The racks are in. The accelerators are powered. What is missing is the small cage at the back of every switch, still empty, waiting for the module that turns electricity into light. A large AI hall has hundreds of thousands of those cages. Until they are filled, the most expensive silicon on earth sits idle.
That is the component Washington is debating whether to restrict. On 10 November 2026, Reuters reported that Microsoft and other companies are pressing the White House to abandon a plan that would bar imports of new models of Chinese optical transceivers (光模块) — the parts that move data between servers, switches and accelerators.
The dispute is not about whether China can build them. It is about what happens to an American buildout that has already paid for the chips.
Key takeaways
- Chinese suppliers held about two-thirds of the global optical transceiver market in 2026, according to Counterpoint Research, cited by Reuters on 10 November 2026.
- Seven of the ten largest module suppliers by 2025 sales were Chinese, and Innolight (中际旭创) ranked first with US$5.33 billion in revenue (LightCounting, May 2026).
- The measure under discussion would target 3.2 Tbps modules, a class Morgan Stanley expects to reach volume production only in 2028.
- China's advantage sits in assembly and packaging; the indium phosphide substrate underneath the laser is a different map.
- TrendForce estimates Chinese firms hold 56% of global module contract-manufacturing capacity in 2026, against 19% for the United States.
A module is a conversion, not a chip
Every port on a switch has to hand data to a fibre. The transceiver does the conversion twice — electricity to light going out, light to electricity coming in. A boring function at an unforgiving speed, repeated hundreds of thousands of times per hall.
The count scales faster than the cluster. Each speed step packs more optical channels into one module, so a faster module consumes more laser die, more lenses and more test time per unit. Demand for the part grows faster than demand for the accelerators it serves.
That is why a shortage here does not degrade a cluster gracefully. It stops it. John Miller, executive vice president and general counsel at the Information Technology Industry Council, told Reuters that cloud companies have already bought the accelerators, and without transceivers those accelerators cannot be connected.
Who actually holds the supply
The concentration is recent, and it is Chinese. LightCounting's ranking of 2025 module sales places seven Chinese firms in the top ten. Innolight's revenue reached US$5.33 billion, up 61%; Eoptolink (新易盛) reached US$3.5 billion, up 189%, overtaking Coherent of the United States.
The filings explain the position. Innolight's revenue for the first nine months of 2026 reached RMB 69.5 billion (≈ US$9.8bn / HK$76.5bn), up 178% year on year, with a 46.6% gross margin in the third quarter alone (Innolight filing, 29 October 2026). That is not a contract assembler's margin.
TrendForce's capacity estimate makes the same point from the other side: 56% of global module contract manufacturing sits in China in 2026, and 19% in the United States. Counterpoint Research has estimated that Western suppliers lack the cleanroom space, packaging lines and yield experience to close that gap within 12 to 24 months.
Why the threshold, not the ban, is the story
Two details matter more than the headline. First, no rule exists yet. The FCC is studying the measure; the White House Office of Science and Technology Policy and the National Security Council met companies about it in September, the same month a bipartisan group of legislators introduced a separate bill covering sensitive federal systems.
Second, the proposed threshold. Morgan Stanley, in an October 2026 report, described a plan that would apply only to modules running at 3.2 terabits per second, a class the bank expects to scale in 2028. That timing is the policy: a threshold set at the next generation removes nothing from a rack being built this winter, and gives American and allied capacity a window to qualify before the volume arrives.
Whether the window is long enough is a separate question. Counterpoint's estimate is about filling today's gap, not about winning the next node — and the next node may not be pluggable at all, with TrendForce expecting co-packaged optics to reach about 8% penetration by 2028.
The substrate is the other map
China's share is counted in modules shipped, not in the material underneath the laser. Indium phosphide (磷化铟) is the substrate on which high-speed laser chips are grown, and its supply is concentrated in the opposite direction: Yole counted the top three makers at more than 90% of the market in 2020, with Sumitomo of Japan at 42% and Beijing Tongmei at 36%.
The raw metal adds a layer. The US Geological Survey estimates China refined 760 tonnes of indium in 2025, about 69% of a world total of 1,100 tonnes, and puts US net import reliance at 100%. China added indium to its export-licensing regime in 2025 (USGS, Mineral Commodity Summaries 2026).
The two maps describe mutual exposure, not one-sided dependency: the United States needs Chinese modules to connect accelerators it has already bought, and Chinese module makers need Japanese and American substrates, laser die and signal-processing chips. Each side can slow the other; neither can replace the other within a product cycle.
Honest limitations
The Reuters report rests on three people familiar with the discussions; Microsoft declined to comment and no agency has published a rule. Market shares differ by metric: revenue, units, capacity and product tier give different answers, and the two-thirds figure is Counterpoint's revenue estimate, not a customs statistic. The Yole numbers are from 2020 and are the most dated here. No module maker discloses customer-level volumes; the dependency is inferred from capacity and public statements. The indium figure is a USGS estimate of refinery output, not of what is available for export.
What readers can do now
- Read the FCC docket and covered list directly: the threshold and the treatment of non-Chinese brands decide who is in scope.
- Map your bill of materials to the two constrained inputs, indium phosphide substrate and signal-processing chips, and ask for qualification lead times, not capacity.
- Track quarterly filings at Innolight and Eoptolink: revenue growth against gross margin shows whether scarcity pricing is still passing through.
