00What is optical interconnect and why AI needs it

Inside an AI accelerator rack, electrical copper is running out of road. At the bandwidth densities Rubin-generation packages require, moving bits with electrons turns into a wall of heat and signal loss. Optical interconnect is the way through.

Optical interconnect is a photonic substitute for electrical wires. A modulator encodes bits onto a beam of light, the beam travels through a glass fibre (or a silicon waveguide etched into a chip), and a photodetector at the far end converts light back to electrical signal. The same job an electrical wire does, done with light instead of electrons. What optical interconnect buys is bandwidth density per millimetre of package edge and power efficiency per bit transmitted. Both matter more the further the signal has to travel and the higher the data rate.

Why AI infrastructure specifically needs it. A modern Nvidia rack (Vera Rubin NVL72) delivers hundreds of petaflops of aggregated compute across roughly 72 GPUs bound together by an interconnect fabric. That fabric has to move terabits per second between GPUs to keep the collective busy. Copper interconnect can hold those bit rates over short distances. The current Vera Rubin NVL72 rack uses passive copper twinax cartridges to bind GPUs together inside a single rack. Its power cost scales badly with data rate and its reach collapses at the higher bandwidths coming with Rubin Ultra in 2027-2028; scale-up traffic beyond the current rack range passes the passive-copper limit entirely. At a certain threshold the electrical interconnect becomes economically difficult to justify against fibre. Switch-level CPO is already in production. Broadcom shipped Tomahawk 5 CPO at 51.2 Tb/s aggregate switch bandwidth in 2023-2024, and Tomahawk 6 CPO at 102.4 Tb/s in October 2025. Both use TSMC COUPE optical engines. Accelerator-side CPO arrives in 2027 with the CPO NVSwitch in Nvidia's Rubin Ultra NVL576 rack. Broader adoption comes when electrical interconnect stops meeting the bandwidth-and-reach requirement at competitive power. The adoption picture across use cases:

CPO adoption across AI infrastructure · 2020 through 2030 Each row is a distinct use case for optics inside an AI datacentre. Bands show where CPO adoption is production (dark), qualification (mid), or early / roadmap (light). Named products anchor the milestones. 2020 2022 2024 2026 2028 2030 TODAY Pluggable optics (400G, 800G, 1.6T faceplate) Volume across every generation Scale-out CPO switch (Ethernet, InfiniBand) Broadcom Humboldt Gen 1 Tomahawk 5 CPO 51.2 Tb/s Tomahawk 6 CPO 102.4 Tb/s, Nvidia Spectrum-6 Scale-up CPO NVSwitch (rack-scale fabric) Established positions Accelerator-adjacent CPO (Passage L200, hyperscaler custom silicon) L200 launch + dev kits Hyperscaler qualification In-package optical I/O (TeraPHY optical chiplet, Passage M1000) Ayar shipments; Passage M1000 UCIe-O + qualification Production Qualification Early / roadmap Broadcom owns the scale-out row. Nvidia in-house is the largest single risk to the two scale-up rows.

How the industry got to CPO. The evolution is a progressive move toward the compute die. Pluggable optics puts the modulator, laser, and detector inside a hot-swappable module on the switch faceplate, connected to the switch silicon over long electrical traces. This has been the dominant approach through the 400G and 800G generations. Near-package optics (NPO) moves the optical engine to a board position immediately adjacent to the switch or accelerator. Co-packaged optics (CPO) integrates the optical engine directly into the switch or accelerator package, sitting millimetres from the compute die. In-package optical I/O (Ayar Labs TeraPHY) or optical interposers (Lightmatter Passage) are the current versions of CPO. Each successive step shortens the electrical path, reduces power, and increases bandwidth density, while increasing packaging complexity and yield risk.

How the industry got to CPO · progressive integration of optics into the compute package Left to right: successive generations bringing optics closer to the compute die. Each step shortens the electrical path and reduces interconnect power; each step increases packaging complexity and yield risk. Pluggable SUBSTRATE COMPUTE DIE PLUGGABLE SerDes trace Optics at faceplate 8-16 SerDes per lane Near-package (NPO) SUBSTRATE COMPUTE DIE NPO Optics adjacent, on board CPO (Ayar chiplet) SUBSTRATE COMPUTE DIE TeraPHY UCIe optical TeraPHY chiplet on substrate UCIe-compliant CPO (Lightmatter interposer) SUBSTRATE COMPUTE DIE Passage M1000 interposer Passage M1000 under die 34 chiplets, 114 Tbps

Where the money is. The datacom optical interconnect market is roughly $14 billion in 2024 and forecast to grow to over $140 billion by 2030 per Yole and adjacent industry estimates. Silicon photonics accounts for approximately two-thirds of the projected 2030 total. CPO specifically is a smaller slice, roughly $15 billion by 2030 per Bank of America, and the scale-up subset that Ayar and Lightmatter primarily address is smaller still, on the order of $6 billion by 2030. The addressable market stays modest relative to the capital already invested; the returns come from the multiple that a strategic acquirer or public market places on a specific supplier position, not from the raw scale of the addressable market alone.

Who competes. Six players across scale-out and scale-up CPO. The two venture pure plays this essay focuses on operate in scale-up CPO, where incumbent positions are less locked in. Full analysis.

The competitive field · who plays whereSix players across scale-out (network switches) and scale-up (accelerator fabric) CPO. Ayar Labs and Lightmatter are the two venture-stage pure plays; the other four are incumbents with adjacent positions.Legend · ▲ scale-out CPO · ● scale-up CPO · ◆ both segmentsBroadcom▲ SCALE-OUT CPO · INCUMBENTTomahawk 6 CPO Ethernet switch shippingproduction since October 2025.READOwns the scale-out row. Not a directcompetitor to Ayar/Lightmatter, but sets thereference for what a mature CPO businesslooks like.Marvell▲ ADJACENT · COHERENT DSP OWNERCoherent optical DSP franchise after 2021Inphi acquisition ($10B, 13.3x sales).READOptical + DSP integration is a comparableexit template. Sets the multiple ceiling fora mature strategic-acquisition scenario.Coherent Corp◆ VERTICALLY INTEGRATED · MERGED ENTITYLasers, optical components, and coherent modulesunder one roof. II-VI + Coherent Inc. merger, 2022.READLightmatter's laser integration play is apartial insourcing of what Coherent sells asa component. Precedent for the verticalshape.Nvidia● SCALE-UP CPO · BIGGEST SINGLE RISKInternal optical capability since 2020Mellanox acquisition. Rubin Ultra scale-upCPO position undisclosed.READThe single most consequential open question.In-sourcing collapses Ayar and LightmatterTAM. Commercial partnership doubles both.FOCUS OF THIS ESSAYAyar Labs● SCALE-UP CPO · VENTURE PURE PLAYUCIe-compliant TeraPHY chiplet.GlobalFoundries 45nm monolithic. Nvidia +Intel + AMD Ventures on cap table.READStrategic-acquisition geometry.Highest-probability exit is chipletintegration by a compute-die owner or OSAT.FOCUS OF THIS ESSAYLightmatter● SCALE-UP CPO · VENTURE PURE PLAYPassage wafer-scale interposer + Guide lightengine. GF Fotonix. Amkor packaging partner.READStandalone or IPO geometry. Broader stackcomplicates strategic bidder, larger blue-skyif wafer-scale becomes a platform.Full analysis of each competitor. The two venture pure plays operate primarily in scale-up CPO, where incumbent positions are less locked in.

01The value creation engine

Co-packaged optics ("CPO") is the migration of the optical interconnect from a pluggable module on the switch faceplate or the accelerator's edge to a die or interposer placed inside the same package as the compute silicon. CPO matters economically because copper interconnect power and reach both collapse at the current densities and bandwidths that AI accelerators require past 2026-2027. Rubin-generation Nvidia scale-up traffic, at more than 200 Tbps per package, is the point at which the power and reach penalty of electrical interconnect becomes economically difficult to justify. CPO is the substitute.

Central tension

CPO can win the technology transition and Ayar Labs and Lightmatter can still lose the economics.

Optics becoming mandatory in AI infrastructure is a physics call and increasingly a consensus one. Which company captures the economics when that transition arrives is a structural call about supplier architecture, Nvidia's in-house build-out, and the packaging leverage that a chiplet integrator or hyperscaler can exert on a merchant photonics vendor. The layers need to clear for either equity story to work: CPO adoption at the architectural level, and supplier value capture at the commercial level. These are separate probability calls with different dependencies.

The value creation engine for a pure-play silicon-photonics vendor is a sequential three-step process. Only revenue at step 1's third milestone (production volume) unlocks the compounding in step 2 and the IP leverage in step 3.

The value creation engine · three-step sequential process for a pure-play silicon-photonics vendorEach step is conditional on the prior step. Step 1 has three distinct milestones inside it. Only revenue at the third milestone (production volume) unlocks steps 2 and 3.STEP 1Design-in and qualification18-24 monthsThe vendor engages with a hyperscaler or Nvidia over an18-24 month cycle. Qualification does not guaranteeproduction allocation. Three distinct milestones:1QualifiedDev-kit access, technical validation,sample supply agreements. Vendor is onthe RFQ list.2SelectedChosen for a specific platformgeneration. Board decision but no volumecommitment.3Production volumeUnits ship. Revenue starts. The path tosteps 2 and 3.STEP 2Position compounds across refreshesEvery subsequent GPU or switch generation inheritsthe interconnect architecture. Revenue per unitrecurs across platform refreshes.Gen NInitial platformGen N+1Refresh inheritsGen N+2Second refreshREVENUE COMPOUNDS ACROSS GENERATIONS3-4 yearstypical platform life per generation1 design-in win= multi-generation revenue streamSubstitution costapproaches full package redesignThis is the “package-lock” economics that gives a merchant vendor a defensible commercial position.STEP 3Patent + licensing leverageIf the vendor owns patents on the specific packagingor fibre-attach approach the platform adopts, licensingrevenue tracks unit volume without new manufacturing.TRACKSUnit volumeof the design-in siliconacross the platform lifeWITHOUTNew capexno factory build, no fabexpansion, no manufacturingAyar: ~170 patents (53 granted). UCIe-O.Lightmatter: multi-domain, focus on interposer + laser + packaging.The IP economics only work if the vendor is at step 3, not merely qualified.MARKET PRECEDENTMarvell paid ~$10B (13.3x EV/Sales) for Inphi in April 2021 to capture this compounding effect at the coherent-DSP layer of 400G/800G pluggables.CPO applies the same pattern one architectural layer inward.

