Next-generation accelerator roadmaps: where Nvidia's Vera Rubin and AMD's Helios stand

The unit of purchase in AI infrastructure is no longer a chip or even a server. It is a liquid-cooled rack that behaves as one machine. Both major accelerator vendors now plan their roadmaps in racks: Nvidia with Vera Rubin and its successors, AMD with Helios. This post lays out what each has said publicly, what has shipped or been confirmed in production, and where the sources disagree or go quiet.

Demand context

Roadmaps matter because the money behind them is large. Nvidia's second-quarter fiscal 2027 results (quarter ended July 26, 2026) reported revenue of $96.2 billion, up 106% year over year, with Data Center revenue of $89.0 billion, up 117%. It guided the third quarter to $108.0 billion, plus or minus 2% ^[4].

Nvidia: Vera Rubin now, Rubin Ultra and Kyber next

What Nvidia states for 2026. The Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6. Nvidia claims it trains large mixture-of-experts models with one-fourth the GPUs of Blackwell and delivers up to 10x higher inference throughput per watt at one-tenth the cost per token ^[1]. The company's technical blog says the rack is "in full production, on track to ship in the second half of 2026" ^[2], and its May 31 Taipei announcement said production shipments are set to begin starting this fall ^[3]. In the August earnings release, CEO Jensen Huang said Vera Rubin is "now in full production" ^[4].

Vera Rubin is also a system, not a single part. The five-rack POD in Nvidia's materials adds a Vera CPU rack of up to 256 CPUs, a Groq 3 LPX inference rack (256 LPU processors, 128 GB of on-chip SRAM), BlueField-4 storage and Spectrum-6 Ethernet with co-packaged optics ^[1][3]. Nvidia lists Dell, HPE, Lenovo and Supermicro among system builders ^[3].

What comes after. Nvidia's technical blog describes a Vera Rubin Ultra NVL576, eight MGX racks of 72 Rubin Ultra GPUs each in one NVLink domain, and Kyber, a rack design that doubles the NVLink domain per rack to 144 GPUs and scales to an NVL1152 system of eight racks using direct optical interconnects. It gives no ship date for either, and no Feynman specifications ^[2]. A Taiwanese outlet's coverage of the GTC keynote places Rubin Ultra and Kyber in 2027 and Feynman in 2028 ^[7]; we could not confirm those years on any Nvidia page we could open, so treat them as reported, not as Nvidia commitments. That article is also internally inconsistent on Kyber's GPU count (576 in one passage, 144 per rack in another), which is a reason for caution.

AMD: Helios and the MI500 promise

At CES on January 5, 2026, AMD presented the Instinct MI455X: 320 billion transistors and 432 GB of HBM4 per accelerator, with AMD claiming up to 10x the performance of the MI355X. The Helios rack combines 72 MI455X-series accelerators with 31 TB of HBM4 and is rated by AMD at up to 2.9 exaflops peak FP4 and 1.4 exaflops peak FP8, per AMD's July 2026 materials ^[9]. At CES it was described as liquid-cooled and weighing nearly 7,000 pounds, with launch announced for later in 2026 ^[5].

AMD also previewed the MI500 series for 2027, on CDNA 6, a 2nm process and HBM4E memory, with a claim of up to 1,000x AI performance versus the MI300X, which was introduced in 2023 ^[5]. AMD's July 23 release repeats that MI500 GPUs are "coming in 2027" and adds MI600 GPUs for 2028 ^[8]. That baseline is two generations back, so the figure says little about competitiveness against current Nvidia parts.

Timing is the open question. AMD's own July 23 release says Helios rackscale solutions are "now in production to be deployed by leading AI companies at gigawatt scale," that systems will come from OEMs including Bull, HPE, Lenovo and Supermicro without dates, and that OpenAI "expects to bring Helios online beginning in the fourth quarter of 2026," with deployments ramping through 2027 ^[8]. The release gives no general availability date for MI455X. The BornCity article of September 30, relaying a September 28 report from dailysynapse.com, says the MI455X ships in small quantities to selected partners in 2026, with a broad market launch expected in the second quarter of 2027 ^[6]. This is a secondary, German-language account of another outlet's report; it is compatible with AMD's "in production" wording (low-volume production and limited deployments first) but is not an AMD statement, so read it as reported.

Side by side

The blanks are the point. Nvidia's primary materials give rack-level interconnect numbers and efficiency ratios but, in the pages reviewed, no per-GPU memory figure, so a direct memory-per-rack comparison cannot be made from sources we can verify.

What to take from it

• Plan for liquid cooling. Helios is liquid-cooled ^[5] and Nvidia's MGX racks can be fully liquid-cooled ^[2]. Facilities without it cannot take either option as-is.

• Treat efficiency ratios as marketing until measured. Figures like 10x throughput per watt compare against Nvidia's prior platform under Nvidia's methodology ^[1].

• Expect staggered availability. Announced, in production, and generally available are three different dates, and the sources above show all three in play.

Practical implication

For most businesses, rack-scale hardware arrives through cloud providers, and the near-term decision is which capacity to reserve and for how long. Ask providers for the specific accelerator generation, delivery dates in writing and independently measured cost per token, rather than relying on vendor ratios.