19.2 terabytes per second. That is the HBM4 bandwidth Nvidia now prints for each Rubin GPU. In February the same table said 22. The gap is 13%, and nobody at Nvidia announced it.

The revision appeared in September product-page edits and the architecture datasheet. HBM4 bandwidth fell from 22 to 19.2 TB/s. NVLink 6 scale-up fell from 3.6 to 3.0 TB/s, a 17% cut. Capacity stayed at 288 GB, and the 50 PFLOPS sparse NVFP4 headline did not move at all.

Read the datasheet as a market signal. The binding constraint on AI infrastructure has shifted from compute to memory bandwidth, and the shipping numbers now say it out loud. We have made this argument for a year; the vendors are only now putting it in a spec table.

First Vera Rubin NVL72 racks are reaching hyperscalers. Oracle, Microsoft Azure, Google Cloud and CoreWeave have all been named on the ramp. Buyers sized order books on the GTC figures, and many are still quoting them.

๐ŸŽฏ
The chip kept its flops and lost 13% of its memory bandwidth.

The eight-GPU NVL8 took the deeper hit: 26% less memory bandwidth and 300 GB less capacity.

Nvidia says the old figures were preliminary targets, and that its third-party tests already ran at 19.2 TB/s.

What changed on the datasheet

The directions are not uniform. HBM4 bandwidth and scale-up links fell across every configuration. Scale-out networking rose. The eight-GPU board lost capacity as well as speed.

ConfigurationPreliminaryShippingChange
Rubin GPU โ€” HBM422 TB/s19.2 TB/sโˆ’12.7%
Rubin GPU โ€” NVLink 63.6 TB/s3.0 TB/sโˆ’16.7%
NVL72 rack โ€” HBM41,580 TB/s1,400 TB/sโˆ’11.4%
NVL72 rack โ€” NVLink260 TB/s216 TB/sโˆ’16.9%
Superchip โ€” HBM444 TB/s38.5 TB/sโˆ’12.5%
HGX Rubin NVL8 โ€” HBM4 capacity2.3 TB2.0 TBโˆ’13.0%
HGX Rubin NVL8 โ€” bandwidth176 TB/s130 TB/sโˆ’26.1%
Scale-out per GPU0.4 TB/s0.45 TB/s+12.5%

Data: NVIDIA Vera Rubin product pages, Wccftech and ServeTheHome comparison tables, Octoberโ€“November 2026.

Nvidia's own pages still disagree with each other in at least one place. That is why these deltas come from comparing archived tables across the product pages, not from a single signed document.

Why bandwidth sets the price

Decode is memory-bound. Generating a token re-reads the model weights and the cached context, so bandwidth, not peak FLOPS, usually caps throughput. At 19.2 TB/s the constraint is the memory bus; the 50 PFLOPS figure sits largely unused on a long-context agentic workload.

The pin math is simple. Eight 2,048-bit HBM4 stacks at roughly 9.4 Gbps per pin land near 19.2 TB/s. ServeTheHome's arithmetic puts the preliminary figure around 10.75 Gbps. TechPowerUp reported the target moving from 11.7 to 9.6 Gbps. Both readings point the same way, and JEDEC's baseline HBM4 target is 8 Gbps per pin.

The specifications published during Vera Rubin's development were preliminary architectural targets. Final specifications on nvidia.com and the Vera Rubin datasheet reflect what the system sustains at qualified, volume production.โ€” NVIDIA representative, statement to Wccftech, 29 October 2026

Nvidia also says its published third-party benchmarks already ran on production silicon, and show 30ร— better performance per watt and 30ร— more tokens per dollar than GB300 NVL72. Artificial Analysis, a separate tester, measured a smaller and more conditional lead: up to roughly 5.7ร— total throughput per megawatt against GB300, concentrated at high interaction speeds.

Two baselines, two headlines. Both can be true at once.

The NVL8 tell: seven stacks, not eight

The eight-GPU HGX Rubin NVL8 absorbed the deepest cut. System HBM4 fell from 2.3 TB to 2.0 TB, and bandwidth from 176 to 130 TB/s. Divide 2.0 TB by eight and you get roughly 250 GB per GPU, which fits seven 36 GB stacks. One stack comes off the board.

Training slipped with it. NVFP4 fell from 280 to 270 PFLOPS, FP8 and FP6 from 140 to 130. Inference stayed at 400 PFLOPS.

That is a supply decision wearing a spec sheet. When a vendor can keep the compute die and cannot keep the memory bin, the stack count is the cheap variable to cut.

