On the evening of August 4th, Reuters released a single sentence. Just one sentence. Before the words could become an official document, they first turned into money.
Four informed sources told Reuters that the Trump administration is drafting a ban. The U.S. Federal Communications Commission (FCC) intends to prohibit the import of new models of optical transceiver modules from China.
In plain terms, what is an optical module? It's those things plugged in everywhere inside data centers. GPUs communicate with GPUs entirely through them translating electrical signals into light. Without them, all that computing power would just be a bunch of mutes.
The meaning of the U.S. officials: announce it within this year, effective immediately upon announcement.
The day the news landed, the pre-market moved first.
Pre-market trading has liquidity as thin as paper. A few buy orders can drive the price up. Applied Optoelectronics, which manufactures optical module systems, saw its pre-market price surge over 21%, settling at a 19.44% gain by the close.
The two companies holding laser chips, Coherent rose 12.35%, Lumentum rose 8.92%; even Corning, an upstream fiber optic glass maker, jumped 9% pre-market.
One sentence, one night, that's the price. Absurd? The question is, why is it worth this price?
Behind this matter stand two calendars:
One is called political time, where documents can be signed overnight. The other is called physical time, measured in years: capacity ramp-up, yield refinement, major customer certification.
First, look at how fast political time moves.
Many people think the FCC is an independent agency that can ban whomever it pleases. I checked the official explanation it posted on July 28th, which clearly states.
According to the "Secure and Trusted Communications Networks Act," the FCC itself has no authority to add items to the Covered List. The determination power lies with an inter-agency group led by the White House; the FCC is only responsible for enforcement.
The list only blocks new models; already approved old models are grandfathered. The determination document also leaves a door open called conditional approval; with a nod from the Department of War (formerly Defense) and the Department of Homeland Security, exemptions can be granted.
Over the past year, two tracks have been operating independently.
One track is at the Department of Defense. On June 8th, Zhongji Innolight was placed on the 1260H list of Chinese military-industrial complex companies. Late on June 10th, Innolight issued an announcement stating that the determination and its basis did not align with objective facts, that it is neither a military enterprise nor a civil-military integration enterprise.
1260H falls under the Department of Defense, separate from the FCC's Covered List; they are two independent mechanisms.
The other track is at the FCC. On July 22nd, the FCC voted to pass new rules: any equipment containing "logic hardware components" (i.e., chips) produced by listed companies cannot receive certification, effective immediately with no transition period.
On July 28th, humanoid robots and power inverters were added to the FCC Covered List, again under the banner of national security.
Two tracks, one direction, both tightening; optical modules sit at the intersection. The document isn't even written yet, but both sides have been moving for a year.
The most critical part of this whole affair is the three words "new model."
What exactly does "new model" refer to? The draft doesn't say. Does the existing stock of certified 800G modules count as new? Do modules produced in Chinese companies' factories in Thailand count as Chinese goods?
What about modules from American brands containing Chinese optical chips?
You see, any answer to these three questions could mean a difference of tens of billions. The insiders who leaked the information themselves admitted: this measure may still be modified or shelved.
Now, look at whose rice bowls are on this table.
Zhongji Innolight holds 27% of the global data center optical transceiver module market, ranking first worldwide, according to Counterpoint data.
I checked its Hong Kong stock listing prospectus. In 2025, 57.26% of its revenue came from the U.S., amounting to 24.2 billion RMB. By Q1 this year, that proportion rose to 61.7%.
Eoptolink is even more outward-facing, with 96.16% of its 2025 revenue coming from overseas.
TFC Optical Communication has 74.35% of its sales from exports. Its largest customer is the Thai contract manufacturing giant Fabrinet, accounting for 63.31% alone.
Some rely more on the domestic market: Accelink has 26.69% overseas revenue, HG Tech has 14%.
I've seen some say this is putting eggs in someone else's basket, while others argue conversely that it's the position Chinese manufacturing has earned with its skill in the global AI wave.
We don't need to rush to judge either viewpoint. I only know one thing: The plumbing for America's AI city was entirely laid by Chinese master craftsmen. Now, the management committee has posted a notice saying they'll be replaced by locals.
The notice is up. Where are the new plumbers? The whole city has no answer.
......
