Blog · Review Essay · Last reviewed June 23, 2026

Chip War and the Compute Substrate of AI

Chris Miller's Chip War: The Fight for the World's Most Critical Technology is a history of semiconductors, but its AI-era value is sharper than a hardware backstory. It explains why AI capability at scale is never only a model. It is a supply chain, a foundry, a lithography tool, a packaging bottleneck, an export license, a power contract, a workforce, and a state strategy.

The compute substrate is the physical, institutional, and legal layer that lets AI systems be trained, served, evaluated, and contested: chips, high-bandwidth memory, interconnects, clouds, data centers, power, cooling, software stacks, capital, export rules, and allocation decisions that decide who can build, audit, reproduce, or depend on them.

The Book

Chip War was published by Scribner on October 4, 2022. Simon & Schuster's publisher page lists the hardcover at 464 pages with ISBN 9781982172008, and describes the book as an account of the decades-long struggle to control microchip technology. The same page notes later paperback material on the CHIPS Act, U.S. export controls on China, and allied technology protection.

Miller is a professor of international history at Tufts University's Fletcher School, where his profile lists research in technology, geopolitics, economics, international affairs, and Russia. That background shapes the book. It is not a narrow engineering chronology. It is a history of how tiny devices became the terrain on which states, firms, militaries, universities, investors, and manufacturing systems compete.

The book's public reception confirms its institutional importance. The Financial Times named Chip War the 2022 Business Book of the Year, emphasizing its account of long, fragile semiconductor supply chains and global dependency on a small number of major manufacturers. The Council on Foreign Relations later awarded it the 2023 Arthur Ross Book Award gold medal for its contribution to understanding international relations.

Those awards are not the reason the book matters here. The reason is that it makes the physical substrate of digital power visible. A public that meets AI through a chat window can easily forget that every answer depends on fabs, masks, chemicals, optics, memory, interconnects, data centers, cooling systems, electricity, logistics, and political bargains that are usually hidden by the finished interface.

Compute Substrate, Defined

Compute substrate means the complete material and institutional base that makes large-scale computation usable. It includes accelerators, high-bandwidth memory, advanced packaging, lithography, EDA software, firmware, servers, networking, storage, data-center sites, power contracts, cooling, cloud identity, export licenses, vendor contracts, and the operators who keep the system available.

The definition matters because a chip sitting in a warehouse is not AI compute in the practical sense. It becomes AI compute only when it is powered, cooled, networked, scheduled, secured, supplied with software, attached to data and workloads, and placed under an authority that decides who may use it. That is why compute is both a technical resource and a governance object.

The word substrate is doing work. It is not a synonym for GPU count. A usable substrate includes the conversion from component to capacity and from capacity to accountability: identity, logs, queueing, attestation, procurement terms, maintenance, incident records, and exit paths. A cluster can be powerful but nearly ungovernable if nobody outside the operator can reconstruct who used it, under what conditions, and with what dependency on vendors or export rules.

This frame keeps the analysis concrete. The question is not whether hardware magically produces intelligence. It is who controls the scarce machinery, which projects get access, which risks trigger oversight, which communities absorb the infrastructure burden, and whether public institutions can test systems without renting all their visibility from the largest platforms.

Current Context

As of June 23, 2026, compute has become an explicit policy category. Article 51 of Regulation (EU) 2024/1689 presumes that a general-purpose AI model has high-impact capabilities when cumulative training computation is greater than 10^25 floating-point operations, while allowing the Commission to update thresholds as technology changes. That is not a full safety test. It is a legal sign that compute scale has entered the governance vocabulary.

U.S. semiconductor policy remains active and unsettled. BIS controls from 2022 onward still frame advanced chips, semiconductor manufacturing equipment, supercomputer end uses, high-bandwidth memory, foreign-direct-product rules, and diversion risks as national-security issues. In May 2025, BIS announced non-enforcement and planned replacement of the AI Diffusion Rule; in May 2026, GAO concluded that the press release's non-enforcement policy was a rule subject to Congressional Review Act submission. In January 2026, BIS moved some license applications for Nvidia H200, AMD MI325X, and similar chips to case-by-case review under specified conditions. BIS's May 31, 2026 guidance clarified that license requirements continue for certain advanced computing items destined for entities headquartered in Country Group D:5 or Macau, even when the entity is located elsewhere; a June 17, 2026 FAQ further clarified that the guidance applies to related ".z" paragraph items, not only 3A090.a and 4A090.a chips.

