International Competition

International Competition: The Compute Race

The contest stopped being about chips alone this week and became an argument about a single model. On July 22 White House science adviser Michael Kratsios said the U.S. had information that Moonshot AI "distilled Anthropic's Fable" to build Kimi K3, using a purpose-built platform to evade detection, and had obtained export-restricted Nvidia GB300 servers — publishing no evidence; Treasury Secretary Scott Bessent, who had said the day before that U.S. "watermarks" were turning up on Chinese models, put sanctions and Entity List designations "on the table." On July 23 the two governments' own evaluators answered a different question and got a different tone: the UK AI Security Institute and Commerce's Center for AI Standards and Innovation jointly reported that Kimi K3 "performs significantly below the most recent frontier cyber-capable models" — 0 of 41 arbitrary-code-execution solves on ExploitBench against an average of 20 for the leaders, step 17 of a 32-step attack range against 28.5 — while noting the U.S. models were tested with safeguards disabled. Beijing answered on July 27, its Ministry of Commerce calling the accusations evidence-free, "a classic case of hegemonic behavior in the AI sphere," and promising "all measures necessary." Nothing has yet been enacted: Commerce had not begun drafting Entity List plans as of July 23, and more than 20 companies including Nvidia, Microsoft, Meta and Palantir — alongside nearly 200 startup founders — pressed the administration against restricting open-weight models. Two older threads also enter the record this cycle. In June the U.S. used export-control authority against a specific frontier model for the first time, ordering access to Claude Fable 5 and Mythos 5 cut off for every foreign national including those inside the United States; and on July 23 Senator Elizabeth Warren invoked the Export Control Reform Act to compel BIS, by August 7, to document a foundry loophole she argues the 2025 rescission of the AI Diffusion Rule left open — the enforcement question underneath a year of loosening.

The Whole Story

The compute race turned into an open confrontation this week. On July 22 the White House science adviser accused Moonshot AI of distilling Anthropic's Fable model to build Kimi K3 and of obtaining export-restricted Nvidia servers, publishing no evidence; Treasury put sanctions and Entity List designations "on the table." A day later the UK and U.S. safety institutes jointly measured K3 and found it "significantly below the most recent frontier cyber-capable models" — the first bilateral state evaluation of a Chinese model, and an answer at odds with the week's rhetoric. Beijing's Commerce Ministry rejected the accusations on July 27 as "a classic case of hegemonic behavior in the AI sphere" and promised countermeasures. Meanwhile the control machinery kept moving in the background: Senator Warren invoked the Export Control Reform Act to compel BIS documents on a foundry loophole she says the 2025 rescission left open, and the record now shows Washington using export-control authority against a specific frontier model — reaching foreign nationals inside the United States.

AI capability has become an instrument of national power, and governments increasingly treat frontier models, the chips that train them, and the talent that builds them as strategic assets — with export controls, restrictions, and state investment to match. This page tracks the competition on observable actions: what governments actually do, what capabilities each bloc demonstrably fields, and how the claims of an 'arms race' hold up against the record.

Continue Reading →

The U.S. accusations "lack any actual evidence, have no legal backing, and effectively amount to a double standard. It is a classic case of hegemonic behavior in the AI sphere."^±

— Chinese Ministry of Commerce · Jul 27, 2026

China's Ministry of Commerce issued the first direct government response to the distillation accusations on July 27, rejecting them outright while arguing that distillation is widespread on both sides — "many U.S. AI companies have distilled from China's models for their own R&D and training" — and that Washington had ignored "the extremely short interval between the release of Chinese enterprises' AI models and U.S. frontier models." It said China "will take all measures necessary to staunchly defend its legitimate and legal rights and interests," and urged the U.S. to honour the consensus reached between the two heads of state. Quoted from CSET's English translation of the Chinese original.

