Broadcom and Nvidia face risk from potential US export restrictions on AI components to China, threatening supply chain access.

The Donald Trump administration's push to ban China's role in the U.S. AI supply chain may have an unintended consequence: making it more difficult and expensive to build the very AI infrastructure U.S. companies are racing to deploy.

Optical transceivers rarely grab headlines, but they are a critical piece of modern AI infrastructure. These devices convert electrical signals into optical signals, allowing the enormous volumes of data generated by AI chips to move rapidly between servers inside data centers. As AI clusters expand, these networking components become just as important as the GPUs performing the computations. Without sufficient optical transceivers, even the most advanced AI chips cannot communicate efficiently across thousands of interconnected servers.

According to BNP Paribas analyst Karl Ackerman, Chinese manufacturers are expected to control more than 60% of the global data center optical transceiver market in 2026, with Innolight and Eoptolink among the largest suppliers serving the U.S. market. At first glance, restricting Chinese suppliers appears to create an opportunity for non-Chinese competitors. However, Ackerman argues the supply chain is far more interconnected than that simple narrative suggests.

Optical transceivers rely on digital signal processors, or DSPs—specialized chips that process high-speed data traveling through fiber-optic networks. BNP estimates these processors account for roughly 40% of a transceiver's bill of materials. Chinese transceiver makers are major customers of DSP suppliers including Broadcom, Marvell, Credo and MaxLinear. If Chinese production were curtailed, these chipmakers could lose a meaningful source of demand.

The effects could extend further. Laser suppliers such as Lumentum Holdings and Coherent also sell components into Chinese transceiver manufacturing, while companies supplying semiconductor materials and manufacturing equipment could see weaker demand if production slows.

Replacing that manufacturing capacity would not be straightforward. U.S. suppliers are unlikely to immediately produce enough 800G and emerging 1.6-terabit transceivers to meet demand as hyperscalers ramp deployments of next-generation AI infrastructure. Instead of strengthening the domestic AI ecosystem, an outright ban could tighten supplies of critical networking equipment, raise infrastructure costs and add further inflationary pressure to already record AI capital spending.

Ackerman notes that optical transceivers produced by Innolight and Eoptolink are widely used in AI server systems built around Nvidia's upcoming Vera Rubin platform and Advanced Micro Devices' MI450 accelerators. The timing is particularly important: BNP forecasts the data center optical transceiver market will grow 41% year over year to roughly $31 billion in 2026, driven by the rollout of Nvidia's Vera Rubin systems, AMD's MI450 platform, Alphabet's TPU8 processors and Amazon's Trainium 3 chips.

For investors, the report highlights an increasingly important reality of the AI buildout: the next supply-chain bottleneck may not come from GPUs, but from the networking hardware that connects them.