How We Got Here: Three Years of Escalation in Four Paragraphs
October 2022: the US barred exports of leading AI accelerators and chipmaking tools to China, betting that compute is the chokepoint for AI capability. The controls tightened annually — 2023 closed the A800/H800 workaround loopholes, 2024 added HBM and more tools, and 2025 delivered whiplash: the H20 (Nvidia's compliant China chip) was banned in April, then un-banned in a summer deal that took an unprecedented 15% cut of China revenue for the US government, even as Beijing discouraged domestic buyers from wanting it.
January 2026 brought the current framework: a BIS rule replacing blanket denial with engineered scarcity — chips below defined performance ceilings (total processing power under ~21,000, memory bandwidth under 6,500 GB/s, roughly H200/MI325X class) became eligible for case-by-case licenses carrying a 25% tariff, a volume cap at 50% of equivalent US domestic shipments, mandatory third-party testing, and know-your-customer obligations. Frontier-class accelerators remain effectively embargoed. The design intent is explicit: keep China one to two generations behind while monetizing, rather than forfeiting, the trailing edge.
China's counter-arsenal matured in parallel. The October 2025 rare-earth export controls — targeting the processed materials and magnets on which Western EVs, drones, robots, and defense systems depend — demonstrated escalation capability so effectively that a November 2025 trade truce suspended them for exactly one year, until November 10, 2026. Beijing has kept the pressure calibrated since: in June 2026 it added ten more US firms, including a rare-earth miner, to its own export-control lists. Both sides now hold loaded chokepoints; neither has fired the full magazine.
The quiet macro fact underneath: China's domestic accelerator ecosystem (Huawei's Ascend line, SMIC's constrained-but-improving 7nm-class output) went from negligible to nationally load-bearing in three years. US controls slowed the frontier gap; they also created the world's most motivated import-substitution program. Nvidia's share of China's AI accelerator market — above 90% pre-controls — has collapsed toward irrelevance, a $50B+ annual market the controls effectively ceded to build the moat.
The Rare-Earth Countdown
The single date every supply-chain office has circled: November 10, 2026, when China's suspended rare-earth controls come up for renewal or reactivation. The leverage math explains why:
- China processes roughly 90% of the world's rare earths and dominates high-performance NdFeB magnet output even more thoroughly. These magnets are irreplaceable on relevant timescales in EV motors, drone and humanoid actuators, wind turbines, hard drives, and precision-guided munitions.
- The 2025 episode proved the weapon works. The October controls produced immediate Western production anxieties (the Nexperia chip-supply crisis around the same period previewed how fast automotive lines feel upstream disruption) and brought Washington to a truce within weeks.
- Western substitution is real but slow: MP Materials' expanding mine-to-magnet chain (with Pentagon equity involvement and price floors), Australian and European processing projects — collectively years from displacing double-digit percentages of Chinese capacity. Stockpiles and dual-sourcing, not independence, are the realistic 2026–2028 posture.
The strategic symmetry is precise: America's chokepoint is at the top of the value chain (chip design, EUV-dependent fabrication), China's at the bottom (processed materials). Chips need magnets in the machines that make and deploy them; magnets need markets. It is mutually assured supply-chain disruption — stable until either side believes it can absorb the hit, which is what both sides' crash localization programs are racing to change.
The Buildout Hedge: Fabs, Subsidies, and the Memory Squeeze
- TSMC's $165 billion Arizona program — multiple leading-edge fabs plus packaging — is the centerpiece of US reshoring, with Taiwan still holding ~90% of sub-5nm capacity: the "silicon shield" remains the single greatest concentration risk in the world economy, which every other line item here exists to hedge.
- Industrial policy got equity teeth. The CHIPS Act era evolved from grants to stakes — the US government's ~10% Intel shareholding (2025) crossed a line unthinkable a decade ago, and the template (state equity in chokepoint firms, as with MP Materials) is now bipartisan practice. Semiconductor policy has fully merged with statecraft.
- The memory market became the surprise battlefield. AI demand for HBM and high-capacity DRAM produced a genuine 2025–2026 shortage — hyperscalers pre-buying supply, consumer RAM/SSD prices spiking, and Meta explicitly citing memory inflation in raising capex guidance. Korea's SK Hynix and Samsung plus Micron control HBM; it is the West's quiet counter-chokepoint, and China's CXMT is years behind on it.
- Everyone else picks lanes: Japan (Rapidus, tool makers) and the Netherlands (ASML's EUV monopoly — still the deepest single-point chokepoint in the entire war) align with US controls while selling everything the rules permit; the Gulf buys access with capital; Europe subsidizes mature nodes and lithography leadership it already owns.
What It Means for the Price of Intelligence
Chip geopolitics reaches your invoice through three channels:
- Compute cost floors. Tariffed, capped, tested chips are more expensive chips; constrained supply of frontier accelerators props up GPU rental rates and slows the token-price collapse below what pure technology trends would deliver. The 25% China tariff and 50% volume cap are, functionally, a tax on global compute clearing prices at the margin.
- A bifurcating stack. The world is heading toward two partially incompatible AI ecosystems — CUDA/US-standards vs. Ascend/Chinese-standards — with everyone in between (the Gulf, ASEAN, India) courted by both and building on open weights to hedge. For builders, multi-vendor portability (and model-layer abstraction) quietly became a geopolitical risk control, not just an engineering nicety.
- Volatility as a planning assumption. The H20's ban-unban-shun arc within nine months, the rare-earth November deadline, election-cycle policy swings — the operative lesson is that any single sourcing dependency (chips, magnets, memory, models) is now a balance-sheet risk. The winning posture for 2026–2028 mirrors the state level at company scale: dual-source, stockpile the irreplaceable, and price the geopolitics into the unit economics rather than treating each episode as a surprise.
The chip war's deepest irony: both superpowers are spending historic sums to make themselves independent of each other in exactly the technologies — AI compute and the materials that embody it — whose economics reward global scale. The bill for that redundancy lands in every token, robot joint, and DRAM stick priced above the pre-war trend line. Efficiency was the peace dividend; resilience is what we're buying now, and it compounds into everything downstream.
About the figures in this guide
Model pricing, context limits and benchmark results move faster than almost anything else in software — sometimes monthly, and rarely with notice.
The figures here reflect what providers published at the time of writing, and they are included to show relative differences and the shape of the trade-offs rather than to serve as a quote. Treat any specific price or score as a planning baseline that needs checking.
Before you size a budget or commit to a model, confirm the current numbers against the provider's own pricing and model documentation. For benchmark results in particular, also check what the benchmark measures — a headline score says little about performance on your workload, which is what your own evaluation set is for.
Spotted something that has since changed, or was wrong to begin with? Let us know and it will be corrected.