The failure works in the opposite direction. If the platform's interconnect architecture shifts before the vendor recovers qualification costs, the economic return on the qualification investment can collapse before sufficient volume accrues to recover it. IP, engineering knowledge, and adjacent customer relationships often survive; the specific design-in revenue stream typically does not. Rockley Photonics filed for Chapter 11 in January 2023 after its silicon-photonics wearable-sensor platform failed to reach commercial deployment inside the 18-24 month window it needed to service its SPAC-vintage debt.

The CPO transmission mechanism · from physics to exit multiple Sequential stages. Each holds an independent probability. Rubin Ultra CPO NVSwitch qualification in H2 2027 is the earliest observable link that would re-rate every downstream stage. Colour progression darkens across the chain from step 1 to step 8. 1 Rubin-era package bandwidth >200 TBPS / PACKAGE Accelerator I/O demand exceeds what copper interconnect can handle at the required distance within the available power range. 2 Copper reach becomes binding CENTIMETRES, NOT METRES Serdes power grows non-linearly with reach. Beyond package boundary at 200 Tbps+, copper is uneconomic. 3 Pluggable to CPO transition H2 2027 ONWARDS Optical engines move into the package. Two viable roadmaps: chiplet-integrated (Ayar) or interposer-scale (Lightmatter). 4 Qualification and design-in 18 TO 24 MONTHS Ayar or Lightmatter qualifies on Nvidia / AMD / hyperscaler custom silicon. Qualification does not guarantee volume. 5 Production allocation at volume RUBIN ULTRA RAMP Selected vendor secures unit volume across a platform generation. Locks in for the platform life (three to four years). 6 Module and IP licensing leverage ARCHITECTURAL LOCK-IN Revenue compounds: module + interposer content per unit, plus IP licensing on packaging or fibre-attach approach. 7 Gross margin expansion TARGET 60 TO 70% IP-plus-hardware mix trends toward high-margin semi IP economics from low-margin optical assembly. 8 Exit multiple re-rating 5-8X TO 15-20X EV/SALES Public and strategic-acquirer multiples move from specialty photonics (Coherent-like) toward IP-plus-hardware (Inphi-like). Every step is conditional on the prior step. Probability compounds downward. Marvell paid ~$10B (13.3x) for Inphi in 2021 to capture this pattern at the coherent-DSP layer of 400G/800G pluggables.

Each stage holds an independent probability. The Rubin Ultra CPO NVSwitch qualification in H2 2027 is the earliest observable link in the chain that would re-rate the implied probability of every downstream step.

The essay covers both Ayar Labs and Lightmatter because they are the two venture-stage pure plays with Nvidia relationships and plausible routes to volume production that hold the most weight for the paired-position analysis. Broadcom, Marvell and Coherent Corp. also have silicon-photonics positions, but as public platform companies they hold different valuation dynamics and are already priced for their optical exposure. Ayar and Lightmatter are the two names where the CPO thesis remains investable on private-market terms.

The paired thesis does not require both companies to win. They compete for structurally different exits. Ayar's IP concentrates at the chiplet layer and its cap table concentrates strategic acquirers (Nvidia, AMD, Intel, GlobalFoundries) that make it a strategic-acquisition candidate. Lightmatter's stack is broader (interposer, CPO engine, laser source) and its cap table concentrates public-market crossovers (T. Rowe Price, Fidelity) plus a Marvell board seat, which makes it the more credible IPO. The two companies can reach similar values through different exits. Underwrite them as a pair with different exit distributions; a winner-take-all race would be the alternative framing.

CPO within the optical interconnect market · segmentation by 2030 Cascade from total DC optical spend down to the Ayar and Lightmatter addressable slice. Sources: Yole (DC optical total), BofA (CPO segment). Scale-up CPO is a ~$6B segment by 2030. Nvidia in-sourcing risk determines remains for Ayar and Lightmatter. DC optical interconnect Total DC optical market, 2030 $144B Coherent · Marvell · Broadcom · Nvidia · others Silicon photonics ~66% of DC optical (Yole) $96B Broadcom + TSMC COUPE · GlobalFoundries · Intel · Amkor · others CPO (all) ~10% of DC optical (BofA) $15B Scale-out: Broadcom Tomahawk 6 · Scale-up: Nvidia + partners CPO scale-up (Ayar/LM addressable) ~4% of DC optical $6B Nvidia in-house · Ayar Labs · Lightmatter · others Ayar and Lightmatter compete for share of the ~$6B scale-up CPO slice by 2030.

02The technology moat

Bandwidth generation ladder · when CPO becomes economically necessary Two rows show pluggable-optics feasibility and CPO adoption at each per-lane rate. LPO extends pluggable life through 800G and 1.6T. CPO becomes economically necessary at 3.2T and above. Pluggable optics feasibility CPO adoption state 400G 2020 Standard pluggable Early demos 800G 2023 Standard pluggable Broadcom TH5 1.6T 2025 Pluggable + LPO Broadcom TH6 · TeraPHY 3.2T 2027 Not economic Passage L200 · Rubin Ultra 6.4T 2029 Not economic CPO only CPO economically necessary

Silicon photonics as a manufacturing platform is roughly a decade old at scale. The important change in 2025-2026 is the move from board-level pluggable transceivers to in-package optical engines placed within millimetres of the compute die. That migration is what "CPO" refers to when the term is used precisely. It requires three things that a pluggable transceiver does not.

  1. Foundry-qualified silicon-photonics process node. Both companies use GlobalFoundries. Ayar runs on GF's monolithic 45 nm CMOS-plus-photonics process. Lightmatter uses GF's newer Fotonix platform, which integrates photonic components with CMOS logic in a single die. Monolithic (Ayar) reduces packaging complexity but caps performance at the older node. Fotonix (Lightmatter) preserves access to newer logic nodes at the cost of a more complex interposer.
  2. Manufacturable light source. Silicon does not lase efficiently, so external DFB laser arrays must be attached or aligned. Ayar's SuperNova is a remote CW-WDM light source sourced from Sivers Photonics. Lightmatter's Guide light engine (Jan 2026, Very Large Scale Photonics platform) integrates hundreds of lasers on a single chip and moves laser manufacturing from manual assembly to foundry production. The Guide DR variant (May 2026) adds liquid cooling and claims 4x rack density.
  3. Package integration path. The optical engine has to survive board-level assembly, thermal cycling, and vibration alongside a compute die dissipating 1 kW-plus. Amkor (packaging) and ASE (test) partner with Lightmatter on a substrate configuration with no volume precedent. Ayar positions TeraPHY as a UCIe-compliant chiplet, which puts the packaging burden on whoever integrates the chiplet (Nvidia, AMD, an OSAT, or a hyperscaler ODM).

Three additional problems the topology diagram does not show:

  1. Yield liability under the die. A failure of the optical engine, fibre attach, or laser source scraps the entire compute assembly. Known-Good-Die testing for optical chiplets remains incomplete industry-wide. Under Ayar's chiplet model, yield-loss allocation falls to whoever integrates the chiplet. Under Lightmatter's interposer model, liability falls to Amkor and Lightmatter jointly.
  2. Fibre-attach and laser reliability. Fibres exit the package alongside a 1 kW die. Neither company has demonstrated the million-hour link reliability figure that Meta disclosed for Broadcom Tomahawk 6 in October 2025. Remote-light-source architectures (Ayar SuperNova, Lightmatter Guide) make lasers field-replaceable, with additional optical-interface complexity at the package boundary.
  3. Standards portability. UCIe defines the electrical die-to-die chiplet interface. Its optical extension (UCIe-O) is what Ayar has bet on and stays unratified. If Nvidia or Broadcom drives a different optical die-to-die specification, Ayar's chiplet position fails to port cleanly. Lightmatter's interposer approach is less exposed to this specific risk but more exposed to substrate decisions made by the hyperscalers.
Where CPO breaks · three failure modes under the die Schematic of the in-package CPO assembly with the three yield / reliability failure modes marked. Each holds different remediation cost for Ayar (chiplet) versus Lightmatter (interposer). IN-PACKAGE CPO ASSEMBLY Package substrate / interposer Compute die 1 kW+ dissipation Optical chiplet Optical chiplet Fibre Fibre Remote laser Blind-mate connector 1 2 3 1 Known-Good-Die gap A wafer probe cannot fully validate an optical die. Industry-wide KGD testing for optical chiplets remains incomplete. AYAR Chiplet integrator handles yield loss LIGHTMATTER Amkor + Lightmatter joint liability 2 Fibre-attach at thermal boundary Fibres exit the package alongside a 1 kW die. Post-package attach risks thermal cycling damage; pre-package attach risks pigtail damage. AYAR No hyperscaler-volume precedent LIGHTMATTER Passage substrate has no volume precedent 3 Laser lifetime and hot-swap DFB lasers degrade. In-package failure without redundancy takes down the placement. Remote laser (SuperNova, Guide) makes replacement possible. AYAR SuperNova: remote CW-WDM from Sivers LIGHTMATTER Guide: hundreds of lasers on chip, foundry produced Neither Ayar nor Lightmatter has demonstrated a million-link-hour reliability figure. Meta disclosed that number in October 2025 for the Broadcom Tomahawk 6 switch only.

