Nvidia CPO switches are in mass production and ramping faster: 15k in 4Q26, 100k in 2027. CPO vs NPO for Rubin Ultra, and why $LITE, Browave, BizLink and $SMTC win while Largan looks priced in.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
Nvidia's co-packaged optics (CPO) switches are in mass production, and the ramp is running ahead of plan. One sell-side update circulating this week forecasts 15,000 Spectrum-X CPO scale-out switches in 4Q26 alone and 100,000 in 2027. The next step is scale-up optics on Rubin Ultra NVL576, where Nvidia is running CPO and near-packaged optics (NPO) side by side. Lumentum, Browave, BizLink and Semtech look like the winners. Largan looks priced for a share it may not get.
This is an update to our co-packaged optics deep dive from June. That post covered what CPO is and who builds it. This one covers what changed in the last three months: volumes, the CPO-versus-NPO fork for scale-up, and which suppliers the new numbers actually help.
👉 Track it live: the optical names sit in the Photonics / CPO basket and the Silicon Photonics & Optical Interconnects trend on Macroplane.
In June, Nvidia ($NVDA) was shipping the first Spectrum-X Photonics switches to a handful of partners. On August 14 it said the platform had entered full production. The first systems went to CoreWeave, Lambda and Oracle, with Microsoft Azure, IBM Cloud and Nebius also named as early adopters (StorageReview).
Nvidia also named the supply chain, which it rarely does:
| Layer | Supplier |
|---|---|
| Silicon photonics wafers | TSMC ($TSM) |
| Wafer- and chip-level packaging and test | SPIL, part of ASE Technology ($ASX) |
| Laser chips | Lumentum ($LITE) |
| Laser module sub-assembly | Suzhou TFC Optical ($300394.SZ) |
| Switch system assembly | Foxconn ($2317.TW) |
Taiwanese press coverage of the same announcement added Browave ($3163.TWO) to the partner list (Dan Nystedt).
The constraint in June was yield on the optical engines. TrendForce flagged optical-engine yield, silicon photonics fab capacity and advanced packaging as the three bottlenecks in July (TrendForce). The new note says yields at key component suppliers have improved enough to put Spectrum-X CPO into genuine volume. That is the main reason the forecast moved up.
The note's forecasts for Nvidia CPO:
| Metric | 4Q26 | 1Q27 | 2Q27 | 2027E | 2028E |
|---|---|---|---|---|---|
| CPO scale-out switches (units) | 15k | 16k | 30k | 100k | – |
| Scale-up optical engines (OEs) | – | – | – | 6.5m | 42m |
Three things stand out.
The 2026 number moved into a single quarter. In August, GF Securities forecast 15,000 Nvidia CPO switches for all of 2026 and 100,000 for 2027 (Seeking Alpha). The new note keeps 100,000 for 2027 but now puts 15,000 in 4Q26 by itself.
The 2027 total looks conservative. 1Q27 and 2Q27 add up to 46,000 units. Hitting 100,000 for the year means about 54,000 across the second half, or roughly 27,000 a quarter. That is below the 30,000 forecast for 2Q27. Either the analyst expects a pause while customers move from Rubin to Rubin Ultra, or the second half is set low. We lean toward the second explanation, because every other CPO data point this year has come in ahead of forecast.
Scale-up is the bigger market. 6.5 million scale-up optical engines in 2027 rising to 42 million in 2028 is a 6.5x jump. That depends on Vera Rubin Ultra NVL576, which links eight 72-GPU racks into one 576-GPU NVLink domain (The Register). Copper cannot span eight racks at NVLink speeds, so the spine layer has to be optical. The note says NVL576 is still on schedule, but it cut its estimates because a later stage of the ramp (it calls it "T4") has slipped to around 1Q28.
For scale-out Ethernet, the question is settled: Spectrum-X uses CPO. For scale-up on Rubin Ultra, Nvidia is building two versions in parallel:
SemiAnalysis expects NPO to reach the market first because the form factor is more mature. The note agrees on timing and still assumes NPO takes the full Rubin Ultra cycle.
