The real bottleneck in AI is no longer the chip. It is the wire. As data centers cram hundreds of thousands of GPUs into a single building, the copper and optics connecting them now burn more power and fail more often than the processors themselves. Nvidia just started shipping the fix, and the numbers are hard to ignore: up to 3.5 times less power, ten times the reliability, and a single switch moving 400 terabits per second.
What Actually Happened
Nvidia has begun shipping its next-generation Spectrum-X co-packaged optics switches to select partners, manufactured in collaboration with TSMC, with broader availability in the second half of 2026. The flagship Spectrum SN6800 delivers up to 409.6 terabits per second of aggregate bandwidth, with 512 ports running at 800 gigabits per second across four ASICs. The headline engineering claim is that integrating the optics directly onto the switch silicon cuts power consumption by up to 3.5x and improves resiliency by 10x compared with conventional pluggable optical modules.
The breakthrough is physical, not just architectural. In a traditional switch, the signal travels from the ASIC across the circuit board to a pluggable transceiver at the faceplate, and every centimeter of that copper path leaks energy and adds points of failure. Co-packaged optics, or CPO, moves the optical engine to sit directly beside the switch ASIC, so the fiber connects almost at the chip itself. That shortened electrical path slashes loss and drops the power needed per optical connection to as low as 9 watts. Multiply that saving across the tens of thousands of links in a single AI cluster and the aggregate figure becomes one of the largest line items a data center operator can attack.
Nvidia is positioning the technology explicitly around what it calls AI factories, the giant single-purpose buildings being constructed to train and serve frontier models. Production capacity is set to expand through the second half of 2026, and the Spectrum-X Ethernet Photonics line is being marketed as the networking backbone for massive generative AI and agentic workloads. The pitch is simple: the compute is ready, the network is the wall, and CPO is how you take the wall down.
Why This Matters More Than People Think
For most of the AI buildout, the public conversation has fixated on GPUs. How many H100s, how many Blackwells, who has the biggest order book. That framing misses where the marginal pain has moved. In a cluster with hundreds of thousands of accelerators, the GPUs spend a punishing share of their time waiting for data to arrive from other GPUs. The interconnect determines how efficiently all that silicon can actually be used, and pluggable optics had become a power hog and a reliability liability. A switch fabric that cuts networking power by 3.5x and failures by 10x does not just save electricity. It raises the effective utilization of the most expensive hardware on Earth.
The power math is the part that should reframe how investors think about the sector. AI data centers are increasingly constrained not by capital or chips but by the megawatts they can secure from the grid. Every watt spent moving bits between switches is a watt not spent on computation. When Nvidia claims it can drop optical networking power to as low as 9 watts per link, it is effectively handing operators back a slice of their power budget to spend on more GPUs. In a world where a single campus can draw as much electricity as a mid-sized city, freeing up power is the same as building capacity, and it is far cheaper than negotiating a new substation.
There is a reliability dimension that matters even more at frontier scale. Training a state-of-the-art model is a single distributed computation that can run for weeks across a hundred thousand chips, and a single failed optical link can stall or corrupt the whole run. Pluggable transceivers, with their connectors and their heat, are a notorious source of those failures. A 10x improvement in resiliency means longer uninterrupted training runs, fewer expensive restarts, and less engineering time spent hunting for the one flaky cable in a building. At the scale of a billion-dollar training cluster, reliability is not a convenience. It is a direct input to how fast a lab can ship its next model.
The cooling angle compounds the power story in a way the headline numbers understate. Pluggable optical modules do not just consume electricity, they dump that energy into the chassis as heat, and removing that heat costs still more power for fans and chilled water. By cutting the optical power draw at the source, CPO reduces both the direct consumption and the cooling overhead that shadows it, so the real facility-level saving is larger than the 3.5x figure on the device itself. In data centers where cooling can rival compute for total energy use, attacking the heat at its origin rather than fighting it downstream is the kind of structural win operators have been chasing for years.
The Competitive Landscape
Nvidia is not the only company that sees co-packaged optics as the next battleground. Broadcom, the dominant merchant supplier of switch silicon, has been shipping its own CPO-enabled Tomahawk platforms and has deep relationships with the cloud providers who design their own networking gear. Marvell, Cisco, and a cluster of silicon-photonics startups are all chasing the same thesis: that the optical interconnect is where the next decade of data center value accrues. Nvidia's advantage is not that it invented CPO. It is that Nvidia sells the GPUs, the switches, and the software as one integrated system, so it can co-design the network and the compute in a way a merchant-silicon vendor cannot.
The historical parallel is the shift from standalone networking gear to vertically integrated systems that defined an earlier computing era. When a single vendor controls the processor, the interconnect, and the orchestration layer, it can extract performance that a best-of-breed assembly of parts cannot match, and it can lock customers into its stack. That is precisely the playbook Nvidia ran with CUDA on the software side, and it is now extending the same logic down into the physical network. The competitive question for Broadcom and the hyperscalers who buy from it is whether an open, multi-vendor optical ecosystem can keep pace with a single company optimizing every layer at once.
The hyperscalers are the wildcard. Google, Amazon, Microsoft, and Meta have the scale and the engineering depth to design their own switches and increasingly want to, precisely to avoid paying Nvidia's integration premium. CPO raises the stakes of that decision, because co-packaging the optics with the ASIC requires advanced packaging capacity at TSMC that is already scarce and increasingly spoken for. Whoever controls that packaging supply effectively gates who can ship CPO at volume, and Nvidia's manufacturing partnership with TSMC gives it a front-of-line position that a hyperscaler building its own switch may struggle to match in the near term.
