere — this book treats intelligence as a given and asks what it takes to run it. The density of case studies, expert quotes, and specific metrics (cap rates, PUE targets, interconnection queue sizes) makes it a reference work as much as a narrative. If you've wondered why data center REITs are trading at 4.4% cap rates or why Nvidia's allocation decisions move markets, this is the book that connects those dots.

Read “Compute Rush” on MixCache.com →