The moat is package-level integration commitment. GlobalFoundries and adjacent fabs sell silicon photonics to anyone. DFB laser arrays are a merchant category, sourced from Sivers Photonics and a small number of other vendors. Value accrues to whichever supplier is already designed into a hyperscaler compute stack or a specific Nvidia platform. Once a hyperscaler has committed its compute die stack to Lightmatter Passage, or a chiplet substrate to Ayar's TeraPHY optical die, substitution cost approaches a full package redesign and, depending on the architecture, a die respin. Both companies entered Nvidia's NVLink Fusion ecosystem on June 2, 2026. Rubin Ultra qualification and production supplier allocation remain undisclosed as of September 2026.

Where CPO plugs into Nvidia Vera Rubin · scale-out versus scale-up The Rubin platform announced at CES 2026 uses CPO in two distinct places. Broadcom owns scale-out (network fabric). Ayar and Lightmatter compete for scale-up (rack-scale accelerator fabric and accelerator-adjacent I/O). SCALE-OUT · NETWORK FABRIC Broadcom incumbent · Ayar/LM absent Nvidia Spectrum-6 Ethernet CPO WITH 200G SERDES · ANNOUNCED CES 2026 Nvidia's scale-out CPO play. Deployed alongside Broadcom silicon in most datacentre stacks. READ Not addressable by Ayar or Lightmatter at 102 Tb/s. Broadcom Tomahawk 6 CPO 102.4 TB/S · SHIPPING SINCE OCTOBER 2025 Bailly CPO ran ~1M port-device hours with zero flaps link flap. Scale-out CPO benchmark. READ Broadcom scale-out lead is defensible through 2028+. SCALE-UP · RACK-SCALE ACCELERATOR FABRIC + XPU I/O Ayar + Lightmatter opportunity NVL576 (Kyber) CPO NVSwitch RACK-SCALE FABRIC · H2 2027 576 GPUs per rack via CPO NVSwitch. Supplier allocation is the single most consequential disclosure event in the paired thesis. READ Ayar, Lightmatter, and Nvidia in-house all in play. Rubin Ultra XPU · in-package optical I/O ACCELERATOR-ADJACENT · H2 2027+ Optical die-to-die interface for accelerator packages. TeraPHY (Ayar UCIe-O chiplet) or Passage (Lightmatter interposer). READ Both entered NVLink Fusion on June 2, 2026. The scale-out row is priced. The scale-up row stays unpriced. Neither Ayar nor Lightmatter has disclosed a scale-out product that competes with Tomahawk 6 at 102.4 Tb/s.

03Ayar Labs · chiplet-level optical I/O

Ayar Labs was founded in 2015 as a joint MIT-Berkeley spinout from the DARPA POEM programme (Photonically Optimized Embedded Microprocessors), led by Vladimir Stojanovic. Mark Wade is CEO. The company is headquartered in San Jose with approximately 255 employees as of July 2026, engineering across Boston, Toronto, Hsinchu, and a new Bengaluru Design Center announced in September 2026.

Funding. Total capital raised is ~$870 million as of September 10, 2026, of which $650 million came in 2026 alone. The Series C (April 2022, $130M) pulled strategic checks from Nvidia, Intel Capital, HPE, AMD Ventures, and GlobalFoundries. The Series D (December 2024, $155M) closed above a $1B valuation. The Series E ran in two tranches: $500M in March 2026 at a $3.75B post-money led by Neuberger Berman, followed by $150M on September 10, 2026 alongside a $225M secondary led by Antero Peak Group at Artisan Partners at pricing that implied a valuation above $5B (per SiliconAngle 10 September 2026).

Strategic backers named at the September 2026 extension. AMD, Alchip Technologies, Intel, MediaTek, Nvidia, and Wiwynn. Wiwynn (Taiwan hyperscale ODM, Wistron subsidiary) joined via strategic investment in the extension round, signalling ODM-side commitment to CPO qualification for hyperscale rack integration alongside the existing chip and IP backers.[HPCwire]

Use of proceeds. Ayar earmarked the extension for the transition to high-volume manufacturing (product development, validation, ecosystem scaling with GlobalFoundries, Amkor, and ASE) and for the opening of a new Bengaluru Design Center in India that complements the existing San Jose, Boston, Toronto, and Hsinchu engineering footprint.[Lightwave]

Ayar Labs product stack · TeraPHY chiplet + SuperNova remote light source Two-part architecture: UCIe-compliant optical chiplet co-packaged with the compute die, fed by an external multi-wavelength CW light source. Foundry: GlobalFoundries 45 nm monolithic CMOS-plus-photonics. SuperNova CW-WDM light source 16 wavelengths CW-WDM MSA compliant 16 fibres One wavelength per fibre 16 Tbps total capacity Feeds up to 256 data channels Sivers DFB laser supplier Merchant array sourced externally Blind-mate connector Field-serviceable optical interface TeraPHY optical chiplet UCIe-compliant · co-packaged with compute die 8 Tbps Bidirectional bandwidth 4 Tbps per direction 8 Optical ports 512 Gbps per port 16 Wavelengths (WDM) 32 Gbps per wavelength ~10 ns End-to-end latency Pluggable-comparable <1e-12 Bit-error rate Uncorrected on-chip 4-8x Power efficiency vs pluggable + SerDes Cumulative shipments: ~15,000 units to select customers through late 2024. High-volume production readiness expected mid-2026. 2026 run-rates not disclosed.

Manufacturing. TeraPHY runs on GlobalFoundries' 45 nm monolithic CMOS-plus-photonics platform, partnership dating to 2015. As host compute dies move to TSMC N3P / N2 with 224G-class SerDes, matching those rates on 45 nm forces disproportionate silicon area for digital control and equalisation, which erodes part of the pJ/bit advantage on the newest nodes. SuperNova uses DFB laser arrays sourced from Sivers Photonics. Cumulative shipments were approximately 15,000 units to select customers through late 2024, with high-volume production readiness expected mid-2026. Specific 2026 run-rates have not been disclosed.

Commercial signals. Named partnerships include Intel (co-packaged photonics tile with FPGAs), HPE (next-generation interconnect on top of Slingshot), and Nvidia NVLink Fusion ecosystem membership (June 2, 2026). Ecosystem alignment and production supplier allocation measure different things. UCIe consortium participation alongside Intel, AMD, Broadcom, Micron, MediaTek and GUC anchors Ayar's chiplet standards position. No named hyperscaler supplier allocation for Rubin Ultra had been disclosed as of early September 2026.

Patents. Approximately 170 patents globally per GreyB Insights, of which 53 have been granted. The most-cited patent (US20180019139A1) has received 17 citations from parties including SanDisk, Disco Corporation, and Inphi. Coverage spans optical chiplet architectures, laser integration, and packaging methods.

Governance and the Antero Peak read. Board additions in January 2024 included Ganesh Moorthy (former Microchip CEO, senior Intel history) and Craig Barratt (former Atheros CEO). Vivek Gupta joined as CSO in October 2025 from Google Cloud, and Sejal Patel Daswani as CPO in June 2026. The September 2026 secondary was led by Antero Peak, a public-equity crossover fund at Artisan Partners. That is a legitimate signal one investor is underwriting a public-market path, but not evidence that an IPO is imminent or that the holding period is defined.

Ayar Labs snapshot

Founded 2015 (MIT/Berkeley, DARPA POEM heritage). Headcount ~255. Capital raised ~$870M (of which $650M in 2026 alone). Latest valuation approximately $5.0B implied by September 10, 2026 secondary. Primary product TeraPHY optical chiplet + SuperNova remote light source. Fab GlobalFoundries 45nm monolithic. Strategic investors Nvidia, AMD Ventures, Intel Capital, HPE, GlobalFoundries. Nvidia relationship NVLink Fusion ecosystem partner (Jun 2, 2026). CEO Mark Wade (co-founder). Public-market signal Antero Peak / Artisan Partners crossover in Sep 2026 secondary.

04Lightmatter · photonic interposer and light-engine stack

Lightmatter was founded in 2017 as an MIT spinout by Nicholas Harris (CEO, PhD EECS/photonics MIT 2017), Darius Bunandar (PhD MIT Physics), and Thomas Graham (MBA MIT Sloan, prior at Google and Google X). The company won the $100K MIT Entrepreneurship Competition in 2017. Headquarters moved from Boston to Mountain View in 2023, with additional offices in Boston, Toronto, and Hsinchu. Headcount as of July 2026 is 363.

Funding. Total capital raised is ~$850 million as of the October 2024 Series D. No subsequent priced round had been disclosed as of September 2026. The Series C ran in two tranches: $154M in May 2023 led by GV, and $155M in December 2023 at a $1.2B post-money. The Series D (October 2024, $400M) closed at a $4.4B post-money led by T. Rowe Price with participation from Fidelity and GV, a roughly 4x step-up from the earlier tranche.

Lightmatter product stack · three-layer photonic architecture Interposer, co-packaged optics, and light-source layer stacked. Foundry: GlobalFoundries Fotonix. Packaging: Amkor. Test: ASE. Passage M1000 INTERPOSER LAYER · 3D ACTIVE PHOTONIC Shipping summer 2025 >4,000 mm² Interposer area Hosts up to 34 chiplets 114 Tbps Total optical (Tx+Rx) Highest disclosed in-package figure 256 fibres Optical bandwidth 448 Gbps per fibre 1,024 Electrical SerDes 56 Gbps NRZ integrated 1.5 kW Integrated power delivery Direct via interposer 8-wavelength WDM Per-fibre channel density Passage L200 CO-PACKAGED OPTICS LAYER Available 2026 32 / 64 Tbps Two variants Different optical port counts >200 Tbps Total I/O per chip Package-level aggregate UCIe D2D Standards interface Die-to-die electrical 5-10x Improvement claim vs existing pluggable solutions up to 8x Training-time speedup On large-model workloads L20 NPO Near-packaged variant Roadmap disclosed, specs TBD Guide + Guide DR LIGHT-SOURCE LAYER Guide Jan 2026 · Guide DR May 2026 Hundreds Lasers per chip Very Large Scale Photonics Foundry-produced Manufacturing shift From manual assembly 8x Bandwidth density Initial disclosure vs baseline First industry Liquid-cooled laser NIC Guide DR variant 4x Rack density DR vs air-cooled OCP workstream Open Silicon Photonics Growing OCP initiative Envise (2021-vintage photonic tensor accelerator) de-prioritised; Lightmatter has consolidated around the Passage-Guide interconnect stack.

