US chip controls may shift from GPUs to design software licenses in China

Baillie Gifford's [Official Research](https://www.bailliegifford.com/en/usa/institutional-investor/insights/ic-article/2026-q1-chinas-chip-gap-10061288) says China is rapidly advancing in AI software while facing a hardware deficit in advanced semiconductors. This is, so far, single-thread reporting — Baillie Gifford only, with no independent confirmation in the packet — and no one in the reported packet is on the record.

For a US export-control lawyer, a chip procurement head, or an AI infrastructure executive, the practical question is whether the control point remains the finished accelerator, or whether it is moving upstream to design software and licensed IP. This article rests on that single document; no outside parties were consulted.

Baillie Gifford’s signal is a two-gap story

The source summary makes two claims that should not be collapsed. First, it says China is “rapidly advancing in AI software, often outperforming US peers.” Second, it says China “faces a significant hardware deficit in advanced semiconductors.” The usual reading is that the second claim cancels out the first: without access to the best chips, China’s AI progress eventually slows.

The more important reading for manufacturers and AI buyers is narrower: software progress and hardware scarcity can coexist long enough to change where governments, vendors, and procurement teams spend their time.

That distinction matters because export controls are usually discussed as if the sale of the restricted chip is the decisive transaction. Baillie Gifford’s framing points to a different regulatory perimeter.

If Beijing is building “domestic chip capacity and a parallel AI ecosystem,” as the source summary says, then the hardest question for US policy is not only which chips can be shipped. It is which upstream tools, reusable design IP, and support relationships remain available to a country trying to close the hardware gap without relying on the same supply routes.

The missing measurement is the whole story

The weakness in the Baillie Gifford signal is that the packet gives no benchmark, hardware baseline, task definition, or reproducibility detail for the claim that China’s AI software is “often outperforming US peers.” Measured against which US peers, on what tasks, and on what hardware?

If the comparison is software performance on constrained hardware, it says one thing about engineering efficiency. If it is model performance in domains where Chinese firms have stronger local data or deployment loops, it says something else. Without that detail, the claim should be treated as directional investor research, not as a settled technical finding.

The hardware side has the same problem. “Significant hardware deficit” is a useful phrase for investors, but it does not specify whether the binding constraint is chip design, manufacturing yield, packaging, memory supply, interconnect, software compatibility, or power availability.

Each constraint creates different work. A manufacturing executive cares whether scarcity means slower factory automation, more local sourcing, or more engineering hours spent adapting AI workloads to less capable infrastructure.

A general counsel cares whether the next regulated item is a physical shipment, a license renewal, or an engineer’s support call. The Baillie Gifford packet does not answer that, which is why the omission is load-bearing.

Finished chips may become the less stable control point The consensus policy read is simple: restrict advanced semiconductor sales and China’s AI ambitions are hobbled. The mechanism fails if China’s response is not to match the most advanced US-linked stack one-for-one, but to build a parallel system that is good enough for domestic deployment and optimized around what it can manufacture, license, or substitute.

Baillie Gifford’s source summary explicitly points to aggressive domestic capacity building and a parallel AI ecosystem; it does not show that this ecosystem can match the frontier, but it does suggest that chip bans alone may not define the limit of Chinese AI capability.

That is why the second-order regulatory consequence is upstream. Export-control pressure can move from “can this accelerator be sold?” to “can this chip be designed, verified, and productized with controlled software or IP?” Electronic design automation software, reusable chip IP, and technical support are less visible than finished GPUs, but they sit closer to the work of making domestic capacity usable.

If controls move there, the burden shifts from logistics and sales compliance toward engineering workflows, license management, and product segmentation.

Analysis: the org chart moves before the factory does The work changes first inside legal, procurement, and engineering operations. A chip company selling into or around China would need tighter coordination between export-control counsel and the teams that manage design tools, customer support, and IP licensing.

An AI infrastructure buyer with China exposure would need to know not only what chips are available, but whether its vendors can keep maintaining the tools and software layers that make those chips useful. The affected employees are not only hardware sales teams; they include application engineers, CAD administrators, compliance reviewers, and procurement managers who sit between policy language and production schedules.

For manufacturing executives, the under-noticed middle is the supplier that is neither a frontier chipmaker nor a cloud AI lab. It is the software vendor, design services firm, contract manufacturer, or systems integrator whose value comes from translating designs into deployable systems.

If direct chip restrictions become less decisive, these intermediaries become more important and more exposed. Their contracts may need clearer territorial rights, termination language, and support boundaries, because the regulated asset may be expertise or licensed design capability rather than a boxed product.

The skeptic’s case is stronger than the headline suggests The obvious counter-read is that moving upstream is harder than restricting a finished chip. Design software and IP licensing are complex, multinational, and deeply embedded in engineering workflows; overbroad controls can also damage US-aligned vendors by pushing customers toward substitutes.

Baillie Gifford’s packet does not prove China can replace the full upstream stack, and it does not identify the mechanism by which domestic chip capacity overcomes the hardest constraints in advanced semiconductors. A regulator can write a rule faster than an ecosystem can replace a toolchain, but a poorly targeted rule can also accelerate the substitution it is meant to prevent.

The near-term test is therefore not whether China announces another AI model or another chip plan. The more revealing signals will be whether US policy language starts naming design software, reusable chip IP, or support services more directly; whether Chinese procurement language emphasizes domestic toolchains rather than only domestic chips; whether global vendors discuss China exposure in terms of licensing and support rather than only hardware shipment; and whether manufacturers begin separating product road maps for China-facing and non-China-facing engineering stacks.

If those signals do not appear, the direct-chip-control thesis remains stronger than the upstream-regulation thesis.

The narrow implication is that the future-of-work story is not mass replacement by Chinese AI or a clean decoupling of supply chains. It is a compliance and engineering labor shift inside the firms that make AI infrastructure usable.

Baillie Gifford gives only a single-thread investor signal, but it is enough to frame a falsifiable claim: as China builds around its chip gap, the scarce work moves from buying restricted hardware to governing the design tools, licenses, and technical knowledge that make domestic hardware useful.