Context: The checkable sub-claim holds: MOFCOM's assertion that "nearly 200 U.S. startups urged the U.S. government not to cut off their access to Chinese open-source models" matches Politico's reporting that nearly 200 startup founders urged the administration that week not to cut off access to Chinese open-weight models, and more than 20 companies including Nvidia, Microsoft, Meta and Palantir published a letter on July 24 opposing "premature restrictions" on open-weight models. The central charge that the U.S. accusations lack evidence is accurate as to the public record — neither Kratsios nor Bessent published any — but it is a claim about what has been disclosed, not proof that no evidence exists; MOFCOM likewise published none for its own counter-assertion that U.S. companies distilled Chinese models.

Two governments measure Kimi K3 — and get an answer the week’s rhetoric did not

Jul 23, 2026 · aisi.gov.uk · nist.gov · x.com

The UK AI Security Institute and the U.S. Center for AI Standards and Innovation, a Commerce Department body, published a joint preliminary assessment of Moonshot AI's Kimi K3 on July 23, the first bilateral state evaluation of a Chinese frontier model. On ExploitBench — a 41-task Carnegie Mellon benchmark over post-2023 vulnerabilities in Chrome's V8 engine — K3 scored 32% against 24% for GLM-5.2, then the most cyber-capable open-weight model; but it achieved arbitrary code execution, the benchmark's highest-severity outcome, on 0 of 41 tasks, where the most cyber-capable models averaged 20 of 41. On "The Last Ones," a 32-step simulated corporate-network attack, K3 reached step 17 on average against 28.5 for the most cyber-capable U.S. models, and completed the range in 1 of 10 attempts against 6/10 and 7/10 for the leaders. The institutes state their own limits: preliminary results, K3's aggregate score estimated from a single benchmark and so carrying a wider confidence interval, a selective task set because of K3's hosting, and — importantly — U.S. closed-weight models evaluated with system-level safeguards disabled to measure maximal capability, which shipping versions do not permit. They also found K3's own safeguards did not stop it attempting exploit development. Set beside the week's other U.S. government artifact, the picture is discordant: on July 22 the OSTP director said Moonshot had run "large scale distillation against U.S. models" to build K3; on July 23 two governments jointly reported that "Kimi K3 performs significantly below the most recent frontier cyber-capable models." The two address different questions — where capability came from, and how much of it there is — and neither resolves the other.

A statutory demand tests whether the 2025 rescission left a foundry loophole open

Jul 23, 2026 · banking.senate.gov · banking.senate.gov

Senator Elizabeth Warren, ranking member of the Senate Banking Committee, invoked the Export Control Reform Act of 2018 — which obliges Commerce to produce "any information obtained at any time" under the export regulations on a ranking member's request — to compel the Bureau of Industry and Security to hand over documents by August 7, 2026. Her argument runs on the machinery this page has tracked: the January 2025 Foundry Due Diligence rule makes foundries such as TSMC presume that advanced chips need a licence unless due-diligence obligations are met, but it works only on top of the AI Diffusion Rule's worldwide licence requirement. Rescinding the latter in May 2025, she writes, "created a foundry loophole that brings the United States back to the very circumstances that allowed Huawei to divert millions of AI chips in the first place." The release also states that the promised AI Diffusion replacement, "more than a year later, still has not been issued" — the narrower January 2026 case-by-case rule for H200-class chips is not a replacement framework. This is a minority-party document: primary as to the demand and its deadline, contested as to BIS's conduct.

“We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.”^?

— Michael Kratsios · Jul 22, 2026

The White House science and technology adviser, OSTP Director Michael Kratsios, posted the allegation on X, adding that Moonshot "developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection," and that the company "has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models" — chips whose export to Chinese entities is restricted. It is the most specific public accusation the U.S. government has made about how a Chinese frontier model was built.

