Nvidia Confirms Copper Isn't Going Anywhere: GTC 2026 Keynote Takeaways for DAC Buyers
Estimated reading time: 7–9 minutes
Key Takeaways
- Nvidia's next-generation Feynman platform will use both copper and co-packaged optics (CPO)—not one instead of the other—for scale-up and scale-out interconnects.
- At GTC 2026, CEO Jensen Huang told the audience the industry needs far more capacity across copper, optics, and CPO to keep pace with AI growth—copper wasn't cut from the roadmap.
- Nvidia networking SVP Gilad Shainer said the real limit on copper isn't cost or supply chain—it's distance. Wherever copper physically reaches, Nvidia defaults to it.
- Copper stays the default for tightly-coupled links like NVLink72 and NVLink144; optics take over once a domain spans multiple racks (1152+ GPUs).
- For buyers: DAC remains the lowest-cost, lowest-power choice for in-rack and short-reach AI server links, and Nvidia's own roadmap backs that up through at least 2028.
What Did Jensen Huang Actually Say at GTC 2026?
During his keynote at GTC San Jose 2026 on March 16, Huang addressed a question that's been circulating since co-packaged optics started grabbing headlines: is Nvidia abandoning copper for its next-generation interconnects?
The answer, in Huang's own words, was no. Speaking about the scale of Nvidia's ecosystem needs, he said: "We need a lot more capacity for copper, for optics, for CPO." He went on to explain that Nvidia is working with its full partner ecosystem—cable makers, connector suppliers, and optics vendors alike—to build out that capacity together, rather than picking a single winner.
That framing matters. Huang wasn't describing copper as a stopgap technology being phased out in favor of optics. He was describing a buildout where both technologies need to scale in parallel, because AI infrastructure is growing faster than either one alone can supply.
This Isn't a New Position for Huang
The GTC 2026 comments echoed something Huang had already told Taiwanese press back in August 2025. In an interview with iNEWS that circulated widely after being shared by semiconductor analysis firm SemiAnalysis, Huang was direct about Nvidia's preference: "We should stay with copper for as long as we can."
Nvidia would only move to silicon photonics "if we must," he added at the time—not because optics are inherently superior, but because copper runs out of reach at certain distances and speeds. Nearly a year later, that logic hadn't changed. It just got a more detailed roadmap attached to it.
Why Copper Still Wins Inside the Rack
Nvidia networking SVP Gilad Shainer put the engineering case plainly at GTC 2026: copper is passive, draws effectively zero power, and has no active components to fail. When a connection fits within copper's reach, Shainer explained, Nvidia uses copper—full stop. The decision isn't about cost-cutting or supply chain hedging; it's simply the more reliable, more efficient option whenever distance allows for it.
That's the same case DAC buyers have been making for years: passive copper twinax costs less per port, draws less power per link, and has no transceivers or lasers to fail. Nvidia's own hardware roadmap now backs that engineering logic at the platform level.
Where the Line Between Copper and Optics Falls
According to Shainer, the deciding factor is domain size. NVLink72 and NVLink144 configurations fit within copper's practical reach, so Nvidia uses copper there. Once a deployment scales to something like an NVLink1152 domain spanning multiple racks, the distance moves beyond what passive copper can support, and co-packaged optics take over.
In other words, the choice isn't "copper vs. optics" as competing philosophies—it's copper where it reaches, optics where it doesn't. That's the same distance-driven logic that governs the DAC-vs-AOC decision on any AI server rack today. See our DAC vs. AOC buyer's guide for the specifics.
What This Means for Your Next AI Server Build
If you've been holding off on DAC cabling because of the co-packaged optics headlines, GTC 2026 is a reason to reconsider. Nvidia's own CEO and networking leadership just confirmed, on stage, that copper is a permanent part of the roadmap through at least the Feynman generation (expected around 2028)—not a technology being sunset.
For most in-rack and short-reach GPU interconnects, the calculus hasn't changed: passive DAC is still the lowest-cost, lowest-power way to connect switches and servers that sit within a few meters of each other. Save optics for the links that actually need the reach.
Frequently Asked Questions
Is Nvidia moving away from copper cabling in AI data centers?
No. At GTC 2026, Nvidia confirmed its next-generation Feynman platform will use both copper and co-packaged optics for scale-up and scale-out interconnects. Copper remains the default wherever distance allows.
What did Jensen Huang say about copper at GTC 2026?
Huang told the GTC 2026 keynote audience that the industry needs "a lot more capacity for copper, for optics, for CPO," framing copper as a growth area alongside optics rather than a legacy technology being phased out.
When should I use DAC instead of optical cabling?
DAC (Direct Attach Copper) is the better choice for short, in-rack links—typically up to about 2–5 meters—where its lower cost and near-zero power draw outweigh the flexibility of optics. Once distance or rack density pushes past copper's practical reach, active optical cables or transceivers take over. See our full DAC vs. AOC comparison for details.
Will copper still be relevant with Nvidia's Feynman architecture?
Yes. Nvidia has said copper will remain part of its interconnect strategy for Feynman, expected around 2028. The company's networking leadership has framed the copper/optics split as a function of distance, not a phase-out timeline.
Sources
This article draws on Nvidia's GTC San Jose 2026 keynote (March 16, 2026) and reporting from SDxCentral, including its earlier coverage of Huang's August 2025 comments on copper and silicon photonics.