A deep dive into co-packaged optics (CPO) — NVIDIA Spectrum-X Photonics switches (SN6810 102.4 Tbps, SN6800 409.6 Tbps, Spectrum-6 ASICs), the ~3.5x power and ~10x reliability gains over pluggable transceivers, why they ship liquid-cooled, the 2025–2026 CPO news timeline, and the full supply chain making them: $NVDA, $AVGO, $TSM, $COHR, $LITE, $FN, $POET, plus the LPO/AEC camp ($CRDO, $ALAB).
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
Co-packaged optics (CPO) move the optical engine out of a plug-in module and onto the same package as the switch chip — and that one change is the biggest shift in data-center networking in a decade. As AI clusters scale past hundreds of thousands of GPUs, the network that connects them is hitting a power and reliability wall that pluggable transceivers can't climb. CPO is the answer the whole industry is now racing to ship. This is a deep dive into what these photonic switches — NVIDIA's Spectrum-X and Quantum-X Photonics, Broadcom's CPO line — actually do, how much better they are, why they arrive liquid-cooled, and — for investors — exactly who is making them.
👉 See it in context: the CPO supply chain lives in the Photonics / CPO basket on Macroplane — $COHR, $LITE, $FN, $POET, $AAOI and the optical names — and the broader Silicon Photonics & Optical Interconnects trend.
🎥 Watch: NVIDIA's New Photonic Technology Explained — a clear visual walkthrough of how co-packaged optics works.
In a conventional AI switch, the chip in the middle is an electrical ASIC, and light only happens at the edges: each port has a pluggable optical transceiver (a QSFP/OSFP module you can pull out by hand) that converts the switch's electrical signals into laser light for the fiber. Every one of those modules carries its own laser, its own digital signal processor (DSP), and its own retimer — and at 800G and 1.6T, those modules are getting hot, power-hungry, and failure-prone.
Co-packaged optics deletes the module. The optical engines are integrated directly onto the switch ASIC's package, so light is generated millimeters from the silicon instead of inches away across a circuit board and a connector. NVIDIA's Spectrum-X Photonics line is the clearest example of the new design:
Both replace the pluggable-transceiver front panel with silicon-photonics co-packaged optics for low-latency, single-mode fiber connectivity. On the InfiniBand side, NVIDIA's Quantum-X Photonics Q3450-LD does the same thing for the back-end GPU fabric: 115.2 Tbps across 144 ports of 800G, four Quantum-X800 ASICs, 18 field-replaceable external laser modules, and a measured 3.95 kW — versus roughly 7 kW for the equivalent pluggable-based switch. The common thread: the optics and the switch silicon now share one package.
The timing matters for investors: all of this was announced at GTC in March 2025, Quantum-X Photonics started shipping in early 2026 at low volume, and NVIDIA began shipping the first Spectrum-X CPO switches to select partners in June 2026, with the production ramp scheduled for the second half of the year.
The gains aren't incremental. By co-packaging the optics, NVIDIA cites four step-changes versus traditional pluggable-transceiver networking:
Two of those deserve a closer look:
Power. A large AI fabric can burn through megawatts just on its optical transceivers. A 1.6T-generation pluggable draws ~25–30 W (the figure behind Jensen Huang's GTC keynote math; today's 800G modules run ~13–17 W), and a 400,000-GPU cluster needs hundreds of thousands of optical connections. NVIDIA's framing is blunt: at hyperscale, CPO saves tens of megawatts of networking power — power that can instead be spent on GPUs. When the binding constraint on AI is electricity (see our data-center power and nuclear coverage), shaving the optics power bill is a direct lever on how many accelerators fit in a given site.
Reliability. The laser is the single most common failure point in optical networking. CPO designs use far fewer lasers, run them as high-power **external light sources** that can be serviced without touching the switch, and keep them off the hot ASIC package. Fewer lasers, better cooled, in a more controlled environment is what gets you to that ~10× resiliency claim — and at the scale of a single training run spanning 100,000+ GPUs, a network that stays up is worth as much as one that's fast.
