CPO's Commercial Year Begins: TSMC COUPE Volume Production and the 200G EML Bottleneck Split 2026 Optical Winners from Losers (W21)

4 hours ago

6 min read

In the same week, TSMC moved CPO from slideware onto a production schedule and Lumentum pulled light-source pricing power into its own hands. CPO's first commercial year hasn't just started; half the winners list is already written.

Introduction

On the afternoon of May 14, at its Taiwan Technology Symposium in Hsinchu, TSMC formally put COUPE (Compact Universal Photonics Engine) on its 2026 volume-production roadmap, and also disclosed that CoWoS 5.5x reticle yield has passed 98% and a 12-HBM configuration is already running on the production line (Focus Taiwan, DIGITIMES). Earlier that week, Lumentum guided FY26 Q3 to $780–830M, up 85% YoY (Lumentum earnings call), while Coherent pushed its D&C share to 75%.

The two threads look independent (TSMC talking packaging, Lumentum talking light sources), but taken together this week was a double starting gun for CPO's commercial year: TSMC has answered "how do you get light into the package," and Lumentum and Coherent have turned "where does the light come from" into their own bargaining chip. CPO will not be a rising-tide story. To take part in the 2026 cycle, you first have to squeeze through these two narrow gates.

TSMC COUPE Volume Production: CPO Goes from Slogan to Scheduled Commitment

Until now, TSMC only talked about "research"

Think back to OFC 2024 and 2025: TSMC's line on silicon photonics was always "we're working on it," with a vague production timeline. COUPE first appeared publicly at the 2024 technology symposium, positioned as "risk production in 2025, production samples in 2026." The market read that as a post-NVIDIA-Rubin event, meaning real customer adoption in 2027.

The May 14 symposium changed that line. TSMC declared COUPE the world's first CPO solution with a 200Gbps microring modulator and pinned volume production to 2026 (Commercial Times (in Chinese)). Alongside it, TSMC disclosed that third-generation CoWoS-S at 5.5x reticle is in production (supporting 12 HBM stacks), a 14x reticle version supporting 24 HBM stacks is coming in 2029, 2nm capacity will grow at a 70% CAGR over 2026–2028, and CoWoS/SoIC capacity will grow more than 80% through 2027.

Put those numbers together: optical engines, HBM stacking and logic process. TSMC laid all three arteries that NVIDIA's Rubin Ultra generation needs on the table at once. Rubin Ultra isn't a 2027 worry; the answers are due in 2026.

Why 200Gbps microring is a watershed

The CPO spec battle has two paths: Mach-Zehnder modulators (the mainstream choice, good signal integrity but area-hungry) and microring modulators (small and power-efficient, but hard to control for thermal drift). NVIDIA Spectrum-X Photonics belongs to the Mach-Zehnder camp, and Broadcom's third-generation 200G/lane CPO platform joined the fight this week; both rely on Mach-Zehnder to secure the first phase.

TSMC's microring bet is a different wager. Running 200Gbps/lane on microrings gives the design freedom of doubling bandwidth per fiber and halving optical connections per GPU. In the Rubin Ultra era, when scale-up domains grow from 72 GPUs to 144 and 288, that directly frees up package area.

The implications for Taiwanese suppliers are concrete: COUPE volume production → ASE's LEAP process benefits directly (ASE earnings call announced a 20% upward revision to 2026 capex), and CoWoS back-end panel-level packaging moves into 310mm fully automated lines. Two-thirds of ASE's capex increase this year is going into new fabs and cleanrooms. That's not a bet on chiplets in general; it's a bet on COUPE plus CoWoS ramping.

Winners and losers

With COUPE in volume production, TSMC becomes the supply-side bottleneck node. It isn't just a process option; it is the de facto standard for CPO in NVIDIA's Rubin Ultra generation. Players who aren't in the COUPE supply chain and haven't secured a spot in TSMC's back-end packaging queue will basically be spectators in the 2026–2027 CPO story.

This also explains why Accton (2345) posted another record month in April, why Luxnet's 800G has entered the Google TPU supply chain, and why PCL Technologies rushed to acquire Pinflex on May 8 to complete its CPO puzzle. Taiwan's optical stocks were already running this narrative this week; TSMC's May 14 symposium just floored the accelerator.

The 200G EML Bottleneck: A Single Point of Failure for 1.6T

Lumentum's +85% guide isn't luck, it's pricing power

On May 8, Lumentum reported FY26 Q2 revenue of $665.5M (+66% YoY) and guided FY26 Q3 to $780–830M, up 85% YoY (Lumentum IR). That far exceeded analysts' prior $720M expectation, and its San Jose InP fab is moving into a major expansion.

Hours later the stock jumped to $1,057, up 17% on the day; AOI rose 24% and Coherent 13% the same day (24/7 Wall St.). That wasn't an ordinary earnings rally. It was the market realizing that light-source supply for 1.6T has a structural constraint.

