Evidence note: This article rests on three kinds of evidence, and they are not interchangeable. First, primary vendor material — NVIDIA’s own announcements, blog posts and developer documentation — which establishes what NVIDIA has said, not whether it is true. Second, second-party engineering reports from Together AI, a company that builds on the hardware in question and published its own kernel numbers; these come from outside NVIDIA but not from a neutral third party. Third, contemporaneous public discussion of the GTC 2025 keynote. No independently reproduced third-party benchmark of Vera Rubin silicon appears anywhere in the material available here. Every specific number below carries its own evidence marker, and those markers are not dropped later in the piece.
Short answer for a team holding a budget decision: the five large multipliers attached to Vera Rubin are all vendor-published and all forward-looking. Every named, concretely described deployment in this record — the agentic coding workflow, the production inference cluster — sits on Blackwell-generation hardware that can be rented now. That asymmetry, not the size of the multipliers, is the thing a capacity decision should turn on.
What NVIDIA Announced: Blackwell Ultra and Vera Rubin at GTC 2025, the Vera Rubin NVL72 at CES
Two announcement moments frame the current generation question.
The first is GTC 2025. Public discussion from the time of that keynote covers the Blackwell Ultra GPU and the Vera Rubin chip (contemporaneous GTC 2025 keynote discussion). Reports that NVIDIA chief executive Jensen Huang personally introduced both parts in that keynote are not corroborated by a primary transcript in this record; the attribution of the introduction to Huang specifically should be treated as unconfirmed, even though the keynote itself is well attested.
The second is CES, where NVIDIA is reported to have launched the Vera Rubin NVL72, a rack-scale AI system positioned as the successor to Blackwell-generation racks, with availability stated for the second half of 2026 (Tom’s Hardware report on the Vera Rubin NVL72 launch). That the launch happened at CES rather than at another venue is reported, not independently confirmed here — a detail that matters less than the date attached to it.
NVIDIA has also published an efficiency argument for the platform aimed specifically at agent workloads (NVIDIA’s own post on Vera Rubin NVL72 efficiency for AI agents).
Five Performance Figures, All of Them NVIDIA’s Own — and a 3.4% Share-Price Dip During the Event
The multipliers in circulation are worth laying out side by side, because reading them as a set makes their common weakness visible.
| Figure | What it is attached to | Source type | Independent check |
|---|---|---|---|
| 3.3x performance | The Vera Rubin chip | NVIDIA statement, as reported | None in this record |
| 5x inference performance vs Blackwell | Vera Rubin NVL72 rack | NVIDIA claim, as reported | None in this record |
| 10x lower cost per token vs Blackwell | Vera Rubin NVL72 rack | NVIDIA claim, as reported | None in this record |
| 30x more work per watt | Vera Rubin NVL72 rack | NVIDIA’s own published claim | None in this record |
| 14x performance improvement | Vera Rubin Ultra, targeted 2027 | NVIDIA statement, as reported | None in this record |
Three points about this table carry the section.
The 3.3x chip-level figure and the 14x Rubin Ultra target are reported statements attributed to NVIDIA and not independently verified; neither the benchmark, the workload, nor the baseline is specified in the material available here, which means the numbers cannot be checked even in principle from this evidence. The 5x inference and 10x cost-per-token figures are vendor claims carried in the launch reporting, again without a disclosed workload definition (Tom’s Hardware). The 30x work-per-watt figure differs in one respect only: NVIDIA states it directly in its own publication, so the fact that NVIDIA claims 30x is established (NVIDIA’s efficiency post) — whether a given workload realises 30x remains unverified by any party other than NVIDIA.
A 3.4% dip in NVIDIA’s share price during the GTC 2025 event is also in circulation. That figure is uncorroborated in this record, and even if the move occurred as described, a single intraday percentage during a keynote establishes nothing about silicon. It is included here only because it is frequently cited alongside the multipliers, and readers deserve to know it has not been checked.
Why it matters: five large numbers from one interested party, with no disclosed methodology, do not constitute a performance picture. They constitute a marketing position that may later prove accurate.
The Second Half of 2026, 2027 for Rubin Ultra: The Dates Are the Part That Touches a Budget
Dates are where vendor claims become decision-relevant, because a date can be falsified by a calendar.
Vera Rubin NVL72 availability is stated for the second half of 2026 (Tom’s Hardware). As of 16 October 2026, that window is roughly three and a half months from closing. Nothing in this record documents general availability, pricing, or a named customer deployment of Vera Rubin NVL72 capacity rentable by a small team — that remains open, not negative; absence of a record here is not evidence of absence in the market.