| CPO switch | Throughput | Ports | Switch ASICs | Power | Form factor |
|---|---|---|---|---|---|
| NVIDIA SN6810 (Spectrum-X) | 102.4 Tbps | 128×800GbE | 1× Spectrum-6 | ~9 W optics per 1.6T port | 2RU, liquid-cooled |
| NVIDIA SN6800 (Spectrum-X) | 409.6 Tbps | 512×800GbE | 4× Spectrum-6 | ~3.5× better than pluggable | liquid-cooled |
| NVIDIA Quantum-X Photonics (Q3450-LD) | 115.2 Tbps | 144×800G InfiniBand | 4× Quantum-X800 | 3.95 kW (vs ~7 kW pluggable-based) | 4U, liquid-cooled |
| Broadcom Bailly (Tomahawk 5 + CPO) | 51.2 Tbps | 128×400G FR4 | 1× Tomahawk 5 | ~70% lower optics power | 4RU, air-cooled (Micas) |
| Broadcom Tomahawk 6 "Davisson" | 102.4 Tbps | 200G/lane optical | 1× Tomahawk 6 | ~70% lower optics power | shipping since Oct 2025 |
*Figures as published by the vendors — each switch name links to its primary source (NVIDIA's engineering blog, Lambda's Q3450-LD teardown, Broadcom's announcements). Treat them as marketing-grade specs, not audited numbers.*
For two decades copper kept up by brute force — every time signaling doubled (NRZ → PAM4, 25G → 50G → 100G → 200G per lane), the channel got a little smarter. But physics is winning. At 200G-per-lane PAM4, a passive copper cable carries a usable signal less than a meter before the loss is unrecoverable. Optics doesn't hit that wall — a single-mode fiber runs hundreds of meters at the same rate. The catch is power: a pluggable optical module spends a big slice of its energy in the DSP that cleans the signal up across the board and the connector. CPO attacks exactly that — by putting the optics on the package, it can drop the heavy DSP and cut the energy per bit by roughly 3× while keeping optical reach.
Here's the same story as a side-by-side, from in-rack copper to co-packaged optics:
| Link type | Energy/bit* | Reach | Latency | Field-serviceable | Where it fits |
|---|---|---|---|---|---|
| Passive copper (DAC) | ~1–2 pJ/bit | <1–2 m | lowest | yes | in-rack, GPU-to-GPU short hops |
| Active copper (AEC) | ~5–8 pJ/bit | ~2–7 m | very low | yes | cross-rack, NVLink-class scale-up |
| Pluggable optics (DSP) | ~15–20 pJ/bit | 100 m–2 km+ | low | yes | today's switch front panels |
| LPO (linear pluggable) | ~8–10 pJ/bit | ~100 m | lower | yes | near-term power cut, DSP removed |
| Co-packaged optics (CPO) | ~5–7 pJ/bit | 100 m–2 km+ | lowest optical | laser-serviceable | highest-radix AI switches |
\Energy figures are rounded, widely-reported industry estimates for direction and magnitude — not audited vendor numbers.*
Look at the power column again: NVIDIA's Quantum-X Photonics packs 3.95 kW into a 4U box, and the Spectrum-X line pushes the same density further. That is an enormous thermal load for a switch — and the harder problem is that co-packaging deliberately puts the heat-sensitive optics right next to the hottest silicon in the box. Air struggles to move that much heat out of that small a volume while keeping the lasers in their happy temperature range (laser wavelength and lifetime both drift with heat). Broadcom's 51.2T Bailly generation still manages with high-efficiency air cooling in 4RU; NVIDIA's denser 102.4T-per-ASIC designs do not.
So at the leading edge, CPO and liquid cooling ship together. The same direct-to-chip and cold-plate liquid loops that the AI rack already needs for its GPUs get extended to the switch. That makes the cooling supply chain part of the CPO story: the precision-cooling and liquid-distribution names — Vertiv ($VRT) and the rest of the Cooling basket — are downstream beneficiaries of every CPO switch deployed, because each one is another high-density, liquid-cooled box in the row. CPO doesn't just change the optics; it pulls more of the rack onto liquid.
This is the part that matters for investors — CPO is a system, and the value is spread across several layers. Here's the map, grounded in the relationships visible on Macroplane.
The companies that actually build the AI switches — and are racing to co-package the optics onto them:
The venture-backed challengers attacking the optical interconnect from a different angle are consolidating fast: Celestial AI (Photonic Fabric) is now inside $MRVL, while Lightmatter (Passage interposer) and Ayar Labs (optical I/O chiplets, fresh off a $500M Series E in March 2026) remain private — and remain who the incumbents are racing.