The key is 200G EML. A 1.6T optical module (8 lanes × 200G PAM4) needs 200G EMLs, and right now Lumentum is the only company in volume production. Coherent showed its 200G InP EML at OFC 2026, but its ramp hasn't arrived. Solutions from China's Accelink and from Sumitomo are still in qualification.

When there's only one supplier and the buyers are Innolight (Q1 +192% YoY), Eoptolink (+106%), AOI, Coherent itself, plus Cisco orders, that bargaining structure decides the fate of optical module BOMs in 1H26. Analysts privately expect Lumentum to raise 200G EML prices by double digits within 2026, with no second source to push back.

Coherent is another answer, but not a substitute

Coherent reported Q3 FY26 revenue of $1.81B this week, +21% YoY, with D&C at 75% (Coherent IR). Its edge isn't a single product line but an IDM structure that controls three light-source paths at once: InP fab, SiPh PIC and GaAs VCSEL. The 400mW high-power CW laser it showed at OFC 2026 is aimed at CPO external laser sources (ELS). That path sidesteps the EML bottleneck, but it also means system designs must change and production comes later.

For buyers, Coherent is "give me the solution for two years from now," and Lumentum is "give me the solution for today." Orders in the 2026 commercial year will go to Lumentum first.

The non-China InP supply chain is the hidden main thread

The non-China InP supply-chain list TrendForce published this week is worth copying in full: CW epitaxy → LandMark Optoelectronics (3081); CW die foundry → Luxnet and Truelight; PD supplier → GCS (Global Communication Semiconductors); PD epitaxy foundry → VPEC (2455); WIN Semiconductors (3105) benefits across both its PA and laser fabs.

At its April 22 earnings call, LandMark announced a five-year InP substrate supply agreement with Sumitomo Electric and an additional NT$1.315B in capex for four more InP MOCVD tools. That agreement effectively locks up non-China InP substrate supply. In an increasingly sensitive geopolitical environment, the value of that lock-up will show up in valuation over the coming quarters.

Read this alongside gallium prices rising another 30% in early 2026 (after a 140% rise in 2025): the compound semiconductor light-source supply chain is becoming a two-factor game of scarcity plus de-risking. Taiwanese suppliers' pricing power comes from stacking process scarcity on top of non-China status.

Other signals worth tracking this week

NVIDIA: FY27 Q1 earnings on May 20, with guidance at $78B; watch Rubin ramp timing, CPO shipment cadence and progress locking up 16-Hi HBM (NVIDIA Newsroom)
Broadcom: launched its third-generation 200G/lane CPO platform, going head-to-head with NVIDIA Spectrum-X Photonics (Broadcom News)
Marvell: at Analyst Day, announced ESUN/UALink sampling in late 2026–2027 and an 800G coherent-lite transceiver at 3.72 pJ/bit (Marvell IR)
Astera Labs: Q1 FY26 revenue of $308.4M (+93% YoY); the Scorpio X-Series 320-lane fabric switch has started shipping, upgrading it from a PCIe retimer player to a rack-scale connectivity platform
Cisco: Q3 FY26 networking revenue +25% YoY to $8.82B; FY26 AI infrastructure order guide raised to $9B (>4x YoY). CEO Chuck Robbins told CNBC the "networking supercycle is only just beginning" (CNBC)
Structural HBM shortage: Q2 2026 DRAM contract prices +58–63% QoQ and NAND +70–75%; the HBM gap is 5.1% (the largest since 2011), and NVIDIA has requested 16-Hi HBM delivery in Q4 2026

STT Take and What to Watch Next Week

This week's takeaway is simple: CPO's commercial year didn't start "in 2026"; it started this week. More importantly, the narrative has a shape. TSMC COUPE volume production defines the winner structure on the system side, and the 200G EML bottleneck defines the bargaining structure on the component side. With both bottleneck nodes tightening at once, players in the 1.6T cycle split into two groups: those with a TSMC back-end packaging slot plus a Lumentum/Coherent light-source allocation, and those without.

Next week, watch NVIDIA's FY27 Q1 call on May 20. Last quarter Jensen already put Blackwell + Rubin at "a cumulative $1 trillion opportunity through 2027" (NVIDIA IR). This time the focus isn't the guide number itself but three details: when Rubin Ultra samples reach hyperscalers, whether CPO uses TSMC COUPE or an in-house design, and which quarter HBM4 16-Hi supply is locked through. Any more specific timeline on these will tighten this week's two narrow gates even further.

The week after brings Computex 2026 and TSMC's North America Technology Symposium, where optical and packaging timelines will be restated. If you can track only one thread, track customer design-in progress on TSMC COUPE: it's the deepest water mark of this cycle.

STT Market Insights | May 18, 2026 | Compiled by the STT editorial team