Two-Thirds of the World's AI Optics Are Chinese. Microsoft Just Told Trump a Ban Would Break the Buildout.

On Nov 10, 2026, Reuters reported that Microsoft and other companies and trade groups have pressed the Trump administration to halt plans to bar imports of new Chinese optical transceivers used in AI data centers. Three people familiar with the situation described the lobbying. Microsoft met officials from the White House Office of Science and Technology Policy and the National Security Council in September 2026 to flag the supply-chain problems any such ban would create. The company declined to comment, and the White House, the FCC and the Chinese embassy did not respond.

The pitch to the White House is that a ban could \"irreparably damage the US edge over China in the AI race\" by slowing the data center rollout. Microsoft has 35 data centers under construction in the US and another 134 planned, according to Cleanview, which tracks data center development. That is the number that matters: a buyer with 169 sites in the pipeline is not making a general argument about national security. It is protecting a construction schedule.

Microsoft spent $39.4 billion of cash on property and equipment in the September 2026 quarter, and told investors capital expenditures including finance leases were $45.2 billion, short of the more than $50 billion guided, because some third-party capacity deliveries it expected in the quarter were delayed. Trailing capital spending is $135.9 billion, or 39% of revenue, against $60.0 billion of free cash flow. Azure still grew 48% on revenue of $92.3 billion, up 19%. Read that combination carefully: a company spending at 39% of revenue is complaining about a component, not about demand.

That last bullet is the whole argument in one number. Microsoft did not lobby because optics are abundant; it lobbied because they are not. When a company running a $170 billion annual capital plan cannot get third-party capacity on the schedule it planned, the marginal value of the next rack falls, and so does the revenue that rack was meant to produce. A rule that removes more than half the supply of the component linking thousands of GPUs inside a building attacks that conversion rate directly. The security argument about data theft and malware is real; the sequencing argument is the one with a price tag on it.

The ban is being built inside the FCC's equipment-authorization system rather than through Commerce export controls. In its Third Report and Order (FCC 26-50, adopted July 22, 2026 and published in the Federal Register on Sept 11, 2026), the Commission closed the \"component-part loophole\": authorization is now denied for devices that incorporate logic-bearing hardware components produced by Covered List entities. The definition sweeps in every digital component of a modern transceiver, because it covers any physical component that generates and uses timing signals above 9,000 cycles per second, including the DSP.

To make that workable the same order requires applicants to file certified hardware and software bills of materials naming every producer, every production location and the percentage of component value attributable to each. It also puts the burden on online marketplaces to display a valid FCC ID at the point of sale, with the harder obligations kicking in six and nine months after publication. For a company assembling modules from Chinese emitters, Japanese packages and American silicon, that paperwork is a cost. For the incumbents it is a moat, because it rewards whoever can document a fully trusted chain end to end.

The congressional version is therefore softer than the agency version, and the agency version is faster. Sponsors of S.5548, introduced on Sept 25, 2026 by Senators McCormick, Gallego, Cornyn and Fetterman, would have the Secretary of Commerce map allied capacity and build a trusted supply chain, with waivers reported to Congress. The FCC can act without any of that. For investors, the filing to watch is not the bill, it is the docket.

Coherent did what the policy bulls wanted. Revenue in its September 2026 quarter was $2.37 billion, up 50% year over year and 56% on a pro forma basis, with GAAP gross margin of 39.1%. Management guided to fiscal 2027 pro forma revenue growth above 60%, a first $3 billion quarter before the end of fiscal 2027, a first $4 billion quarter in fiscal 2028, and the first quarter of non-GAAP gross margin above 45% before the end of calendar 2028. The operating story is exactly what a decoupling bull case requires.

The filings show what it costs to chase that growth. Capital spending was $775.5 million in the quarter and $1.78 billion over the trailing twelve months. Trailing operating cash flow is negative $59.1 million and trailing free cash flow negative $1.84 billion, with inventory rising $737 million in the quarter alone. The company that would be the first beneficiary of a ban is funding the capacity behind it out of its balance sheet, one quarter at a time.

The market has already tried to price this, and it paid up for the wrong asset. On Nov 2, 2026, Marvell Technology agreed to acquire Lumentum in all-stock terms with an implied enterprise value of $115 billion, or $1,132 per Lumentum share — 4.5 Marvell shares for each Lumentum share, a 23.9% premium to the Oct 28 close, more than $1 billion of run-rate cost synergies, and closing targeted for the first half of calendar 2028. Against Lumentum's trailing revenue of $3.74 billion and 45.9% gross margin, that is roughly $31 of enterprise value for every dollar of trailing sales.