CPO is built on a photonic integrated circuit, and someone has to fab it:
The light-making and light-shaping layer — and the part of the chain that's most levered to CPO volume:
Not everyone thinks CPO wins immediately. A cohort is betting that linear-drive pluggable optics (LPO) and active electrical cables (AEC) bridge the gap — keeping optics pluggable but stripping out the power-hungry DSP, or replacing short optical links with smarter copper:
The honest read: CPO and LPO/AEC coexist for now — CPO for the highest-radix switches where power and density are binding, pluggables/LPO/AEC for the rest. The investable point is that all of these roads run through the same optics and connectivity suppliers.
If you want the one-paragraph version of why 2026 is the year CPO stopped being a roadmap slide, it's this sequence:
For the live feed of what's moving the names, the Silicon Photonics & Optical Interconnects trend page tracks co-packaged optics news against the companies it touches, daily.
Until August 2026 co-packaged optics was three vertically integrated stacks. NVIDIA built its own optical engines for its own switches. Broadcom built its own on TSMC COUPE. If you were a hyperscaler who wanted CPO, you bought one vendor's entire architecture.
The Open Silicon Photonics for AI Systems workstream is the attempt to break that. Formally launched inside the Open Compute Project on 13 August 2026, it pairs a roughly 300-page architectural blueprint with a target of first specifications in Q4 2026. The engineering goal is to scale OCP Modular Hardware System and Open Rack v3 environments from 72 nodes to more than 1,024 while keeping multi-vendor interoperability.
The membership list is the interesting part, because it tells you who expects to sell into a merchant CPO market: Celestica ($CLS), Dell ($DELL), Flex ($FLEX), Foxconn Interconnect Technology ($6088.HK), Global Unichip ($3443.TW), Hyve Solutions, Keysight ($KEYS), Lightmatter (private), Qualcomm ($QCOM) and Quanta Cloud Technology ($2382.TW), across ten founding and nine new members.
Two consequences for investors:
Component suppliers get addressable. In a single-vendor architecture, the optical engine supplier is whoever NVIDIA or Broadcom chose. In a specified one, laser, amplifier and interposer suppliers can sell to anyone building to the spec. That is what makes the smaller names in layer two of the photonics stack investable rather than hopeful.
Test moves upstream. You cannot field-replace an optical engine sealed inside a switch package, so the yield and validation burden shifts into manufacturing. Keysight's presence as a founding member is the tell, and Celestica's hyperscaler CPO switch award, ramping in 2027, is the manufacturing side of the same trade.
The date to watch is Q4 2026, when the first specifications are due.
CPO is a classic picks-and-shovels theme: the switch vendors get the headlines, but the durable exposure is in the components every design needs regardless of who wins.
Co-packaged optics integrates the optical engine — the part that converts electrical signals to laser light — directly onto the same package as a switch or accelerator chip, instead of housing it in a separate pluggable transceiver module. This shortens the electrical path, cuts power, and improves reliability, which is why it's being adopted for the highest-bandwidth AI networking switches.
At 800G and 1.6T per port, pluggable transceivers consume too much power and fail too often to scale to clusters of hundreds of thousands of GPUs. The optical transceivers in a large AI fabric can burn megawatts on their own. CPO addresses both: NVIDIA cites roughly 3.5× better optical power efficiency and ~10× better network resiliency versus pluggables — power and uptime that translate directly into more usable GPUs.
Four step-changes versus pluggable transceivers, per NVIDIA: ~3.5× better optical power efficiency, ~63× better signal integrity (the ASIC-to-optics electrical loss drops from ~22 dB to ~4 dB), ~10× better network resiliency from 4× fewer, externally mounted lasers, and ~1.3× faster deployment. Broadcom's production data at Meta backs the power claim at roughly 65–70% optics power savings.
Quantum-X Photonics (InfiniBand) began shipping in early 2026 at low volume after slipping from late 2025. The first Spectrum-X Photonics Ethernet switches (SN6810, SN6800) started reaching select partners in June 2026, with the production ramp in the second half of 2026. Broadcom's CPO line has been shipping longer — Bailly since 2024–25 and Tomahawk 6 "Davisson" since October 2025.
The switch/integrator layer is led by NVIDIA ($NVDA), with Broadcom ($AVGO), Marvell ($MRVL), and Cisco ($CSCO) competing. The photonics are fabbed by foundries led by TSMC ($TSM), with GlobalFoundries ($GFS) and Tower ($TSEM). The optical engines, lasers, modulators, and assembly come from Coherent ($COHR), Lumentum ($LITE), Fabrinet ($FN), POET ($POET), Applied Optoelectronics ($AAOI), MACOM ($MTSI), and Lightwave Logic ($LWLG), plus private players like Ayar Labs. Fiber and connectors come from Corning ($GLW) and TE Connectivity ($TEL).