Context: No evidence was published with the post, and Politico reported that Kratsios "did not provide evidence to back up the claims" and that the post was not coordinated with the Commerce Department, which holds Entity List authority. Moonshot denies the allegation; China's Ministry of Commerce called it evidence-free on July 27. The technical premise drew scepticism from within the administration's own orbit — one person close to the White House told Politico the model was benchmarking ahead of Fable in places and "You can't just distill your way ahead of somebody." Corroborating context, not proof of this claim: Anthropic reported in February 2026 that DeepSeek, Moonshot and MiniMax generated over 16 million exchanges with Claude through roughly 24,000 fraudulent accounts — a report that predates Kimi K3 and does not name it.

“We are finding watermarks of our US large language models on many Chinese models, and that’s unacceptable.”^?

— Scott Bessent · Jul 21, 2026

Treasury Secretary Scott Bessent, in a Fox Business interview on July 21, said the administration was examining whether Chinese open-source models were built on stolen U.S. intellectual property — "a very technical AI word for it called distillation, but you and I would call it theft" — and named no company. The next evening he set out the policy consequence on X: "We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table." The watermark assertion is the only evidentiary claim either official offered this week.

Context: No supporting evidence was published: neither Treasury nor Commerce released watermark findings, sample models, or a methodology, and Bessent identified no company. The claim is not settled by the one government measurement published this week — the UK AISI/CAISI assessment of Kimi K3 evaluated cyber capability, not provenance. The threat itself remained unexecuted as of this cycle: Politico reported on July 23, citing a person familiar with the department's thinking, that Commerce had not begun drafting Entity List plans for Moonshot or other Chinese AI companies, and that any decision would depend on conclusive Commerce findings.

An executive order lets the Pentagon point AI at its own supply chain — and the headline outruns the text

Jul 20, 2026 · whitehouse.gov · defensescoop.com

Executive Order 14415, "Securing America's Defense Supply Chains and Ensuring Domestic Acquisition of Critical Materials," orders the Department of War to require prime contractors and subcontractors to trace every component back to raw-material origin, and to vet suppliers for foreign ownership, control or influence. Its artificial-intelligence content is one permissive clause in Section 3(d): the department shall map national-security vulnerabilities "using any tools and technologies to include artificial intelligence to assist in doing so." DefenseScoop headlined the order "Trump authorizes use of AI for defense supply chain mapping"; its own body text is more careful, noting the edict "noted that the Pentagon may use AI." Readers can check both: the order neither mandates AI nor funds it, and the deadlines it sets — 180 days for implementation guidance, 90 days for a testing-and-qualification strategy — attach to the sourcing requirements, not to the AI clause.

Export control reaches inside the United States: a frontier model cut off from foreign nationals

Jun 12, 2026 · anthropic.com · gtlaw.com · nature.com

Citing national security authorities, the U.S. government issued an export-control directive requiring Anthropic to suspend all access to its Claude Fable 5 and Mythos 5 models by any foreign national — not only abroad, but inside the United States, including the company's own foreign-national employees. Unable to verify nationality in-product, Anthropic disabled both models for every customer. What is new here is the instrument, not the outage: the export-controls practice at Greenberg Traurig called it "the first known U.S. use of export control authorities to regulate a particular AI frontier model on a national security basis," and noted that because it bit on deemed exports it reached API calls, embedded agents and internal tooling wherever a foreign national could reach them. Anthropic's contemporaneous filing records that the directive "did not provide specific details of its national security concern"; access was restored at the end of June. The chip-control regime this page has tracked since 2022 now has a model-level counterpart.

No AI-chip export policy to China can be “permissive, implementable, enforceable, and protective of US national security” at once.^?

— Chris McGuire · Jan 14, 2026

Writing for the Council on Foreign Relations, former NSC technology official Chris McGuire argued the January 2026 case-by-case rule proved that a permissive chip-export policy cannot simultaneously be enforceable and protect national security — and warned the same logic could be stretched to license far more capable Blackwell/GB300 chips. A direct rebuttal to the dependency theory, casting the loosening as strategically incoherent.

