Theses›Datacentre & AI compute›Optical interconnect

Optical interconnect

Through 2028 the AI/datacenter optical-interconnect transition runs pluggable → LPO/LRO → CPO, not pluggable → CPO.

  • Closing
  • $162B by 2035

The call, a draftDemand is winning: AI-cluster optics is over $10B in 2026 heading toward about $100B by 2030. Co-packaged optics ships inside NVIDIA's and Broadcom's own systems, but merchant Ethernet volume is a 2028-or-later event. The window is late.

D1D2D3D4R1R2R3C1C2C3foragainst D1D2D3D4R1R2R3C1C2C3foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the drivers.
  1. D1AI clusters are scaling into supernodes of 64 or more chips, and copper at 200 gigabits a second only carries about a metre, so the links between racks and chips go optical.
  2. D2Lasers are a quarter to half the cost of an optical link and scarce, and NVIDIA paid $4B to secure laser capacity.
  3. D3Silicon photonics rises from 43% to 76% of transceivers by 2030, whichever packaging wins.
  4. D4Linear-drive pluggables are about 15% of the market in 2026, and NVIDIA co-designed a new pluggable with Cisco for the next speed generation.
  5. R1The listed optical names trade at 10-27x sales after 4-11x runs, so being right is already in the price.
  6. R2Fibre-to-chip coupling for co-packaged optics is still an unsolved manufacturing step, with the lens route costing 2-5 dB of loss and heavy manual assembly.
  7. R3Value at the component layer keeps going to strategic acquirers, with European photonic component companies repeatedly bought by US chip incumbents.
  8. C1Marvell's co-packaged optics run-rate reaches $500M by the end of its fiscal 2028.31 Mar 2028
  9. C2A volume AI accelerator shipping over 100k units adopts third-party optical chip-to-chip links from Ayar Labs, Lightmatter or Marvell.31 Dec 2027
  10. C3NVIDIA's Spectrum-X Photonics Ethernet switch reaches general volume shipment.30 Jun 2027
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
202720282029today123456789predictionswatch 202720282029today123456789predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Nvidia Spectrum-X Photonics (Ethernet CPO) does NOT reach general/volume customer shipment before mid-2027 — it slips ≥6 months past the announced 2H-2026 target.Jun 2027 · open
  2. 2No merchant Ethernet CPO switch reaches 1M+ cumulative units shipped before end-2027 — i.e.Dec 2027 · open
  3. 3No volume AI accelerator (>100k units shipped) integrates third-party optical chip-to-chip I/O before end-2027.Dec 2027 · open
  4. 4Whether NVIDIA's Spectrum-X Photonics switch reaches volume shipment by mid-2027.Dec 2027 · open
  5. 5Whether any merchant Ethernet co-packaged switch outside NVIDIA passes 1M units.Dec 2027 · open
  6. 6Marvell's CPO revenue run-rate is below $500M as of the end of its FY28 — vendor reality undershoots the $8–9B-by-2030 analyst TAMs.Mar 2028 · open
  7. 7Pluggable optical transceiver revenue (all speeds) in CY2028 still exceeds CY2025 — CPO is additive at the top tier, not a wholesale replacement.Dec 2028 · open
  8. 8By end-2028, LPO+LRO ports outnumber CPO ports in cumulative 1.6T-generation datacenter shipments — the intermediates win the generation.Dec 2028 · open
  9. 9Whether pluggable transceiver revenue in 2028 still beats 2025, and linear-drive ports outnumber co-packaged ports at 1.6T.Dec 2028 · open
Expressionslisted, private, technologies
Private · 11 private companies, 10 of them in the US.

Celero Communications · CScale · Polaris Electro-Optics · Avicena · Photon IP · Mixx Technologies · Google OCS (Apollo / Palomar) · Ayar Labs · Lightmatter · Celestial AI · Lumilens

15 more private companies tracked, not shown

Technologies
What changedthe record, newest first
  • 16 Oct 2026The man who sells the switches says copper stays inside the rack and the pluggable keeps the volume until at least 2028. Co-packaged optics is a keynote, not yet a product line.
  • 15 Oct 2026The best-funded new optical-interconnect company of the year raised over $700 million, and the first thing it shipped to a hyperscaler was a plug-in transceiver. Co-packaged optics is on its roadmap, after an intermediate step. That is the order I expected: the plug-in module keeps selling while the more integrated options arrive in stages.
  • 15 Oct 2026Nvidia, the company most associated with putting optics right next to the switch chip, has just co-designed a new plug-in optical module with Cisco for the next speed generation. The plug-in transceiver is not dying on schedule, because the customers who have to swap failed parts in the field still want it.
9 earlier
  • 15 Oct 2026Nvidia's co-packaged optical switch is in production with a named five-company supply chain and six cloud customers lined up. I had bet it would slip well past its date, and that bet now looks likely to lose. The broader call, that optics-next-to-the-chip stays inside Nvidia's and Broadcom's own systems for a few more years, is untouched.
  • 21 Jul 2026A number for the layer that matters: the laser is a quarter to half the cost of an optical link, confirmed by two separate companies. That is why the value sits in the light source rather than the system, and why Nvidia paid four billion to secure its supply.
  • 20 Jul 2026Frontier models are scaling out into supernodes of 64-plus chips, and the connections between those chips are increasingly light, not copper. The physics is doing the marketing.
  • 19 Jul 2026Co-packaged optics is about half a percent of modules this year, heading for a third of them by 2030. The gap between how much it is discussed and how much of it ships is the whole story. Packaging is where the margin sits at the changeover, and then it drifts back to silicon.
  • 23 Jun 2026AI-cluster optics is already a $10B+ market in 2026 on a path to ~$100B by 2030; the consensus now calls 2026 the first year of co-packaged-optics volume, but that volume is proprietary-fabric, and general-market CPO is still the deferred bet.
  • 22 Jun 2026Co-packaged optics has crossed from research into early production: copper has hit its power wall for chip-to-chip data movement, and the optical-interconnect transition is now a build, not a forecast.
  • 22 Jun 2026Everyone says co-packaged optics is about to replace pluggables; the real near-term path runs through LPO/LRO, and the durable value is the silicon-photonics content and the lasers, not the packaging fight.
  • 19 Jun 2026The AI optical-interconnect winners are real, but the listed pure-plays now trade at 10-27x sales after 4-11x runs — the thesis being right is fully priced, so the edge is gone until a capex drawdown resets the multiple.
  • 17 Jun 2026Multi-wavelength co-packaged optics is far harder than single-wavelength (16 precisely-spaced on-chip lasers), which both slows volume merchant CPO and concentrates the value in laser supply/integration rather than the packaging format.
Reading
Show the reading