The previous seven gates, from AI chips and HBM and CoWoS all the way through to data centers and power, were all about “how the thing gets made.” This gate is a little different. It’s more like a layer of rules sitting over everything that came before: who can obtain how many advanced chips, and who can run an advanced process, is largely decided by each country’s export controls.
This piece is a neutral compilation of the AI chip war and export controls: which tools the US uses, how the rules shifted in 2026, how far China’s homegrown effort has come, and where Taiwan fits. Let’s be clear up front: this article only states each party’s public measures and the facts; it makes no political commentary. This is the deep-dive version of Gate 8 in The AI Hardware Supply Chain, End to End.
What the Chip War Is
The “chip war” is the media’s umbrella term for a string of back-and-forth moves: the US–China export controls and tech competition around advanced AI chips, semiconductor manufacturing equipment, compute, and AI models.
The starting point is usually dated to October 2022: the US Commerce Department began restricting advanced computing chips and related equipment from flowing to China, and over the following years the rules were repeatedly adjusted — tightened, then fine-tuned. It moves more than just the two countries; it also affects the equipment makers caught in the middle (such as the Netherlands’ ASML), chip designers (such as NVIDIA), and foundries (such as TSMC). Because what it shapes is “who can obtain compute,” it is seen as one of the most pivotal geopolitical variables of the AI era.
The US Control Toolkit: In Plain Terms
On the US side, enforcement runs mainly through the Bureau of Industry and Security (BIS) under the Commerce Department. The tools sound numerous, but in plain terms they come in three kinds.
The first is lists, the most famous being the Entity List: a company placed on it requires US businesses to apply before selling to it, and approval is usually hard to get. The second is rules and licenses, where the Export Administration Regulations (EAR) set which chips and equipment are restricted and whether a license is needed before a sale. The third is the review posture, of which two are common: “presumption of denial” amounts to a near-blanket no, while “case-by-case review” allows individual approvals when conditions are met. As you’ll see below, the key shift in 2026 was certain chips loosening from the former toward the latter.
The 2026 Update: A Conditional, Narrow Door Opened for the H200
The most closely watched turn in early 2026 involved NVIDIA’s H200. The H200 isn’t the newest Blackwell generation, but it remains a high-end AI accelerator.
The new rule in January 2026 changed the “direct US export” of H200-class chips (and AMD’s equivalents) to China and Macau from “presumption of denial” to “case-by-case review.” But the bar isn’t low, and it comes with a long list of conditions: performance must sit below a certain ceiling (for example, total processing performance, TPP, below 21,000, and memory bandwidth below 6,500 GB per second), total export volume is capped by ratio, customer vetting is required (KYC — confirming the buyer’s identity and intended use first), and there must be third-party testing in the US. Reselling the goods onward to a third country (re-export) or transferring them within China mostly remains tightly controlled.
At the same time, two things that often get conflated need to be kept separate: one, the US has publicly noted a fee of about 25% on shipments to approved Chinese customers (this is the revenue-sharing arrangement on H200 sales to China); two, in January 2026 the White House separately issued a Section 232 notice (Section 232 is the legal basis under which the US adjusts imports on national-security grounds), levying a 25% tariff on some advanced chips imported into the US, with carve-outs for US domestic data centers, R&D, and the like. These two 25% figures are not the same thing. As of May 2026, Reuters reported that the US had approved roughly ten Chinese companies to purchase the H200, though delivery had not yet happened, and actual transactions still hinge on both sides’ policies and approvals.
The Equipment Gate: EUV Still Banned, DUV Tightening
Beyond chips, manufacturing equipment is another battleground, and this part is closely tied to the ASML gate.
In short: the most advanced EUV lithography machines (the exposure equipment used for advanced processes) have seen no sign of easing for China; the focus in 2026 has turned to whether to further restrict older DUV immersion tools and servicing. The Netherlands already expanded equipment export controls in early 2025, and members of the US Congress have proposed further tightening restrictions on equipment makers. The logic mirrors that of controlling chips: limiting access to equipment and servicing affects the speed of capacity expansion, yield maintenance, and upgrades for advanced processes.
On China’s Side: Homegrown Chips and Saving Compute
After being placed under controls, China has on one hand pushed homegrown chips and on the other found ways in software to save compute.
On hardware, Huawei’s Ascend series (such as the 910C and 910D) is seen as the flagship of China’s homegrown AI chips. In 2026 there were reports that companies like ByteDance and Alibaba had an intent to order after testing. On software, some Chinese AI companies reduce compute pressure through model architecture, long-context efficiency gains, quantization, and distillation (a compression technique where a large model “teaches” a smaller one); for example, a new DeepSeek model was reported to be adapted for the Huawei Ascend ecosystem.
A neutral note is in order: the US has alleged that some Chinese models obtained their capabilities through disputed methods such as “distillation,” while China and the companies involved have denied or pushed back. This part is a dispute where accounts differ; this article only records that the contention exists and makes no judgment. Most industry analysis still holds that China’s homegrown AI chips trail the most advanced products in single-chip performance, interconnect, and software ecosystem; what to watch next is how that gap changes and how mature the alternatives become.
The “Silicon Shield” and Taiwan’s Position
When the conversation turns geopolitical, you often hear the term “silicon shield.” First, to be clear: it is an analytical concept, not a legal guarantee. It refers to Taiwan’s pivotal position in the global semiconductor supply chain, which may raise the incentive for the various parties to maintain regional stability; it is often used to discuss geopolitical risk, but it does not represent any security commitment.
On the institutional side, in early 2025 the US rolled out an “AI Diffusion” framework (a framework managing cross-border access to compute and model weights, unrelated to the diffusion in image generation), tiering different countries for compute access, with Taiwan placed in a less-restricted category (often called Tier 1). However, after May 2025 this framework entered a state of “announced as not enforced, but the formal rescission process not yet complete,” with the regulatory text still on the books. In other words, the rules themselves are still in flux, so mind the point in time when reading.
Do the Controls Work? A More Neutral View
Finally, the effects — which is also where it’s easiest to overstate.
The more neutral research view is this: export controls have raised the cost for China to obtain advanced compute, lengthened the timeline, and shifted supply chains and investment directions, but a full block is hard to achieve. The reasons include the gray zone around older equipment, smuggling and third-country transshipment, domestic substitution, and using software efficiency gains to save compute. The controls carry costs of their own too: a survey by the think tank CSIS found that many semiconductor and IT firms reported export-license reviews averaging over 180 days, with more than thirty percent exceeding even 300 days, which knocks on to normal business. Add in countermeasures like rare earths and critical metals, and this back-and-forth is more complex than “who banned whom.”
Key Takeaways for This Gate
After working through export controls, first fix its place: it’s more like a layer of rules sitting over the previous seven gates, deciding who can obtain compute and who can run an advanced process.
The key shift in 2026 is that H200-class chips bound for China loosened from “presumption of denial” to “conditional, case-by-case review,” alongside tariffs and fees; on the equipment side, EUV remains banned and DUV is tightening. China, for its part, develops alternatives through homegrown Ascend chips and software efficiency gains. The more neutral view: the controls raise cost and delay but are hard to turn into a full block, and this back-and-forth will keep shifting.
To see the most tightly controlled gate — equipment — go back and read What Is ASML; to see how all eight gates string together, head back to the supply-chain overview.