NVIDIA Rubin is no longer a roadmap slide. The first Vera Rubin NVL72 systems are now generally available at hyperscalers, NVIDIA has begun shipping the platform into full production, and — in the same breath — the company quietly revised several of Rubin's headline specifications downward in a September datasheet update. Memory bandwidth dropped from 22 TB/s to 19.2 TB/s per GPU, NVLink scale-up bandwidth fell from 3.6 TB/s to 3.0 TB/s, and networking bandwidth crept upward across the rack. None of this changes the 288 GB of HBM4 per GPU or the 50 PFLOPS of FP4 inference compute, but it does change the arithmetic for anyone planning AI infrastructure around the new generation.
This guide puts the shipping Rubin hardware next to the two generations it replaces — Blackwell and Hopper — with the revised numbers, explains what each change actually means for real workloads, decodes NVIDIA's 30x/10x/35x performance claims, and lays out AMD's answer in the form of the Instinct MI400 series and the Helios rack. Whether you rent cloud GPUs, build models, or just want to understand where the AI hardware race stands in November 2026, this is the complete picture — verified against sources published between October 26 and November 2, 2026.
TL;DR — NVIDIA Rubin in November 2026
| What | Vera Rubin — NVIDIA's AI GPU generation after Blackwell Ultra (GB300/B300) |
| Status | In full production and shipping; first NVL72 systems generally available at hyperscalers (Nov 2026) |
| Key revision | HBM4 bandwidth cut 22 → 19.2 TB/s per GPU (~13% lower); NVLink 6 cut 3.6 → 3.0 TB/s; scale-out networking raised 0.4 → 0.45 TB/s |
| Unchanged | 288 GB HBM4 per GPU; 50 PFLOPS NVFP4 inference; 35 PFLOPS NVFP4 training |
| Competition | AMD Instinct MI400 (MI455X: 432 GB HBM4, up to 40 PFLOPS FP4) in the Helios rack, volume deployments expected 2H 2026 |
| Biggest cloud bet | Google Cloud A5X instances on Vera Rubin, GA November 2026, planned scale-up to 960,000 GPUs |
| For gamers | Nothing yet — no new consumer GeForce generation in 2026; Rubin is data-center only |
What Actually Changed: NVIDIA's Quiet Spec Revision
The revisions surfaced in the September 2026 update of NVIDIA's architecture datasheet, where the publication IT Home spotted the new figures; ServeTheHome reported the story on October 31, 2026, and the timing is significant — the corrected numbers describe hardware that is actually shipping, not a future preview. A second report on November 1, 2026, relaying an NVIDIA representative's comments via Wccftech, covered four full configurations: the Rubin GPU, the Vera Rubin Superchip, the Vera Rubin NVL72 rack, and the HGX Rubin NVL8.
Here is the reported earlier-vs-revised comparison, with the figures as published:
| Configuration | Specification | Earlier figure | Revised figure |
| Rubin GPU | HBM4 capacity | 288 GB | 288 GB (unchanged) |
| Rubin GPU | HBM4 bandwidth | 22 TB/s | 19.2 TB/s |
| Rubin GPU | NVLink 6 bandwidth | 3.6 TB/s | 3.0 TB/s |
| Rubin GPU | Scale-out networking | 0.4 TB/s | 0.45 TB/s |
| Vera Rubin Superchip | HBM4 bandwidth | 44 TB/s | 38.5 TB/s |
| Vera Rubin Superchip | NVLink bandwidth | 7.2 TB/s | 6 TB/s |
| Vera Rubin Superchip | Networking bandwidth | 0.8 TB/s | 0.9 TB/s |
| Vera Rubin NVL72 | HBM4 capacity | 20.7 TB | 20.7 TB (unchanged) |
| Vera Rubin NVL72 | HBM4 bandwidth | 1,580 TB/s | 1,400 TB/s |
| Vera Rubin NVL72 | NVLink bandwidth | 260 TB/s | 216 TB/s |
| Vera Rubin NVL72 | Networking bandwidth | 28.8 TB/s | 32.4 TB/s |
| HGX Rubin NVL8 | HBM4 capacity | 2.3 TB | 2 TB |
| HGX Rubin NVL8 | HBM4 bandwidth | 176 TB/s | 130 TB/s |
| HGX Rubin NVL8 | NVLink Switch bandwidth | 28.8 TB/s | 24 TB/s |
| HGX Rubin NVL8 | Networking bandwidth | 1.6 TB/s | 1.8 TB/s |
| HGX Rubin NVL8 | NVFP4 training | 280 PFLOPS | 270 PFLOPS |
| HGX Rubin NVL8 | FP8/FP6 training | 140 PFLOPS | 130 PFLOPS |
The pattern is deliberate, not random: per-chip memory and interconnect bandwidth came down, while cluster-level networking went up. That trade tells you how NVIDIA sees Rubin being used — not as 72 isolated monsters, but as one giant networked system where the fabric between nodes matters more than squeezing the last drop out of each GPU.
