Lightmatter set out to build a computer that does math with light. Nine years later, its most important products don't compute anything. They move data between chips, and the company that once pitched a photonic AI processor is now one of the best-funded bets on the copper wall inside AI data centers.
That shift is the whole story. A $400 million Series D at a $4.4 billion valuation in October 2024 paid for a product line called Passage, a laser line called Guide, and a push into Nvidia's ecosystem. The hard part, turning demos and evaluation kits into volume orders, is still ahead.
This guide covers what Lightmatter actually sells, how each Passage product works, who builds it, how it stacks up against Ayar Labs and Marvell's Celestial AI unit, and where the claims outrun the evidence. Last updated: November 18, 2026.
Key Takeaways
- Lightmatter raised a $400 million Series D at a $4.4 billion valuation in October 2024, its last disclosed round.
- The Passage M1000 is a 3D photonic interposer rated at 114 Tbps of total optical bandwidth.
- Passage L200 and L200X are 3D co-packaged optics engines at 32 Tbps and 64 Tbps.
- Lightmatter told Taipei Times NPO ships in 2027 and high-volume CPO starts in 2028.
- No hyperscaler has publicly named itself a Passage customer; Fierce reports five customer CPO tape-outs in 2026.
Lightmatter at a glance
Here's the short version before the detail. Every figure in the table comes from Lightmatter's own releases or from reporting on its funding, so treat specs as vendor claims until customers ship systems.
| Item | Detail |
|---|---|
| Founded | 2017, by Nick Harris, Darius Bunandar and Thomas Graham (MIT) |
| Headquarters | Mountain View, California |
| CEO | Nick Harris, Ph.D. |
| Total funding | About $850 million |
| Last valuation | $4.4 billion (Series D, October 2024) |
| Series D lead | T. Rowe Price, with Fidelity and GV |
| Core product family | Passage (M1000, L200, L200X, L20, L20 CPX) |
| Laser product family | Guide, Guide DR |
| Fab and packaging partners | GlobalFoundries, TSMC, Tower, Amkor, ASE |
| Ecosystem memberships | NVLink Fusion, Open CPX MSA, UALink, OCP, UCIe |
Source: MIT News on the founders; Lightmatter homepage for the $850 million total funding, $4.4 billion valuation, partners and standards bodies; TechCrunch on the Series D investors.
From Envise to Passage: why Lightmatter stopped selling compute
Lightmatter's first headline product was Envise, a photonic AI accelerator that ran matrix math in light. The Envise product page is still live, but the company now frames compute as research and interconnect as the business.
The research is real. In April 2025, a 40-plus author Lightmatter team published "Universal photonic artificial intelligence acceleration" in Nature. The processor ran ResNet, BERT and a DeepMind-style Atari reinforcement learning agent at near-electronic precision.
Lightmatter's own write-up gives the hardware numbers: six chips in one package, 50 billion transistors, 1 million photonic components, and 65.5 trillion 16-bit block floating-point operations per second at 78 watts electrical plus 1.6 watts optical.
That's impressive for a photonic system. It's also far behind current GPUs in raw throughput, and the same blog post says the problem to solve today is interconnecting millions of chips. That sentence is the pivot.
The shoreline problem Passage is built to fix
Every accelerator talks to the outside world through SerDes lanes on the edge of its die, the "shoreline." Chips have hit the reticle limit, so the edge can't grow much. Lightmatter's pitch is to put optical I/O under or on top of the chip instead of only along its edges.
The company calls this "edgeless I/O." According to the L200 launch release, 3D integration lets SerDes sit anywhere on the die and gives each L200 the bandwidth of 40 pluggable optical transceivers. For the wider context, Inside Deep Tech's silicon photonics guide covers the building blocks.
