🌐 The Global AI Arms Race · Linh Truong
🌐 Personal notes · May 2026

The Global AI Arms Race

My working map of the contest now reorganizing world power: who controls compute, chips, energy, and frontier models — and whether the race for advanced AI ends in managed competition or runaway escalation. This is the strategic backdrop behind every other note in this series.

The question behind this note: is AI a normal technology race that markets and diplomacy can absorb — or a security competition where the perceived prize (decisive economic and military advantage) is so large that restraint feels like surrender? What follows is my read of the players, the chokepoints, the timeline, and the levers that decide whether this stays a race or becomes a confrontation.
📍 Axis: United States ↔ China ⚙️ Prize: Compute · Models · Autonomy 📅 Horizon: 2017 → 2030 🧭 Stance: Race is real; outcome is a choice ✍️ By: Linh Truong
2 poles
The US and China define a bipolar frontier — most other states pick a stack, supply a chokepoint, or buy access to one side
~$300B+
Annual AI-infrastructure capex by the largest US hyperscalers alone (2025) — multi-year build-outs now run to the trillions
3 chokepoints
EUV lithography, leading-edge foundries, and high-bandwidth memory — narrow gates that turn supply chains into leverage
2 endings
Managed competition with guardrails, or an unbounded race that pressures safety, exports surveillance, and risks conflict
01 · What I'm tracking

The arms race in one page

AI is being treated by the major powers as a strategic technology — like nuclear, aerospace, or semiconductors before it — because it promises compounding advantage in economic productivity, military capability, and information dominance simultaneously. That triple payoff is why investment behaves like a security competition rather than a normal market: nobody wants to be second, and "second" is assumed to be permanent.

Why it's a race

Winner-take-most fear

Leaders believe AI advantage compounds — better models attract more compute, capital, and talent, which build still-better models. The fear of a runaway leader makes restraint feel unilaterally disarming.

What's contested

The full stack

The fight runs from energy and chips at the bottom to frontier models and autonomous systems at the top. Whoever controls the narrow layers — fabs, lithography, memory — holds leverage over everyone above.

What's at stake

The shape of world order

The race is reorganizing alliances, trade, and deterrence. It can entrench a bipolar tech-bloc world, export surveillance and autonomy, and erode the norms that keep great-power competition below the threshold of conflict.

🎯

What I keep coming back to: the dangerous part of an arms race is rarely the technology itself — it is the logic it imposes. Each side reads the other's investment as a threat, accelerates to stay ahead, and treats safety, transparency, and arms-control as costs that slow it down. The strategic task is to change that logic: make verification cheaper than mistrust, make chokepoints serve stability rather than just denial, and keep enough cooperation alive that the worst tail risks — accidents, proliferation, autonomous escalation — are managed jointly even by rivals. The race is real. The ending is still a choice.

02 · The Big Picture

The AI arms-race system — how inputs become strategic power

A narrow set of capability inputs (compute, energy, data, capital) is poured into arenas of competition (frontier research, the chip supply chain, military autonomy, standards), producing forms of national power (economic, military, informational) that reshape the strategic order — polarity, alliances, and stability. Two feedback loops decide the ending: a race loop that accelerates, and a restraint loop that stabilizes.

FIG 1

Capability inputs → arenas of competition → forms of power → strategic order

1 · CAPABILITY INPUTS 2 · ARENAS OF COMPETITION 3 · FORMS OF POWER 4 · STRATEGIC ORDER Compute & Chips GPUs · accelerators · clusters Energy & Power grid · gas · nuclear · siting Data & Talent corpora · researchers · labs Capital & Capex hyperscalers · states · funds Frontier Models research · scaling · agents Chip Supply Chain EUV · fabs · memory Military & Autonomy drones · C2 · decision support Standards & Diffusion rules · exports · adoption Economic Power productivity · markets · jobs Military Power deterrence · speed · autonomy Information Power narrative · influence · soft power Polarity uni- ↔ bi- ↔ multipolar Alliances & Blocs chip coalitions · spheres Strategic Stability deterrence vs. accident Tech Sovereignty autonomy vs. dependence + Restraint loop: verification, controls, confidence-building, shared safety − Race loop: threat perception → more capex → faster capability → more fear ⚖ Governance gateway export controls · treaties · compute accounting
Read it left→right: the same inputs can produce either outcome. The governance gateway — export controls, treaties, and compute accounting — is the valve that decides whether power concentrates and escalates (red race loop) or stays bounded and verifiable (green restraint loop). Today the red loop is dominant; the green loop is thin but not closed.
03 · The Board

The players — who holds which part of the stack

Two superpowers set the frontier, but the race runs through a handful of others who hold a chokepoint (Taiwan, the Netherlands, South Korea), supply capital and energy (the Gulf), or are trying to build sovereign capacity (EU, India, UK, Japan). Read this as a power-distribution map, not a ranking — positions move quarter to quarter.

