NVIDIA's Kyber rack, the system Rubin Ultra NVL576 ships in, has slipped more than a year. NVIDIA disputes the framing without denying the date.
Here's the part that actually matters for anyone renting GPUs this year: this isn't really Rubin Ultra news. It's GB300 NVL72 and Vera Rubin NVL72 news. Every month Kyber slips is a month those two racks stay the fastest thing a cloud customer can actually book, which changes the math on reserving one now versus waiting.
TL;DR: Is Rubin Ultra NVL576 Delayed to 2028?
- Cause: A 78-layer PCB midplane mounting Kyber's 144 GPUs has manufacturability problems spanning signal integrity, power delivery, thermal design, and layer count.
- NVIDIA's line: "Our roadmap is intact," a spokesperson told CNBC, while confirming base Vera Rubin NVL72 still ships in fall 2026.
- What's unaffected: GB300 NVL72 and Vera Rubin NVL72 stay the practical ceiling through 2027 at minimum, both open for reservation on Spheron now.
- Market reaction: NVIDIA shares fell 1.4% the day the report circulated, 7 July 2026.
What Kyber Actually Is and Why Rubin Ultra Needs It
Kyber is the rack chassis NVIDIA built specifically for Rubin Ultra. It isn't a chip, it's the physical system that houses the chips, and it's a genuine departure from the tray design NVIDIA has used since GB200. GB200, GB300, and the current Vera Rubin NVL72 all share a horizontal compute tray layout, 72 GPU packages connected through an NVLink switch backplane. Our Vera Rubin NVL72 spec guide covers that architecture, and it's the system Kyber replaces at the high end once Rubin Ultra ships.
Kyber switches to vertical trays, and NVL144 scale means twice the GPU packages of an NVL72 rack in roughly the same floor footprint. NVIDIA showed the Kyber architecture publicly at GTC in spring 2026, months before the delay report surfaced, so the design itself wasn't a secret. What's reportedly in trouble is the 78-layer PCB midplane that physically mounts and connects all 144 GPU packages in those vertical trays. A midplane at that layer count is a genuinely hard board to manufacture, and SemiAnalysis's reporting describes problems touching signal integrity, power delivery, thermal design, and the achievable layer count itself, not one isolated defect.
NVL576 is the step beyond Kyber itself: it links eight Kyber racks together over co-packaged optics into a single larger domain.
On interconnect: NVIDIA has stepped up an NVLink generation with each rack-scale family so far, NVLink 4 on Hopper's H100/H200, NVLink 5 on Blackwell's B200/B300, and NVLink 6 on Vera Rubin NVL72 at 3.6 TB/s per GPU. If that pattern holds, Kyber would be expected to carry NVLink 7, though NVIDIA hasn't confirmed a number for Rubin Ultra's fabric.
The Delay: What CNBC and SemiAnalysis Reported, and NVIDIA's Response
It would be consistent with "R200" already being the more common 2026 industry shorthand for a dual-die Rubin package, but a dual-die base Rubin chip and a dual-die Rubin Ultra package are not confirmed to be the same decision.
NVIDIA's public response has been narrow. Asked about the report, a company spokesperson told CNBC, "Our roadmap is intact." That's a statement on roadmap integrity without addressing specific delays.
The market read it as real news: NVIDIA shares fell 1.4% on 7 July 2026, the day the report circulated.
Rubin Ultra NVL576 Specs That Still Apply Once It Ships
Independent of the delay, NVIDIA's own roadmap disclosures put Rubin Ultra at roughly double Vera Rubin's FP4 compute, about 100 PFLOPS of FP4 per GPU against Vera Rubin's 50 PFLOPS, with 1 TB of HBM4e memory per GPU, deployed in NVL576 "Kyber" racks, as detailed in our Rubin GPU specs guide. Those figures were disclosed assuming the originally planned package design. If the two-die scale-back SemiAnalysis describes turns out to be accurate and NVIDIA confirms it, the per-GPU compute figure is the one most likely to move, since it's a direct function of die count. The memory and rack topology numbers are less obviously affected by a die-count change and more likely to hold.
Treat this section the way NVIDIA treats its own preliminary tables: directionally right, not a number to build a procurement plan around until NVIDIA publishes a confirmed spec sheet closer to launch.
Vera Rubin Ultra Availability: Where Things Stand After the Delay
Vera Rubin Ultra availability, in plain terms, now means 2028 at the earliest for the first Kyber racks to leave NVIDIA's own production line, and that's before accounting for who gets allocated first. First Rubin Ultra access goes to hyperscalers and frontier labs, the same pattern base Vera Rubin followed with its eight named H2 2026 launch customers. Specialist and neocloud providers come after.
That gap isn't small. Our Vera Rubin NVL72 cloud availability tracking found the gap between first-cohort hyperscaler access and neo-cloud access has historically run 6 to 12 months across recent NVIDIA generations. Apply that same pattern to a 2028 Kyber date and a realistic window for broader cloud marketplace access to Rubin Ultra lands somewhere in late 2028 to 2029, not 2028 itself. A reader planning around "Rubin Ultra in 2028" as a rentable product is really planning around 2029.