The Passage product family, explained
Passage isn't one chip. It's a family that runs from a socketed near-package module to a full photonic interposer, so customers can pick how deep into their package they let optics go.
| Product | Type | Bandwidth | Key spec | Availability |
|---|---|---|---|---|
| Passage M1000 | 3D active photonic interposer | 114 Tbps total | 4,000+ mm², 256 fibers, 1.5 kW | Reference platform, 2025 |
| Passage L200 | 3D CPO engine | 32 Tbps | UCIe die-to-die, under 5 pJ/bit optical | Announced for 2026 |
| Passage L200X | 3D CPO engine | 64 Tbps | 106/112 Gbps PAM4 | Announced for 2026 |
| Passage L20 | NPO / OBO optical engine | 6.4 Tbps each direction | 200G lanes, BiDi, 30 W max | Sampling late 2026 |
| Passage L20 CPX | Open CPX module | 12.8 Tbps total | Over 500 m reach, 212.5 Gbps PAM4 | Eval kits Q1 2027 |
| Guide DR | Laser NIC (light source) | Drives up to 51.2 Tbps | 200 mW per fiber, 64 fibers | Launch 2027 |
Source: M1000 release; L200 release and L200 product page; L20 release; L20 CPX release; Guide DR release; Guide DR timing from Taipei Times.
Passage M1000: the photonic interposer
The M1000 is the most ambitious part. Announced on March 31, 2025, it's a multi-reticle active photonic interposer of more than 4,000 square millimeters that carries 114 Tbps of total optical bandwidth across eight tiles.
Customer dies stack on top, and a reconfigurable waveguide network routes light to wherever I/O is needed. The M1000 release lists 1,024 electrical SerDes, 256 edge-attached fibers at 448 Gbps each, and 1.5 kW of power delivery. It's built on GlobalFoundries' Fotonix platform with Amkor packaging.
Here's the catch: the M1000 is a reference platform. It shows what a customer design could look like. Harris told Fierce in October 2026 that Lightmatter is working on two interposer designs with customers, not that any are in production.
Passage L200 and L200X: 3D co-packaged optics
The L200 family stacks an electronic IC on the photonic chip using chip-on-wafer bonding, then attaches to an XPU or switch over UCIe. The L200 release rates the L200 at 32 Tbps and the L200X at 64 Tbps, with 16 WDM wavelengths per fiber for 800 Gbps to 1.6 Tbps per fiber.
The electrical side originally came from Alphawave Semi, now part of Qualcomm. In March 2026, Lightmatter said a sampled Passage CPO chiplet with Qualcomm's 112G SerDes hit 1.6 Tbps per fiber, which it claims is up to 8x the bandwidth per fiber of existing NPO and CPO.
If you track co-packaged optics broadly, Inside Deep Tech's CPO full guide covers the wider CPO market, from switch packages to GPU packages and the September 2026 Lightmatter CPX news.
Passage L20 and L20 CPX: the near-term on-ramp
The L20 is the product most likely to ship first, and it's the least exotic. It sits near the package or on the board rather than inside it, uses standard 224G-class electrical interfaces, and moves 6.4 Tbps in each direction with a 30 W maximum TDP.
Its trick is bidirectional optics: each fiber carries transmit and receive on two wavelengths. Lightmatter says that halves fiber management needs versus standard DR optics.
The L20 CPX packages the same engine into an Open CPX MSA module. Lightmatter's model of a 512-GPU pod claims the BiDi design cuts fiber from about 130,000 strands to 65,000 and saves roughly 15% of scale-up interconnect cost. That's the company's own modeling, not a customer result.
Guide: the laser business hiding in plain sight
Every CPO and NPO engine needs an external light source, and lasers are the part of optics that fails most often. Guide is Lightmatter's answer: a "very large scale photonics" chip with many tunable lasers on one die, so a spare can be tuned to replace a failed channel.
The May 2026 Guide DR release moves that light source off the front panel into a liquid-cooled card sized to OCP NIC 3.0. Each module supplies 200 mW per fiber across up to 64 fibers, and four modules can feed 204.8 Tbps of CPO switching in a 1RU tray.
This matters more than it looks. Harris told Fierce there are "simply not enough lasers" for every company that wants optical scale-up. A vendor with its own laser line has a supply-chain lever that pure optical-engine startups don't.
