Before a Laser Reaches a CPO Switch, How Many Times Is It Tested? Who Gets Paid?

@carrioresearch · 2026-10-12 19:09

An indium phosphide laser can't be fully tested while it is still on the wafer. Its two mirrors don't exist yet. They only appear when the wafer is cleaved into bars, which is why a 2005 patent now owned by Coherent ($COHR) states plainly that edge-emitting lasers "cannot be tested in wafer form." So every laser headed for a CPO switch gets tested one at a time, after cleaving: powered up, measured, baked for hours, measured again, and only then mounted in a laser module. Each of those steps takes a machine and a time slot. In Part 1 we looked at whether there will be enough InP lasers for CPO. This piece follows one laser through the five main screens: what each one catches, and why higher power makes each one harder to pass. Then it follows the money, from the laser makers to $FORM and $TER to the companies that build the burn-in racks. [Figure 1: Main process steps and five screening points for a CPO laser] 1. Why a laser can't be tested on the wafer It comes down to where the light comes out. A CPO external light source uses an edge-emitting InP laser: the light leaves through a facet on the side of the chip. That facet is made by cleaving. Lumentum ($LITE) and Meta ($META) described their own high-power CW laser for CPO in a 2022 paper: after thinning, polishing and cleaving, the front facet gets an anti-reflection coating below 0.1% and the back facet a high-reflection coating above 95%. Until that happens, there is no laser to test. A VCSEL emits from its top surface, so it can be probed while the wafer is still whole. That is one of its long-standing advantages over edge emitters. But VCSELs are gallium arsenide parts for short-reach multimode links. The lasers in shipping CPO switches are InP CW lasers that only need to stay on and stay steady; NVIDIA ($NVDA) does the modulation on the silicon photonics engine with micro-ring modulators, and keeps the lasers in pluggable external laser source modules. There is one exception worth knowing. MACOM ($MTSI) makes InP lasers with etched facets, so the mirrors exist before the wafer is cut and the lasers can be tested at wafer scale. It says more than 100 million of them are in the field, and its 75 mW and 100 mW CW lasers are aimed at 1.6T silicon photonics pluggables. [Figure 2: Three laser types compared by mirror formation and wafer-level testability] Here's a question you can carry to any optical part: can it be tested on the wafer? If not, every unit has to pass through a test station on its own, and test capacity has to grow in step with output. 2. Five screens, and what each one catches Laser diodes degrade in three ways, according to a review in Comptes Rendus Physique: rapid, gradual and catastrophic. The five screens are built around catching the first and third before a laser ships, and putting a number on the second. [Figure 3: Five screening points for an InP CPO laser: screen 1, epi wafer inspection] Screen 1: the epitaxial wafer. Rapid degradation usually shows up in the first hundred hours of operation, as threshold current climbs quickly. The cause is crystal defects that grow under current and heat into dark lines and dark spots, eventually destroying the active region. Those defects start in the substrate and the epitaxial layers, so the first check happens before any laser is patterned: thickness, composition and defects across the wafer. InP has an edge here. NASA data show defects grow much more slowly in InGaAsP/InP than in AlGaAs/GaAs, though dark-spot defects are still common in degraded InP lasers. [Figure 4: Five screening points for an InP CPO laser: screen 2, bar test] Screen 2: the bar test. Right after cleaving and coating, each laser is powered for the first time and swept for its LIV curve: light output against current and voltage, threshold current, slope efficiency and peak wavelength. Chroma's 58620 system is built for this on edge-emitting lasers, ficonTEC sells high-power bar and chip-on-submount testers that measure whole bars in parallel, and MPI's VEGA probers handle edge emitters at the wafer, package and die level. This screen also has to deal with the most sudden failure. Catastrophic optical mirror damage happens when the facet melts from current concentration and light absorption, and it is the main limit on how much power a laser diode can put out. The Jiménez review calls it silent and hard to screen. InP starts with a big