Why NVIDIA Says the Numbers Changed
According to the NVIDIA representative quoted by Wccftech on October 29, 2026, the earlier figures were preliminary architectural targets, while the revised figures are the specifications for systems that sustain qualified, volume production. In plain language: the launch numbers were engineering goals, and the shipping numbers are what the factory can actually guarantee at scale.
The representative added an important framing point: individual component specifications do not determine application performance, which NVIDIA treats as a data-center-scale result involving compute, networking, memory, software, and models together. That is partly marketing, but it is also technically fair — a 13% memory-bandwidth revision does not translate into 13% slower AI training, because real workloads are shaped by the balance of all these factors, plus the software stack running on top.
Two timeline facts anchor the story. First, NVIDIA announced on May 31, 2026, that Vera Rubin was ramping into full production. Second, in the later Wccftech statement, the representative said Vera Rubin NVL72 was in full production and shipping. These are separate statements at different points in the production timeline — the May announcement was the ramp, the October/November confirmation is the volume reality. And on the software side, CUDA Toolkit 13.4, released September 9, 2026, added preview support for Rubin at compute capability SM_107 — explicitly marked as not for production or benchmarking, but enough for developers to start compiling against the new architecture now.
Vera Rubin NVL72: The System That Is Actually Shipping
The NVL72 is the unit that matters. It is a single rack-scale system containing 72 Rubin GPUs linked by NVLink 6, with 20.7 TB of HBM4 in aggregate and — per the revised figures — 1,400 TB/s of memory bandwidth, 216 TB/s of NVLink bandwidth, and 32.4 TB/s of networking bandwidth across the rack. The Vera Rubin Superchip at its heart pairs each Rubin GPU with NVIDIA's Vera CPU, and the whole assembly is designed to behave as one coherent compute domain for trillion-parameter-class models.
The first NVL72 systems are now generally available at hyperscalers, per the October 31 ServeTheHome report — meaning the first cloud customers are taking delivery this fall, exactly on the H2 2026 schedule NVIDIA held across its roadmap all year. Rubin was first announced at Computex 2024 and confirmed at GTC 2026 with that H2 2026 delivery window for the first cloud cohort; nothing since has moved the date. Broader availability is expected through 2027.
One useful piece of naming trivia for anyone reading spec sheets: R100 and R200 are alias names you will see in some sources for Rubin variants, and Rubin CPX — an earlier long-context inference variant — was replaced in NVIDIA's lineup. When a cloud provider or benchmark mentions "R100," it is talking about Rubin.
Rubin vs Blackwell vs Hopper: The Full Generation Comparison
Three NVIDIA GPU generations are active in production cloud right now. Hopper (2022, H100/H200) is the mature workhorse with the lowest rental prices. Blackwell (2024, B200/B300) is the current production standard for serious inference. Rubin (2026) is the new arrival. The table below uses per-GPU figures; where Rubin numbers have been revised, the shipping figures are shown, with the older preview figures noted.