TBL 1

National scorecard — strengths, dependencies, and posture

PlayerFrontier modelsCompute & chipsEnergyGovernance postureCore role / dependency
United StatesLeadingDesigns · cloudConstrained gridMarket-led, pro-deployment, export-control leverOwns design (Nvidia), EDA, cloud; depends on Taiwan for fabrication
ChinaFast followerBuildingAbundantState-directed, self-reliance push, "AI+" strategyStrong on energy, data, deployment; throttled on EUV & HBM
European UnionCatching upImport-reliantMixedRights-based regulation (AI Act); sovereignty driveRule-maker & market; thin on compute & frontier labs
TaiwanMinorFoundry kingTightStrategic ambiguity; "silicon shield"TSMC fabricates most leading-edge AI chips — the single biggest chokepoint & flashpoint
NetherlandsMinorEUV monopolyOKAligned with US controls (contested)ASML is the sole maker of EUV lithography — no advanced chips without it
South KoreaEmergingMemory leaderMixedAllied; industrial-policy heavySK Hynix & Samsung dominate HBM — the memory that feeds AI accelerators
Gulf (UAE, Saudi)Buying inAcquiringCheap & abundantSovereign-fund led; courting both polesCapital + energy + land; building giant clusters (G42, HUMAIN); access depends on US sign-off
United KingdomResearch-strongThinConstrainedSafety-institute pioneer; pro-innovationHome of DeepMind (US-owned) & AISI; talent-rich, compute-poor
India · JapanBuildingNicheMixedAligned-ish; sovereign-AI ambitionsIndia: data & talent scale. Japan: materials & tools (a quiet chokepoint)
Legend: Decisive Constrained/Minor Medium Strong-ish Abundant. The strategic insight: no single country holds the whole stack. The US leads in design and models but cannot fabricate at the frontier domestically; China has energy and scale but is gated on lithography and memory. Interdependence is the leverage — and the risk.
FIG 2

Frontier capability vs. compute & supply-chain control — where players sit

Compute & supply-chain control → Frontier AI capability (research + models) → SUPPLY-CHAIN CHOKEPOINTS hold the stack · don't build models FULL-STACK SUPERPOWER capability + control DEPENDENT / ASPIRANT CAPABILITY-RICH, COMPUTE-LIGHT United States China Taiwan · TSMC Netherlands · ASML S. Korea · HBM Japan · tools European Union UK · DeepMind Gulf · capital India Positions are directional. The US sits top-right because it pairs frontier models with control over design, EDA, and cloud — even though fabrication lives in Taiwan.
The contested space is the upper band: the US wants to stay top-right; China is climbing from the center toward it; the chokepoint holders (top-left) are the prizes both poles court or coerce. The bottom-right (UK, EU) shows the recurring trap of the West-aligned world — strong research, weak compute sovereignty. Bubble size ≈ relative weight in the race, not GDP.
04 · The Decisive Terrain

The compute war — chips, chokepoints, and export controls

If AI is the race, advanced semiconductors are the track. The leading edge depends on a long, brittle supply chain with a few near-monopoly steps. Since 2022 the United States has weaponized those steps with export controls — turning the chip supply chain into the primary instrument of the arms race. This is where strategy is most concrete.

FIG 3

The semiconductor chokepoint chain — and where the valves are

From software to silicon to a frontier model — every advanced AI chip passes through all of these EDA software design tools US (Synopsys, Cadence) Chip design + IP GPUs · accelerators US/UK (Nvidia, AMD, Arm) EUV lithography prints finest features NL — ASML (sole) Fab tools + materials etch · deposit · resist US · Japan · NL Leading-edge fab 3–5nm manufacturing Taiwan (TSMC) · Korea HBM memory feeds the accelerator Korea (SK Hynix) · US AI clusters → frontier models US · China · Gulf ⚑ CHOKEPOINT ⚑ CHOKEPOINT ⚑ CHOKEPOINT ⛔ US-led export controls act as a valve on the chain By denying China advanced chips, EUV tools, and chip-making equipment, the US tries to cap rivals' compute — the "small yard, high fence" doctrine Each chokepoint is a point of leverage — and a single point of failure. A blockade or accident at any one cascades across the whole race.
The chain is the strategy. Three steps are near-monopolies: EUV lithography (ASML), leading-edge fabrication (TSMC), and HBM memory (SK Hynix/Samsung). Control any one and you can throttle a rival's frontier compute. This is why Taiwan is simultaneously the world's most important factory and its most dangerous flashpoint — the "silicon shield" cuts both ways.
TBL 2

The export-control escalation — how the "chip war" tightened

MoveRoughly whenWhat it didThe countermove
First broad controlsOct 2022US restricts advanced AI chips (e.g. A100/H100 class) and chip-making tools to ChinaNvidia ships cut-down parts (A800/H800) designed to sit just under the line
Tool & ally tightening2023Netherlands & Japan align on equipment limits; US closes the "just-under-the-line" loopholeChina accelerates SMIC capacity and domestic tool development; stockpiling
The China-specific chip2023–24Nvidia designs the H20 for the China market; later sales repeatedly restricted, paused, taxedHuawei's Ascend line (910B/910C) positioned as the domestic substitute
Tiered "diffusion" frameworkJan 2025Outgoing US rule sorts the world into tiers, capping where chips and model weights can flowAllies and the Gulf object to being "tiered"; framework becomes a bargaining chip
The pivot & the deals2025New US administration rescinds the tiered rule; shifts to case-by-case mega-deals (e.g. Gulf clusters) with security stringsChina leans into self-reliance & efficiency (see DeepSeek); courts the Global South
The pattern is a ratchet-and-adapt cycle: every restriction triggers a workaround (cut-down chips, smuggling, domestic substitutes, or algorithmic efficiency), which triggers a tighter restriction. The open strategic question — unresolved as of mid-2026 — is whether controls actually slow China's frontier or mostly accelerate its drive for self-sufficiency while straining allies. Dates and thresholds shift often; verify current rules before citing.
Doctrine

"Small yard, high fence"

The US aim isn't to decouple everything — it's to deny a narrow set of strategic capabilities (frontier compute) behind a very high wall, while keeping the rest of trade open. Easier said than enforced.