Context: Forward-looking enforceability claim tied to the January 2026 rule; still open, but this cycle added the first on-the-record challenge to the “implementable, enforceable” half. On 2026-07-23 Senator Elizabeth Warren, ranking member of Senate Banking, invoked the Export Control Reform Act to compel BIS to document what she calls a foundry loophole: the January 2025 Foundry Due Diligence rule presumes advanced chips need licences, but rests on the AI Diffusion Rule’s worldwide licence requirement, and she argues the May 2025 rescission removed that predicate — citing reporting that it “may be allowing an unknown number of high-end semiconductors to reach Beijing.” That is an allegation from a minority-party document, not a finding; BIS’s response is due 2026-08-07 and is the next checkable milestone. Not resolvable before the 2027-12-31 horizon.

Sell China the fourth-best chip so “their developers get addicted to the American technology stack.”^?

— Howard Lutnick · Jul 15, 2025

Explaining why the administration reversed course and let Nvidia resume H20 sales, Commerce Secretary Howard Lutnick said the U.S. does not sell China its best, second-, or third-best chips — and that the point of selling the ‘fourth one down’ is to keep Chinese developers reliant on U.S. technology: ‘You want to sell the Chinese enough that their developers get addicted to the American technology stack.’ A rare on-record statement of the dependency theory behind the policy — the opposite of the denial logic the controls were built on.

Context: Forward-looking strategy claim; still open, but this cycle produced the first substantial evidence running against it. The dependency Lutnick described was on the American stack; by July 2026 the reported dependency ran partly the other way — nearly 200 U.S. startup founders urged the administration not to cut off access to Chinese open-weight models, arguing smaller American firms rely on them because they cannot afford frontier compute (Politico, 2026-07-23), and more than 20 companies including Nvidia, Microsoft, Meta and Palantir signed a letter against “premature restrictions” on open-weight models (CNBC, 2026-07-25). Earlier evidence still cuts the other way too: BIS warned industry against Huawei Ascend chips and Chinese buyers were reported pursuing domestic substitutes. Not resolvable before the 2027-12-31 horizon; revisit against adoption data on both sides.

Stanford’s 2026 AI Index: the U.S.–China frontier gap narrows to 2.7%

Apr 1, 2026 · hai.stanford.edu

Stanford HAI’s 2026 AI Index reported that the measured performance gap between the top U.S. and Chinese models had effectively closed — the two traded the lead repeatedly through 2025, and by March 2026 the leading U.S. model led by just 2.7% on the Index’s benchmark composite. The U.S. still produced more top-tier models, hosted 5,427 data centers (more than ten times any other country), and drew $285.9 billion in private AI investment (over 23 times China’s), while China led in publication volume, patents, and industrial-robot installations — a split scorecard, each half tied to a named metric rather than a single ‘who’s winning’ rank.

The controls loosen: BIS moves H200-class chips from denial to case-by-case review

Jan 13, 2026 · public-inspection.federalregister.gov

On January 13, 2026, BIS published a final rule shifting the license-review policy for advanced chips bound for China and Macau from a presumption of denial to case-by-case review for parts below a set ceiling (Total Processing Performance under 21,000 and DRAM bandwidth under 6,500 GB/s) — a band that includes Nvidia’s H200 and AMD’s MI325X. Approval carries conditions: China-bound shipments capped at 50% of a firm’s U.S. volume, end-use and physical-security certifications, no model-weight transfer, and third-party lab review of every shipment. It codified the Trump-era turn from tiered denial toward conditioned access.

The Gulf corridor opens: U.S. approves advanced-chip exports to the UAE’s G42

Nov 20, 2025 · g42.ai

On November 20, 2025 the White House authorized exports of advanced AI chips to the UAE’s G42, governed by a G42-built ‘Regulated Technology Environment’ compliance framework vetted by Commerce and BIS. The approval unlocked Stargate UAE — a 1-gigawatt cluster G42 is building for OpenAI with Oracle, Cisco, Nvidia, and SoftBank — within a larger planned 5-gigawatt U.S.–UAE campus, extending U.S.-controlled compute into the Gulf under a matched-buildout model.

