Plan on the Shipping Datasheet

A datasheet is a promise with footnotes. ServeTheHome reported on November 1 that NVIDIA has locked down the shipping specifications for the Rubin GPU, and the final figures are not the figures the industry spent a year planning around. The changes appeared quietly in September, on product pages and in the architecture datasheet, where IT Home first spotted them. No keynote walked them back. The numbers simply moved — and every capacity model built on the earlier numbers moved with them.

We read spec revisions the way auditors read restated earnings: calmly, and in full. A revision is not a scandal; NVIDIA labelled the earlier figures preliminary, with all values up to and subject to change. A revision is, however, a work order. Every cluster plan, serving budget and procurement comparison drawn up against the preliminary table now carries a footnote of its own, and the professional habit is to re-baseline before the hardware arrives, not after.

The headline is simple enough. The memory holds. The compute holds. The bandwidth — the measure that decides how fast a model can actually be fed — is what gave ground.

01 · The quiet edit

ServeTheHome traces the change to September, when NVIDIA posted updated specifications for the Rubin GPU and the products built from it. The timing matters. The revision lands as the first Vera Rubin NVL72 systems become generally available at hyperscalers, which means the new table describes hardware customers can actually order and receive, rather than a target the engineering teams were still chasing. Preliminary numbers belong to the announcement season; shipping numbers belong to the purchase order.

There is a craft lesson in the delivery method. The figures were not corrected in a launch event or a blog post; they were edited in place, on the pages practitioners read while sizing real systems. Teams that archive the datasheets they plan against — dated, versioned, kept beside the model that consumed them — catch a change like this in a diff. Teams that planned from keynote slides are still quoting the old table.

02 · What actually changed

Per GPU, the memory system keeps its size and loses some speed. Capacity stays at 288 GB of HBM4 across eight stacks. Memory bandwidth steps down from 22 TB/s to 19.2 TB/s, because each pin now runs at 9.375 Gbps instead of the preliminary 10.75 Gbps — a reduction ServeTheHome puts at roughly 11 percent. NVLink, the scale-up fabric that lets many GPUs behave as one machine, steps down from 3.6 TB/s to 3.0 TB/s per GPU, a 17 percent reduction. Scale-out networking, the path between nodes, moves the other way, from 0.4 to 0.45 TB/s.

At rack scale the arithmetic compounds. A Vera Rubin NVL72 now lists 1,400 TB/s of aggregate memory bandwidth against 1,580 TB/s in the preliminary table, and 216 TB/s of NVLink against 260 TB/s, while scale-out rises to 32.4 TB/s and capacity holds at 20.7 TB. The eight-GPU HGX Rubin NVL8 absorbs the deepest revision: bandwidth falls from 176 to 130 TB/s, capacity from 2.3 TB to 2.0 TB, and NVFP4 training from 280 to 270 PFLOPS.

03 · Why bandwidth is the budget line

Compute figures sell accelerators; bandwidth figures run them. When a large model generates text, each new token requires the machinery to read the model weights and the accumulated context from memory. The wider and faster that memory path, the more tokens a system serves per second, and the cheaper each answer becomes. A memory-bandwidth revision of roughly a tenth therefore lands directly on the serving economics of long-context workloads — the exact workloads context engineering keeps making larger.

That is the operational meaning of the revision. A context budget, a batching plan and a cost-per-answer model are all, underneath, bandwidth budgets. Since December 2022, our context engineering work at Ovia has treated the context as a designed resource with a cost curve, and the curve is drawn by hardware numbers like these. When the datasheet moves, the honest response is to re-run the model, restate the assumptions and say plainly which line in the budget changed and why. AI operations, practised as a craft, is mostly that: verified numbers, versioned assumptions, no surprises billed later.

Procurement deserves the same treatment. Comparisons drawn during the preliminary season set one side's targets against the other side's documents. The shipping table is the first fair basis for those comparisons, and it cuts both ways: ServeTheHome notes that even the revised pin rate sits above the HBM4 specification's own target rate and above the rate listed for AMD's MI455X, so the revision narrows a lead rather than surrendering it.

04 · The fine print

Three caveats keep the story honest. First, NVIDIA flagged the original figures as preliminary from the start; the revision is a normal part of bringing a new memory generation to volume, not a broken promise. Second, NVIDIA has not explained the change. ServeTheHome offers the working assumption that memory controller speeds, stack speeds or yields set the limit — reported there as an assumption, and worth treating as one until the company speaks. Third, headline compute is untouched, and not every workload is bandwidth-bound; training and dense compute work will feel this revision far less than high-volume inference serving will.

One more caution belongs to the reader rather than the vendor. Secondary coverage has rounded the same change to several different percentages, depending on whether the writer measured pin rates, per-GPU bandwidth or whole-system bandwidth. The figures in this piece follow ServeTheHome's comparison table, preliminary against shipping, line by line. When numbers matter, the table beats the summary — including this summary.

05 · What to watch

Watch three things from here. Watch whether the shipping table holds as volumes rise, or moves once more when the next revision of the hardware reaches customers. Watch the first independent benchmarks on production NVL72 systems, which will show how much of the paper bandwidth serving stacks actually capture. And watch the rival documents: competing platforms will publish against the revised table now, and the comparisons will sharpen.

The durable lesson predates this product cycle and will outlast it. Plan on the shipping datasheet. Archive it, version the assumptions drawn from it, and re-baseline without embarrassment when it changes. Keynote numbers are weather; the datasheet is climate. Studios that plan on climate build systems — and budgets — that hold.