margin: NASA puts the damage threshold for InGaAs(P)/InP at tens of MW/cm2, against less than 1 MW/cm2 for AlGaAs/GaAs. CPO eats into that margin. $AAOI says its CPO laser runs at 300 mW against 100 mW for a 1.6T pluggable, on a chip about six times the size. Lumentum's CPO family goes up to 400 mW. [Figure 5: Five screening points for an InP CPO laser: screen 3, burn-in] Screen 3: burn-in. This is the heaviest of the five. Lasers are mounted on carriers and run at high temperature and high current for hours, so the weak ones fail in the factory instead of in a data center. An ILX Lightwave white paper puts typical burn-in at under 100 hours; one published GR-468 qualification recipe runs lasers at 85°C for 48 hours. Newport's guidance is to reject parts whose threshold current, power or slope efficiency shifts by 5% to 30%. Burn-in is where machines and hours pile up. Every laser occupies a slot for the whole run, so the slot count sets the ceiling on output. Chroma's 58604 burn-in racks drive up to 500 mA per channel at up to 125°C, and its 58605-HD drives up to 2,500 mA for high-power parts. In its July 30 earnings deck, Chroma showed a new 58604-HS laser burn-in system in its CPO test flow, with up to 2,688 channels per system at up to 1 A per channel. Suzhou Lianxun Instruments (SSE: 688808) lists a chip-on-carrier laser burn-in system in its first-half report and says shipments to leading international customers keep growing. [Figure 6: Five screening points for an InP CPO laser: screen 4, spectral test] Screen 4: the spectrum. After burn-in, each laser is checked for wavelength and side-mode suppression: does it lase at the right wavelength, in a single mode? Yokogawa's AQ6370E optical spectrum analyzer has a dedicated SMSR mode for exactly this. Broadcom's ($AVGO) CPO light module carries eight CWDM DFB lasers, each on its own wavelength, and this screen gets tighter as the spacing narrows. $AAOI said in August that for next-generation DWDM external light sources, yield loss "will be easily 40%-50%" compared with 1310 nm lasers. [Figure 7: Five screening points for an InP CPO laser: screen 5, reliability qualification] Screen 5: reliability qualification. This one isn't run on every laser. When a new product is introduced, samples go through Telcordia GR-468 accelerated aging: at least 2,000 hours at 70°C or 85°C, with testing continued to 5,000 hours for information. Damp heat at 85°C and 85% humidity runs 1,000 hours. The output of all this is a single number. Lumentum says its ultra-high-power laser fails at 20 FIT at 400 mW, meaning 20 failures per billion device-hours, measured on thousands of production lasers that went through standard burn-in. An NVIDIA Quantum-X Photonics switch has 18 external laser source modules with 8 lasers each, or 144 lasers. At 20 FIT, 144 lasers running all year come to about 0.025 failures per switch per year, or roughly one laser failure for every 40 switches a year. That is why NVIDIA put the lasers in field-replaceable modules: a failed laser gets swapped, not the switch. A 20 FIT laser doesn't come straight off the line; it is what's left after screens 1 to 4. There is a second path. When the laser is grown or bonded onto the silicon photonics wafer itself, it can't be burned in one by one; the whole wafer is burned in. That is Aehr Test's FOX-XP. In August, Aehr announced another order from its lead silicon photonics customer and said that "identifying latent defects and stabilization of the integrated lasers before packaging improves manufacturing yields." External laser sources follow the traditional route of laser-by-laser burn-in after cleaving; integrated lasers go through wafer-level burn-in. 3. Who gets paid The first layer is the laser makers themselves, since screening capacity sits inside their own plants. Coherent was clear in August that test is not its bottleneck: "indium phosphide capacity continues to be our primary constraint... we're not constrained in the assembly and test capacity right now." That also tells you where the next orders go. Coherent's internal InP output was set to double year over year by September and more than double again by the end of 2027, and every new wafer has to be matched by test and burn-in slots. Lumentum's proxy statement, filed October 6, says it "rapidly scaled assembly and test capacity, both internally and through strategic partnerships with contract manufacturers." $LITE $COHR $AAOI are buying this capacity. The second layer is the test-platform