Energy and siting are now part of the same story. The International Energy Agency's April 2026 Key Questions on Energy and AI update reports that global data-center electricity demand grew 17% in 2025, while AI-focused data centers grew 50%; its central projection roughly doubles total data-center electricity consumption from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. Those figures cover data centers broadly, not only AI, but AI is now one of the forces pushing campuses, interconnection queues, local power negotiations, cooling decisions, and public infrastructure conflict into the open.

Supply-chain transparency has also become more concrete. CISA and G7 partners released Software Bill of Materials for AI - Minimum Elements on May 12, 2026; ANSSI describes the AI SBOM as a mapping of an AI system's supply chain, deployed components, and dependencies. That matters for this book because the hardware layer is no longer just a background input. It is part of the record that buyers, auditors, regulators, and public institutions need if they are going to understand what kind of machine they have inherited.

Chips Before Models

The useful correction in Chip War is simple: there is no placeless AI capability. The model that appears to speak from nowhere depends on machines built somewhere, by companies and workers operating inside specific jurisdictions, capital markets, supply chains, and security regimes.

This makes the book a companion to The Stack, A Prehistory of the Cloud, the data-center civic machine, and AI factory industrial policy. Each text breaks the illusion that computation is a smooth surface. Miller's contribution is to show the chip layer as a strategic layer: not only a component market, but a control point for military capacity, economic growth, cloud scale, and AI capability.

OECD's 2025 work on AI infrastructure makes the connection explicit: training and deploying AI systems relies on a complex, capital-intensive global supply chain, with chips, data centers, cloud computing, power, cooling, and network infrastructure all forming part of the AI compute ecosystem. In that context, Chip War reads less like a pre-AI business history than like a map of the terrain under frontier AI.

This matters for safety because a risk register that begins with model behavior begins too late. Failure can arrive through chip provenance, queue priority, cloud tenancy, sanctions exposure, supply diversion, grid congestion, data-center outage, opaque managed-service logs, or a regulator's inability to reproduce what a dominant platform can run. The substrate is not background. It is where many permissions and many failure modes are set.

Auditability is therefore partly an infrastructure property. A model card may describe intended use, but the ability to check that description depends on logs, reproducible environments, retained artifacts, model-weight custody, cloud records, and enough independent compute to rerun or stress the system. When those pieces are missing, accountability becomes a negotiation with the infrastructure owner.

The implication is uncomfortable for any model-centered theory of progress. Better algorithms matter. Better datasets matter. But capability also depends on who can buy accelerators, secure high-bandwidth memory, reserve advanced packaging capacity, schedule cloud clusters, finance data-center buildout, and keep export-controlled tools inside trusted channels. AI does not float above political economy. It condenses it.

The Border Inside the Machine

Miller's book clarifies why compute governance has become a border problem. If advanced AI depends on advanced chips, then export control, foundry geography, equipment chokepoints, cloud access, and allied coordination become ways of governing model capability before the model exists.

The U.S. Bureau of Industry and Security made that logic visible in October 2022, when it announced controls on advanced computing and semiconductor manufacturing items involving China. The rules added controls on certain semiconductor manufacturing equipment, created license requirements for PRC facilities meeting specified fabrication thresholds, restricted some U.S. person support for relevant chip development and production, and phased in advanced computing and supercomputer controls.

BIS then kept revising the boundary. Its December 2024 package added controls on semiconductor manufacturing equipment, software tools, high-bandwidth memory, foreign-direct-product pathways, red-flag guidance, and many Entity List actions. The January 2026 license-review revision and May-June 2026 guidance show the policy problem in miniature: a compute border must track chips, related items, parent-company control, remote access, supply assurance, third-party testing, and the incentives created by each technical line.

That policy history belongs beside the compute-border problem. The border is not only a line around territory. It is also a line through fabrication tools, chip designs, cloud accounts, training clusters, model weights, packaging capacity, software stacks, and professional knowledge. A state trying to govern AI eventually discovers that the machine's body is distributed across allies, vendors, campuses, ports, data centers, and standards.

It also discovers that technical thresholds become political incentives. Draw a line around a chip parameter, and designers search for a compliant workaround. Restrict a tool, and firms reroute procurement. Limit on-premise chips, and cloud access becomes more valuable. Regulate model weights, and open weights, distilled models, remote inference, and cross-border services become part of the next policy argument. The technical system answers the rule, and the rule answers the technical system.

Good governance therefore has to keep policy instruments distinct. Chip-shipment controls, cloud-access controls, model-weight controls, procurement rules, research-compute grants, and public-sector audit powers each create different incentives and different civil-liberties risks. Treating them as one generic "AI control" invites both leakage and overreach.