Rubin Ultra, carrying the 14x target, is placed in 2027. That timing is reported and not independently confirmed, and a 2027 part is outside any current-cycle budget decision in practical terms: a team cannot rent it, cannot benchmark it, and cannot plan a migration against an unpriced future rack.
For a solo developer or a small team, the practical reading is narrow. A second-half-2026 rack launch typically reaches hyperscale and large-enterprise buyers before it reaches hourly or small-cluster rental markets. The gap between "launched" and "available to a two-person team at a defensible price" is itself unquantified in this record.
The One Measurement From Outside NVIDIA: Together AI’s NVFP4 Kernel From 42% of Roofline to Over 22 PFLOPS
One engineering report in this record contains concrete numbers produced by someone other than NVIDIA.
Together AI reports porting its ThunderKittens kernel library to the Vera Rubin NVL72, and reports that rebuilding its NVFP4 general matrix multiply around the new hardware moved it from 42% of roofline to over 22 PFLOPS (Together AI’s own engineering report). Those figures are a second-party measurement: produced outside NVIDIA, published by the team that wrote the kernel, and not independently reproduced. The 42% starting point and the 22 PFLOPS result both carry that status; neither is a neutral benchmark.
Together AI also reports the rebuilt kernel is competitive with cuBLAS and with the CuTe DSL implementations. That comparison is the vendor’s own characterisation of its own kernel against NVIDIA’s libraries, uncorroborated here, and "competitive" is not a number.
This section carries the article’s most load-bearing claim, so the reservation is stated plainly: this single report is the closest thing in the record to evidence that Vera Rubin hardware does real work at high throughput, and it is one team’s account of one kernel. It is meaningfully more than a keynote slide. It is meaningfully less than a reproduced benchmark.
Why the Port Was a Rebuild and Not a Recompile: ISA Changes on the Vera Rubin NVL72
The detail that generalises beyond Together AI’s specific numbers is the shape of the work, not its result.
The Vera Rubin NVL72 is reported to introduce instruction set architecture changes relative to prior NVIDIA hardware (Together AI). That characterisation is uncorroborated by NVIDIA documentation in this record, and the specific instructions involved are not enumerated here.
If it holds, the implication is concrete for anyone carrying hand-written kernels: the move from Blackwell to Rubin reads as a rebuild rather than a recompile. Together AI’s account describes restructuring a kernel around the hardware, not retargeting a compiler flag. A team whose performance depends on framework-level code — standard PyTorch, a served inference engine, a hosted API — absorbs that cost through its vendor. A team with its own fused kernels absorbs it directly, and should budget engineering time, not just rental hours, for a generation jump. Readers tracking the surrounding efficiency argument may find the companion piece on Vera Rubin and Blackwell Ultra performance per watt useful alongside this one.
Where the Large Agentic Coding Workload Actually Runs Right Now
Against five forward-looking multipliers, it is worth asking what named hardware the headline agentic coding workload actually sits on.
NVIDIA’s own developer publication describes scaling agentic coding workflows with Meta’s Muse Spark Open model on the GB300 NVL72 — a Blackwell-generation rack, not a Rubin one (NVIDIA’s developer post on the workflow). The fact established here is that NVIDIA documents this pairing on Blackwell hardware; independent confirmation of the deployment is open.
Per the same NVIDIA publication, Muse Spark Open is described as a coding-optimised mixture-of-experts model with 816 billion total parameters, 42 billion active parameters, and a 1 million token context window. Those three figures are as stated in NVIDIA’s own developer documentation and not independently verified; no third-party evaluation of the model at those dimensions appears in this record.
The asymmetry is the finding. The largest coding model in this cycle’s announcements is documented running on the generation available today, while the 5x and 10x figures attach to the generation that is not.
The Production Inference Cluster Three Companies Actually Built Is Blackwell B300
The second named deployment in the record points the same way.
Together AI reports building a dedicated inference cluster with IBM Cloud and NVIDIA for running open models at production scale, on Blackwell B300 hardware (Together AI’s capacity announcement). The three-party arrangement and the B300 generation are as reported by Together AI and not independently corroborated here; capacity, pricing and utilisation figures are not stated in this record and remain open.
So the one company in this record that has published hands-on Vera Rubin kernel numbers is, by its own account, expanding its production enterprise capacity on Blackwell. That is not a contradiction — early silicon access and production fleet buildout run on different clocks — but it is a data point about which clock a customer-facing service runs on.