The new plumbers actually exist; the three companies whose shares surged pre-market are them. But the market is pricing them based on their expected state three years from now. Right now, these companies are still apprentices.
First, look at the roles of these three companies.
Applied Optoelectronics makes optical module systems, the business most similar to Chinese manufacturers, giving it the greatest elasticity, hence the biggest surge.
Coherent and Lumentum hold the laser diodes and high-speed optical chips. NVIDIA invested $2 billion in each of them in March this year, precisely to lock in capacity.
They sell chips and also make systems. Corning is an indirect beneficiary. The logic in one sentence: if new Chinese modules can't enter, the money from North American cloud providers will have to go to someone else.
Can that someone else handle the money?
I reviewed the documents Applied Optoelectronics filed with the U.S. SEC. In Q1 this year, revenue was $151 million, net loss was $14.3 million. The 800G modules the market went crazy for brought in how much revenue in Q1? $4.6 million, accounting for 5.6% of its data center revenue.
Capacity? By the end of Q1, nearly 100,000 units per month for 800G.
Commitments? Its own press release states: by the end of this year, a combined 500,000+ units per month for 800G and 1.6T; by the end of 2027, 930,000 units.
This July, it just broke ground on two new factories in Pearland, Texas.
Pre-market pricing reflects promises; what's delivered from the factory floor is spot goods.
Another detail: the same press release states that this "face of American manufacturing" has its engineering and manufacturing spread across three locations: Texas, Taipei (China), and Ningbo (China).
Part of the American substitute's capacity is in China.
This creates a very interesting mess of accounts: if the ban is really drafted, how is "Chinese goods" defined?
Based on place of production, do modules from the Ningbo factory count as Chinese goods? Based on brand, it's a NASDAQ-listed company, headquartered in Texas, a bona fide American brand. A blanket cut might first hit its own people.
The Chinese master's factory is in Thailand; the American apprentice's production line is in Ningbo. They've swapped places; drawing a red line anywhere becomes awkward.
The two chip companies have genuine expertise, but the volume is genuinely insufficient.
Just at the beginning of this month, at the RAISE summit in Paris, Lumentum's CEO told the truth himself: "Our two companies combined cannot meet the demand that NVIDIA and other customers are asking of us."
How big is the gap? Over 30%. Coherent's order backlog stretches to 2028.
On the demand side, the building keeps getting taller every day.
One GPU roughly requires 6 high-speed optical modules. 800G demand was 24 million units in 2025 and is projected to surge to 63 million in 2026.
The capex of North American cloud providers: Google just raised its 2026 guidance to $195-205 billion, Meta is at $130-145 billion, Amazon around $220 billion.
Huatai Securities' assessment on August 4th: the sector's previous adjustment was more due to trading and sentiment; the expansion trend of AI computing power spending remains unchanged.
The apprentices can't handle the orders, the building is still being topped off, and these two calendars are destined to collide head-on one day.
......
When will they collide? I pulled out the construction schedules from both sides.
Look at the U.S. first. Forget about building factories; just getting through the certification door takes two to three quarters. Referring again to that Applied Optoelectronics press release: in March this year, it announced securing a major 1.6T order from a super cloud provider, worth over $200 million.
When does shipment begin? Q3. From order placement to delivery, a minimum of two quarters, and that's for a repeat customer familiar with older products.
Once inside the door, skill is another hurdle.
For the 200G EML lasers inside 1.6T modules, a McKinsey analysis gives a yield range from 15% to 50% — a more than threefold difference, depending on wafer generation and design difficulty.
In plumber's terms, for every three pipes the apprentice makes, one and a half end up in the scrap heap. The piles of wasted pipes in the corner are all tuition fees.
Now look at the construction schedule for new factories:
Coherent's 6-inch indium phosphide production line in Texas: internal capacity doubling by end of this year, more than doubling again by end of 2027.
Lumentum's wafer fab in Greensboro acquired from Qorvo won't reach full production until early 2028. The ban aims to be effective this year; the earliest the new plumbers will graduate is the year after next.
Political time demands 2026; physical time demands 2028.
At this point, Zhiyuan has to pull back half the statement. The one in debt isn't just the American side.