Manufacturing. Silicon photonics runs on GlobalFoundries Fotonix. Packaging partnership with Amkor was announced November 2024, with expanded coverage in August 2025. ASE (an Amkor subsidiary) provides testing and assembly. High-volume manufacturing is engineered jointly across GlobalFoundries, Amkor, and ASE.

Commercial signals. No named production customer for Passage M1000 or L200 had been disclosed as of September 2026. NVLink Fusion ecosystem membership (June 2, 2026) aligns the Passage roadmap with Nvidia's interconnect architecture. Nvidia disclosed at GTC Taipei 2026 that NVLink Fusion will extend to photonic interconnects, positioning Lightmatter alongside Ayar as ecosystem partners. In August 2026 Lightmatter launched the Open Silicon Photonics for AI Systems initiative, now an official OCP workstream that has grown from ten founding members to nineteen companies.

Patents. Multi-domain portfolio across optical computing devices (G06E), optical modulators (G02F), semiconductor packaging (H10W), solid-state devices (H01L), and telecommunications switching (H04Q). Notable focus on photonic packaging, programmable silicon and glass interposers, thermal management, and electrical-signal-integrity in mixed electronic-photonic assemblies. Total patent count and grant rate not publicly disclosed.

Governance and revenue. Board additions in July 2024 included Richard Beyer (30+ years semis/telecom, on Micron's board) and Robin Washington (Alphabet, Honeywell, Salesforce). Jason Zander joined in March 2025 (leads Microsoft's Strategic Missions and Technologies). Raghib Hussain (President of Products and Technologies at Marvell) holds an earlier-round board seat. Simona Jankowski joined as CFO in July 2024 from Nvidia. The board composition reflects public-market and strategic-partner heritage more than Ayar's, whose cap table is weighted toward semi-industry strategic investors. Lightmatter's 2024 ARR was approximately $116 million per Getlatka disclosure; at the $4.4B Series D post-money the post-money-to-ARR multiple was approximately 38x. No 2025 or 2026 revenue has been disclosed, and Lightmatter has not priced a round in nearly two years. If public markets in 2027 discipline high-growth hardware at 15x forward ARR (below the 20-30x range the Bull case assumes), Lightmatter would need approximately $300M+ in clean recurring ARR to defend even the Series D reference. The $4.4B reference is a stale anchor; the arithmetic sensitivity to a public-market multiple compression stays outside the paired scenario tree model.

Lightmatter snapshot

Founded 2017 (MIT). Headcount 363. Capital raised ~$850M as of Oct 2024 Series D. Latest valuation $4.4B (Oct 2024, T. Rowe Price led). Primary products Passage M1000 interposer (shipping summer 2025) + Passage L200 CPO (2026) + Guide light engine (Jan 2026) + Guide DR liquid-cooled NIC (May 2026). Fab GlobalFoundries Fotonix + Amkor + ASE. Strategic board Microsoft (Jason Zander), Marvell (Raghib Hussain), plus public-market crossovers T. Rowe Price and Fidelity. CFO Simona Jankowski (ex-Nvidia). 2024 ARR approximately $116M (Getlatka). Implied valuation/ARR (Oct 2024) approximately 38x (post-money-to-ARR; not a formal valuation/ARR without disclosed net debt / cash).

Ayar Labs and Lightmatter · funding trajectory and implied valuation Post-money valuation at each disclosed round (log scale). Series letters indicate round type. Sep 2026 marker shows Antero Peak secondary implied valuation for Ayar. $50M $100M $500M $1B $5B A B C ($130M) D ($155M, >$1B) E-1 ($500M) E-2 + secondary A B C-1 ($154M) C-2 ($155M) D ($400M) 2018 2020 2021-05 2022-04 2023-05 2023-12 2024-10 2024-12 2026-03 2026-09 Ayar Labs Lightmatter Ayar raised $650M in 2026 alone; Lightmatter last priced Oct 2024

05Ayar versus Lightmatter · head-to-head

Head-to-head capability matrix · Ayar Labs versus Lightmatter Nine dimensions rated 1-5 on relative advantage. Higher bar = stronger position on that dimension. Editorial rating, not counting quantitative benchmarks. Peak per-package bandwidth Ayar 3 LM 5 Ayar ~8 Tbps per chiplet; Lightmatter 114 Tbps interposer Standards portability (UCIe) Ayar 5 LM 3 Ayar drop-in via UCIe; Lightmatter is interposer-specific Cap table capital raised Ayar 5 LM 5 Both $850-870M cumulative Volume production readiness Ayar 4 LM 3 Ayar 15,000+ units shipped by 2024; Lightmatter Passage L200 shipping in 2026 Product breadth (stack depth) Ayar 3 LM 5 Ayar chiplet+laser; Lightmatter interposer+CPO+laser+NIC Strategic acquirer alignment Ayar 5 LM 3 Ayar cap table concentrates Nvidia/AMD/Intel/GF; Lightmatter more crossover-heavy IPO-track alignment Ayar 3 LM 5 Lightmatter has Marvell+Microsoft+T Rowe+Fidelity board mix Named hyperscaler design-in disclosure Ayar 3 LM 3 Neither has publicly disclosed a Rubin Ultra design-in as of Sep 2026 Nvidia NVLink Fusion partnership Ayar 4 LM 4 Both joined June 2, 2026 AYAR LABS LIGHTMATTER

The two companies compete for the same qualification budgets at the same hyperscalers on the same Nvidia platform generation. The head-to-head moves past "who has better silicon photonics" (both use GlobalFoundries and both have functional laser sources); it is on architectural placement inside the accelerator package and the resulting form of the commercial commitment a customer makes.

DimensionAyar LabsLightmatter
Architectural placementChiplet on package (UCIe-compliant electro-optical chiplet)Interposer under chiplets (Passage 3D photonic interposer) or CPO die adjacent to XPU (L200)
Where the customer buysChiplet drop-in; integration into the customer's UCIe-based chiplet substrateFull interposer or CPO engine; customer builds the compute die stack onto Lightmatter's interposer
Yield-liability ownerChiplet integrator (Nvidia, AMD, hyperscaler ODM)Lightmatter + Amkor jointly (interposer product)
Peak per-package bandwidth~8 Tbps bidirectional per TeraPHY chiplet (multiple chiplets stack)114 Tbps total optical (Passage M1000, up to 34 chiplets); 32/64 Tbps (Passage L200 CPO)
Light sourceSuperNova external CW-WDM (Sivers DFB arrays)Guide integrated foundry-produced laser engine (Jan 2026); Guide DR liquid-cooled variant (May 2026)
Manufacturing partnerGlobalFoundries 45nm monolithicGlobalFoundries Fotonix + Amkor + ASE
Total capital raised~$870M (through Sep 2026)~$850M (through Oct 2024)
2026 fundraising activity$650M raised in 2026 (Series E + secondary)No priced round disclosed since Oct 2024 Series D
Latest implied valuation~$5.0B (Sep 2026 Antero Peak secondary)$4.4B (Oct 2024 Series D)
Board strategic tiltNvidia + AMD + Intel + GlobalFoundries semi-industry heavyMarvell + Microsoft + T. Rowe Price + Fidelity crossover heavy
Disclosed 2024 ARRNot disclosed (industry estimates $50-150M range)~$116M (Getlatka)
NVLink Fusion statusEcosystem partner (Jun 2, 2026)Ecosystem partner (Jun 2, 2026)
Recent volume claim~15,000 units shipped to select customers by late 2024; production-ready mid-2026Passage M1000 shipping summer 2025; L200 available 2026; no unit-count disclosure

The difference matters commercially.

Ayar's chiplet-level architecture is more portable, but less defensible per placement. A UCIe-compliant chiplet offers a more portable integration path across hyperscaler and Nvidia substrates that support the relevant interface and optical requirements, which reduces qualification friction. It also means the customer can dual-source or substitute a competing UCIe chiplet if one emerges. Ayar's moat depends on being the first credible optical-UCIe chiplet at production volumes; the moat weakens if a second UCIe-compliant optical chiplet reaches parity by 2028. The substitution risk holds a specific technical caveat. Electrical UCIe D2D is standardised and interoperable; the optical extension (UCIe-O) is still in draft with no ratified specification for fiber alignment, wavelength allocation, or optical PMD. Drop-in substitution at the optical tier requires re-qualifying the substrate, the optical connector interface, and the remote laser supply. Ayar's package-level lock-in in the 2026-2028 window is therefore closer to Lightmatter's than the pure chiplet-versus-interposer framing implies. The substitution risk becomes real once UCIe-O ratifies and a second production-volume optical chiplet reaches parity.

Lightmatter's interposer approach is less portable, but more defensible per placement. Once a hyperscaler's compute die is engineered onto Passage, the substitution cost can approach a full package redesign and, depending on the architecture, a die respin. That is worth more per placement than Ayar's chiplet position, but it requires each placement to survive an extended qualification cycle in which Lightmatter and the customer jointly own yield risk.

Ayar is trying to make itself easier to adopt. Lightmatter is trying to make itself harder to remove. Ayar's advantage on qualification friction is Lightmatter's disadvantage on qualification cycle. Ayar's exposure to standardised UCIe competition is Lightmatter's protection against substitution. They are solving different commercial problems. Ayar favours portability. Lightmatter favours embeddedness. An integrator-heavy customer favours Ayar; an architecture-heavy customer favours Lightmatter.

What the difference means

Ayar and Lightmatter overlap only partially; full substitution stays limited. Their architectures create different commercial positions and different likely exits. Ayar's chiplet-level position makes it a natural bolt-on for a chiplet-integrator acquirer (Nvidia, AMD, potentially Broadcom for scale-up expansion). Lightmatter's broader interposer-plus-CPO-plus-laser stack makes it more coherent as a standalone photonics company, which is the IPO precedent. The distinction matters more than which company has better silicon photonics.