The Taiwan problem

One island, most of the chips

Concentrating leading-edge fabrication in Taiwan makes the whole world's AI ambitions hostage to one strait. "Friend-shoring" fabs to Arizona, Japan, and Germany is the hedge — slow and expensive.

The efficiency surprise

DeepSeek's lesson

China's DeepSeek showed strong models can be trained with far less top-tier compute than assumed. If algorithms outpace hardware denial, the entire export-control theory of victory weakens.

05 · The Combatants

The frontier labs — the corporate front of the race

States set the terrain, but a small number of frontier labs actually build the models. The striking feature of this race is that much of the cutting edge sits inside private companies backed by hyperscalers — making the competition simultaneously commercial, geopolitical, and (for some) explicitly safety-driven. Below is the rough order of battle.

TBL 3

The frontier labs at a glance

LabBlocFlagship lineBacking / computeDistinct angle
OpenAIUSGPT & o-series (reasoning)Microsoft + Stargate build-outScale + product reach; AGI mission with commercial arm
AnthropicUSClaudeAmazon & Google; AWS computeSafety-forward framing; enterprise & agentic focus
Google DeepMindUS/UKGeminiAlphabet's TPUs & cashFull vertical: own chips, data, distribution, deep research bench
Meta AIUSLlama (open weights)Meta's massive GPU fleetOpen-weight strategy — commoditize the model, win the ecosystem
xAIUSGrokOwn "Colossus" mega-clusterRapid compute build; integrated with a social platform
DeepSeekChinaV-series & R-seriesQuant-fund roots; constrained computeEfficiency breakthrough; open weights; shook markets in early 2025
Alibaba · othersChinaQwen (+ Zhipu, Moonshot, etc.)Chinese cloud giantsStrong open-weight models; domestic-chip migration underway
MistralEUMistral / MixtralEU capital; sovereignty narrativeEurope's flagship bid for frontier independence; open-ish weights
Legend: US-aligned China EU. Two cleavages matter as much as the US–China line: open weights vs. closed (Meta, DeepSeek, Mistral, Qwen vs. OpenAI/Anthropic/Google) and safety-forward vs. ship-fast. Open weights diffuse capability globally — undercutting export controls and accelerating the race from below. Model names evolve fast; treat the "flagship" column as the product family, not a specific version.
Structure

Hyperscaler dependency

Frontier labs are fused to cloud giants for compute and capital. The real capex — data centers, chips, energy — sits with Microsoft, Amazon, Google, Meta, blurring the line between company and national asset.

Tension

Mission vs. market

Several labs were founded on safety or AGI missions yet must now win a commercial race. The pressure to ship can erode the very caution that justified building them — a live governance problem.

The open-weight wildcard

Diffusion from below

Powerful open-weight models (often from Meta, Mistral, and Chinese labs) put near-frontier capability in anyone's hands. Great for access and competition — but it erodes any control regime built on guarding the frontier.

06 · The New Bottleneck

Energy & infrastructure — the constraint that surprised everyone

The race has quietly migrated from a chip story to a power-and-land story. Training and serving frontier models consumes electricity at industrial scale; the binding constraint in many regions is no longer "can we buy the GPUs?" but "can we power and cool the cluster, and connect it to the grid?" This reshapes who can compete — and reopens energy geopolitics.

The demand shock

Data centers as a load center

AI data centers are becoming gigawatt-scale consumers, straining grids and stretching interconnection queues for years. Power availability is now a siting decision — clusters chase cheap, abundant, fast electricity.

The nuclear pivot

Hyperscalers buy firm power

To secure clean, always-on supply, the largest players are signing nuclear deals — restarting reactors, contracting output, and funding small modular reactors (SMRs). Energy strategy is now AI strategy.

The geopolitical opening

Why the Gulf matters

States with cheap energy, capital, and open land — especially the UAE and Saudi Arabia — can host mega-clusters others can't. That turns energy-rich states into AI players, and access into a bargaining chip with Washington.

The friction

Grids, water, and politics

Local communities, water use, ratepayer costs, and climate targets create real resistance. Whether democracies can build infrastructure fast enough — without burning public trust — is itself a competitive variable.

Why this matters strategically: energy converts the race from a pure technology contest into a physical, industrial, and environmental one. China's abundant, fast-built power capacity is a genuine advantage; the West's slower permitting and constrained grids are a genuine drag. Whoever can pair frontier chips with cheap, clean, reliable power at scale holds an edge that no algorithm fully erases — and the climate cost of an unbounded build-out is a shared bill nobody is fully accounting for.

07 · The Hard Edge

Military AI & autonomy — where "arms race" is literal

Most of the AI race is economic, but its sharpest edge is military. Defense establishments now treat AI as the next revolution in military affairs — compressing the decision cycle, enabling mass autonomy, and integrating sensing-to-shooting. Ukraine has become the live laboratory; the Pacific is the scenario everyone games. This is where racing dynamics become most dangerous.