Chinese optical modules may have their pipes laid across the globe, but the heart isn't in their own hands. I checked Zhongji Innolight's annual report; it states clearly: 200G and higher-speed EML lasers primarily rely on imports, and suppliers for high-speed electrical chips are also mainly overseas.
Translating to the plumber's ledger: the pipes were laid by the master, but the valves are held in someone else's hands.
Even more interestingly, both sides hold a clock belonging to the other.
In February 2025, China's Ministry of Commerce and the General Administration of Customs placed indium phosphide, trimethylindium, and triethylindium on the dual-use items export control list, requiring licenses for all exports.
AXT, one of the global top three indium phosphide substrate suppliers based in the U.S., has its main production capacity in its Chinese subsidiary, Beijing Tongmei. Once the export controls landed, its shipments to overseas customers had to pass through an additional gate.
Note the timing: this control measure came a full year and a half before the rumors of the optical module ban. It's a card both sides have long had tucked in their pockets.
Part of the interest in the upstream of America's AI machine is calculated by China.
America carries another debt on its head: electricity. A Financial Times report in April stated that nearly 40% of U.S. data center projects originally planned for this year face delays.
Delivery times for high-power transformers, according to industry media statistics, have stretched from the past two or three years to around five years. Before the plumbers even graduate, the city's main water valve can't be turned on.
Looking at both calendars together, how do institutions judge?
The consensus from CICC, CITIC, and others on August 4th: optical modules only perform photoelectric conversion, don't store business data, have weak security logic, so the probability of the policy being more thunder than rain is not low.
Counter-evidence is also on the table: Innolight's own announcement on June 10th regarding the 1260H list, that was published in black and white by the U.S. Department of Defense.
Both sides have their evidence. I only know that in 2028, debts on both sides come due together. Who catches their breath first? Nobody knows now. What both sides are doing right now is the same thing: moving house.
......
This moving house is happening in two directions simultaneously.
One direction: moving capacity outwards. Eoptolink's Thailand factory Phase II started production as early as early 2025.
TFC Optical Communication's Thailand base is accelerating production ramp-up this year. HG Tech's overseas factories are spread across Thailand, Vietnam, and Hungary.
The other direction: moving markets towards non-U.S. regions.
Saudi Arabia's HUMAIN aims to build 1.9 GW of AI data center capacity by 2030. The UAE is building a 5 GW campus, the largest outside the U.S. mainland.
Domestically in China, according to the NDRC's stance, the "15th Five-Year Plan" computing power network construction involves 4 trillion RMB in new direct investment. Newly built intelligent computing centers are to prioritize domestic high-speed optical interconnect.
The cushion is real, but so is the inability to fill the gap.
How big is the table of North American cloud providers? The account was tallied earlier: combined 2026 capex guidance exceeds $700 billion. It's hard to set up a second table of that size in the short term.
There's another twist in moving house. The July 22nd component-level ban has clear logic: it doesn't matter where you assemble; it looks at where the chips come from. Can factories in Thailand circumvent it? Nobody can say for sure now; it depends on how the document is written.
Think again about Ningbo. Even the American substitute's own capacity is partly located in China. These pipes are intertwined; the hand holding the scissors will also tremble.
Some say, simply switch tracks to CPO (Co-Packaged Optics), integrating optical devices directly next to the chip.
Switching to this track also requires queuing up.
A Goldman Sachs report from April 17th makes it clear: even if CPO penetration in NVIDIA's new platform reaches 29% by 2028, the market for pluggable optical modules will still expand tenfold. Interconnection across racks relies 100% on pluggable modules.
At this stage of the game, the picture looks like this:
The master craftsman packs up his tools to find work in other cities. The apprentice stays in the city, working overtime to hone his skills. The city's water consumption rises every day.
That unwritten document will eventually be written. The day it's written, it all comes down to how "new model" is defined, how large the exemptions are, and whether a transition period is granted.
Before these three words are settled, all the wild surges and plunges are just political time setting the price. The bill from physical time will be presented in 2028.
Core Data Sources:
[1]. Reuters, FCC official documents, U.S. SEC disclosure filings, various company annual reports and prospectuses, data from LightCounting, Counterpoint, and other institutions, brokerage research reports; This article does not constitute any investment advice.
This article is from WeChat public account: Wang Zhiyuan , Author: Wang Zhiyuan