06Nvidia's CPO plan through 2027

Nvidia's CPO plan · Jan 2026 through H2 2027 Key events on the path from Rubin scale-out CPO (already shipped) to Rubin Ultra scale-up CPO (H2 2027). Jan 2026 CES 2026 Vera Rubin NVL72 announced Spectrum-6 CPO switch inside rack (scale-out CPO shipped) Jun 2 2026 GTC Taipei NVLink Fusion ecosystem Ayar Labs + Lightmatter named partners Sep 10 2026 Ayar Series E $150M extension + $225M Antero Peak secondary implied $5B valuation Q1 2027 GTC 2027 keynote Rubin Ultra CPO NVSwitch supplier disclosure (single biggest catalyst) H2 2027 Rubin Ultra Kyber NVL576 with CPO NVSwitch ships in volume Supplier allocations locked Past events in vermillion; today marker in bright vermillion; future milestones in muted vermillion. H2 2027 Rubin Ultra design-in allocation is the single most consequential unresolved event.

Nvidia's public disclosures at CES 2026 and GTC 2026 established a specific timeline for CPO adoption inside the Rubin-generation platform.

Event When What it is Ayar / Lightmatter relevance
Vera Rubin NVL72 launch CES 2026 Rack-scale system combining six co-designed chips. The Spectrum-6 Ethernet switch uses CPO with 200G SerDes at 102 TB/s switching capacity, powering east-west traffic across the rack. Nvidia claims 10x reliability, 5x uptime, 5x power efficiency versus traditional pluggable optics. First Nvidia platform with CPO inside the rack; scale-out only. Zero. The switch CPO is Nvidia's own silicon.
NVLink Fusion opens to third parties GTC Taipei, 2 Jun 2026 Nvidia opens the NVLink Fusion ecosystem to third-party partners including Ayar Labs and Lightmatter. Both companies' optical engines are electrically and optically compatible with Nvidia's SerDes and optics. Partner-selection signal only. Supplier allocation comes at a later stage.
Vera Rubin Ultra ships H2 2027 NVL576 rack (codename “Kyber”) ships with a CPO NVSwitch that replaces pluggable optical transceivers used on prior rack-scale designs. Integrating optics into the NVSwitch package shortens the electrical path between switch SerDes and fibre. Copper in-rack, CPO out-of-rack. First scale-up design-in in the Rubin family that CPO can win.
Commercial progression: NVLink Fusion membership to repeat-commercial win Economic value rises sharply at the last stages. AYAR + LIGHTMATTER PUBLIC POSITION NOT DISCLOSED AS OF SEP 2026 1 NVLink Fusion membership Ecosystem signal Weak revenue read 2 Evaluation / dev-kit access Weak revenue read 3 Qualification Transition zone 4 Named platform / supplier allocation Economic value zone 5 Volume production Economic value zone 6 Repeat commercial win Economic value zone SIGNAL STRENGTH FOR REVENUE Weak Transition Strong Both companies are visibly at stage 1. Stages 4-6 for the Rubin Ultra generation stay undisclosed as of September 2026.

NVLink Fusion membership is only the first step. Economic value rises sharply at stages 4-6; ecosystem membership and early evaluation are much weaker signals of revenue. Both Ayar and Lightmatter have publicly reached stage 1. Steps 4-6 for the Rubin Ultra generation stay undisclosed as of September 2026.

Observable milestones in 2027 · in order of likely occurrence

# Milestone What to watch
1 Rubin Ultra CPO NVSwitch supplier disclosure Nvidia names the NVL576 CPO NVSwitch supplier: in-house, Ayar, or Lightmatter.
2 Hyperscaler qualification programme completion Ayar or Lightmatter platforms qualified in a hyperscaler programme tied to Rubin Ultra design-ins.
3 First pluggable-optics revenue displacement First hyperscaler AI training deployment in which CPO revenue displaces incumbent pluggable-optics revenue.

For Ayar and Lightmatter, the question is how much of that CPO position Nvidia keeps in-house. Nvidia acquired Mellanox in 2020, giving it InfiniBand plus optical DNA. Nvidia has not disclosed the extent of internal CPO development for Rubin Ultra, but the historical pattern in comparable Nvidia acquisitions (Mellanox for interconnect, Cumulus for switching software) suggests that when Nvidia decides an interconnect layer is strategic it prefers to own it; partnering stays the alternative. The NVLink Fusion ecosystem opening is compatible with two interpretations. First, Nvidia has decided partners will fill non-critical positions and Nvidia-internal CPO will occupy the critical scale-up design-ins. Second, Nvidia is testing partner solutions across Rubin Ultra and the winners of that qualification will become the Rubin Ultra design-in-holders. As of September 2026, the public evidence is insufficient to distinguish between these two interpretations. That ambiguity drives most of the valuation range.

07The full competitive field

The CPO opportunity attracts multiple credible competitors. Reducing the analysis to Ayar-versus-Lightmatter overstates the venture-stage share of the addressable market and understates the incumbent positions that have already been won or that are actively being built. This section covers the four material incumbents alongside the two pure plays.

Broadcom · scale-out CPO incumbent

Broadcom has invested through three generations to hold the scale-out CPO position. Humboldt (2021) was the industry-first CPO platform. Tomahawk 5 / Bailly (51.2 Tbps) integrated eight silicon-photonics engines at ~70% lower power and ~8x silicon-area efficiency versus pluggables. Tomahawk 6 / Davisson has been shipping at 102.4 Tb/s since October 2025 using TSMC COUPE. Broadcom disclosed approximately one million cumulative 400G CPO port-device hours with zero link flaps on the prior-generation Bailly CPO platform, a track record that informed Tomahawk 6-Davisson qualification. Neither Ayar nor Lightmatter offers a directly competing scale-out product. The scale-up expansion path is the joint veto for both pure plays: Broadcom has the hyperscaler relationships, TSMC COUPE packaging expertise, and capital to move into accelerator-adjacent CPO.

Signal to watch: Broadcom scale-up product announcement at OFC 2027 or investor days. Any credible timeline compresses the addressable market for both pure plays.

TSMC COUPE · the foundry-integration threat

TSMC is positioning COUPE (Compact Universal Photonic Engine) as a scale-up packaging option for Nvidia, AMD, and hyperscaler custom-silicon programs. The technical roadmap integrates the electronic IC (EIC) and photonic IC (PIC) at the wafer level using SoIC and CoWoS-adjacent flows. The specific integration risk is TSMC bringing COUPE optical engines directly into the CoWoS-L reconstituted-wafer flow alongside HBM4 and the compute dies, so the scale-up interconnect is sourced natively from the foundry packaging line; the current GlobalFoundries + Amkor / ASE two-step stays the legacy path. If that reaches production inside the Rubin Ultra + 1 or + 2 window, hyperscalers acquire a single foundry-plus-packaging pipeline. That collapses the supply-chain complexity that helps merchant vendors defend against Nvidia in-sourcing.

Signal to watch: TSMC OIP (Open Innovation Platform) reference-flow updates on COUPE for scale-up. A TSMC-supplied scale-up CPO reference flow available directly to Nvidia and hyperscaler custom-silicon teams would compress the merchant-vendor opportunity even without a Broadcom product launch.

Marvell · coherent DSP incumbent with potential scale-up entry

Marvell acquired Inphi in April 2021 for ~$10B (cash plus stock) at ~13x EV/Sales. Critically, Inphi had $600M+ in profitable, high-gross-margin run-rate revenue at close; the multiple applied to audited GAAP cash flows, not to a qualification pipeline. Ayar (~15k cumulative units through 2024) and Lightmatter (Passage L200 pre-production) are still venture-stage bets. Marvell also brought Inphi's coherent optics drivers, optical PHYs, COLORZ transceivers on silicon photonics, and Canopus coherent DSPs. Marvell today competes in coherent optics and pluggable transceivers with substantial datacom and long-haul optical DSP positions. Marvell has two obvious routes on scale-up CPO: acquire a position (Lightmatter is the most plausible target given the board relationship via Raghib Hussain) or build one internally on top of coherent DSP. There is no public indication which route Marvell prefers.

Signal to watch: Marvell board changes at Lightmatter, or a Marvell scale-up CPO product announcement. Either move materially restructures Lightmatter's exit distribution.

Coherent Corp. · vertically integrated laser + optics

Coherent Corp. is the merged entity formed when II-VI Incorporated acquired Coherent Inc. in July 2022 for ~$7B (cash plus stock) and subsequently rebranded under the Coherent name. It is vertically integrated across lasers, optical modulators, silicon photonics, and packaging. Its public communication is oriented toward industrial and laser markets; AI accelerator placement stays a different vector. It has the technology components to build a CPO product; the question is whether Coherent enters scale-up CPO or remains a merchant supplier of components to whoever wins. If Coherent supplies the components without competing head-on, both Ayar and Lightmatter could source lasers or optical components from Coherent (modest positive on cost).

Signal to watch: Coherent Corp. product announcement or partnership disclosure that moves it from component supplier toward integrated CPO competitor.

Nvidia in-house · the elephant in the room

Nvidia acquired Mellanox in April 2020 for ~$6.9B, bringing InfiniBand plus optical DNA in-house. Since then Nvidia has built the Spectrum-X Ethernet line (CPO integrated into the Spectrum-6 generation announced at CES 2026) and has committed publicly to a CPO NVSwitch shipping in H2 2027 as part of Vera Rubin Ultra (NVL576 / Kyber). Nvidia has not disclosed whether the Rubin Ultra CPO NVSwitch is built on internal silicon photonics or on a partner-supplied optical engine. The NVLink Fusion ecosystem opening in June 2026, which named Ayar and Lightmatter as partners, is compatible with two interpretations: partners fill non-critical positions while Nvidia-internal CPO takes the critical placements, or Nvidia is testing partner solutions across Rubin Ultra and qualification winners become the design-in-holders. That choice drives most of the valuation range. Nvidia does not need Ayar or Lightmatter to prove that CPO works; it needs multiple credible suppliers to prevent any one of them from controlling a strategic bottleneck. The ideal merchant CPO supplier from Nvidia's perspective may be the one that is technically indispensable at the qualification bar but economically replaceable at the price negotiation.