makers, $FORM and $TER. One distinction matters here. In CPO, both mainly test silicon photonics wafers and optical engines, not the burn-in of individual InP lasers. Laser screening and optical engine test are two separate pools of money, and both grow with CPO. $FORM's CPO product is Triton, developed with Advantest and TEL. It tests silicon photonics wafers from one side and handles both edge coupling and surface coupling. On IHP's 200 mm wafers, FormFactor measured coupling loss as low as 1.51 dB per facet, better than cleaved fiber arrays. In the June quarter, Systems revenue rose 44.1% to $48.5 million, "primarily driven by sales of Triton." In December FormFactor bought Keystone Photonics for $20.6 million net of cash, whose optical probes are used to test silicon photonics and CPO devices. HBM is the other engine: DRAM probe card revenue was $85.0 million, up 49.0% and 32.9% of total revenue, and HBM accounted for about 75% of the increase. Guidance for the September quarter is $270 million, plus or minus $10 million. $TER is first a compute test company. Memory was $212.3 million in the June quarter, 18.9% of semiconductor test revenue. On the July call it said "CPO alone will be a 300 to 700 million market by 2028," with the low end of 2027 "in the $200 million range." It bought PIC test specialist Quantifi Photonics in May 2025 for $127.2 million, and its Photon 100, built on UltraFLEXplus, covers wafer, optical engine and co-packaged module test. Management also says that after more than a decade of fab equipment outgrowing test, the trend reversed in 2024 and test is now outpacing fab equipment. September-quarter guidance is $1.2 billion to $1.3 billion. [Figure 8: FormFactor and Teradyne compared on CPO products, latest quarter and valuation] At Friday's close, $TER traded at 34.6 times 2027 consensus earnings with 21.8% expected revenue growth, and $FORM at 36.9 times with 17.5%. HBM is today's business for both; optical test is the next leg. The third layer is the burn-in and laser test equipment makers, the closest to the laser itself. Chroma (2360 TT) has the broadest laser burn-in and LIV lineup, and its new CPO burn-in system is already in its earnings deck. Its semiconductor and photonics test segment was NT$4.39 billion in the June quarter, 39% of sales and up 99% year over year; photonics isn't broken out separately. Lianxun Instruments sells chip-on-carrier laser burn-in, laser chip sorting and silicon photonics wafer test systems, with first-half revenue up 208.7%. ficonTEC, a wholly owned subsidiary of Robo-Technik (SZSE: 300757), makes bar and chip testers and the assembly equipment around them. $AEHR owns the wafer-level path: fiscal 2026 revenue was $50.0 million, effective backlog $100.6 million, and its fiscal 2027 guidance is $130 million to $150 million. [Figure 9: Burn-in and test equipment makers in the InP laser chain, with evidence in their own filings] My call: each CPO laser carries three to four times the power of a pluggable laser and feeds several channels at once, so each one is worth more screening. Through 2027 and 2028, as InP output doubles and then doubles again, burn-in and test capacity has to be added alongside every new wafer. On the optical engine side, that favors $FORM and $TER. Closer to the laser, Chroma and Lianxun Instruments are the two names with direct burn-in products in their own filings. If lasers move onto the silicon wafer, that's $AEHR's business. What I'm watching next FormFactor's third-quarter report on October 28: whether Systems keeps growing on Triton, and what it says about optical probing. AXT's third-quarter report on October 29: the substrate side of the same chain. Aehr's first-quarter report on November 2, the first quarter under its new June fiscal year: follow-on orders from the lead silicon photonics customer, and any mention of integrated lasers. Teradyne's third-quarter report, date not yet announced: whether the $200 million low end for 2027 CPO test still holds. Chroma's next earnings call: whether the 58604-HS shows up as orders. Lumentum and Coherent's next calls: whether high-power laser shipments catch up, and whether test capacity gets its own mention. OFC in March 2027: whether etched-facet lasers move up into the CPO power class, which would bring wafer-level testing to the CPO laser itself. In Part 3 we follow the light out of the laser and into the fiber: fiber arrays and alignment. Stay tuned.

曝光 3318 · 评论 2 · 点赞 8 · 书签 6 · 曝光/时 294.20135951132534