| Spec | H100 (Hopper) | H200 (Hopper) | B200 (Blackwell) | B300 (Blackwell Ultra) | Vera Rubin |
| Architecture | Hopper | Hopper | Blackwell | Blackwell Ultra | Rubin |
| VRAM | 80 GB | 141 GB | 192 GB | 288 GB | 288 GB |
| Memory type | HBM3 | HBM3e | HBM3e | HBM3e | HBM4 |
| Memory bandwidth | 3.35 TB/s | 4.8 TB/s | 8 TB/s | 8 TB/s | 19.2 TB/s (revised; was 22) |
| FP4 dense | N/A | N/A | 9 PFLOPS | 15 PFLOPS | 50 PFLOPS (inference) / 35 PFLOPS (training) |
| FP8 dense | ~1,979 TFLOPS | ~1,979 TFLOPS | 4,500 TFLOPS | ~4,500 TFLOPS | ~17,500 TFLOPS* |
| Transistors | 80 billion | 80 billion | 208 billion | 208 billion | 336 billion |
| TDP | 700 W | 700 W | ~1,200 W | 1,400 W | ~2,300 W |
| NVLink generation | 4 | 4 | 5 | 5 | 6 |
| NVLink bandwidth | 900 GB/s | 900 GB/s | 1.8 TB/s | 1.8 TB/s | 3.0 TB/s (revised; was 3.6) |
| Networking | ConnectX-7 (400G) | ConnectX-7 (400G) | ConnectX-8 (800G) | ConnectX-8 (800G) | ConnectX-9 (2x 800G, up to 1.6T) |
*The Rubin FP8 figure (~17,500 TFLOPS) is derived by analysts from NVIDIA's GTC 2026 NVL72 system-level spec (1,260 PFLOPS FP8/FP6 across 72 GPUs) — it is an approximate per-GPU figure pending final NVIDIA-published per-GPU specs, not an NVIDIA-confirmed number. Treat it as a working estimate. Similarly, a projected Rubin FP16 figure of ~8,000 TFLOPS circulates in analyst write-ups but is not NVIDIA-confirmed. One genuine footnote: the B300's FP4 figure is 15 PFLOPS per NVIDIA's developer blog, though math from the DGX B300 page yields roughly 13–13.5 PFLOPS — a minor variance worth knowing about if you are comparing datasheets line by line.
The single biggest leap in the table is memory bandwidth: 8 TB/s on Blackwell to 19.2 TB/s on Rubin. That 2.4x jump is the largest memory-bandwidth increase across any NVIDIA generation transition, and it matters more than the raw compute numbers for the workloads Rubin is built for.
Memory Explained: HBM3 to HBM3e to HBM4
Rubin's 288 GB of HBM4 is organized as 8 stacks of 36 GB each, running at 9.375 Gbps per pin to deliver the revised 19.2 TB/s. To understand why the revision still leaves Rubin in a different league, follow the memory story across generations:
Hopper's H100 shipped with HBM3 at 3.35 TB/s. The H200 was essentially the same GH100 die with HBM3e pushing bandwidth to 4.8 TB/s — no new CUDA cores, no new Tensor Core generation, just better memory. That single change was enough to let a 70B model fit in FP16 on one H200 where the H100 needed two-way tensor parallelism. Blackwell's B200 doubled H100's bandwidth to 8 TB/s on the same HBM3e type, and the B300 (Blackwell Ultra) reached 288 GB of capacity while keeping bandwidth at 8 TB/s — more stacks do not automatically mean more bandwidth once the memory bus is saturated. The practical upshot: B300 fits larger models without multi-GPU sharding, but the bandwidth per token is the same as B200.
Rubin breaks the pattern. For long-context inference, the KV cache — the stored attention state that grows with every token of context — is the dominant bandwidth consumer at high token counts. A request with 128K context on a 70B model generates a KV cache that, at 8 TB/s, limits how fast the GPU can stream attention weights; at 19.2 TB/s, that same cache takes roughly 40% of the time. Models that currently require four H100s or two B200s for throughput at long context may fit and serve efficiently on a single Rubin GPU. That is the workload the entire architecture is aimed at: not higher peak FLOPS for their own sake, but memory bandwidth that keeps the compute units fed during long-context inference.
NVLink 6 and Networking: The Trade NVIDIA Made
Each generation has roughly doubled NVLink bandwidth: 900 GB/s on NVLink 4 (Hopper), 1.8 TB/s on NVLink 5 (Blackwell), and — even after the revision — 3.0 TB/s on NVLink 6 (Rubin). NVLink is the fabric that ties GPUs together inside one system, and each doubling roughly halves the communication-to-compute ratio for fixed model sizes, directly improving tokens per second at scale.
Here is the interesting part of the revision: NVIDIA lowered NVLink scale-up bandwidth (3.6 → 3.0 TB/s) while raising scale-out networking (0.4 → 0.45 TB/s per GPU, 28.8 → 32.4 TB/s at NVL72 level, 1.6 → 1.8 TB/s on the HGX NVL8). The design is leaning on cluster-level networking — ConnectX-9 at up to 1.6T — rather than on making each individual GPU-to-GPU link faster. That fits the large NVL72 deployments now reaching general availability: when your unit of compute is a 72-GPU rack, the network between racks starts to matter more than the links inside one GPU.