What This Means If You're Deciding Between GB300 NVL72 Now and Waiting for Rubin Ultra
If your roadmap has you weighing a GB300 NVL72 reservation against holding out for Rubin Ultra, the delay settles the question for anyone shipping before 2029. GB300 NVL72 and Vera Rubin NVL72 are, by a wide margin, the fastest rack-scale systems a cloud customer can actually reserve for the next two-plus years. Our GB300 NVL72 vs GB200 NVL72 pricing and availability breakdown has the real rack price spread and who's renting each today, including the finding that GB300 NVL72 is quote-only or reserved-capacity at every provider that post could verify, including CoreWeave, Azure, Lambda, and Together AI.
The power and cooling picture backs this up independently of the schedule. NVIDIA's MGX 1.2 rack spec, built around a 250kW ceiling, comfortably covers GB300 NVL72 and Vera Rubin NVL72's reported 190kW to 230kW draw, but it doesn't cover Kyber, reported at roughly 600kW, as our MGX 1.2 breakdown of what 250kW racks mean for GB300 lays out. Facilities aren't built to house a Kyber rack yet even where the silicon exists, which is a second, independent reason Rubin Ultra capacity won't show up at scale on short notice once NVIDIA does ship it.
"Wait for the faster rack" only pays off if the wait is short. A wait that now runs through 2028, plus the historical hyperscaler-to-neocloud gap on top, is not a short wait. For nearly any workload with a 2026 or 2027 timeline, reserving GB300 NVL72 or Vera Rubin NVL72 capacity today is the rational move, not a consolation prize while the real thing is still in the PCB fab.
How to Reserve GB300 or Vera Rubin NVL72 Capacity on Spheron Today
Spheron lists both GB300 NVL72 and Vera Rubin NVL72 on its pricing page as reservation items rather than instant on-demand instances, Vera Rubin NVL72 specifically marked as shipping H2 2026 with specs confirmed at launch. Neither carries a published hourly rate yet, which is honest: NVIDIA itself hasn't shipped either rack to a broad customer base, so there's no live market rate to quote. Spheron sources that capacity from its data center partner network and typically turns a reservation request into a sourced quote within 24 to 48 hours.
Reserve a slot through Spheron's Vera Rubin page with your expected GPU count, workload, and timeline, and the team follows up once allocation opens. The same reservation path applies to GB300 NVL72.
For anything that can't wait on a reservation cycle, B200 and B300 are both bookable on Spheron right now with per-minute billing and no commitment. On-demand rates run $9.95/hr per GPU per hour for B200 and $5.40/hr for B300, as of 17 Sep 2026.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Sep 2026; GB300 NVL72 and Vera Rubin NVL72 have no published hourly rate anywhere in the market, Spheron included, since neither has shipped broadly. See the Spheron docs for how to get started, and confirm live on-demand rates in the app before committing.
Rubin Ultra NVL576 is worth tracking, and worth revisiting once NVIDIA confirms a Kyber ship date it's willing to put a number on. It isn't worth holding a 2026 or 2027 roadmap hostage to.
Kyber's slip means GB300 NVL72 and Vera Rubin NVL72 stay the fastest rack you can actually reserve through 2028. Lock in capacity before the next allocation window closes.
Quick Setup Guide
Open the Spheron pricing page and locate GB300 NVL72 or Vera Rubin NVL72. Both are listed with a Reserve action rather than a live hourly rate, since neither sells on-demand yet.
Fill in the reservation form with how many GPUs you need, what you're running, and when you need capacity live.
Spheron sources rack-scale capacity from its data center partner network and responds with pricing and setup details, typically within 24 to 48 hours of the request.
Frequently Asked Questions
Yes. SemiAnalysis reported in July 2026 that NVIDIA's Kyber rack, the system Rubin Ultra NVL576 ships in, slipped more than 12 months from its original 2027 target to 2028. CNBC confirmed the delay the same week, citing manufacturing problems with the rack's circuit board. NVIDIA has not confirmed a delay publicly and told CNBC its roadmap is intact, but it has not disputed the 2028 date either.
Kyber is the rack architecture NVIDIA designed to carry Rubin Ultra GPUs at NVL144 scale, with NVL576 linking eight Kyber racks together over co-packaged optics. It replaces the NVL72 tray layout used by GB200, GB300, and Vera Rubin with a vertical tray design connected through a dense PCB midplane, the component reported to be causing the delay.
Yes. Both are listed on Spheron's pricing page as reservation items, since neither ships with a published hourly rate yet. Spheron sources rack-scale capacity from its data center partner network and typically responds to a reservation request within 24 to 48 hours. B200 and B300 are available on-demand today with per-minute billing for workloads that cannot wait on a reservation process.
No, based on NVIDIA's own statements. Vera Rubin NVL72, the non-Ultra system that uses NVL72 racks rather than Kyber, is reported to remain in full production, with fall 2026 deliveries planned to eight cloud customers. The delay applies specifically to the Kyber rack architecture that Rubin Ultra NVL144 and NVL576 depend on.