Most coverage of this choice focuses on reliability. For investors, the more useful question is how much each supplier sells per switch under each design. CPO puts more fiber, more shuffle hardware and more precision assembly inside the switch. NPO uses more sockets and more discrete TIAs and drivers on the board. The table sums it up:
| Supplier type | Gains more under CPO | Gains more under NPO |
|---|---|---|
| Fiber harnesses, shuffles, internal cabling ($3163.TWO, $3665.TW) | Yes, higher content per switch | |
| Linear TIAs and drivers ($SMTC) | Yes, discrete parts sit next to each socketed engine | |
| Lasers and external light sources ($LITE) | Yes | Yes, both designs need them |
| Fiber-array units ($3363.TWO, $300394.SZ, $3008.TW) | Yes | Yes, but share is contested |
The design Nvidia leans toward won't change the total amount of optics shipped. It will change which suppliers get the margin.
Feynman goes further. For the generation after Rubin Ultra, the note says Nvidia is designing scale-up CPO with optical engines of up to 6.4T each on QM6 switch packages. That doubles today's leading 3.2T engine class. Nvidia has said Feynman, due in 2028, will support both copper and CPO NVLink (The Register). If CPO wins then, the suppliers with CPO-heavy content get a second, bigger cycle.
Lumentum is the only company Nvidia named as the laser chip supplier for Spectrum-X, and it is exposed to both designs. The note sees CPO and NPO revenue at 12% of Lumentum's sales in CY27 and 26% in CY28.
The company's own guidance supports that. On its August earnings call, Lumentum said ultra-high-power (UHP) laser revenue for CPO should reach about $50M a quarter by the end of 2026 and pass $100M a quarter in the March 2027 quarter. It said it was "very much further behind relative to our ability to supply" UHP lasers. It also booked an additional multi-hundred-million-dollar UHP order for scale-out shipping in 1H27, plus its first external light source (ELS) module order for delivery in 2H27 (Semiconductor Today). Management called its "breadth of NPO engagements" the first sign that optics is moving inside the rack.
The stock is not cheap at $914.85 and a $68.9B market cap. The argument for it is that the CPO share of revenue can roughly double in a year while the core transceiver-laser business is still growing. Our LITE teardown covers the customer concentration risk. Our external laser source explainer explains why ELS modules carry higher ASPs.
(The note lists "Broadcom (3163 TT)". 3163 is Browave's ticker, so we read that as Browave.)
A CPO switch replaces a front panel of pluggable modules with hundreds of fibers routed inside the chassis, from the optical engines to the faceplate connectors. That internal fiber is new content that pluggable switches never had. This is why the note prefers these two over NPO-leaning names: CPO carries more fiber content per switch than NPO.
Semtech is on the list for the opposite reason. If NPO wins scale-up, as the note assumes, each socketed engine needs linear TIAs and modulator drivers placed next to it. Semtech sells exactly those parts. On September 21 it launched 224G linear octal TIAs (GN1838L, GN42T380) and an octal MZM driver (GN42M380) sized for 1.6T to 12.8T NPO and CPO engines (Electronics Weekly). It added indium phosphide lasers through the HieFo acquisition in March.
The financials are catching up. Revenue for the July quarter was a record $341.9M, up 33% year on year (Semtech 8-K). Guidance for the October quarter is $410M, with data center revenue up 45% sequentially (Seeking Alpha). The May quarter, the last one on Macroplane's SMTC page, was $291.0M. The note calls supply "tight," and Semtech's own commentary on InP laser capacity says the same. Tight supply keeps pricing firm, but it also caps how fast volume can grow.
