A report from chip analyst firm SemiAnalysis landed this week with a claim that rippled through the AI trade: Nvidia’s next flagship rack system, code-named “Kyber” and built around its Rubin Ultra chips, will not ship until 2028 because of manufacturing problems. The stopgap plan Nvidia considered to bridge the gap was shelved, according to the report, because customers objected. Nvidia responded the same day: the chip roadmap, it said, is fully intact.
CNBC and Tom’s Hardware followed the report with their own coverage, and within days the delay claim was a fixture of sell-side notes. The stakes are clear. Data-center chips and systems account for the overwhelming majority of Nvidia’s revenue, and investors have built the company into the most valuable chip maker on earth on the promise of an uninterrupted product cadence: Blackwell this year, Rubin next, Rubin Ultra after that.
SemiAnalysis, a firm whose supply-chain reporting has moved markets before, said the Kyber system’s problems center on manufacturing, according to people who have read the report. The interim product meant to fill the gap would have required data centers to accommodate a different form factor, and customers pushed back, the people said. Nvidia declined to comment beyond its statement that the roadmap is intact.
The back-and-forth matters beyond one product. Market participants read the report as a potential signal that AI infrastructure investment, the biggest capital-spending wave in technology history, may be approaching its peak. If Nvidia’s product cycle slips, the logic runs, hyperscalers’ procurement plans slip with it, and the trillion-dollar chip complex loses its engine.
That reading may be premature, analysts said. Even a delayed rack system would not shrink demand for AI accelerators, which has exceeded supply for two years. The question was never whether Nvidia can sell everything it makes, one analyst said; the question is whether customers will keep buying everything Nvidia makes at the prices Nvidia charges.
Nvidia has been here before. Reports of delays, overheating, and supply shortfalls have accompanied every major launch in recent years, and each time the company has shipped, sold out, and raised prices. The difference this time is the size of the bet. Rubin Ultra is designed for the largest AI training clusters on earth, the kind that consume as much electricity as a small city, and hyperscalers have signed contracts based on its specifications.
The customer-objection detail is the one that most caught investors’ attention. If the stopgap was rejected, it suggests buyers are less willing to accept workarounds than they were during the pandemic-era chip shortage, when they took whatever they could get. It also suggests the product cycle is becoming the customer relationship: the rack, not the chip, is what locks a data center into a vendor.
Rubin Ultra is the second generation of Nvidia’s Rubin family, named for astronomer Vera Rubin, and the successor to the Blackwell architecture that currently powers most large AI training runs. The Kyber rack system is the product Nvidia designed to make Rubin Ultra easy to deploy: preassembled, liquid-cooled, and wired for the power densities of the next generation of models. A slip there would ripple through every customer plan written around it.
For competitors, a delay would open a window. AMD, Google’s in-house chips, and Amazon’s custom silicon all target the same AI training market, and all have struggled to dent Nvidia’s share while Nvidia ships on schedule. A slip in Kyber would give them a year of undivided attention from buyers, though none has yet shown it can match Nvidia’s software stack and system engineering.
Suppliers would feel it too. Nvidia’s systems depend on advanced packaging capacity at TSMC and on high-bandwidth memory from SK Hynix, Samsung, and Micron. Analysts who track those supply chains said a delayed rack system would free up packaging capacity for other customers, a small but real shift in an industry that has been sold out for years.
Part of the muted reaction reflects fatigue with the delay narrative, but part reflects a genuine debate about the industry’s trajectory. Bullish analysts argue that even a 2028 ship date would leave Nvidia with two full years of selling the current architecture, while bears read any slip as the first crack in a buildout whose scale has no precedent. Both sides agree the Kyber decision, whenever it is announced, will be read as a statement about how long the AI boom can run.
None of this is certain. Nvidia’s denial was categorical, and SemiAnalysis has been wrong before, as its own analysts have acknowledged. The market’s reaction this week, muted for a report that would once have moved the stock sharply, said something else: investors have heard the delay story before, and they are waiting for the earnings call.
Until Nvidia reports, the Kyber question stays open. The report, true or not, has done its job: it put the durability of the AI buildout on the table, where it will sit until the company’s next guidance.