TBL 4

How AI enters the military — capability, status, and the risk it adds

ApplicationWhat it doesMaturityRisk it introduces
Autonomous drones & swarmsCheap, attritable mass; loitering munitions; coordinated swarmsFieldedLowers the cost of force; proliferates to non-state actors
ISR & targetingFusing sensor feeds to find and prioritize targets at machine speedMaturingCompresses the human's time to judge; accountability gaps
Decision support / C2AI-assisted command, planning, and battle managementMaturingAutomation bias; opaque recommendations under stress
Cyber & electronic warfareAutomated intrusion, defense, and spectrum operationsActiveFaster-than-human escalation; attribution problems
Logistics & predictive maintenanceSustainment, readiness, and supply optimizationAdoptingLower-risk, high-payoff — the "boring" force multiplier
Nuclear C2 interfaceAI touching early-warning / decision pipelinesRed lineThe gravest risk — most states verbally commit to human control of nuclear use
Legend: Already fielded Maturing/active Adopting Declared red line. The defense AI ecosystem now blends traditional primes with newer software-first firms; commercial labs have also loosened blanket bans on national-security work. The recurring debate at the UN (the CCW process) is over lethal autonomous weapons — whether "meaningful human control" can be defined and enforced before the capability is ubiquitous.
Speed

The compressed decision cycle

AI's promise is acting faster than the adversary. Its danger is the same: removing humans from the loop to gain speed, in a crisis where a wrong call escalates in seconds.

Mass

Attritable autonomy

Cheap autonomous systems flip the cost equation — many expendable platforms beating a few exquisite ones. Initiatives to field thousands of low-cost drones aim squarely at this, with the Pacific in mind.

Stability

The arms-control vacuum

Unlike nuclear weapons, military AI has no mature treaty regime, no verification, and no agreed thresholds. The technology is diffusing faster than the diplomacy meant to govern it.

08 · How We Got Here

Timeline — from research milestone to security competition

The arms race didn't begin with a model; it began when states recognized AI as strategic and started instrumenting the supply chain, the budget, and the diplomacy around it. A compressed view of the inflection points.

FIG 4

Milestones of the AI arms race, ~2017 → 2026

2017 2019 2020 2022 2023 2024 2025–26 China's national plan AI as state priority (2017) Scaling era begins large models prove scale (2020) ChatGPT moment AI goes mass-market (late 2022) Stargate & mega-capex trillion-scale build-outs (2025) First chip controls US export bans (Oct 2022) Bletchley Summit first safety summit (2023) EU AI Act in force first big rulebook (2024) DeepSeek shock efficiency upends assumptions (Jan 2025)
Two storylines converge: a capability curve (scaling → ChatGPT → reasoning models → agents) and a statecraft curve (national plans → export controls → summits → rulebooks). Around 2022 they fused — and AI stopped being a research field and became an instrument of national power. Dates are approximate anchor points.
2017–2021 · IGNITION

Recognition

States publish national AI strategies; scaling laws prove that more compute reliably buys more capability. The race is framed, not yet armed.

2022–2023 · ARMING

Weaponizing the chain

ChatGPT triggers mass awareness; the US opens the chip-control front; the first safety summit and executive actions appear. AI becomes statecraft.

2024–2025 · ESCALATION

Capex & rulebooks

Trillion-scale infrastructure plans; the EU AI Act; tiered export frameworks; DeepSeek's efficiency shock; energy becomes the binding constraint.

2026 → · CONTEST

Agents & autonomy

The frontier shifts to agentic systems and embodied/military autonomy. The open question moves from "who leads?" to "can this be governed at all?"

09 · The Race to Regulate

Governance — three models and a thin global layer

Running alongside the capability race is a race to set the rules. Three regulatory philosophies are crystallizing — and whoever's template the world adopts gains durable advantage. Above them sits a thin, fragile layer of multilateral coordination that has so far produced declarations more than enforcement.

EU · Rights-based

Regulate the risk

The AI Act classifies uses by risk, bans some outright, and imposes obligations on high-risk and frontier systems. Bets that the "Brussels effect" exports its rules globally — at the cost of speed and homegrown capability.

US · Market-led

Win, then guardrail

After an early safety-and-rights push, US policy pivoted toward deployment, deregulation, and competitiveness — treating leadership itself as the safety strategy, with controls aimed mainly outward at rivals.

China · State-control

Steer and secure

Rules emphasize content control, algorithm registration, and alignment with state priorities, paired with aggressive industrial support. AI serves stability and national strength first.