Largan is the world's best smartphone lens maker, and it is trying to become a CPO supplier. It is building fiber arrays (FA), micro-lens arrays and eventually full fiber-array units (FAUs), the precision parts that couple light from the silicon into fiber. It has bought land or factory space seven times in 2026 for this (DigiTimes). Its automated FA pilot line was scheduled to start in September, and mass production could begin as early as mid-2027, possibly for Rubin Ultra (TrendForce).
The note's objection is price, not technology:
Largan's own management has said as much. Chairman Lin En-ping said in July that an existing supplier "will definitely retain a certain share, and likely still a relatively higher share." He added that the one volume customer wants FA parts only, not full FAUs, and is multi-sourcing. On margins: "ASP will definitely be higher, but if yields are poor, the damage is also severe." Largan's claimed sub-0.3-micron FA precision (against an industry range of 0.5 to 0.8 microns) is a real advantage. But the stock at NT$5,995 already prices it in, before a single volume shipment.
Yes. Nvidia said on August 14, 2026 that Spectrum-X Ethernet Photonics CPO switches had entered full production, with first shipments to CoreWeave, Lambda and Oracle. The Quantum-X InfiniBand CPO switches started shipping in low volume in early 2026.
Sell-side forecasts have been rising. GF Securities forecast 15,000 in 2026 and 100,000 in 2027 as of August. A September update keeps 100,000 for 2027 but puts 15,000 in 4Q26 alone, followed by 16,000 in 1Q27 and 30,000 in 2Q27.
Co-packaged optics (CPO) bonds the optical engines permanently onto the switch package. Near-packaged optics (NPO) mounts them in sockets right next to the package, so a failed engine can be replaced. CPO is denser and shortens the electrical path. NPO is easier to service and uses a more mature form factor. Nvidia is developing both for Rubin Ultra NVL576 scale-up.
Lasers: Lumentum ($LITE). Silicon photonics: TSMC ($TSM). Packaging: ASE/SPIL ($ASX). Laser sub-assembly: TFC ($300394.SZ). Internal fiber and harnesses: Browave ($3163.TWO) and BizLink ($3665.TW). TIAs and drivers for NPO: Semtech ($SMTC). Fiber-array units: FOCI ($3363.TWO), with Largan ($3008.TW) trying to enter.
The CPO earnings the market expects (about NT$97 per share by CY28 in one sell-side estimate) depend on Largan winning share from incumbent FAU suppliers whose yields are improving. The stock already trades around 20x those CY28 earnings. This is research and education, not financial advice.
Yes. Nvidia said on August 14, 2026 that Spectrum-X Ethernet Photonics CPO switches had entered full production, with first shipments to CoreWeave, Lambda and Oracle. The Quantum-X InfiniBand CPO switches started shipping in low volume in early 2026.
Sell-side forecasts have been rising. GF Securities forecast 15,000 in 2026 and 100,000 in 2027 as of August. A September update keeps 100,000 for 2027 but puts 15,000 in 4Q26 alone, followed by 16,000 in 1Q27 and 30,000 in 2Q27.
Co-packaged optics (CPO) bonds the optical engines permanently onto the switch package. Near-packaged optics (NPO) mounts them in sockets right next to the package, so a failed engine can be replaced. CPO is denser and shortens the electrical path. NPO is easier to service and uses a more mature form factor. Nvidia is developing both for Rubin Ultra NVL576 scale-up.
Lasers: Lumentum ($LITE). Silicon photonics: TSMC ($TSM). Packaging: ASE/SPIL ($ASX). Laser sub-assembly: TFC ($300394.SZ). Internal fiber and harnesses: Browave ($3163.TWO) and BizLink ($3665.TW). TIAs and drivers for NPO: Semtech ($SMTC). Fiber-array units: FOCI ($3363.TWO), with Largan ($3008.TW) trying to enter.
The CPO earnings the market expects (about NT$97 per share by CY28 in one sell-side estimate) depend on Largan winning share from incumbent FAU suppliers whose yields are improving. The stock already trades around 20x those CY28 earnings. This is research and education, not financial advice.