The supply-chain implication runs deeper than most coverage acknowledges. Co-packaged optics fuses two industries that used to operate separately: the photonics makers who built pluggable transceivers and the advanced-packaging lines that assemble high-end chips. Folding the optical engine into the switch package means the optics now compete for the same scarce packaging slots as the GPUs themselves, which is a zero-sum fight inside TSMC's most constrained facilities. That collision is why control of packaging capacity, not transistor design, is becoming the chokepoint of the entire AI hardware stack, and why a company that can reserve that capacity wields leverage over rivals that no benchmark can capture.
Hidden Insight: The Networking Bill Is Quietly Becoming the Compute Bill
The number nobody outside the data center world appreciates is what fraction of an AI cluster's cost and power now goes to moving data rather than processing it. As clusters scale, the interconnect grows faster than linearly, because connecting more nodes requires disproportionately more switching. The result is that networking has crept from a rounding error to a large share of both the capital budget and the power draw of a frontier cluster. Nvidia shipping CPO is the clearest public admission yet that the industry has hit the point where the wires are as much of a problem as the chips.
This reframes Nvidia's whole strategy in a way the GPU-centric narrative misses. The company is not defending a chip business. It is defending a systems business, and the moat is shifting from the accelerator to the fabric that ties accelerators together. Once an operator commits to Nvidia's CPO switches, its Spectrum-X Ethernet, and its software, the cost of mixing in a competitor's GPU or a third-party switch rises sharply, because the performance comes from the integration. The interconnect is becoming the lock-in, and CPO is the most physical, hardest-to-rip-out expression of that lock-in yet.
There is a forward signal here about where the constraint moves next. If CPO solves the power and reliability of the switch-to-switch network, the next wall is the connection between the GPU and its memory, and between racks at building scale. Nvidia has already telegraphed silicon photonics moving deeper into the system, eventually toward optical links inside the rack and perhaps onto the accelerator package itself. The trajectory points toward a data center where light, not copper, carries data at every level except the last few millimeters. The company that owns that transition owns the physical layer of AI, and Nvidia is moving to own it before anyone else can build the manufacturing base.
The bear case, however, deserves a clear hearing. Co-packaged optics has been promised as imminent for the better part of a decade, and it has repeatedly slipped because the manufacturing is genuinely hard. When the optics are soldered next to the ASIC, a single failed laser can mean replacing an entire expensive switch rather than swapping a cheap pluggable module, which is the exact opposite of the serviceability data center operators have spent years optimizing for. Skeptics point out that the 3.5x and 10x figures are vendor numbers measured under favorable conditions, and that real-world yield, repairability, and field reliability at volume remain unproven. The risk is that CPO ships to a handful of marquee partners, generates impressive benchmarks, and then runs into the unglamorous reality that operators do not want to throw away a 400-terabit switch because one laser died.
What to Watch Next
The first thing to track over the next 90 days is which partners actually take delivery and put these switches into production clusters rather than test labs. Nvidia says it is shipping to select partners, and the identity of those partners matters. If the major AI labs and a hyperscaler or two deploy CPO at scale, it validates the technology has crossed from demo to dependable. If shipments stay confined to controlled pilots through year-end, it suggests the manufacturing and serviceability concerns are still being worked out behind the scenes.
By the 180-day mark, watch TSMC's advanced packaging capacity and how it is allocated. CPO volume is gated by packaging, and any sign that Nvidia is locking up that capacity tells you how aggressively it intends to push competitors out of the optical switch market. Watch Broadcom's response too, since its CPO roadmap and its relationships with hyperscalers building custom silicon are the most credible counterweight. A public CPO deployment by a hyperscaler using non-Nvidia switches would be the strongest evidence that the optical interconnect will be a contested market rather than another Nvidia monopoly.
The deepest indicator is power. Watch the disclosed power-usage figures from the next wave of AI data center buildouts and whether networking's share of the total starts falling. If operators begin reporting that they fit more compute into the same megawatt envelope, CPO will have delivered on its core promise and the technology will become table stakes within two years. If the power figures do not budge, it will mean the savings looked better on a slide than on a utility bill, and the wire will remain the wall a while longer.
In the AI buildout the chip was never the whole story. The company that wins is the one that figured out the bottleneck moved into the wires, and started shipping light before anyone else admitted the copper had run out.
Key Takeaways
- Nvidia began shipping Spectrum-X co-packaged optics switches to select partners, built with TSMC, with broad availability in H2 2026.
- The Spectrum SN6800 moves up to 409.6 Tb/s across 512 ports at 800 Gb/s, spanning four ASICs in one switch.
- CPO cuts networking power up to 3.5x and improves reliability 10x by placing optical engines beside the ASIC, with links as low as 9 watts.
- Power, not chips, is now the binding constraint on AI data centers, so every watt saved on the network becomes capacity for more GPUs.
- Manufacturing risk is real: CPO serviceability and yield at volume remain unproven, and TSMC packaging capacity gates how fast it can scale.
Questions Worth Asking
- If the interconnect is becoming the lock-in, does Nvidia's real moat live in the GPU at all, or in the fabric around it?
- Can an open, multi-vendor optical ecosystem keep pace with one company co-designing the chip, the switch, and the software together?
- When power is the true constraint on AI, how should you value a company by the megawatts it can secure rather than the chips it can buy?