An unprecedented cut: Nvidia and AMD to hand the U.S. 15% of their China AI-chip revenue

Aug 11, 2025 · apnews.com

On August 11, 2025, Nvidia and AMD agreed to remit 15% of their revenue from AI-chip sales to China — Nvidia’s H20 and AMD’s MI308 — to the U.S. government in exchange for export licenses. President Trump confirmed the terms, saying he had wanted 20% before Huang negotiated him to 15%. Framing a national-security export control as a revenue share for the Treasury was without clear precedent, and drew objections from China-hawks in both parties who called it a tax on selling China the very chips the controls were meant to deny.

Reversal: the U.S. clears Nvidia to resume H20 sales to China

Jul 15, 2025 · cnbc.com

Nvidia said on July 14–15, 2025 that Washington would let it resume H20 shipments to China after CEO Jensen Huang met President Trump, with export licenses to follow. Nvidia had said it could have sold roughly $8 billion of H20s in the quarter before the April halt. The whiplash — restrict in April, permit in July — set up the revenue-sharing arrangement that followed. Commerce Secretary Lutnick’s stated rationale is recorded, and tagged as a prediction, as a separate statement.

Frontier labs enlist: the Pentagon awards up to $200M each to Anthropic, Google, OpenAI, and xAI

Jul 14, 2025 · defensescoop.com

On July 14, 2025 the DoD’s Chief Digital and Artificial Intelligence Office awarded contracts worth up to $200 million each to Anthropic, Google, and xAI — joining OpenAI, which had received the first such award on June 17 — to develop ‘frontier AI’ and agentic workflows for national-security use. The awards followed the labs’ June rollouts of government-tailored models (Anthropic’s ‘Claude Gov,’ ‘OpenAI for Government’), formalizing the frontier labs as national-security contractors.

Trump’s Commerce Department rescinds the AI Diffusion Rule

May 13, 2025 · bis.gov

On May 13, 2025 — two days before it would have taken effect — the Trump Commerce Department announced it was rescinding Biden’s AI Diffusion Framework, with Under Secretary Jeffery Kessler ordering enforcement staff to stand down while a replacement was drafted. In its place BIS issued guidance warning industry against using Huawei Ascend chips, cautioning that letting U.S. chips train Chinese models ‘may’ carry consequences, and advising firms to guard supply chains against diversion — a shift from a rules-based tiering system toward case-by-case discretion.

Even the compliance chip gets caught: U.S. requires a license for Nvidia’s H20

Apr 15, 2025 · techcrunch.com

Nvidia disclosed on April 15, 2025 that the U.S. government would require an indefinite license to export its H20 — the chip Nvidia had designed specifically to stay under the 2023 thresholds — citing the risk it could be used in a Chinese supercomputer. Nvidia booked a $5.5 billion charge. The move showed how far the goalposts had shifted: a chip built expressly to be exportable was no longer exportable.

DeepSeek’s models are “broadly comparable to the best U.S.-developed AI models introduced in mid-2024.”^✓

— Center for Strategic and International Studies · Mar 7, 2025

CSIS assessed that DeepSeek’s V3/R1 were broadly comparable to leading mid-2024 U.S. models, naming OpenAI’s o1 and Anthropic’s Claude 3.5 Sonnet as the reference points — a bounded, model-to-model comparison rather than a single-axis ranking. It implied a capability lag of months, not years, and became a touchstone for arguing the frontier gap had narrowed.

Context: Corroborated by later independent measurement: Stanford HAI’s 2026 AI Index found the top U.S. and Chinese models traded the lead through 2025 and sat within 2.7% by March 2026. The mid-2024-comparability assessment held up as the gap continued to close.

