Why copper refuses to die, and who gets paid while it lives
One research note calls copper dead; the same month Credo is stretching copper links to 30 meters and Marvell says it plans to intercept copper-based solutions. The disagreement is the trade story.
The argument inside the rack
Two things were said about copper within weeks of each other, and they cannot both be the whole truth. In August, a research note recorded that Astera Labs -- whose retimers clean up high-speed signals inside AI servers -- declares copper dead and moves towards NPO/CPO technology, shorthand for near-package and co-packaged optics that put light sources next to the chip.1
Then in September, Credo Technology ($CRDO), a connectivity chip and cable supplier, said on its FY2027 Q1 call that it is developing active link cables using MicroLED technology to extend length to 30 meters and to address reliability and power efficiency issues in scale-up opportunities.2 That is not a company planning copper's funeral. That is a company extending its reach.
The resolution is simple once you stop treating copper as one market. Inside the rack, over short runs, electrical links stay cheap, low-latency and already qualified. The fight is over the distance where copper runs out of road -- and the whole industry is spending to move that line.
Why the seam is the business, not the metal
The money is not in copper or optics. It is in the engineering that keeps a signal intact as data rates climb. Astera Labs' retimer became an essential part of every AI server by solving high-speed signaling over copper.5 That is a position earned on copper's limits, which is exactly why a company sitting on it can also push towards optics without contradicting itself.
Credo attacks the same seam from the cable side. MicroLED-based active cables reaching 30 meters sit between passive copper and full optical links, aimed at scale-up -- connecting accelerators into one larger machine -- where reliability and power are the binding constraints.2
So the question for an investor is not "copper or light". It is who owns the transition: the component that makes the old medium go further, and the one that makes the new medium affordable.
Marvell is explicit about it
Marvell Technology, a custom silicon and data-center connectivity supplier, put the strategy in plain words on its FY2027 Q2 call in August: it is developing UAL switching and plans to intercept electrical or copper-based solutions.6 "Intercept" is the tell -- the copper installed base is the target market, not an obstacle.
The scale behind that ambition shows up in its own guidance. Marvell's Data Center revenue is expected to grow more than 60% year-over-year in fiscal 2028, driven by strong growth across all key data center businesses, on the company's own stated expectation.7 If that holds, it is a company funding a switching push from a growing base rather than a standing start.
Marvell also disclosed an expanded commercial agreement and associated warrant with a key hyperscaler for custom silicon programs.8 A warrant attached to a customer agreement is a commitment signal: the buyer is tying itself to the roadmap, not sampling it.
The platform owners set the rules
Interconnect choices are not made in a vacuum; they are made inside whichever platform the accelerators live in. NVIDIA added NVHBM to NVLink Fusion, a custom high-bandwidth memory that moves the memory controller into the base die.9 When the platform owner redraws where the controller sits, every link and every retimer around it gets re-specified.
Broadcom is characterized as an AI Chips play in the 2026-2027 AI buildout phase, in the same research framing that places Credo Technology as a Photonics play for that period.3,10 Those two labels, set against each other, are the copper-to-light story in miniature: the chip that generates the traffic, and the link that carries it further.
In our data, NVIDIA leads, Broadcom is second and Astera Labs is third. Read that as where our scoring sees the strongest position today across this seam -- the platform, the custom-silicon franchise, and the signal-integrity layer -- not as a view on who wins the physics argument.
The number that frames everything
Capex-flow estimates measure how much 2026 AI capital spending our model expects to land at one company. For NVIDIA, that estimate is about $265.2B, in a range of $192.9B to $308.6B, as of July 2026. It is a flow to that single company, not a sector total, and names below medium confidence are left out of these estimates entirely.
That figure is the budget the interconnect fight is being waged inside. Links, retimers and cables are a slice of a spend of that order, which is why both copper extension and optics can be funded at once -- the buyer does not have to choose a single answer this cycle.
The constraint is physical, and NVIDIA named it: its entire supply chain is challenged and running flat out.4 In that condition, a design win is only worth what the chain can actually ship, and qualified, already-manufacturable copper solutions have an advantage that slideware does not.
What to watch from here
Three concrete markers would settle the argument faster than any narrative. First, whether Credo's 30-meter MicroLED cables move from development to shipping volumes, since the company frames them around reliability and power in scale-up.2 Second, whether Marvell's UAL switching actually intercepts copper-based sockets, which would show up as data-center mix rather than as a slogan.6
Third, watch Credo's own spending. It reported capex up 53% quarter-over-quarter for the quarter ending August 1, 2026, flagged as a semi-spend surprise above the 25% threshold.11 Companies do not step up capital spending that sharply for a product line they expect to be short-lived; it is a readable, dated signal of internal conviction.
The constructive way to hold this: copper is not a dying asset class, it is a shrinking distance. Each generation of chips pushes the crossover point closer to the package, and the companies that get paid are the ones selling the bridge at whatever distance the crossover happens to sit this year. That is a recurring revenue pattern, not a one-time switch.
What would change our view
- Credo's 30-meter MicroLED active cables slipping out of development, or arriving without the reliability and power gains the company describes2
- Marvell's data center growth tracking away from the more-than-60% year-over-year expectation it stated for fiscal 20287
- Evidence that NPO/CPO adoption moves faster than the copper-extension roadmap, validating the view that copper is finished inside the rack1
- A supply chain that stays challenged and flat out long enough to delay new interconnect designs rather than old ones4
In your own deployments, where does copper actually stop being the cheaper answer -- inside the rack, at the row, or somewhere else entirely?
Answer on XSources
- 1Industry research note, research note, Aug 13, 2026.
- 2Earnings call CRDO FY2027 Q1, transcript, Sep 1, 2026.
- 3Industry research note, research note, Aug 22, 2026.
- 4Earnings call NVDA FY2027 Q2, transcript, Aug 26, 2026.
- 5Industry research note, research note, Sep 16, 2026.
- 6Earnings call MRVL FY2027 Q2, transcript, Aug 27, 2026.
- 7Earnings call MRVL FY2027 Q2, transcript, Aug 27, 2026.
- 8Earnings call MRVL FY2027 Q2, transcript, Aug 27, 2026.
- 9Industry research note, research note, Sep 24, 2026.
- 10Industry research note, research note, Aug 22, 2026.
- 11Credo Technology financial statements, filing data, Sep 3, 2026.
For information only. Not investment advice.
Related research
All research →
The gas pipes behind AI power: what this week's reports tell you to watch
The latest gas reports show two clocks running at once. Majors and producers are posting strong quarters now, while new pipelines are being planned for 2030 and later.
Co-packaged optics is being scheduled, not debated
Six companies across memory, networking, lasers and foundry have now put dates on moving light into the package. The timeline runs from qualified lasers today to a 2028 foundry ramp.
Welding is getting busier: what the latest calls say about process tooling
The latest reports from welding and dispensing suppliers show volume and orders starting to do more of the work than price. Here is what Lincoln Electric, Graco, ITW and Nordson said, and what to watch next.