Key points
- Research firm SemiAnalysis says Nvidia's Kyber NVL144 rack, the cabinet meant to hold 144 of its 2027 Rubin Ultra chips, is delayed more than 12 months to 2028, blamed on a 78-layer circuit board that is proving hard to manufacture.
- Nvidia's backup plan, a back-to-back rack called NVL72x2, has reportedly been scrapped after cloud providers rejected the design, and the report says Nvidia has no proven way to scale up Rubin Ultra until its Feynman chips arrive, an opening for AMD's MI500X.
- A separate SemiAnalysis report from June 30 says Nvidia is cutting Rubin Ultra from four compute dies to two, roughly halving its real-world performance versus the original design.
- NVDA stock barely reacted, trading up as much as 1.7% on the day, because the same research note also said Nvidia's second-half fiscal 2027 data center revenue would run 20% above Wall Street's estimate.
- The photonics trade splits on the news: AMD, Broadcom (AVGO), Marvell (MRVL), Astera Labs (ALAB), and Credo (CRDO) look like winners from a longer copper era, Applied Optoelectronics (AAOI) and Ciena (CIEN) are neutral, Coherent (COHR) and Lumentum (LITE) face a slower payoff on Nvidia's $4 billion CPO investment, and POET Technologies (POET) has the most direct exposure.
On July 6, a single research report did more to move the Nvidia story than Nvidia itself did. SemiAnalysis published research saying Nvidia's next rack-scale AI system has slipped by more than a year. The backup plan for that exact scenario has reportedly been scrapped too, and the flagship 2027 chip has already been quietly downsized to compensate. Nvidia hasn't confirmed any of this. The stock barely blinked anyway. Here's what SemiAnalysis is claiming, and why it didn't move the needle.
The rack that slipped a year
Kyber NVL144 is Nvidia's next rack-scale design, a cabinet meant to wire 144 of its most advanced chips together into one giant computer. Jensen Huang showed it off at GTC 2026, Nvidia's big developer conference, as the rack meant to house the 2027 Rubin Ultra chip lineup. SemiAnalysis says that timeline is gone: Kyber now slips to 2028, more than 12 months late. The cause is specific rather than vague. A 78-layer printed circuit board sits in the middle of the rack, the "midplane" every chip's signals and power have to cross, and SemiAnalysis says it keeps running into problems with signal integrity, power delivery, and heat that Nvidia's manufacturing partners haven't cracked yet.
This is a hardware problem, and hardware problems at this scale take years to fix, not weeks. A board with 78 layers has to carry enormous power and extremely fast signals through the same block of material without the layers bleeding into each other. Getting that to work at commercial volume is a manufacturing challenge in its own right, separate from whatever chip eventually sits on top of it.
The backup plan got cancelled too
Nvidia reportedly had a fallback ready in case Kyber slipped: NVL72x2, a design that bolted two current-generation Oberon racks back-to-back with plain copper connections instead of the newer architecture. SemiAnalysis says that plan is dead too. Cloud providers who would have had to actually run the thing pushed back hard, calling it awkward to operate and too expensive to deploy at scale. Nvidia dropped it.
That leaves a gap. NVL576, the larger system that links eight racks together with co-packaged optics instead of copper, faces the same delay risk or a limited initial rollout, per the report, because CPO manufacturing isn't ready either. The optical switch that system needs, CPO NVSwitch, won't be ready until Nvidia's Feynman generation, two full product cycles from now, according to SemiAnalysis. So for the time being, Nvidia has no proven way to scale Rubin Ultra past a single rack. SemiAnalysis names AMD's upcoming MI500X as the rival best positioned to exploit that gap. Nvidia's stopgap, per the report, is volume: sell more of the current Oberon Rubin and Oberon Rubin Ultra racks to make up the difference.
Rubin Ultra itself got smaller
A separate SemiAnalysis report from June 30, picked up by Tom's Hardware, adds one more piece. Nvidia's original Rubin Ultra packed four compute dies and 16 HBM4E memory stacks onto a single package, a full terabyte of memory on one chip. SemiAnalysis says Nvidia dropped that design for a smaller two-die, eight-stack version after running into manufacturing execution problems with the advanced packaging. The new chip gives up a lot to get there: SemiAnalysis pegs it at roughly half the compute performance and half the memory of the original four-die design, traded for something Nvidia can actually build at volume. Nvidia hasn't said a word about this publicly, and SemiAnalysis's reports have been disputed before, so this is worth treating as a serious claim rather than a confirmed fact until Nvidia weighs in.