One ceiling survives all three generations: the NVLink switch fabric still tops out at 72 GPUs on GB200, GB300, and Vera Rubin NVL72 alike. The domain did not get bigger — the pipes inside it got wider, and the pipes leaving it got wider still.
Compute: 50 PFLOPS of FP4 and the 400-Billion-Parameter Math
Rubin's headline compute number is 50 PFLOPS of NVFP4 inference performance per GPU (35 PFLOPS for training) — 5.6x the B200's 9 PFLOPS and 3.3x the B300's 15 PFLOPS. FP4 is the key precision story of this generation: Blackwell introduced native FP4 Tensor Core hardware, and Rubin scales it dramatically. Hopper, notably, has no FP4 support at all — FP4 workloads on H100/H200 fall back to FP8 or worse, because the minimum Tensor Core precision on Hopper is FP8.
Here is why the FP4 number matters in practice. A 400-billion-parameter model at FP4 precision needs roughly 200 GB of VRAM for its weights (half a byte per parameter). Vera Rubin's 288 GB of HBM4 holds that with room left over for the KV cache — and the 19.2 TB/s of bandwidth means the attention computation does not become the bottleneck even at 128K context lengths. In other words: single-GPU inference for 400B-parameter models in FP4, a job that currently requires multi-GPU sharding at FP8. That is the concrete capability the spec sheet is selling, and it survived the revision untouched.
For training, the 35 PFLOPS NVFP4 figure and the NVL8-level 270 PFLOPS (revised from 280) describe the same story at rack scale. And a caution carried over from the table notes: third-party FP8/FP16 per-GPU derivations for Rubin are estimates, not NVIDIA-published specs — the numbers NVIDIA actually stands behind are the NVFP4 figures and the revised bandwidth specs.
Decoding NVIDIA's 30x, 10x and 35x Performance Claims
NVIDIA's product pages and briefings throw around big multipliers, and the November 1 report helpfully separated them — because they use different metrics, different baselines, and different configurations. Here is what each claim actually compares:
The 30x claims: NVIDIA's representative told Wccftech that third-party agentic benchmarks use production specifications, and cited 30x more performance per watt and 30x more tokens per dollar versus the GB300 NVL72. Both use GB300 NVL72 (Blackwell Ultra) as the comparison system, but they are two different metrics — efficiency and cost-effectiveness — not the same measurement stated twice.
The 10x claims: NVIDIA's product page describes separate comparisons against the GB200 NVL72 (regular Blackwell, an older baseline): one-tenth the cost per million tokens and up to 10x more tokens per megawatt. The cost-per-million-tokens comparison is specific — it names Kimi-K2-Thinking with 32K input and 8K output sequence lengths. Change the model or the sequence lengths and the number changes.
The 35x claim: the same page says Vera Rubin NVL72 paired with Groq 3 LPX can deliver up to 35x more throughput per megawatt for trillion-parameter models. That figure is tied to that specific hardware pairing and that model class — it is not the same comparison as the 30x figures versus GB300 NVL72.
The honest reading: Rubin is a large generational step on every axis NVIDIA measures, but the multipliers are not interchangeable. A 30x tokens-per-dollar claim against GB300 is a different statement from a 10x tokens-per-megawatt claim against GB200, and neither tells you what your workload will do. The revised shipping specs — not the launch-day targets — are the numbers to plan against.
Google Cloud A5X: The 960,000-GPU Plan
The biggest single Rubin commitment disclosed so far came on the Wednesday before last from Google Cloud, which alongside NVIDIA unveiled A5X instances built on the Vera Rubin architecture. General availability is slated for November 2026 — this month — and the plan calls for scaling the infrastructure across multiple sites to as many as 960,000 Rubin GPUs.
That number deserves its full context, because the reporting was careful about it: 960,000 GPUs describes an intended build-out, not capacity already running in a data hall. It is a statement of how far Google intends to take Rubin deployment, not a measurement of what exists today. The distinction matters when judging how much of the plan has actually been de-risked — but even as a plan, it is the largest single Rubin deployment figure any cloud provider has put on paper, and it signals that Google sees Rubin as a multi-year platform, not a stopgap.
For renters, the practical takeaway is simpler: Vera Rubin capacity will first appear inside the big clouds' managed instance families (Google's A5X being the first named one), with broader availability — including the spot and on-demand markets where smaller teams actually buy — expected through 2027.
DSX POD: Data Centers Assembled in 14 Days?