TBL 5

The multilateral layer — lots of talk, little enforcement

Venue / instrumentWhat it producedReachHonest limitation
Bletchley Summit (2023)First multilateral statement on frontier-AI risk; signed by US, China, EU & othersBroadDeclaratory; no binding commitments or verification
Seoul Summit (2024)Frontier-safety commitments from leading labs; network of safety institutesLabs + statesVoluntary; commitments self-defined and self-reported
Paris AI Action Summit (2025)Pivot toward "action," investment, and inclusionSelectiveSome key states declined to sign — coordination fraying
EU AI Act (2024→)First comprehensive binding rulebook, phased inEU marketRegional; enforcement & frontier-model rules still bedding in
UN bodies / scientific panelGlobal advisory reports; an international scientific assessment processUniversalAdvisory only; no teeth, slow, consensus-bound
Legend: Wide Partial. The uncomfortable truth: the governance that exists is mostly voluntary, regional, or declaratory. There is no IAEA-equivalent for AI — no body with inspection rights, agreed thresholds, or enforcement. Closing that gap, even partially and even among rivals, is the central diplomatic task of the next few years.
10 · The Ladder

The escalation ladder — from competition to conflict

"Arms race" implies a spectrum, not a single state. The relationship between the poles can sit anywhere from cooperative coexistence to open conflict. Mapping the rungs clarifies where we are, which way we're sliding, and which moves are off-ramps versus accelerants.

FIG 5

Rungs of the AI competition — and the off-ramps at each level

▲ escalation cooperation 1 · COOPERATION Joint safety research · shared standards · scientific exchange · crisis hotlines Off-ramp: build verification & trust 2 · COMPETITION Subsidies · capex race · talent poaching · standards rivalry — normal, mostly healthy Off-ramp: keep markets & science open 3 · RESTRICTION Export controls · sanctions · investment screening · partial decoupling — where we are now Off-ramp: "small yard," not whole economy 4 · CONFRONTATION Cyber sabotage · covert ops · supply-chain coercion · proxy pressure on Taiwan Off-ramp: red lines & back-channels 5 · CONFLICT Kinetic clash over chips/Taiwan · autonomous escalation · AI in nuclear pipelines Off-ramp: human-in-loop · de-escalation
As of mid-2026 the relationship sits around rung 3 (restriction), with recurring touches of rung 4 (cyber, coercion). The strategic goal isn't to eliminate competition — rung 2 is normal and even productive — but to prevent the slide to rungs 4–5 and keep enough of rung 1 alive that the worst tail risks are managed jointly. Off-ramps exist at every level; they require deliberate use.
11 · What To Do

The strategy stack — racing without crashing

If you accept that the race is real and won't be wished away, the strategic question becomes: how do you compete hard while keeping the tail risks bounded? I think of it as a stack — from the things any serious player must do to stay in the race, up to the cooperative moves that keep the race from becoming a catastrophe.

Layer 1 · Compete

Secure the inputs

  • Compute: guarantee access to chips and clusters; diversify fabrication away from a single strait.
  • Energy: treat power as strategic infrastructure — permit, build, and decarbonize fast.
  • Talent: attract and retain researchers; immigration is industrial policy.
Layer 2 · Protect

Defend the chokepoints

  • Targeted controls: a genuinely "small yard, high fence" — not maximalist decoupling that backfires.
  • Supply resilience: friend-shore fabs, stockpile critical inputs, secure the toolchain.
  • Security: protect model weights and labs as national-security assets.
Layer 3 · Govern

Build domestic guardrails

  • Capacity: fund safety institutes and public evaluation of frontier models.
  • Standards: transparency, incident reporting, and testing before deployment.
  • Resilience: protect elections, infrastructure, and the information ecosystem from misuse.
Layer 4 · Stabilize

Cooperate where it counts

  • Red lines: keep humans in control of nuclear and other catastrophic decisions.
  • Verification: develop compute-accounting and inspection ideas — the "chips-for-peace" agenda.
  • Channels: crisis hotlines and shared scientific assessment, even with rivals.
🧭

The organizing idea: the bottom of the stack (compete, protect) is where states naturally spend their energy — it feels urgent and unilaterally rational. The top (govern, stabilize) is chronically underfunded because it looks like a brake. But history's arms races were survived precisely by investing in the top while competing at the bottom — deterrence plus arms control, not one or the other. The winning move is to do both at once, and to make verification cheap enough that even rivals prefer it to mutual suspicion.

12 · What Could Break

Risk map — the failure modes of a fast race

An accelerating, lightly-governed race generates a characteristic set of risks. I sort them by likelihood and severity — not to predict, but to prioritize attention and to separate the loud-but-survivable from the quiet-but-catastrophic.

FIG 6

Likelihood × severity of arms-race failure modes (directional)

Severity → Likelihood (next ~5 years) → SEVERE BUT LESS LIKELY · watch closely SEVERE & LIKELY · top priority MINOR & RARE · monitor MANAGEABLE BUT COMMON · absorb Safety race-to-bottom Taiwan / chip-war conflict Autonomous-weapon accident Capability proliferation AI-enabled bio/cyber misuse Surveillance export Market concentration Energy / climate strain Bloc fragmentation Positions are judgment calls, not measured probabilities — meant to triage attention, not to forecast.
The cluster that should worry strategists most is the top-right: a safety race-to-the-bottom (each side cutting corners to stay ahead) and capability proliferation (open weights + cheap training spreading dangerous capabilities). The top-left holds the low-probability/high-severity tails — a Taiwan conflict or an autonomous-weapons accident — that justify keeping cooperation alive even amid rivalry. The bottom-right risks are real but absorbable.
The defining risk

Competition erodes caution

The core danger of any arms race: the pressure to not be second pushes every player to cut safety margins. The faster the race, the thinner the testing — and the higher the chance of a shared accident.