Signal to watch: GTC 2027 (Q1) Rubin Ultra CPO supplier disclosures. Anyone waiting for that entry enters at higher valuations; anyone entering before it accepts disclosure risk.

Ayar Labs and Lightmatter · the venture-stage pure plays

Detailed treatment and . In this competitive field their position is: two venture-stage pure plays with Nvidia relationships (NVLink Fusion, June 2, 2026) and plausible routes to volume production. Combined addressable market is scale-up CPO minus Nvidia in-house share minus any Broadcom scale-up expansion, which is a smaller share of the ~$6B 2030 scale-up CPO estimate than a first read of the two companies would suggest.

Second-tier and failure cases · Ranovus and Rockley

Ranovus delivered the industry's first 6.4 Tbps CPO with integrated laser at OFC 2024. Its Odin CPO 3.0 claims ~50% reduction in power, footprint, and cost versus prior solutions. Total funding is ~$87.7M as of August 2025, roughly one-tenth of Ayar or Lightmatter cumulative capital. Credible on technology; well behind on capital and commercial scale. Rockley Photonics is the base-rate failure case: SPAC listing at ~$15.33/share in August 2021, Chapter 11 in January 2023, delisted at ~$0.19/share. Rockley's silicon-photonics wearable-sensor platform failed to reach commercial deployment inside the 18-24 month runway available to service its SPAC-vintage debt. The cautionary lesson is about capital structure and deployment window: silicon-photonics companies with unproven commercial deployment paths fail at a non-trivial base rate.

Competitive-field summary

Scale-out CPO is Broadcom's for this generation, and no hyperscaler has signalled a switch. Marvell is the strategic-acquirer wildcard for Lightmatter and could either buy or build. Coherent Corp. is a merchant-supplier optionality that could enter as a competitor. Nvidia in-house is the largest single risk to Ayar and Lightmatter's Rubin Ultra design-in allocation. Ayar and Lightmatter compete for the residual scale-up CPO share that Nvidia does not in-source and Broadcom does not enter. Ranovus is a distant third pure-play; Rockley is the failure-mode reference.

08The bull-bear frame

Five framings across three polarities. Bear and Bull dominate the underwriting; the two intermediate framings deserve less. The tree below names each framing, the mechanism that activates it, and the qualifier that limits it.

The bull-bear frame · five framings across three polarities Bear (Broken CPO, Value Trap) and Bull (Strategic-acquirer re-rating) dominate. The two intermediate framings deserve less. Each terminal card names the mechanism and the qualifier that limits it. CPO thesis BEAR Broken CPO or Value Trap Load-bearing BETWEEN Intermediate framings Lower weight BULL Strategic-acquirer re-rating Load-bearing 1 Broken CPO LPO extends past 2028; hyperscalers accept higher-power pluggables through the Rubin generation. Both companies run out of runway before volume CPO revenue. QUALIFIER Optical reach becomes binary at 3.2T per lane; hyperscalers face full CPO or split-cluster, not a graceful slowdown. 2 Value Trap Technology works, but incumbents capture the economic value: Nvidia in-house on the critical NVSwitch placement, Broadcom on scale-out, TSMC COUPE on foundry-integrated packaging. QUALIFIER Ayar's cap-table concentration in semi-industry strategic investors compounds the risk. Most likely exit path and parties most incentivised to 3 Mispriced pivot Market values these as photonics-module suppliers when the real business is IP licensing plus differentiated architecture. QUALIFIER Current valuations already reflect IP-plus-hardware premia. Additional mispricing needs Ayar to $200M+ ARR or Lightmatter to $300M+ ARR by 2027. 4 Architecture ahead of its time Credible next-generation IP with commercial deployment 12-24 months later than current valuations imply. QUALIFIER Less likely given the NVLink Fusion partnership timing that specifically positions both companies for the Rubin generation. 5 Strategic-acquirer re-rating Upside comes from acquisition by Nvidia, AMD, Broadcom or Marvell rather Standalone product revenue stays smaller. Marvell paid ~$10B for Inphi (~13x EV/Sales) to capture the compounding effect at the coherent-DSP layer. QUALIFIER Highest-probability outcome for Ayar (chiplet integration). Less likely for Lightmatter, where broader stack complicates integration and public-market Root → polarity → terminal framing. Numbered 1-5 top to bottom. Each qualifier is where the framing is stress-tested against its counter-argument. Framings 1, 2, and 5 dominate the essay's underwriting. Framings 3 and 4 are noted but weighted lower.

09The scenario-weighted exit model

Paired scenario tree · probability-weighted exit EV Ayar Labs and Lightmatter side-by-side. Same CPO thesis, different geometries: Ayar sells strategic, Lightmatter goes standalone or IPO. Exit distributions differ even though core exposures overlap. AYAR LABS Each scenario holds a probability, exit-EV range, and 2Y IRR. Weighted expected EV uses these probabilities as our current confidence read. Reference EV: ~$5.0B. Probability-weighted 2Y expected EV: $8.25B. TODAY $5.0B Sep 2026 25% BEAR Exit EV $1.5-2.5B 2Y IRR -55% to -40% Nvidia in-sources; strategic bid capped; hyperscaler design-in delayed 40% BASE Exit EV $6-9B 2Y IRR 10% to 35% Nvidia partial adopter; strategic exit at $6-9B; hyperscaler qualification lands 25% BULL Exit EV $10-14B 2Y IRR 40% to 65% Multiple hyperscaler design-ins; strategic bidder auction; Nvidia commercial 10% BLUE-SKY Exit EV $15-20B 2Y IRR 70% to 100% Ayar becomes de-facto scale-up standard; Intel or Broadcom strategic WEIGHTED EXPECTED EV $8.25B +65% vs reference · 2Y IRR +28% CONFIDENCE READ Weights reflect our current CPO thesis read. Nvidia in-sourcing risk and hyperscaler qualification timing are the two levers that move Ayar toward bear or bull. LIGHTMATTER Each scenario holds a probability, exit-EV range, and 2Y IRR. Weighted expected EV uses these probabilities as our current confidence read. Reference EV: ~$4.4B. Probability-weighted 2Y expected EV: $9.31B. TODAY $4.4B Sep 2026 25% BEAR Exit EV $1.5-3B 2Y IRR -60% to -35% Passage adoption stalls; wafer-scale interposer economics too fragile 35% BASE Exit EV $6-10B 2Y IRR 15% to 55% Hyperscaler custom-silicon design-in; standalone growth; IPO in 2028 30% BULL Exit EV $10-16B 2Y IRR 55% to 90% Multi-customer wafer-scale platform; Passage becomes photonics interconnect standard 10% BLUE-SKY Exit EV $16-25B 2Y IRR 95% to 140% IPO at wafer-scale premium; Nvidia commercial partnership; blue-sky IP position WEIGHTED EXPECTED EV $9.31B +112% vs reference · 2Y IRR +45% CONFIDENCE READ Lightmatter distribution is fatter than Ayar. Standalone or IPO geometry means downside is worse and upside is larger. Passage design-in confirmation is the single lever.

The valuation model for both companies is exit-scenario weighted; the DCF approach fits public companies with visible revenue with visible revenue. The scenarios differ between Ayar and Lightmatter because their exit distributions differ. The probabilities are our underwriting judgements, not market-implied probabilities. The decisive supplier-allocation information stays non-public.

Ayar Labs · probability-weighted exit EV

Scenario Probability Exit EV (midpoint) Contribution
Bear25%$2.0B$0.50B
Base40%$7.5B$3.00B
Bull25%$12.0B$3.00B
Home run10%$17.5B$1.75B
Probability-weighted total100%$8.25B
Current implied secondary (Sep 2026)$5.0B · 61% of PW

At $5B secondary, buyers accept a 25% probability of 40-55% enterprise-value drawdown against a 65% probability of 10-100% upside. The 39% gap between current secondary and probability-weighted EV is the compensation for the bear-case probability of substantial loss.

Common vs enterprise value

Ayar has raised approximately $870M in preferred capital. The seniority and liquidation preference of the September 2026 securities stay undisclosed. In a $1.5-2.5B bear exit, preferred capital pays ahead of common; common equity and early-round holders can take a materially steeper drawdown than the headline enterprise-value decline suggests. The bands above are enterprise-value bands. Common-equivalent bands are lower in bear scenarios; anyone modelling a specific secondary transaction needs the full waterfall.

Lightmatter · probability-weighted exit EV

Scenario Probability Exit EV (midpoint) Contribution
Bear25%$2.25B$0.56B
Base35%$8.0B$2.80B
Bull30%$13.0B$3.90B
Home run10%$20.5B$2.05B
Probability-weighted total100%$9.31B
Current post-money (Oct 2024 Series D)$4.4B · 47% of PW

Lightmatter's expected value is more sensitive to the exit-multiple assumption than Ayar's. The model uses high-growth semi-plus-IP multiples. Public markets in 2027-2028 that assign hardware multiples instead would compress the PW EV materially. Lightmatter's bear case takes longer to develop than Ayar's. Passage holds no immediate Nvidia in-sourcing risk, and Broadcom scale-up expansion would take multiple years. The bull case holds more optionality on IPO market timing.

What the model says

Ayar prices at approximately 61% of probability-weighted expected value; Lightmatter at approximately 47%. The probability-weighted values are decision inputs under uncertainty; fair-value estimates would require different inputs. The largest single input is the probability assigned to each company owning a substantive commercial position at Rubin Ultra, which stays beyond direct observation from public information today. The arithmetic is precise. The probability distribution behind the arithmetic stays imprecise. That is the limitation of the model. Use it to frame the decision, not as a fair-value estimate. Neither is an obvious mis-pricing at current secondaries. The paired position is the cleaner expression of the CPO thesis: Ayar has the cleaner strategic-exit case; Lightmatter has the cleaner IPO case. Their exit-level differences (different acquirer set, different capital-market conditions) reduce the probability that both go to zero simultaneously. Buying equal amounts of both at today's reference valuations still leaves substantial common exposure to Rubin Ultra, Nvidia in-sourcing and CPO timing, and the total falls materially below the modelled probability-weighted exit value across both companies, while retaining substantial common CPO exposure to the underlying Rubin Ultra adoption and Nvidia in-sourcing risks.