What This Changes for a Small Team Deciding Where to Rent Capacity This Year
The reader question was whether 3.3x, 5x, 14x, 30x per watt and 10x lower cost per token justify delaying a capacity decision. On this evidence, a usable criterion looks like this.
Commit to currently available Blackwell-generation capacity if the workload is inference or agent-driven coding that runs on framework-level code; if the need is inside the current budget cycle; and if the alternative is idle months. Both named deployments in this record — the agentic coding workflow and the production inference cluster — are Blackwell-generation, which means the reference configurations, the tuning advice and the vendor support exist for that generation today.
Consider holding only if all of the following are true: the commitment is a multi-year rack-scale purchase rather than hourly rental; the workload is dominated by custom kernels, so a rebuild is coming regardless and may as well land once; and there is capacity to absorb a slip, because a stated second-half-2026 window with no documented general availability as of 16 October 2026 is a schedule, not a delivery.
Do not treat the multipliers as a reason to wait. None has a disclosed workload definition, none has an independent benchmark, and the one outside figure in the record — 22 PFLOPS on a rebuilt NVFP4 kernel (second-party, not reproduced) — describes a microbenchmark on a single operation, not end-to-end cost per token on a real serving workload. A 10x cost-per-token improvement, if realised, would change rental economics decisively; it is also precisely the claim with no outside check behind it.
The pattern of announcement generation running ahead of shipping-integration generation is not unique to compute racks. It showed up this cycle in networking as well, as covered in the piece on BlueField-4’s announcement versus shipping BlueField-3 integrations, and in agent tooling, where the harness that placed second at KDD Cup 2026 got smaller rather than larger.
The Follow-Up That Has Not Been Run: A Proposed Same-Model, Two-Generation Cost-Per-Token Comparison
No experiment was performed for this article, and the comparison below is a proposal, not a result. Nothing here should be read as a measurement.
The missing number is the one that would actually settle the reader question: cost per million output tokens for one fixed model, served identically, on a Blackwell-generation rack and on a Vera Rubin NVL72, at matched concurrency and matched output quality. A proposed design would hold the model, the serving stack version, the quantisation format, the batch policy and the prompt distribution constant; vary only the rack generation; and report tokens per second per rack, tokens per second per watt, and billed dollars per million tokens, alongside a measured quality check to confirm the two configurations are producing comparable output rather than trading accuracy for speed.
Two things block it. Rentable Vera Rubin NVL72 capacity accessible to a small team is not documented in this record, so one arm of the comparison has no environment. And the vendor 10x cost-per-token claim is a ratio with no stated baseline, which means there is no published configuration to reproduce or contradict.
The honest status is therefore: the decisive measurement does not exist here, the design above has not been executed, and readers should discount any ratio — including every figure in the table above — until someone with access to both generations publishes a method alongside the number. For a related example of why a system’s own self-reported win count is a weak substitute for an outside measurement, see the analysis of 303 agent runs across three Roblox games.
Comments
7 responses to “Vera Rubin Arrives in the Second Half of 2026 — and the 816-Billion-Parameter Coding Model Scaling Today Runs on a Blackwell Rack”
[…] published from it before that date understated the providers it scored. PlayAgit’s review of efficiency claims against Vision’s agent logs ran into the same class of problem from the other […]
[…] Vera Rubin NVL72 Efficiency Claims and What Vision’s Agent Logs Assess […]
[…] applies to vendor efficiency and performance claims, which is the approach taken in the review of Vera Rubin NVL72 efficiency claims against agent logs, and in the accounting of timing, changes and evidence gaps across a set of thumbnail and […]
[…] Thirteen cross-checked claims out of 392 is a cross-check rate in the low single digits. That is the honest ceiling on how much of any Claude-related reporting on this site has been independently confirmed rather than recorded — and it is the reason the vendor claims above are attributed rather than asserted. The same discipline applied to a hardware vendor’s efficiency numbers in this site’s assessment of Vera Rubin NVL72 claims. […]
[…] Vera Rubin NVL72 Efficiency Claims and What Vision’s Agent Logs Assess […]
[…] Vera Rubin NVL72 Efficiency Claims and What Vision’s Agent Logs Assess […]
[…] publishes enough method detail for someone else to reason about it. That is the same problem an earlier piece on this site about reading a vendor’s own published multipliers for one hardware… worked through: a single-vendor number is not useless, but what it is comparable to has to be […]