The Recursive Supply Chain

The strongest AI-era reading of Chip War is recursive. Chips make AI systems possible. AI systems increase demand for chips. That demand changes capital allocation, data-center planning, energy contracts, export controls, national industrial strategies, and the design priorities of chip firms. Those changed conditions then shape which AI systems can be trained next.

This loop is not abstract. NIST's CHIPS for America materials describe semiconductor devices as critical components in AI, quantum computing, and other advanced technologies. Its March 2024 fact sheet says the CHIPS and Science Act of 2022 provides the Department of Commerce with $52.7 billion over five years to boost semiconductor manufacturing and research in the United States while investing in workers, including $39 billion for incentives and $11 billion for domestic R&D ecosystem development.

That is a feedback loop between prediction and production. Governments subsidize chip capacity because they expect compute to be strategic. Firms build data centers because they expect model demand to grow. Model developers design larger or more inference-heavy systems because infrastructure appears available. The resulting demand becomes evidence that the original infrastructure bet was necessary.

The loop is also allocative. As AI demand rises, foundry slots, high-bandwidth memory, advanced packaging, power capacity, cloud reservations, and engineering talent are reprioritized around the actors most able to pay or to claim strategic importance. Scarcity then becomes a design constraint, a market signal, and a political fact at the same time.

The same recursive pattern appears in smaller forms. A benchmark rewards more compute. Labs buy more compute. More compute produces stronger benchmark results. Investors fund the labs. Cloud providers build for the labs. Regulators begin to use compute as a proxy for risk. Model capability becomes easier to imagine as a function of hardware scale, and hardware scale becomes a governance category.

What AI Adds

Chip War was published as generative AI was about to become a mass interface. Its newer paperback context points toward AI, but the book's central narrative still comes from a broader semiconductor history: Cold War electronics, consumer devices, manufacturing specialization, Japan, Korea, Taiwan, China, the United States, lithography, military systems, and globalization.

AI adds three pressures to that history.

First, it makes advanced chips socially legible. Most people never cared which node, package, memory stack, or accelerator architecture mediated their digital life. Generative AI changed that. GPUs, NVIDIA, TSMC, CUDA, high-bandwidth memory, and advanced packaging became public-policy terms because they now appear to govern who can train, rent, or deploy AI capability at scale.

Second, AI turns chips into institutional capacity. A hospital, school, city, laboratory, startup, military command, or regulator that lacks compute may still use AI through vendors. But it does not control the capacity on which it depends. It rents AI capacity through someone else's infrastructure, logs, terms, models, update schedule, and failure modes.

Third, AI turns infrastructure scarcity into social sorting. If advanced compute is limited, somebody allocates it. The allocation may happen through markets, national-security priorities, cloud contracts, research grants, public compute programs, export licenses, procurement rules, or platform partnerships. The result decides which actors can experiment, audit, compete, comply, or resist.

The allocation question is where the site's recurring themes become material. The interface presents a single answer, but upstream access rules decide whose systems can exist, whose failures are studied, whose evidence is preserved, and whose neighborhood absorbs the power, water, land, and tax choices required by the machine.

Governance and Safety

A useful Chip War reading should produce operational questions, not just geopolitical awe. Compute is powerful because it is countable enough to govern, physical enough to locate, expensive enough to concentrate, and distributed enough to evade simple control.

Compute governance can improve safety when it gives regulators, researchers, and affected institutions enough capacity to inspect systems independently. It can also deepen risk when thresholds, licenses, and cloud chokepoints centralize power without transparency, appeals, privacy limits, competition safeguards, and protected access for public-interest testing.

Keep four ledgers. A serious institution should maintain a capacity ledger for chips, clusters, regions, energy, cooling, and queue limits; an access ledger for who can run which workloads and under what authority; a dependency ledger for vendors, tools, logs, model custody, and exit paths; and an externality ledger for land, power, water, labor, tax incentives, resilience, and emergency-service burden. The ledgers are not paperwork after the fact. They are how oversight reaches the substrate before the interface becomes normal.

Start with an AI bill of materials. A model or service should be understood through its chips, memory, packaging, cloud provider, data-center region, model provider, software stack, data sources, power assumptions, vendor dependencies, and export-control exposure. The site's AI bill of materials frame turns the hidden substrate into an auditable map, and the 2026 G7/CISA AI SBOM guidance is a useful marker because it treats infrastructure as part of the AI supply-chain record.

Separate safety compute from capability compute. Red-teaming, evaluations, privacy tests, robustness checks, incident review, provenance work, and independent audits require compute too. A governance regime that funds training but treats evaluation as overhead will systematically underbuild safety capacity.