Europe’s answer: InvestAI mobilizes €200 billion and four ‘gigafactories’

Feb 11, 2025 · luxembourg.representation.ec.europa.eu

At the Paris AI Action Summit on February 11, 2025, Commission President Ursula von der Leyen launched InvestAI, an initiative to mobilize €200 billion for AI investment across the EU, including a €20 billion fund for four AI ‘gigafactories’ — each envisioned with around 100,000 latest-generation chips, roughly four times the scale of the AI factories already being stood up on €10 billion of existing public support. Europe’s bid to avoid dependence on U.S. and Chinese compute.

Stargate: a pledged $500 billion for U.S. AI infrastructure

Jan 21, 2025 · openai.com

A day after the inauguration, OpenAI announced the Stargate Project — a new company intending to invest up to $500 billion over four years in U.S. AI data centers, with $100 billion pledged ‘immediately.’ SoftBank, OpenAI, Oracle, and MGX are the initial equity backers, with SoftBank chair Masayoshi Son as chairman and buildout beginning in Texas. It is the private-capital counterpart to the public compute race — the amounts are commitments, not deployed spending. (Tracked in depth on The Money and Infrastructure.)

DeepSeek-R1: a Chinese model at the frontier, and a $600 billion jolt to the arms-race story

Jan 20, 2025 · csis.org

On January 20, 2025 — inauguration day — the Chinese lab DeepSeek released its R1 reasoning model, days after its V3 base model. Built despite export controls, R1 performed comparably to leading U.S. reasoning models and, within a week, DeepSeek’s app had overtaken ChatGPT atop the U.S. App Store while Nvidia shed more than $600 billion of market value in a single session. Widely dubbed a ‘Sputnik moment,’ R1 became the central exhibit in the debate over whether controls were slowing China or spurring it to do more with less. (DeepSeek’s own claim that its final training run cost only ~$5.6M is recorded, and fact-checked, as a separate statement.)

The Biden endgame: an ‘AI Diffusion’ framework and a foundry crackdown

Jan 13, 2025 · federalregister.gov · bis.gov

In its final days the Biden administration published two rules. The ‘Framework for Artificial Intelligence Diffusion’ (effective January 13, 2025) established a global three-tier system of country caps and license exceptions governing exports of advanced chips and, for the first time, closed-weight AI model weights. A companion rule (January 15) tightened foundry and packaging controls — requiring an approved-designer attestation or verified transistor count to ship advanced chips — and added 16 PRC and Singapore entities, including AI firm Sophgo, to the Entity List. Most compliance provisions were set to bite May 15, 2025.

DeepSeek says its V3 model’s final training run used 2,788,000 H800 GPU-hours — about $5.6 million.^±

— DeepSeek · Dec 26, 2024

DeepSeek’s own V3 technical report stated the model’s final pretraining run consumed 2.788 million Nvidia H800 GPU-hours, which at an assumed $2/GPU-hour it valued at roughly $5.576 million — a figure widely repeated as evidence that frontier models could be trained cheaply despite export controls. The number describes only the last successful run, not the total cost of building the model.

Context: The $5.6M figure is DeepSeek’s own and, per CSIS, covers only the final successful pretraining run — excluding prior experiments, post-training, staff, and the underlying hardware. SemiAnalysis estimated DeepSeek/High-Flyer operates ~50,000 Hopper-class GPUs and spent ~$1.6 billion on GPU-server capex. So the marginal-run figure may be accurate while the ‘trained for $5.6M’ framing understates true cost by orders of magnitude.

The October 2023 update: closing the workaround chips and going worldwide by headquarters

Oct 17, 2023 · cset.georgetown.edu

On October 17, 2023, BIS overhauled the 2022 rule. It redefined controlled chips by Total Processing Performance and Performance Density (dropping the interconnect-speed parameter Nvidia had engineered its China-specific A800/H800 chips around), created a ‘Notified Advanced Computing’ license exception for less-capable datacenter chips, extended controls to 43 more countries, and — for the first time — imposed a worldwide license requirement keyed to an end user’s headquarters location. A new ‘red flag’ obliged foundries such as TSMC to scrutinize orders combining 50-billion-plus-transistor designs with high-bandwidth memory.