Why NVDA barely moved
Given three separate delays and cancellations in one report, you'd expect Nvidia's stock to sell off hard. It didn't. NVDA traded up as much as 1.7% on the day, because the same SemiAnalysis note carried a second, far more bullish claim buried in it: Nvidia's data center revenue in the second half of fiscal 2027 is tracking about 20% above what Wall Street currently expects, thanks to a cleared supply bottleneck and a faster-than-expected Vera Rubin ramp. Investors picked the nearer-term number over the 2028 problem. It also helps that Nvidia wasn't exactly riding high going into this. The stock is up only about 4.6% year to date, coming off its worst six-month stretch since late 2022, so a decent amount of bad news was probably priced in already.
AMD moved more. Shares jumped as much as 2.5% in premarket trading, recovering roughly half of a sharp drop from the prior session, in a move several outlets tied directly to the scale-up gap this report describes. AMD hasn't shipped anything yet to prove it can fill that gap, but a rival with no proven high-end scaling path for a year or more is the kind of headline that gets bought first and questioned later.
What this means for the photonics trade
That distinction is doing a lot of work here. Kyber, NVL72x2, NVL576, and CPO NVSwitch are all scale-up hardware, and scale-up is exactly what just got delayed. The gear that connects whole racks to each other, rather than chips to each other, runs on Ethernet and InfiniBand instead, a layer nobody in this report claimed is broken. Sort the photonics names by which side of that line they actually sit on and you get three genuinely different stories, not one.
Start with who benefits outright. AMD, Broadcom (AVGO), Marvell (MRVL), Astera Labs (ALAB), and Credo (CRDO) all sell the retimers, gearboxes, and switch silicon that scale-up traffic runs through right now, and every extra year copper sticks around is billable time for that group. A market-tracking account that broke this news down by ticker put all five in its bullish column, for exactly that reason.
Applied Optoelectronics (AAOI) and Ciena (CIEN) are a different animal. Their revenue comes from the 800G and 1.6T pluggable transceivers and metro gear on the scale-out side, the same backlog behind AAOI's June selloff. Whether CPO NVSwitch ships in 2027 or 2029 doesn't show up on either company's income statement, which is likely why that same account called AAOI neutral rather than bullish or bearish; Ciena's business sits close enough to that layer to earn the same shrug. If there's an upside for either name, it's secondhand: Nvidia says it needs to sell a lot more Oberon Rubin and Oberon Rubin Ultra racks to cover for the smaller chip, and every additional rack in a data center needs its own scale-out ports. That's a real number. It just doesn't trace back to Kyber specifically.
Coherent (COHR) and Lumentum (LITE) can't claim the same distance, because Nvidia's money is already inside their businesses. On March 2, 2026, Nvidia wired $2 billion apiece into both companies, four billion combined, earmarked for co-packaged optics: supply commitments, capacity rights, and in Coherent's case, access to Nvidia's own CPO designs. Their pluggable-transceiver revenue is fine either way. What isn't quite as fast is the payback clock on that $4 billion, now that the CPO NVSwitch piece it was funding points to the Feynman generation instead of the next one. Their current business is untouched. Their next one just got pushed out.
POET Technologies (POET) doesn't have that cushion. It's a much smaller, earlier-stage silicon photonics company with little existing pluggable business to hide behind, so its stock lives or dies on how soon the market believes photonic integration is actually coming. That showed up in the numbers on June 9, when SemiAnalysis first floated the CPO delay: POET dropped about 12%, worse than Lumentum's roughly 8% and Ciena's roughly 7%, despite being the smallest name in the group by a wide margin. A single-thesis stock takes a harder hit from bad thesis news than a company with somewhere else to point.
This isn't the first time this shop has moved this sector
July 6 wasn't the first morning SemiAnalysis rearranged this trade. Back on June 9, the firm argued that full-scale CPO production was really a 2028 or 2029 story, later than the industry itself had been signaling, and optical stocks got hit hard that day. AAOI lost about 17%, and Coherent, Lumentum, and Ciena each gave up somewhere in the 7% to 11% range. A rival analysis pushed back within days, picking apart the yield math behind that call and pointing to roughly 100 million laser units already under order as evidence the ramp was moving faster than SemiAnalysis had modeled. We wrote about that stretch when it happened, and the underlying AI optics demand never actually cracked, whatever the charts did that week. Two research notes, about four weeks apart, both moved the same handful of stocks before anyone outside SemiAnalysis could check the work.
What to watch next
Nvidia hasn't said a public word about any of this, so the first thing to watch is whether that changes, on the next earnings call or sooner. From there, track AMD's MI500X through 2027 for actual hyperscaler orders, not just a favorable headline. And watch the Rubin systems already shipping to AWS, Microsoft Azure, and Google Cloud this fall: if that ramp keeps accelerating, a 2028 Kyber slip will matter a lot less by the time it arrives than it looks like it should today. None of this is investment advice. What ships in the next four quarters moves a stock more than a rack that's two years out, and Nvidia is currently saying the near-term part is on track.