The diffusion risk

The frontier won't stay fenced

Open weights and falling training costs mean near-frontier capability spreads to states and non-state actors regardless of controls. Containment buys time, not permanence.

The structural risk

A fractured world

The race is splitting the globe into tech blocs with incompatible standards, supply chains, and information spaces — raising costs for everyone and shrinking the common ground needed to manage shared risks.

13 · Where It Could Go

Scenarios — four futures for the race

Not predictions — a span of plausible endings, to pressure-test strategy against more than one future. Each turns on two variables: how concentrated capability becomes, and how much cooperation survives the competition.

Scenario A · Managed Competitionbest plausible

Rivalry stays below the threshold of conflict. The poles compete hard on capability but sustain red lines, crisis channels, and a thin but real verification regime. Chokepoints are used for stability, not just denial.

  • Off-ramps at rungs 3–4 are used deliberately
  • Some shared safety science survives the rivalry
  • Diffusion is broad but the worst misuse is contained
Scenario B · Bipolar Lock-Inmost likely path

A durable two-bloc world. Separate stacks, standards, and information spaces harden around the US and China. Most states pick a side or hedge. Stable-ish, but brittle, expensive, and short on shared guardrails.

  • Decoupling deepens; supply chains bifurcate
  • The Global South becomes contested terrain
  • Tail risks are managed poorly across the divide
Scenario C · Multipolar Diffusionopen-weight world

Capability spreads faster than control. Open weights, efficient training, and the Gulf's capital democratize the frontier. Power disperses to many players — more competition and access, but far harder to govern and easier to misuse.

  • Export controls lose bite
  • Many credible AI powers, no clear hegemon
  • Misuse risk rises with the number of actors
Scenario D · Race-to-the-Brinktail risk

The logic of the race wins. Perceived proximity to decisive advantage triggers corner-cutting, sabotage, or a scramble over chips/Taiwan. Safety is sacrificed for speed; an accident or miscalculation escalates.

  • Cooperation collapses to rung 4–5
  • Safety margins vanish under time pressure
  • A shared catastrophe forces belated, costly rules
🔭

My read: the default trajectory points at Scenario B (bipolar lock-in) with a strong undertow toward C (diffusion) as open weights and efficiency erode any one bloc's monopoly. The work of strategy — and of every guardrail in §11 — is to bend toward A (managed competition) and keep firm distance from D. None of these is fixed; the levers that move us between them are mostly political choices, not technical inevitabilities.

14 · What To Watch

Indicators — the dashboard I keep

A handful of signals tell you which scenario you're sliding toward faster than any headline. These are the gauges I check to read the state of the race.

TBL 6

Leading indicators of the race's direction

IndicatorWhat it signalsToward stability ↘ or escalation ↗
Frontier compute gapLargest training clusters, US vs. ChinaWidening gap → pressure to race / proliferate ↗
Export-control posture"Small yard" vs. maximalist decouplingTargeted & allied → stable ↘ · maximalist → backfire ↗
Taiwan tensionMilitary activity & rhetoric around the straitThe single highest-severity gauge — rising ↗
Open-weight frontierHow close open models track the closed frontierCloser → diffusion & harder control ↗
Safety-institute fundingPublic capacity to evaluate frontier systemsRising & networked → stable ↘
Military-AI normsProgress on human control / autonomous-weapons limitsAgreed thresholds → stable ↘ · vacuum → escalation ↗
Energy build-outPower secured for clusters; clean vs. fossilFast & clean → competitive + sustainable ↘
Crisis channelsActive US–China hotlines & shared assessmentOpen → off-ramps available ↘ · closed → blind ↗
The two gauges I'd watch above all: Taiwan tension (highest severity) and the open-weight frontier (highest leverage on whether the whole control regime holds). If those two move the wrong way together, every other indicator gets harder to manage.
15 · The Short Version

Ten things to hold onto

If you remember nothing else from this note, remember these.

1 · It's a race because the prize compounds

Leaders fear AI advantage snowballs — so investment behaves like a security competition, not a normal market.

2 · The terrain is the supply chain

Compute, chips, and energy — not models alone — decide who can compete. Control the narrow layers, control the race.

3 · Three chokepoints hold everyone hostage

EUV (ASML), leading-edge fabs (TSMC), and HBM (Korea). Each is leverage — and a single point of catastrophic failure.

4 · Export controls ratchet, then get adapted around

Every restriction breeds a workaround — cut-down chips, smuggling, domestic substitutes, or sheer efficiency.

5 · Efficiency can beat denial

DeepSeek showed strong models on less compute. If algorithms outrun hardware controls, the whole denial strategy weakens.

6 · Energy is the surprise constraint

The binding limit is increasingly power and land, not just GPUs — turning AI into an industrial and climate question.

7 · The military edge is the dangerous edge

Autonomy compresses decision time and removes humans from the loop — with no mature treaty regime to govern it.

8 · Governance is thin and mostly voluntary

There is no IAEA for AI. Summits produce declarations, not inspections — closing that gap is the central diplomatic task.

9 · Open weights diffuse power from below

Near-frontier capability spreads regardless of controls — great for access, corrosive to any containment regime.

10 · The ending is a choice, not a forecast

Compete at the bottom of the stack, cooperate at the top. Deterrence and arms control — that's how arms races are survived.

