Next-Gen Nvidia, AMD GPUs May Slip to 2028 as AI Wins Chips

What's being reported

A rumor now circulating in PC hardware circles holds that Nvidia's next GeForce generation, widely expected to be called the RTX 60 series, and most of AMD's next-generation Radeon lineup will not reach store shelves until 2028. The claim traces to a single hardware leaker, Kepler_L2, posting on the AnandTech forums, and it was picked up almost simultaneously by a wide swath of hardware outlets in mid-September 135910111213.

According to Kepler_L2, only one new GPU die is currently on track for a 2027 launch: an AMD part codenamed AT2, reportedly built on the RDNA 5 architecture 39101213. Everything else — the rest of AMD's RDNA 5 stack, described in various reports as AT0, AT3 and AT4, and the entirety of Nvidia's Rubin-based consumer lineup — would reportedly wait until 2028 9101112. On the Nvidia side, that lineup is said to include as many as five dies, GR202, GR203, GR205, GR206 and GR207, spanning a range from a flagship RTX 6090 down to mainstream parts 91214.

The leak matters because it contradicts earlier, more optimistic timelines. Some previous reporting, including from leaker kopite7kimi, had pointed to a second-half-2027 launch window for Rubin-based gaming cards, and other rumors floated an even more aggressive early-2027 date 1314. Kepler_L2 directly disputed those claims when asked about them, saying they were incorrect 13.

Why outlets say this is happening

TechSpot's framing is the most pointed among the coverage: consumer graphics cards are reportedly taking a back seat because "AI pays better" 35. That thesis lines up with a separate, and factually distinct, thread of reporting about Nvidia's actual AI hardware business. Nvidia's official announcements around its Rubin data-center platform, unveiled at CES in January and expanded at GTC in March, describe a six-to-seven-chip system built around a new Rubin GPU, Vera CPU, NVLink 6 switching and related networking silicon, claimed to cut inference cost per token by up to 10x and cut the GPU count needed to train mixture-of-experts models by 4x compared with Blackwell 151617. Those are Nvidia's own performance claims, not third-party lab results, and they describe a data-center architecture — not a confirmed consumer product.

Separately, Bloomberg reported, and Reuters, Tom's Hardware and other outlets corroborated, that Nvidia has told major customers to expect AI server prices to rise more than 15% for systems shipping from early 2027, with both Grace Blackwell and Vera Rubin configurations affected 181920. The stated driver is memory, not the GPU die itself: soaring HBM and server DRAM costs are reported to be pushing GPU silicon's share of total server cost down from roughly 80% in prior generations to around 40% in next-generation systems 20. TrendForce data cited alongside that reporting projects server DRAM contract prices climbing 13-18% quarter-over-quarter, with 2027 supply looking tighter still 19.

Put together, the gaming-GPU delay rumor and the AI-server price increase reporting describe the same underlying resource squeeze from two different angles: advanced packaging, HBM and DRAM capacity are scarce, and AI infrastructure is a far more profitable use of that capacity than another round of GeForce or Radeon cards.

Where the reporting agrees

Every outlet covering the GPU delay traces the claim back to the same single source — Kepler_L2's AnandTech forum post — and every one of them is careful to note that neither AMD nor Nvidia has confirmed any 2028 timeline 13910111213. There's also broad agreement on the shape of the leak: AT2 as AMD's lone 2027 exception, and the bulk of RDNA 5 and all of Rubin-based GeForce landing in 2028 910111213. Multiple outlets independently note that this contradicts earlier, more bullish 2027 rumors for Nvidia's Rubin gaming cards 111314. On the AI-infrastructure side, Bloomberg's reporting on 15%-plus server price increases driven by memory costs is corroborated by Reuters, Tom's Hardware, igor'sLab and Value Add Pulse, all pointing to the same underlying driver and the same early-2027 timeframe 181920.

Where it doesn't

The most notable tension is around motive. TechSpot explicitly attributes the delay to AI demand outcompeting gaming GPUs for capacity 35. Other outlets, including Guru3D, VideoCardz and Techtroduce, decline to offer a cause at all, sticking strictly to what Kepler_L2 said and stressing that he did not explain why the schedule looks this way 91112. Notebookcheck adds a wrinkle the others don't dwell on: it notes that Moore's Law Is Dead had already been floating a similar 2028 timeline and that Kepler_L2's claim aligns with, rather than originates, that theory — and it offers its own speculative explanation, tying AMD's single 2027 exception to the RDNA 5-based PS6 console, expected late next year 13. No other outlet raises the console angle.

There's also a dating discrepancy worth flagging. Several aggregator-style writeups (shattered.io, Techtroduce, Guru3D) date the leak and its first pickup by Wccftech to mid-September 2026, while the underlying event — a forum comment — has no independent verification of timing beyond what these secondary sites report 101112. VideoCardz, one of the more established outlets in this space, presents the same information without independently dating the original post, which suggests the September 2026 date may itself be circulating from a single point of attribution rather than being separately confirmed 9.

On specifications, verdictbits.com stands somewhat apart by publishing detailed rumored VRAM, bus width and die-to-tier mappings for an RTX 6090, 6080 and 6070, while explicitly labeling most of those figures as single-source and provisional 14. Guru3D and Techtroduce, by contrast, decline to publish any specification table beyond die names, treating anything more granular as unconfirmed 1112. This isn't a factual contradiction so much as a difference in editorial risk tolerance — some outlets are willing to publish leak-grade numbers with heavy caveats, others aren't.

Finally, it's worth being precise that the Rubin referenced in the GPU-delay rumor and the Rubin detailed in Nvidia's own CES and GTC announcements are not confirmed to be the same silicon. Nvidia's press materials describe a data-center Rubin GPU rated at 50 petaflops of NVFP4 inference compute, deployed in 72-GPU NVL72 racks 1517. Nothing in Nvidia's official materials confirms that a GeForce RTX 60 card will use identical Rubin silicon, memory or packaging — a distinction Guru3D makes explicitly and that the research underlying this story treats as an open question rather than a settled fact 12.

The read that holds up

The evidence best supports treating this as two separate, well-corroborated stories that happen to share a name and a plausible causal link, rather than one confirmed narrative. The 2028 gaming-GPU delay is real as a leak — it is genuinely being reported by a wide range of outlets, all citing the same source, all appropriately hedged. It is not real as a confirmed roadmap, and outlets that present it as near-fact rather than rumor are getting ahead of what Kepler_L2 or anyone else has actually established.

The AI-infrastructure side of the story rests on firmer ground. Nvidia's Rubin platform and its inference-cost claims are on the record from the company itself, and the 15%-plus AI-server price increase is corroborated across multiple financial and trade outlets citing consistent memory-cost mechanics. That reporting doesn't prove AI demand is the reason consumer GPUs are being delayed — Kepler_L2 never said that — but it does establish, independently, that the same memory and packaging capacity gaming GPUs would need is under real strain from AI infrastructure demand. TechSpot's "AI pays better" framing is a reasonable inference from that evidence, not a confirmed fact, and readers should treat it accordingly until AMD or Nvidia says anything at all about 2027 or 2028.