Because they pack extreme power density — NVIDIA's Quantum-X Photonics is 3.95 kW in a 4U chassis — and co-packaging puts heat-sensitive optics next to the hottest silicon. Air cooling struggles to move that much heat while keeping the lasers in their stable temperature range, so NVIDIA's CPO switches ship with direct liquid cooling, extending the same loops the GPUs already use (Broadcom's lower-density 51.2T Bailly systems remain air-cooled). That makes cooling names like Vertiv ($VRT) downstream beneficiaries.
CPO removes the pluggable module entirely. LPO (linear-drive pluggable optics) keeps a pluggable module but removes its power-hungry DSP. AEC (active electrical cables) replaces short optical links with signal-boosted copper. CPO wins on power and density for the largest switches; LPO and AEC ($CRDO, $ALAB) compete for shorter reaches and more conservative deployments. They coexist today.
It's one of the most direct ways to play AI networking, but it's a component-supplier theme with real cyclicality and customer-concentration risk (a few hyperscalers and NVIDIA drive demand). The diversified approach is to track the layer — the Photonics / CPO basket — rather than betting on a single design win. This is not financial advice.
Co-packaged optics integrates the optical engine — the part that converts electrical signals to laser light — directly onto the same package as a switch or accelerator chip, instead of housing it in a separate pluggable transceiver module. This shortens the electrical path, cuts power, and improves reliability, which is why it's being adopted for the highest-bandwidth AI networking switches.
At 800G and 1.6T per port, pluggable transceivers consume too much power and fail too often to scale to clusters of hundreds of thousands of GPUs. The optical transceivers in a large AI fabric can burn megawatts on their own. CPO addresses both: NVIDIA cites roughly 3.5× better optical power efficiency and ~10× better network resiliency versus pluggables — power and uptime that translate directly into more usable GPUs.
Four step-changes versus pluggable transceivers, per NVIDIA: ~3.5× better optical power efficiency, ~63× better signal integrity (the ASIC-to-optics electrical loss drops from ~22 dB to ~4 dB), ~10× better network resiliency from 4× fewer, externally mounted lasers, and ~1.3× faster deployment. Broadcom's production data at Meta backs the power claim at roughly 65–70% optics power savings.
Quantum-X Photonics (InfiniBand) began shipping in early 2026 at low volume after slipping from late 2025. The first Spectrum-X Photonics Ethernet switches (SN6810, SN6800) started reaching select partners in June 2026, with the production ramp in the second half of 2026. Broadcom's CPO line has been shipping longer — Bailly since 2024–25 and Tomahawk 6 "Davisson" since October 2025.
The switch/integrator layer is led by NVIDIA ($NVDA), with Broadcom ($AVGO), Marvell ($MRVL), and Cisco ($CSCO) competing. The photonics are fabbed by foundries led by TSMC ($TSM), with GlobalFoundries ($GFS) and Tower ($TSEM). The optical engines, lasers, modulators, and assembly come from Coherent ($COHR), Lumentum ($LITE), Fabrinet ($FN), POET ($POET), Applied Optoelectronics ($AAOI), MACOM ($MTSI), and Lightwave Logic ($LWLG), plus private players like Ayar Labs. Fiber and connectors come from Corning ($GLW) and TE Connectivity ($TEL).
Because they pack extreme power density — NVIDIA's Quantum-X Photonics is 3.95 kW in a 4U chassis — and co-packaging puts heat-sensitive optics next to the hottest silicon. Air cooling struggles to move that much heat while keeping the lasers in their stable temperature range, so NVIDIA's CPO switches ship with direct liquid cooling, extending the same loops the GPUs already use (Broadcom's lower-density 51.2T Bailly systems remain air-cooled). That makes cooling names like Vertiv ($VRT) downstream beneficiaries.
CPO removes the pluggable module entirely. LPO (linear-drive pluggable optics) keeps a pluggable module but removes its power-hungry DSP. AEC (active electrical cables) replaces short optical links with signal-boosted copper. CPO wins on power and density for the largest switches; LPO and AEC ($CRDO, $ALAB) compete for shorter reaches and more conservative deployments. They coexist today.
It's one of the most direct ways to play AI networking, but it's a component-supplier theme with real cyclicality and customer-concentration risk (a few hyperscalers and NVIDIA drive demand). The diversified approach is to track the layer — the Photonics / CPO basket — rather than betting on a single design win. This is not financial advice.