US export controls on AI chips aren’t a list of banned products. They’re a formula. A chip crosses two calculated numbers — Total Processing Performance (TPP) and Performance Density (PD) — or it doesn’t, and that crossing point decides whether a license is needed to sell it abroad. Because the rule is a threshold rather than a blacklist, chip designers can engineer a chip specifically to land on the legal side of it — which is exactly what a new wave of Chinese AI chip startups is now trying to do.
The rule that decides which chips need a license
The US Bureau of Industry and Security (BIS), part of the Commerce Department, administers the Export Administration Regulations — the body of law that governs which “dual-use” technologies can be sold to which countries. In October 2023, BIS created a specific classification for advanced AI chips, Export Control Classification Number 3A090, and defined it not by brand or country of design but by two measurable technical properties. Any chip that clears the bar needs a license to leave the US for large parts of the world; which license tier applies also depends on the destination’s country group.
Two numbers: TPP and performance density
Total Processing Performance measures a chip’s raw peak compute, calculated from the number of multiply-accumulate operations it can do per second and the bit-width of those operations. Performance Density divides that number by the chip’s physical die area — compute per square millimeter, in effect.
Under the current rule, a chip needs a worldwide license if its TPP reaches 4,800, or if its TPP reaches 1,600 and its performance density reaches 5.92. A second, less restrictive tier applies at lower combinations of the two numbers, requiring a license only for sales to specific restricted country groups.
BIS added performance density deliberately, not as an afterthought. Early versions of the rule judged chips on raw TPP alone, which meant a company could, in theory, hit the same total compute as a banned chip simply by wiring together several smaller, less dense ones. Performance density closes that loophole by penalizing chips that pack a lot of compute into a small area — which is precisely the design trait that makes a single chip powerful.
Designing a chip that stays under the line
Because the rule rewards low performance density, a chip that spreads its compute across a larger die, or that leans on an older manufacturing process instead of the most advanced one, can post strong real-world numbers while its calculated density stays under the threshold that triggers the toughest license tier.
That is the strategy behind Dongfang Suanxin, a Shanghai-based startup led by Tsinghua University semiconductor veteran Wei Shaojun, which came out of stealth in mid-2026 with its DF1000 accelerator. Rather than chase the cutting-edge manufacturing node and high-bandwidth memory that top-tier AI chips depend on — both squarely inside what US controls try to restrict — the company built the DF1000 on a mature 14-nanometer process using a 3D-stacked “near-memory computing” layout, stacking compute and memory logic on top of each other instead of relying on external HBM. The company says the design delivers 520 teraflops of BF16 performance and 6.4 terabytes per second of memory bandwidth without the components export rules were written to catch.
This is a legal strategy, not a workaround of the law itself — a company that stays under the threshold owes no license, by design of the rule. It’s a different phenomenon from chip smuggling, where already-controlled hardware is illegally diverted to a restricted buyer after the fact.
Why it matters
A threshold defined by a formula is, by nature, a moving target. BIS has already rewritten the numeric bar behind AI chip controls several times since 2022, each round responding to a way the previous version could be routed around. If chip architectures built to sit just under today’s line become common and genuinely competitive, expect the thresholds — or the metrics used to calculate them — to move again. For now, the contest between what regulation restricts and what engineering can route around it is a defining, ongoing feature of the global AI chip market.
In the news
See our brief on Dongfang Suanxin’s DF1000 launch for more on the chip driving this story.
FAQ
Is designing a chip to fall under the threshold illegal? No. If a chip’s calculated TPP and performance density fall below the control levels, it isn’t subject to the license requirement — that’s how the rule is written to work.
Does staying under the threshold mean a chip can be sold anywhere? Not automatically. Chips in the lower tier can still require a license for sales to specific restricted country groups, and separate rules — such as the Entity List or the Foreign-Produced Direct Product Rule — can restrict a transaction even when the chip itself is under the performance bar.
Why not just ban chips by brand or country of origin? A formula based on measurable performance applies uniformly regardless of who designs or manufactures the chip, which is also why companies outside China study the same numbers when deciding how to design their own products.
Sources: US Bureau of Industry and Security, Export Administration Regulations (Commerce Control List, ECCN 3A090, effective October 2023); reporting on Dongfang Suanxin’s DF1000 from the South China Morning Post and TrendForce.