The loss compounds at factory scale. Nvidia's own 100 MW reference build lists 40,000 Rubin GPUs and 2 ZFLOPS of sparse NVFP4 inference, about 2.5 kW of site power per GPU. Strip 13% from the memory bandwidth feeding those 40,000 GPUs and the flop figure survives on the slide while the tokens-per-second curve bends downward. At that size, a single-digit bandwidth cut is a procurement problem, measured in megawatts and delivery quarters.

This is why the tokens-per-watt metric has crowded out peak FLOPS in every serious procurement conversation. Power is fixed by the site, not the datasheet. Whatever bandwidth the memory can supply becomes the ceiling on revenue from that power.

AMD leads on paper, Nvidia leads on the rack

On vendor paper AMD's MI455X now looks stronger. It lists 432 GB of HBM4 per GPU at 23.3 TB/s, with Helios aggregating about 31 TB and 1.67 PB/s across a 72-GPU rack. Rubin's shipping 19.2 TB/s and 288 GB sit below that line.

Paper ignores the software and the domain. Nvidia's 72-GPU NVLink 6 domain and CUDA remain the reason clusters get bought, not the per-GPU bandwidth column. Still, the column moved.

As we wrote in October, SK hynix says HBM orders run at more than twice available supply and its DRAM and HBM output is effectively sold out through 2029. Our October analysis of the HBM4 supply squeeze laid out why the memory makers, not the chip designers, set the pace of this buildout.

What to re-run before signing

Capacity models built on 22 TB/s are now optimistic by roughly one-eighth on the memory-bound part of a workload. That is not a rounding error at rack scale. A cluster sized for a target tokens-per-second figure at 22 TB/s lands near 87% of it at 19.2, before any software tuning.

The gap widens where decode dominates. Artificial Analysis reports that 97% of a request's input tokens, on average, repeat from the previous request in agentic sessions. Prefix caching and KV-cache offload carry much of the load, and most of its Rubin configurations push KV cache to CPU memory over NVLink-C2C at 1.8 TB/s per superchip. Bandwidth you lose on the HBM side has to be absorbed somewhere else in the stack.

Three checks matter more than the headline delta. Rerun any memory-bound benchmark at 19.2 TB/s, not 22. Treat the HGX NVL8 as a different product class, because 2.0 TB and 130 TB/s is a 26% cut and the seven-stack layout leaves memory capacity uneven across the lineup. Compare on tokens per megawatt at the interaction speed you actually serve, since the vendor's 30ร— claim and the independent ~5.7ร— result diverge as per-user speed rises.

None of this makes Rubin slow. It makes the buying math less forgiving. Re-baseline first, argue with the sales team second. Every procurement model should now carry 19.2 on its own line.

What has to be true by mid-2027

๐Ÿ”ฎ
By Q3 2027, at least one shipping Rubin configuration will carry a higher per-GPU memory-bandwidth figure than 19.2 TB/s โ€” either a validated HBM4 speed bin or a Rubin Ultra part. Horizon: 12 months.

Probability: 60% โ€” SK hynix and Samsung both have HBM4E samples with customers, and Nvidia rarely leaves a gap like this open for a full generation.

๐ŸŸข Yields recover (30%)

HBM4 speed bins hold at 10 Gbps and above, Rubin Ultra ships at the original target, and the NVL8 returns to eight stacks.

Consequence: the 19.2 TB/s figure becomes a short-lived footnote.

๐ŸŸก 19.2 holds through 2027 (50%)

Supply stays tight, buyers re-baseline capacity planning on the shipping figure, and AMD keeps the paper lead in memory bandwidth.

Consequence: tokens-per-watt becomes the only honest comparison metric.

๐Ÿ”ด Supply tightens again (20%)

HBM4 allocation falls further, NVL8 capacity drops again, and agentic deployments slip into 2028.

Consequence: rack-level throughput gains get eaten by lead times.

The datasheet is now the most honest document Nvidia publishes. It is the only one whose numbers have survived contact with production. The next revision deserves the same scrutiny, and the same skepticism.

NVIDIA Vera Rubin NVL72 โ€” shipping specifications
The primary table behind every figure above: 288 GB HBM4 at 19.2 TB/s per Rubin GPU, 20.7 TB and 1,400 TB/s per NVL72 rack, 216 TB/s NVLink.
Checked against the vendor page directly, not a re-quote, because several outlets still print the old 22 TB/s.
NVIDIA updates Rubin GPU specs as first production hardware arrives
Side-by-side preliminary and shipping tables, plus the pin-rate arithmetic behind the 19.2 TB/s number and the seven-stack read on HGX NVL8.
The clearest comparison of old and new figures at per-GPU, pack, and rack scale.
NVIDIA's official statement on the Vera Rubin spec change
Nvidia calls the earlier numbers preliminary architectural targets and says its third-party benchmarks were measured on production specifications.
The vendor's own framing, quoted in full โ€” worth reading against the datasheet deltas.