16 · References & Sources

Where the ideas in this note come from

Annotated bibliography behind the arms-race system map, the player scorecard and quadrant, the semiconductor chokepoint chain, the export-control timeline, the frontier-labs table, the energy and military-AI sections, the governance comparison, the escalation ladder, the strategy stack, the risk quadrant, the scenario set, and the indicator dashboard. Section tags (e.g. §04) show where each source is used. Diagrams and the strategy/scenario framing are my synthesis unless noted.

Scope. Synthesis of AI-policy research, semiconductor and export-control analysis, defense-technology studies, and AI-governance sources (as of May 2026). Hero-strip figures (e.g. "~$300B+" hyperscaler capex, "3 chokepoints") are directional and compress several reporting sources — treat them as orders of magnitude, not audited totals. US export-control rules, chip thresholds, and model versions changed repeatedly through 2024–25; verify the current state before citing specifics. Not legal, policy, investment, or national-security advice.

Citations are numbered continuously [1]–[n] within this section.

The race framing, national strategy & the big picture (§01–§02, §15)

  1. Stanford HAI, AI Index Report (2024–2025 editions). Compute, investment, model, and policy trends — backbone for the hero stats, §02 inputs, and §05 lab landscape. hai.stanford.edu/ai-index — hero, §02, §05.
  2. Aschenbrenner, L., Situational Awareness: The Decade Ahead. 2024. Influential (and contested) statement of the "race to AGI as national-security competition" thesis — §01 winner-take-most framing and §13 race-to-the-brink scenario. situational-awareness.ai — §01, §13.
  3. Kissinger, H., Schmidt, E., & Mundie, C., Genesis: Artificial Intelligence, Hope, and the Human Spirit (and The Age of AI, 2021). AI as a force reshaping power, statecraft, and strategic stability — §01 stakes cards and §02 forms-of-power column. — §01, §02.
  4. National Security Commission on Artificial Intelligence (NSCAI), Final Report. 2021. US framing of AI as strategic competition with China; compute, talent, and microelectronics recommendations — §01 race logic and §11 strategy stack. nscai.gov — §01, §11.
  5. State Council of the PRC, New Generation Artificial Intelligence Development Plan. 2017. China's national ambition to lead AI by 2030 — §08 timeline ignition point and §03 China posture. — §03, §08.

Players, compute geopolitics & the semiconductor supply chain (§03–§04, FIG 2–3, TBL 1–2)

  1. Miller, C., Chip War: The Fight for the World's Most Critical Technology. Scribner, 2022. Definitive account of semiconductor geopolitics and chokepoints — FIG 3 chain, TBL 1 dependencies, and §04 "small yard" doctrine. — §03, §04, FIG 3.
  2. Center for Security and Emerging Technology (CSET), Georgetown. Research on compute, chips, talent, and the US–China technology balance — TBL 1 scorecard, FIG 2 positioning, and §04 controls analysis. cset.georgetown.edu — §03, §04.
  3. US Bureau of Industry and Security (BIS), Department of Commerce — advanced-computing & semiconductor export rules. 2022–2025. The successive control packages described in TBL 2; verify current text. bis.gov — §04, TBL 2.
  4. Center for a New American Security (CNAS), AI & export-control program. Analysis of the "diffusion" framework, the H20 episode, and allied coordination — §04 escalation table and §11 protect layer. cnas.org — §04, §11.
  5. Epoch AI — compute, training cost, and frontier-model trends. Data behind cluster-scale and compute-gap claims — §04 efficiency surprise, §14 frontier-compute-gap indicator. epoch.ai — §04, §14.
  6. DeepSeek, technical reports for the V3 and R1 model families. Late 2024–early 2025. Source for the "efficiency shock" — §04 DeepSeek card, §05 labs table, §08 timeline. deepseek.com — §04, §05, §08.
  7. SIA / TSMC / ASML / SK Hynix corporate & industry reporting. Foundry, EUV-lithography, and HBM market structure — TBL 1 chokepoint holders and FIG 3 near-monopoly steps. semiconductors.org — §03, §04, FIG 3.

Frontier labs, capex & energy (§05–§06)

  1. OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, Alibaba — model cards, system cards & announcements. 2023–2025. Source for TBL 3 flagship lines, backing, and the open- vs. closed-weight cleavage. — §05.
  2. Stargate / hyperscaler capital-expenditure announcements and earnings disclosures. 2024–2025. Microsoft, Amazon, Google, Meta data-center capex and the OpenAI–Oracle–SoftBank infrastructure venture — hero "~$300B+", §06 demand shock, §08 timeline. — hero, §06, §08.
  3. International Energy Agency (IEA), Electricity / Energy and AI analyses. 2024–2025. Data-center electricity demand and grid strain — §06 demand shock and §14 energy indicator. iea.org — §06, §14.
  4. Reporting on hyperscaler nuclear & SMR agreements (reactor restarts, power-purchase deals, SMR investments). 2024–2025. §06 nuclear-pivot card. — §06.
  5. Coverage of Gulf AI infrastructure: UAE's G42 and Saudi Arabia's HUMAIN / PIF initiatives. 2024–2025. Energy-plus-capital entry into the race and US security conditions on chip access — §03 Gulf row, §06 geopolitical-opening card. — §03, §06.