Interactive model · /cpo-model/

Run your own scenario weights, exit-EV ranges, and correlation assumptions across both companies. 10,000-draw Monte Carlo. Reference private valuations, per reported round pricing, at approximately $5.0B (Ayar) and $4.4B (Lightmatter).

Open the paired CPO model →

10Vetoes and tripwires

Ten thesis-breaking events, grouped by who they affect. Each holds an observable signal.

Vetoes and tripwires · ten thesis-breaking events One shared existential risk (top). Three Ayar-specific vetoes (left column). Three joint vetoes affecting both (middle). Three Lightmatter-specific vetoes (right). Each card names the mechanism and the observable signal. ! COMMON MACRO TRIPWIRE TSMC COUPE · foundry-integrated packaging If TSMC provides an end-to-end optical wafer-level packaging flow (SoIC / CoWoS with integrated EIC / PIC) directly to hyperscalers for custom ASICs, no one has an incentive to manage the fractured GlobalFoundries + Amkor / ASE supply chain. Broadcom already uses COUPE for Tomahawk 6. A TSMC-supplied scale-up flow at 2028 WATCH: OFC 2027 · TSMC OIP reference-flow updates · any TSMC-supplied CPO reference disclosure AYAR-SPECIFIC VETOES V1 Nvidia in-sources Rubin Ultra CPO NVSwitch If NVL576 (Kyber) CPO NVSwitch is built on internal silicon photonics (not a partner engine), the strategic-acquirer thesis loses its highest-conviction leg. Ayar retains the NVLink Fusion ecosystem role but SIGNAL TO WATCH Nvidia Rubin Ultra disclosure at GTC Q1-Q2 2027 V2 Second optical-UCIe chiplet reaches parity If Intel, Broadcom, or a stealth competitor demonstrates a UCIe-compliant optical chiplet at Ayar's bandwidth and power efficiency in 2027, Ayar loses first-mover moat. SIGNAL TO WATCH OFC 2027 or hyperscaler qualification disclosure V3 Founder or CTO exit / large secondary sale Mark Wade (CEO, co-founder) and Vladimir Stojanovic (CTO) are load-bearing on Ayar's technology roadmap execution. Either exit signals a leadership shift at maximum execution risk. SIGNAL TO WATCH Form 4-equivalent private-market disclosures on JOINT VETOES (both) V1 LPO extends its useful life past 2028 If LPO reliability at 3.2T optics reaches production and hyperscalers accept the higher-power pluggable, both companies' commercial deployment window slips right. SIGNAL TO WATCH Hyperscaler procurement disclosures on LPO deployment V2 Broadcom expands from scale-out to scale-up Broadcom has hyperscaler relationships, TSMC COUPE packaging expertise, and capital to move into scale-up CPO. A product announcement compresses the addressable market for both. SIGNAL TO WATCH OFC 2027 and Broadcom investor days V3 CPO commoditisation · succeeds too well If optics becomes mandatory across AI infrastructure, more vendors invest in making it manufacturable. Optical engine becomes a component category; specialty IP fades. Technology winners sit apart from the economic SIGNAL TO WATCH Rapid vendor entry at Rubin Ultra +1 / +2, or margin LIGHTMATTER-SPECIFIC VETOES V1 Passage L200 secures no hyperscaler design-in by end-2027 Passage L200 is the CPO product where hyperscaler qualification is the credibility test. Without a named design-in by Q4 2027, the IPO narrative loses its anchor. SIGNAL TO WATCH re:Invent 2027, Ignite 2027, Google Cloud Next 2027 V2 Marvell moves scale-up CPO internally Marvell holds a board seat via Raghib Hussain. If Marvell builds scale-up CPO in-house on coherent-DSP expertise or exits the Lightmatter position, Lightmatter loses one of its most valuable strategic-acquirer SIGNAL TO WATCH Marvell board changes at Lightmatter or Marvell V3 Silicon-photonics IPO market fails to price the premium Lightmatter's IPO thesis depends on public markets paying silicon-photonics premium multiples. If a comparable IPO in 2026-2027 prices at low-double-digit EV/Sales, below the 20-30x bulls model, the IPO window SIGNAL TO WATCH Any comparable silicon-photonics IPO priced in Ayar-specific Joint · affects both Lightmatter-specific Ayar vetoes concentrate on Nvidia + moat · Lightmatter on hyperscaler design-in + public markets A paired position needs to model the joint vetoes as correlated across both, not as independent risks.

What would change the thesis

The vetoes above are asymmetric between the companies by design. Ayar's vetoes concentrate around Nvidia in-sourcing and technology-portability erosion. Lightmatter's concentrate around hyperscaler design-in confirmation and public-market receptivity. Joint vetoes (LPO extension and Broadcom scale-up expansion) affect both. Anyone underwriting a paired position needs to model the joint vetoes as correlated across both positions; independent-risk framing understates the joint exposure.

Cap-table geometry · why Ayar goes strategic and Lightmatter goes IPO Named strategic investors + board members visible on each cap table. Concentration by investor type predicts exit structure. Ayar Labs · strategic-acquisition tilt Strategic (semi/hyperscaler) Crossover / public-market Public-market executive Nvidia Series C, D, E AMD Ventures Series D, E Intel Capital Series A through E GlobalFoundries Series A (fab partner) HPE Series C MediaTek Series E Lockheed Martin Ventures Series A, C Neuberger Berman Series E lead Antero Peak (Artisan) Sep 2026 secondary ARK Invest Series E Qatar Investment Authority Series E Sequoia Global Equities Series E Lightmatter · IPO tilt Strategic (semi/hyperscaler) Crossover / public-market Public-market executive T. Rowe Price Series D lead Fidelity Series D GV (Google Ventures) Series C lead Marvell (Hussain, board) board seat Microsoft (Zander, board) board Mar 2025 Board: Beyer (Micron) board Jul 2024 Board: Washington (Alphabet+) board Jul 2024 CFO: Simona Jankowski ex-Nvidia, Jul 2024 Ayar concentrates strategic acquirer optionality (12 named investor entries, 9 strategic). Lightmatter concentrates public-market and IPO governance signals (8 entries, 5 public-market or crossover). The cap-table structure drives the asymmetric-outcomes thesis.

11The investment call

Layers. First, valuation bands: verdict per implied private-market valuation for each company. Second, trigger-based calls: what disclosures move the verdict.

Ayar Labs valuation bands. Reference: $3.75 billion Series E post-money (March 2026), approximately $5.0 billion implied by Antero Peak secondary (September 10, 2026).

BandVerdictTrigger
Below $3BAttractive if accessible; typically requires a down-round or distressed secondaryBuy if Rubin Ultra CPO NVSwitch confirmed and Nvidia has not in-sourced; watch for corporate discount to strategic buyer
$3-4BAccumulate; below current secondary, matches Series E post-moneyLONG on NVLink Fusion → dev-kit / qualification milestones; source institutional secondary if available
$4-5BConsistent with current secondary; requires acquisition or IPO within 24 months to earnHold; not a compelling entry above $4.5B unless a specific supplier position confirmation is disclosed
$5-7BRequires Rubin Ultra CPO NVSwitch win plus visible scale-up design-in allocationBuy discipline evaporates; requires the strategic-acquirer framing to be the base case
Above $7BPrices in successful IPO or hostile acquisition outcome that has not been disclosedAvoid; asymmetric downside if either Nvidia in-sources or the IPO market fails to price photonics premium

Lightmatter valuation bands. Reference: $4.4 billion Series D post-money (October 2024). No priced round disclosed since. Implied secondary-market pricing available on Forge / EquityZen but not systematically disclosed in public data as of September 2026.

BandVerdictTrigger
Below $3BAttractive; would represent down-round from Oct 2024 post-moneyBuy aggressively if secondary access exists; likely requires a broader photonics correction
$3-4.4BFair to attractive; step-down from Series D and matches broader tech correctionBuy; broader corpus + Marvell/Microsoft board strength supports the IPO trajectory
$4.4-6BCurrent Series D and modest upcycle; the IPO-candidate premium reflectedHold; entry acceptable if the buyer has a 3-year IPO-holding-period model
$6-9BPrices in successful IPO within 12-18 months at 25-40x ARRSelective; requires ARR growth to $300M+ and a specific Passage L200 hyperscaler design-in disclosure
Above $9BPrices in a successful IPO at premium multiples and continued scale-up design-in winsAvoid unless IPO filed and pricing confirmed; comparable to Inphi acquisition-multiple peak

Trigger-based calls. Treat this as a trigger-based call. Entry points depend on specific disclosures. The bands above are underwriting ranges, not transaction prices. Private-market “$X post-money” does not equal $X of economic equity value: liquidation preferences, share-class differences, information rights, and future dilution all reduce the effective economic value to a common-equivalent holder. Anyone modelling a specific transaction needs the actual term sheet.

Ayar prices at approximately 61% and Lightmatter at approximately 47% of probability-weighted expected exit value with materially different exit structures. Owning both is the cleaner way to express the CPO thesis but requires twice the capital and does not remove the common exposures (CPO timing, Rubin Ultra, Nvidia in-sourcing, hyperscaler adoption, Broadcom scale-up, TSMC COUPE, photonics multiples).

Ayar Labs.