Set those multiples side by side and the contradiction is plain. Coherent is the larger company by revenue, the only US supplier of scale, and the name policy bulls own, and it trades near six times sales while Marvell Technology is paying nearly five times that multiple for a smaller business. The decoupling premium has been concentrated in the asset being bought out rather than in the capacity being built, which is not a stable equilibrium: it holds only if the ban actually lands.

The quiet beneficiary sits upstream of the module makers. Sumitomo Electric's device division sells the optical devices — the emitters and packages — into the transceiver manufacturers themselves, says most of its customers are overseas, and describes engaging them on designs from the start. Group revenue was ¥5.11 trillion in the fiscal year ended March 31, 2026, up 9.2%, with trailing operating income of ¥484.7 billion. Somebody has to supply the laser and the package whoever ends up assembling the module, and a Japanese supplier of both is paid by whichever maker scales.

Since Aug 4, 2026, the trade in Chinese optics names has been a story about Chinese earnings rather than about Washington. Zhongji Innolight reported revenue of CN¥27.77 billion in the September 2026 quarter, against CN¥6.67 billion in the March 2025 quarter, and lifted trailing revenue to CN¥82.78 billion at a 43.0% gross margin. Eoptolink grew fiscal 2025 revenue 187% to CN¥24.84 billion and carried trailing revenue to CN¥40.19 billion at a 47.9% gross margin. Both are named directly in S.5548, and both are compounding through the policy debate.

Meanwhile the names that were supposed to benefit de-rated. Coherent closed at $214.66 on Nov 11, 2026, about 51% below its 52-week high of $440, and Marvell Technology sits roughly half below its 52-week high of $329.88 after falling 13% on the day the Lumentum deal was announced. The market did not abandon the AI buildout; it re-underwrote the timing and cost of decoupling while the Chinese leaders kept growing into the gap.

The mirror image makes that unwinding concrete. On Sept 23, 2026, the Financial Times reported — and Reuters relayed — that China's State-owned Assets Supervision and Administration Commission is surveying the use of Broadcom switches in state-controlled data centers, with preliminary findings that Broadcom equipment could account for as much as 90% of the gear in use. Washington is restricting Chinese optics at the door of American data centers; Beijing is auditing American switch silicon inside its own, where locally made gear is the likely replacement.

The asymmetry in exposure is worth sizing. Broadcom generated $89.10 billion of revenue over the trailing twelve months at a 68.8% gross margin, with fiscal 2025 revenue of $63.89 billion against $51.57 billion a year earlier, and its shares trade near 18 times enterprise value to trailing sales. A procurement shift inside Chinese state clouds reaches that revenue base. Arista Networks, the other US-listed networking supplier in the chain, generated trailing revenue of $11.89 billion at a 62.9% gross margin and sells into the same hyperscaler buildout that an optics shortage would slow. One name faces a policy hit from Beijing, the other faces a delivery hit from Washington.

The near-term calendar is thin. The FCC has published no text, no docket number and no date, and the White House, the FCC and the Chinese embassy all declined to comment. What is dated is the trade truce, which runs toward Jan 10, and the statute, S.5548, which needs a five-year transition and renewable waivers whenever no trusted alternative exists. Watch the Covered List and the equipment-authorization dockets rather than the headlines: a listing would do more damage to Chinese optics than the bill ever will.

Over one to three years the question is who can build capacity fast enough to matter. Coherent's targets — a $3 billion quarter before the end of fiscal 2027, a $4 billion quarter in fiscal 2028 and non-GAAP gross margin above 45% before the end of calendar 2028 — are the bull case for a US supplier scaling into the gap. Lumentum's handover to Marvell Technology in the first half of 2028 is the bull case for owning the gap outright. Both assume the gap is worth owning, which is precisely what Microsoft is telling the White House it is not.

Microsoft is not arguing that Chinese optics are safe, that the security case is weak, or that decoupling is wrong in principle. It is arguing that a ban applied at the wrong layer, before trusted capacity exists, raises the cost of every AI rack and slows the rollout. That is a statement about sequencing, and sequencing is the one part an investor can act on. The scarce asset is the ability to ship modules at scale, not the rule that might create the shortage.

So the position follows the supply chain rather than the headline: suppliers of the component with a documentable, trusted chain and a balance sheet to expand — Coherent, Lumentum, Sumitomo Electric — rather than valuations that depend on a specific rule landing on a specific date, such as Marvell Technology paying 31 times trailing sales for a deal that closes in 2028. And caution on the buyers whose capital conversion depends on hardware they cannot yet source, Microsoft chief among them, while the named Chinese champions keep compounding until a rule actually bites.