Use compute thresholds with humility. Thresholds can trigger notice, evaluation, reporting, or regulator attention, but they are not a theory of intelligence. Better algorithms, data quality, tool use, post-training, inference-time methods, model distillation, and deployment context can all move risk outside a simple training-FLOP line.

Govern cloud access as well as chip shipment. Export controls that stop a box at customs can still leak through rented compute, offshore subsidiaries, managed training jobs, remote inference, or cloud accounts. That pushes governance toward customer due diligence, scoped identities, logging, privacy limits, appeals, and clear rules for public-interest researchers.

Build public capacity. If only hyperscalers and heavily financed labs can afford serious compute, independent evaluation, open science, public-sector AI, incident review, and local adaptation become dependent on corporate permission. Public compute programs, secure research environments, and procurement terms are therefore safety infrastructure, not merely research support.

Make local costs visible. Data centers draw on land, substations, water or air cooling, tax incentives, labor, emergency services, and political consent. A national AI strategy that counts chips but ignores communities near the infrastructure is not complete governance.

Where the Book Needs Friction

The book's strength is its state-firm-industrial narrative. That is also its limit. It is excellent at showing why chips matter to geopolitical competition, but readers should pair it with books that foreground labor, extraction, environmental cost, repair, public accountability, and the people who live near the infrastructure. Atlas of AI, Feeding the Machine, Data Driven, and The Costs of Connection pull those questions forward.

Readers should also resist turning compute into a master variable. Training FLOPs, chip shipments, memory bandwidth, and data-center power are real constraints, but they do not by themselves explain model behavior, institutional judgment, or social harm. Deployment context, data rights, tool permissions, human review, business incentives, and legal remedies still decide how the system acts on people.

It also should not be read as a simple brief for national self-sufficiency. Semiconductor supply chains are deeply specialized for a reason. No country can easily reproduce every layer of design, fabrication, equipment, chemicals, packaging, testing, logistics, software, and talent without enormous cost and delay. Sovereignty in this domain often means managed dependence, not pure independence.

Finally, the national-security frame can crowd out democratic governance. A chip shortage, export-control dispute, or AI arms-race story can make public scrutiny feel like delay. But the systems being built will shape work, energy, surveillance, education, science, war, and public memory. The fact that chips are strategic does not make them too important for democratic argument. It makes the argument more urgent.

What This Changes

Chip War changes the AI question from "What can the model do?" to "What substrate lets the model exist, and who controls that substrate?"

For AI governance, that means compute policy cannot be a side topic. Model audits, safety cases, licensing, incident reports, procurement rules, and transparency registers all sit downstream from the material capacity to train and serve models. If only a few firms and states can secure that capacity, governance becomes entangled with market concentration before any chatbot reaches a user.

For public institutions, the lesson is to track dependency. Which systems require rented cloud compute? Which vendors control the chips, software stack, and logs? Which workloads can be paused, audited, moved, or shut down? Which public services would fail if model access, data-center capacity, or export policy changed?

The practical test is blunt: could the institution explain the system's compute dependency to a regulator, a procurement officer, a local utility, an affected person, and its own incident-response team? If each audience receives a different story, the substrate is not yet governed. It is merely rented.

For readers, the lesson is simpler. When an interface feels weightless, ask where its weight went. The answer will include chips, power, water, workers, fabs, tools, ports, permits, treaties, subsidies, export controls, and balance sheets. The machine's apparent intelligence is partly the visible tip of a hidden industrial arrangement.

Miller's book belongs on an AI reading shelf because it forces that arrangement back into view. The future is not being generated only in prompts. It is also being etched, packaged, cooled, shipped, financed, licensed, and defended.

Source Discipline

This review separates book facts, current policy facts, and interpretation. Publisher, university, award, and official government pages carry the bibliographic and policy claims. Interpretive claims about compute as substrate, border, and allocation are readings built from those facts, not predictions that any particular AI outcome is inevitable.

For current compute claims, the important habit is to name the unit and boundary: training compute, inference compute, evaluation compute, chip type, memory bandwidth, data-center load, energy use, export-control threshold, or allocation rule. A vague claim about "more compute" is not enough for governance.

Policy claims also need status words. A statute, interim final rule, press release, non-enforcement policy, GAO legal decision, guidance document, FAQ, voluntary standard, and research paper do different kinds of work. This review dates the legal and policy posture to June 23, 2026 because export controls, AI SBOM guidance, and energy forecasts are moving targets.

This review makes no claim that any AI system is conscious, divine, or AGI. When it says "AI capability at scale," it refers to deployed machine-learning capacity as a social and infrastructural phenomenon, not to personhood or metaphysical status.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books