The United States has seen “no evidence” that China can produce advanced 7nm chips at scale.^±

— Gina Raimondo · Sep 19, 2023

Commerce Secretary Gina Raimondo said in September 2023 that the U.S. had no evidence China could manufacture advanced 7nm chips in large volumes — a remark that landed just after TechInsights confirmed SMIC 7nm silicon in the Huawei Mate 60 Pro. The narrow claim (‘at scale’) proved more durable than the broad reading: SMIC did field 7nm parts, but independent analysts reported the process ran at limited yield without EUV, so high-volume leading-edge production remained constrained.

Context: SMIC demonstrably produced 7nm chips (TechInsights teardown, Sep 2023), so the existence of the capability was already contradicted. But Raimondo’s specific claim was about volume — producing at scale — and reporting through 2024–2025 indicated SMIC’s 7nm yields remained low without EUV, limiting high-volume output. True on capability existing, defensible on scale.

A 7nm chip in the Huawei Mate 60 Pro: proof China cleared a barrier the controls aimed to hold

Sep 4, 2023 · techinsights.com · csis.org

A TechInsights teardown of Huawei’s Mate 60 Pro found its Kirin 9000S processor fabricated by SMIC on a 7nm (N+2) process — physical evidence that China had reached 7nm without extreme-ultraviolet (EUV) lithography, using multi-patterned deep-ultraviolet tools. TechInsights vice chair Dan Hutcheson called it a demonstration of ‘the technical progress China’s semiconductor industry has been able to make without EUV,’ and predicted ‘even greater restrictions than what exist today’ — a forecast the following years bore out. Independent analysts noted the achievement came at limited yield and did not resolve China’s dependence on imported lithography.

Pentagon launches ‘Replicator’ to counter China’s mass with attritable autonomy

Aug 28, 2023 · defensescoop.com

Deputy Defense Secretary Kathleen Hicks unveiled the Replicator initiative on August 28, 2023, at the NDIA Emerging Technologies for Defense Conference, setting a goal to field thousands of attritable, autonomous systems across multiple domains within roughly 18–24 months to offset the numerical advantage of China’s military buildup. It marked the U.S. defense establishment’s move to translate AI and autonomy into fielded capability at scale.

The October 7 controls: the U.S. moves to choke off China’s access to AI compute

Oct 7, 2022 · federalregister.gov · bis.gov

On October 7, 2022, BIS issued the interim final rule that became the foundation of the AI chip-control regime. It created new control classifications (ECCNs 3A090/4A090) for high-performance chips headed to China and two new Foreign Direct Product rules; it set fab-level thresholds (logic at 16/14nm FinFET or below, DRAM at 18nm half-pitch, NAND at 128 layers or more) with a presumption of denial for PRC-owned facilities; and, for the first time, it barred ‘U.S. persons’ from supporting advanced-chip production at Chinese fabs. The rule took effect in phases through October 21, 2022.

CHIPS and Science Act becomes law: $52B to reshore semiconductor manufacturing

Aug 9, 2022 · hai.stanford.edu

President Biden signed the CHIPS and Science Act on August 9, 2022, authorizing roughly $52 billion in subsidies and tax credits for domestic semiconductor manufacturing and about $200 billion in research authorizations across AI, quantum, and robotics — an industrial-policy bet, framed explicitly against China, to keep leading-edge chip fabrication (and the compute AI runs on) inside the United States.

U.S. tightens the Huawei net: Entity List additions and a broadened Foreign Direct Product Rule

Aug 17, 2020 · federalregister.gov

Effective August 17, 2020, the Commerce Department's Bureau of Industry and Security added 38 more non-U.S. Huawei affiliates to the Entity List, ended the Temporary General License, and amended the Foreign Direct Product Rule so that any foreign-made item produced with U.S. technology or software requires a license when a listed Huawei entity is a party. It was the template — controlling foreign production through U.S. tooling — that the later chip-wide controls would scale up.