Who builds it and who it plugs into
Lightmatter is fabless, so its credibility depends on partners. The homepage lists TSMC, GlobalFoundries, Tower Semiconductor, Amkor and ASE as manufacturing partners. ASE assembles one CPO engine with detachable fiber arrays in Taiwan, per Taipei Times.
- Nvidia NVLink Fusion (June 2026). Lightmatter will build CPO and NPO products compatible with Nvidia's optical and SerDes tech, so custom XPUs can reach Nvidia switch silicon. The release does not make Nvidia a customer.
- Open CPX MSA (September 2026). A multi-vendor module standard for optics next to the ASIC. Lightmatter's L20 CPX is its first bidirectional module.
- OCP and Flex (October 2026). Lightmatter showed a 204.8 Tbps CPO switch tray and a 102.4 Tbps four-accelerator XPU tray built to Open Rack specs, with Flex on the switch tray.
- UALink and UCIe. Memberships keep Passage usable outside Nvidia's stack. Inside Deep Tech's UALink guide covers the open scale-up fabric those trays target.
The pattern is consistent: standards first, named customers later. For a component supplier selling into hyperscaler roadmaps, that's normal. It also means outsiders can't verify volume.
Funding and valuation versus the photonics field
Lightmatter's Series D quadrupled its valuation from a mid-2023 round. No IPO filing or new funding round had been announced as of November 18, 2026, so that 2024 price is still the only public marker.
Meanwhile the field got crowded and more expensive. Ayar Labs closed a $500 million Series E at a $3.75 billion valuation in March 2026, bringing it to $870 million raised, with Nvidia, AMD and MediaTek investing. Marvell closed its roughly $3.25 billion Celestial AI acquisition in February 2026.
| Company | Core product | Approach | Funding or deal | Headline spec |
|---|---|---|---|---|
| Lightmatter | Passage, Guide | 3D CPO, NPO, photonic interposer, lasers | About $850M raised; $4.4B valuation (2024) | 114 Tbps M1000; 64 Tbps L200X |
| Ayar Labs | TeraPHY, SuperNova | UCIe optical I/O chiplet plus external laser | $870M raised; $3.75B valuation (2026) | 8 Tbps per chiplet |
| Celestial AI (Marvell) | Photonic Fabric | Optical scale-up chiplets and memory fabric | Acquired for about $3.25B (2026) | Marvell targets $500M run rate |
Source: Lightmatter homepage and M1000 and L200 releases; Ayar Labs Series E release and TeraPHY page; Marvell on the Celestial AI close and revenue targets; SDxCentral on the $3.25 billion price.
The Marvell release is the most useful benchmark in the table. It expects Celestial AI revenue to start in the second half of fiscal 2028 and reach a $500 million annualized run rate in Q4, doubling to $1 billion a year later. That's a public company putting a date on optical scale-up revenue. Lightmatter hasn't.
Inside Deep Tech's optical I/O guide goes deeper on TeraPHY and Photonic Fabric if you're comparing chiplet designs.
What ships when: Lightmatter's 2026 to 2029 timeline
Lightmatter's own dates have firmed up over 2026. Here's what the company has said on the record.
- Late 2026: Passage L20 sampling for NPO and on-board optics.
- Q1 2027: L20 CPX evaluation kits to customers.
- 2027: NPO engines and Guide DR laser NICs ship, plus initial low-volume CPO, per Taipei Times.
- Early 2027 to 2029: NPO ramps to high volume in early 2027, with CPO following by the end of 2028 or early 2029, per Fierce.
Those two interviews, two weeks apart, give slightly different CPO dates: high volume starting in 2028 in one and "end of 2028 or early 2029" in the other. Read that as a 2028 to 2029 window, not a fixed launch.
What Lightmatter means if you plan AI infrastructure
You probably won't buy a Passage chip directly. You'll meet it inside an accelerator, a switch tray or a rack design from an ODM. That changes which questions are worth asking now.
- If you design custom XPUs: Passage L200 attaches over UCIe, and Lightmatter says custom-order clients secure their own foundry capacity while it builds on top, per Taipei Times. Ask early who owns the wafer allocation.