Military AI, autonomy & strategic stability (§07, §10, TBL 4, FIG 5)

  1. Scharre, P., Army of None: Autonomous Weapons and the Future of War and Four Battlegrounds: Power in the Age of Artificial Intelligence. W. W. Norton, 2018 / 2023. Foundational on autonomy, the decision loop, and AI-military competition — TBL 4 and §07 cards. — §07, §10.
  2. US Department of Defense — CDAO, Responsible AI strategy, and the Replicator initiative. 2023–2025. Doctrine for attritable autonomy and data/AI integration — TBL 4 maturity column and §07 mass card. ai.mil — §07.
  3. UN Convention on Certain Conventional Weapons (CCW), Group of Governmental Experts on Lethal Autonomous Weapons Systems. Ongoing. The "meaningful human control" debate — §07 arms-control-vacuum card and §10 off-ramps. unoda.org — §07, §10.
  4. Political Declaration on Responsible Military Use of AI and Autonomy (US-led, 2023). Voluntary norms including human control of nuclear use — §07 nuclear red line and §10 rung-5 off-ramp. state.gov — §07, §10.
  5. Horowitz, M. & others — analyses of AI & nuclear stability and the Ukraine drone war. 2022–2025. Battlefield evidence and escalation risk — §07 ISR/cyber rows and FIG 5 escalation ladder. — §07, §10.

Governance models, summits & the multilateral layer (§09, TBL 5)

  1. European Parliament & Council, Regulation (EU) 2024/1689 (AI Act). 2024. Risk-tiered binding rulebook, phased 2025–2027 — §09 EU model and TBL 5. eur-lex.europa.eu — §09.
  2. Bradford, A., Digital Empires: The Global Battle to Regulate Technology. Oxford University Press, 2023. The three-model framing (EU rights-based, US market-led, China state-driven) — §09 governance cards. — §09.
  3. UK Government, Bletchley Declaration (2023) and Seoul Declaration / frontier-safety commitments (2024). The summit arc — TBL 5 and §08 timeline. gov.uk — §08, §09.
  4. AI Action Summit, Paris (2025) — official outputs. The pivot toward action and the fraying of consensus — TBL 5 and §12 bloc-fragmentation. elysee.fr — §09, §12.
  5. White House, Executive Order 14110 (2023) and subsequent 2025 AI policy / AI Action Plan. The US safety-to-deployment pivot — §09 US model and §08 timeline. whitehouse.gov (verify latest) — §09.
  6. Cyberspace Administration of China, generative-AI and algorithm-registration measures. 2022–2024. China's content-control governance — §09 China model. cac.gov.cn — §09.
  7. UN High-Level Advisory Body on AI, Governing AI for Humanity (2024) and follow-on scientific-panel process. The thin global layer — TBL 5 and §11 stabilize layer. un.org/ai-advisory-body — §09, §11.
  8. International AI Safety Report (Y. Bengio, chair), 2025. Shared cross-government scientific assessment — §11 verification agenda and §12 tail-risk framing. internationalaisafetyreport.org — §11, §12.

Strategy, verification, risk & scenarios (§11–§14)

  1. RAND Corporation — AI, compute governance, and US–China technology-competition research. Compute-accounting and verification ideas behind the §11 stabilize layer and §14 indicators. rand.org — §11, §14.
  2. "Computing Power and the Governance of AI" (Sastry, Heim, et al.) and related compute-governance / "chips-for-peace" proposals. 2024. The case for using compute as a governance lever — §11 verification and §04 controls. governance.ai — §04, §11.
  3. Allison, G., Destined for War: Can America and China Escape Thucydides's Trap? Houghton Mifflin Harcourt, 2017. Great-power-rivalry framing behind the §10 escalation ladder and §13 scenarios. — §10, §13.
  4. Schelling, T., Arms and Influence and The Strategy of Conflict. Yale / Harvard UP. Classic arms-race, deterrence, and arms-control logic underpinning §10 off-ramps and the §11 "deterrence + arms control" thesis. — §10, §11.
  5. Brookings, Carnegie Endowment, and CSIS commentary on tech-bloc fragmentation and AI competition. 2023–2025. §12 structural-risk card and §13 bipolar/multipolar scenarios. csis.org — §12, §13.

Author synthesis & companion notes

  1. Truong, L., The Global AI Arms Race — personal working notes. May 2026. Original diagrams (FIG 1–6, TBL 1–6), strategy stack, escalation ladder, scenario set, and indicator dashboard. LinhTruong.com — all sections.
  2. Truong, L., companion notes: Global Political Transformation in the AI Era, Global Economic Transformation in the AI Era, and Geopolitics Transformation in the AI Era. The political-power, economic-diffusion, and geopolitical-bloc frames this arms-race map sits beneath. Same author collection. — §02, §09, §12.
📎

Before you quote externally: FIG 2 (player quadrant), FIG 5 (escalation ladder), and FIG 6 (risk quadrant) are judgment-based positioning diagrams, not measured data — they triage attention, not forecast outcomes. The FIG 3 chokepoint chain is a simplified schematic of a far more complex supply chain. Capex and compute figures are orders of magnitude. US export-control thresholds, the status of specific chips (e.g. the H20), and model versions shifted repeatedly across 2024–2025 — verify the current state of any specific rule, deal, or model before citing it.

↑ Top