Ayar Labs · trigger-based investment call Verdict by implied private-market valuation. Each zone is activated by a specific disclosure trigger. Current reference (Sep 2026 Antero Peak secondary): approximately $5.0B. STRONG LONG $0B · $3.75B LONG $3.75B · $4.5B HOLD $4.5B · $5.5B AVOID $5.5B · $8B $0B $2B $3.75B $4.5B $5B $6B $8B Series E $3.75B Mar 2026 Current $5.0B secondary Antero Peak, Sep 2026 STRONG LONG TRIGGER IF Nvidia discloses Ayar as Rubin Ultra CPO NVSwitch supplier at high-volume placement OR hyperscaler discloses Ayar for custom-silicon 2027 design-in with volume commitment. ACTION Enter selectively up to ~$5B secondary (subject to terms review). Increase conviction on drawdown below $4B. LONG TRIGGER IF NVLink Fusion partnership converts to specific dev-kit or qualification disclosure by end-2026, without a design-in commitment yet. ACTION Underwrite selectively in the $3.75-4.5B range, subject to secondary-terms review. HOLD TRIGGER At current $5.0B implied. No specific supplier position confirmation disclosed. ACTION Not a compelling entry above $4.5B. Current secondary pricing reflects fair value on probability-weighted expected value. AVOID / SHORT-EQUIVALENT TRIGGER IF Nvidia in-sources the Rubin Ultra CPO NVSwitch AND no hyperscaler custom-silicon design-in disclosed. ACTION Value-trap scenario. Write down or exit secondary. Anyone holding at $5B should exit now and avoid holding. Bands are enterprise-value bands. Common-equity waterfall (liquidation preferences, share-class rights) can sit lower in the bear zone. Any specific secondary transaction needs full terms review.

Lightmatter.

Lightmatter · trigger-based investment call Verdict by implied private-market valuation. Each zone is activated by a specific disclosure trigger. Current reference: $4.4B Series D post-money (Oct 2024). No priced round disclosed since. STRONG LONG $0B · $3B LONG $3B · $4.4B HOLD $4.4B · $6B SELECTIVE / AVOID $6B · $10B $0B $2B $3B $4.4B $6B $9B $10B Current $4.4B Series D reference T. Rowe Price, Oct 2024 $9B AVOID threshold no design-in STRONG LONG TRIGGER IF Passage L200 secures named hyperscaler design-in (Meta, Microsoft, Google, or Amazon) with disclosed unit volumes and production commitment (not ACTION Enter selectively up to ~$6B secondary (subject to terms review). Increase conviction on drawdown into $4-5B range. LONG TRIGGER IF no priced Series E has cleared by end-2026 AND Series D pricing at $4.4B remains the reference AND secondary pricing is available at or below Series ACTION Enter with staged sizing; avoid concentrated. Series D reference is stale; multiple-compression risk if public markets discipline HOLD TRIGGER At current $4.4B Series D reference. If Series E clears at $6-9B: upgrade to LONG only on named design-in disclosure. Downgrade to HOLD if round ACTION Not a compelling entry above $6B without specific Passage L200 design-in. Watch for Marvell strategic moves via board seat. SELECTIVE / AVOID TRIGGER IF Series E clears above $9B without Rubin Ultra CPO NVSwitch or Passage L200 hyperscaler design-in confirmation. ACTION Pre-IPO overpricing scenario. Wait for the drawdown. Defending $9B needs $600M+ ARR at 15x public-market discipline, or $300M+ ARR at Bands are enterprise-value bands. Stale-anchor risk: Series D reference is nearly two years old. Any secondary transaction needs full terms and multiple-compression review.

Paired position discipline. Ayar holds the cleaner strategic-acquisition geometry; Lightmatter holds the cleaner standalone-and-IPO geometry. A paired position gives exposure to two different exit paths but does not eliminate the common CPO risks. Public data at this granularity stays limited to estimate the correlation between their exits, so we treat it as moderate, not independent. Both depend on the underlying CPO thesis being right and both are exposed to CPO timing, Rubin Ultra, Nvidia in-sourcing, hyperscaler adoption, Broadcom, TSMC COUPE, and photonics multiples. Different acquirer sets and different capital-market conditions limit full correlation; do not model the two exits as independent.

12Watch list · Q4 2026 through Q4 2027

Events that would change the valuation.

TimingEventImpact
Q4 2026Any Lightmatter priced round or major secondary transactionEstablishes updated Lightmatter reference; also signals continued institutional demand or its absence
Q4 2026Hyperscaler custom-silicon roadmaps at re:Invent (Amazon Trainium), Ignite (Microsoft Maia), Google Cloud Next (TPU roadmap)Any named CPO partnership with Ayar or Lightmatter for 2027 custom silicon would trigger STRONG LONG
Q1 2027Nvidia GTC 2027 keynote disclosures on Rubin Ultra CPO NVSwitch supplierThe key event: Nvidia's Rubin Ultra CPO supplier disclosure. Determines whether Ayar/Lightmatter get the Kyber design-in or Nvidia in-sources.
Q1 2027OFC 2027 disclosures on competing UCIe optical chiplets from Intel, Broadcom, or stealth playersAny credible parity announcement compresses Ayar's chiplet moat
Q2 2027Broadcom investor day / OFC scale-up CPO product announcementA credible Broadcom scale-up product would compress addressable market for both companies
Q2 2027Any silicon-photonics IPO pricing (comparable or adjacent)Establishes public-market receptivity for silicon-photonics premium multiples; affects Lightmatter IPO thesis directly
H2 2027Vera Rubin Ultra volume shipment; Meta/Microsoft/Google production disclosures on CPO adoptionShows whether CPO is scaling into production; determines whether the addressable market ramps as modelled
Q4 2027Lightmatter S-1 filing (if IPO track) or hyperscaler Passage L200 design-in disclosureThis would be the clearest confirmation of the Lightmatter IPO case. S-1 disclosure would put Passage L200 revenue on the public record
OngoingAny Nvidia, AMD, Broadcom, Marvell strategic acquisition announcement in silicon photonicsReprices strategic-acquirer thesis for whichever of Ayar/Lightmatter stays outside the target set

Most of the important disclosures should arrive in H1 2027, when Rubin Ultra supplier allocations become visible. Anyone holding either position through 2027 should be sized appropriately for a Q1-Q2 2027 information shock.

13Glossary of terms used

  • CPO (Co-Packaged Optics). The migration of the optical interconnect from a pluggable module on the switch faceplate or accelerator edge to a die or interposer within the same package as the compute silicon. Reduces electrical trace length and lowers interconnect power at the cost of higher packaging complexity and joint yield risk.
  • TeraPHY. Ayar Labs' electro-optical chiplet. UCIe-compliant. Approximately 8 Tbps bidirectional per chiplet across 8 optical ports using 16-wavelength WDM.
  • SuperNova. Ayar Labs' external CW-WDM light source, sourced from Sivers Photonics DFB laser arrays. Delivers up to 16 wavelengths and 256 data channels.
  • Passage M1000. Lightmatter's 3D active photonic interposer. Approximately 4,000 mm² area, up to 34 chiplets, 114 Tbps total optical bandwidth, 256 optical fibres at 448 Gbps each. Announced March 2025, shipping summer 2025.
  • Passage L200. Lightmatter's co-packaged-optics die adjacent to XPUs. Two variants at 32 and 64 Tbps. Uses UCIe D2D interface. Available in 2026.
  • Guide. Lightmatter's integrated laser engine platform, based on Very Large Scale Photonics technology. Foundry-produced hundreds of lasers per chip. Guide DR (May 2026) is the liquid-cooled laser NIC variant.
  • UCIe (Universal Chiplet Interconnect Express). Open industry standard for die-to-die communication between chiplets on a package. Ayar's TeraPHY is a UCIe-compliant optical chiplet; Lightmatter's Passage L200 uses UCIe D2D interface at the CPO die boundary.
  • NVLink Fusion. Nvidia's ecosystem programme extending NVLink to third-party partners. Ayar Labs and Lightmatter both joined June 2, 2026. Ecosystem membership is a partner-selection signal but does not constitute a specific supplier allocation.
  • Vera Rubin (NVL72). Nvidia's 2026 rack-scale AI platform. Uses Spectrum-6 CPO Ethernet switch inside the rack for east-west traffic.
  • Vera Rubin Ultra (NVL576 / Kyber). Nvidia's 2027 next-generation rack. Ships with a CPO NVSwitch. The Kyber CPO NVSwitch supplier allocation is the single most consequential milestone for Ayar and Lightmatter through 2027.
  • Broadcom Tomahawk 6 / Davisson. Broadcom's 102.4 Tb/s CPO Ethernet switch, shipping since October 2025. Uses TSMC COUPE optical engines. Broadcom disclosed approximately one million cumulative 400G CPO port-device hours with zero link flaps on the prior-generation Bailly CPO platform (October 2025 disclosure at OCP).
  • LPO (Linear-drive Pluggable Optics). Lower-power evolution of pluggable optics that reduces DSP power overhead. Being adopted at 800G pluggable generation. LPO delays but does not eliminate the CPO transition; LPO cannot reach bandwidth densities required past 3.2T optics.
  • Scale-up vs scale-out CPO. Scale-out CPO addresses network switching (Broadcom Tomahawk 6, Nvidia Spectrum-6). Scale-up CPO addresses accelerator-adjacent interconnect and NVSwitch (Rubin Ultra Kyber, hyperscaler custom silicon). Ayar and Lightmatter primarily compete in scale-up.
  • GlobalFoundries Fotonix. GlobalFoundries' silicon-photonics platform integrating photonic components with CMOS logic in a single die. Lightmatter's Passage stack runs on Fotonix; Ayar's TeraPHY runs on GlobalFoundries' older 45 nm monolithic CMOS-plus-photonics process.
  • Marvell / Inphi acquisition (April 2021). Approximately $10 billion cash-plus-stock transaction; approximately 13.3x EV/Sales at the time. The canonical comparable for CPO exit multiples in this essay.
  • Rockley Photonics (Chapter 11, January 2023). SPAC-vintage silicon-photonics wearable-sensor company that failed to reach commercial deployment inside the runway available. Cited in the essay as the base-rate failure case for silicon-photonics execution risk.