- If you run an Nvidia-heavy fleet: NVLink Fusion compatibility is the bridge. The June 2026 release is about semi-custom XPUs talking to Nvidia switches, so it matters most for mixed fleets.
- If you buy racks to OCP specs: the Flex switch tray and the four-accelerator XPU tray are the closest thing to a reference design you can evaluate today.
- If you own the fiber plant: bidirectional optics is the near-term win. Halving strands per link matters more in a fiber shortage than any headline Tbps number.
One practical rule: put optics where copper can't reach, which today means scale-up domains that span more than one rack. Inside a single rack, copper NVLink-style backplanes still do the job, and the business case for photonics is weaker.
Honest limits: where the Lightmatter story is thin
Lightmatter has strong engineering and the right partners. The public evidence still has gaps you should weigh.
- No named volume customer. Harris said custom-order clients are mostly industry giants, including hyperscalers, per Taipei Times, but none are named. Nvidia and Qualcomm are partners, not disclosed buyers.
- Specs are vendor-measured. The L20 CPX release says its figures come from internal testing or modeling. The 15% cost saving and fiber counts are Lightmatter's own pod model.
- Supply chain is the gating factor. Harris told Fierce laser and fiber shortages will take a couple of years of factory build-out to fix. That cuts against every vendor's timeline, including his.
- Copper isn't dead. Futurum's Brendan Burke told Fierce copper keeps an enduring role and optics wins mainly in multi-rack scale-up and large-model inference.
- Serviceability is still unproven at scale. Burke named laser failures inside trays as the main adoption barrier. Guide's redundant lasers are a smart answer, but fleet data doesn't exist yet.
- Valuation is stale. The $4.4 billion figure is from October 2024, before Ayar Labs' raise and the Celestial AI deal reset the comparables.
Inside Deep Tech's take
FAQ
What does Lightmatter make?
Lightmatter makes silicon photonics interconnect chips (Passage) and laser light sources (Guide) that link AI accelerators and switches with optical fiber instead of copper. Its earlier Envise photonic processor is now a research line, per its Nature paper write-up.
Is Lightmatter a public company?
No. Lightmatter is private. Its last disclosed round was a $400 million Series D at a $4.4 billion valuation in October 2024.
What is the Lightmatter Passage M1000?
It's a 3D photonic interposer reference platform with 114 Tbps of total optical bandwidth. Customer chips stack on top and send data through light anywhere on its surface.
Does Nvidia use Lightmatter?
Not as a disclosed customer. Lightmatter joined the NVLink Fusion ecosystem in June 2026 to build CPO and NPO products compatible with Nvidia's switch silicon.
How is Lightmatter different from Ayar Labs?
Ayar Labs sells a UCIe optical I/O chiplet rated at 8 Tbps, while Lightmatter spans NPO modules, 3D CPO and a full photonic interposer, plus its own lasers.
When will Lightmatter products ship in volume?
NPO engines and Guide DR lasers are due in 2027, with high-volume CPO from 2028, per Taipei Times. Evaluation kits for the L20 CPX are due in Q1 2027.
What to watch next
- Whether any hyperscaler or OEM names Passage in a 2027 system after the L20 CPX kits ship.
- A new funding round or IPO filing that resets the $4.4 billion valuation.
- Volume claims for Guide DR, since lasers are the industry bottleneck Harris flagged to Fierce.
- Whether Marvell's Celestial AI hits its fiscal 2028 revenue start, which sets the bar for every photonics startup.
- Interposer customer designs. Inside Deep Tech's optical circuit switching guide shows how fast hyperscalers adopt optics once a design is qualified.
Lightmatter's next year is about proof, not specs. If you're planning optical scale-up for 2027 or 2028, start with the standards and form factors Passage targets, then compare engines. Inside Deep Tech's co-packaged optics full guide is the best place to start.
Cover image: optical fiber glowing blue from upconversion fluorescence, NTNU Faculty of Natural Sciences and Technology (Per Henning), CC BY 2.0, via Wikimedia Commons, blurred with title overlay.