NVIDIA's answer to the slowest part of AI infrastructure is DSX POD, a modular reference design for Vera Rubin AI data centers developed with partners and presented the same week as the A5X announcement. The pitch: factory-integrated units that could cut provisioning at prepared sites from six months to as little as 14 days, with worldwide shipments through partners targeted for early 2027.
The caveat is built into the claim and worth stating plainly: the 14-day figure applies to prepared locations, not to greenfield construction. NVIDIA is promising faster delivery of integrated systems — the racks, the cooling, the networking, pre-integrated at the factory — not a shortcut for building an entire facility from scratch. Until partner shipments begin in early 2027, the compressed timeline remains a projected capability rather than an operating standard anyone has reached. The next hard checkpoint is that partner delivery schedule.
Running costs got attention the same week through NVIDIA's DSX AI Factory Platform, presented with European adoption in focus. Per company figures, the software could cut electricity use by up to 30% at the same token output, while closed-loop cooling is designed to bring water consumption for facility cooling close to zero. Both are targets rather than fleet-wide results — but paired with DSX POD, they sketch a two-sided effort: speed up deployment, then trim what it costs to keep the racks running.
Rubin Ultra and the Kyber Delay to 2028
Not everything in the Rubin family is on schedule. Rubin Ultra — the larger NVL144/NVL576 rack system built on NVIDIA's new Kyber rack design — has slipped from a 2027 target to 2028 over circuit-board manufacturing issues, according to industry reporting. The base Vera Rubin NVL72 is explicitly not affected by that delay; it is the Ultra-scale configuration that moved.
This is a normal pattern in rack-scale computing: the bigger the domain, the harder the manufacturing. A 72-GPU rack is a solved packaging problem at this point; a 144- or 576-GPU Kyber rack is a new one. For buyers, the practical effect is that 2027's flagship will be Vera Rubin NVL72, not Rubin Ultra — and anyone whose roadmap assumed Ultra-scale systems next year should re-plan around NVL72 or wait.
AMD's Answer: Instinct MI400 and the Helios Rack
NVIDIA does not have the data-center GPU market to itself, and AMD's counterpunch is already launched. AMD formally launched the Instinct MI400 series at its Advancing AI event on July 23, 2026, built on the new CDNA 5 architecture — and like NVIDIA, AMD has moved from selling accelerators to selling integrated AI racks.
The headline part is the MI455X: 432 GB of HBM4 (twelve stacks, 1.5x the MI355X's 288 GB), up to 23.3 TB/s of memory bandwidth, and up to 40 PFLOPS of peak FP4 / 20 PFLOPS of peak FP8 per AMD's figures. Note the trade versus Rubin: the MI455X carries more memory (432 GB vs 288 GB) and slightly more bandwidth (23.3 vs 19.2 TB/s), but less FP4 compute (40 vs 50 PFLOPS). It is a memory-capacity-first design against Rubin's compute-and-bandwidth balance. A second part, the MI430X for HPC and sovereign AI, is expected in 2027.
The MI455X does not ship as a card you can buy — it is a component of Helios, AMD's 72-GPU rack-scale system: 72 MI455X GPUs plus 18 EPYC "Venice" CPUs per double-wide rack, 31 TB of HBM4 in aggregate, and up to 2.9 exaFLOPS of FP4 inference per AMD. AMD's product page says the Helios reference design "is being shared with partners now, with volume deployments expected in 2H 2026" — this half — and Reuters reported AMD saying Helios was in full production and would begin shipping "in the coming months."
Availability is the honest caveat. AMD has not announced public MI455X pricing, a broad OEM availability list, or a detailed calendar for customer deliveries. What it does have is commitments: reported customer plans include OpenAI (a 6-gigawatt multi-year agreement, with Helios expected online starting Q4 2026), Anthropic (up to 2 gigawatts of MI400-class GPUs), Microsoft Azure (Helios deployment), Oracle (large GPU cluster commitments), plus Meta and xAI. Until Helios reaches customers at volume, the MI355X — AMD's current shipping flagship with 288 GB of HBM3e — is what hyperscalers actually buy. AMD's data-center/Instinct revenue stood at $6.7B as of early August 2026, and the company reports quarterly results on November 3 — the next checkpoint for whether the commitments are converting into revenue.
What About Gamers? Why Rubin Matters Even If You Never Rent a Cloud GPU
Here is the short version: Rubin changes nothing on store shelves this year, and everything about why the shelves look the way they do. Rubin is data-center only — there is no GeForce version, no retail card, no gaming driver. NVIDIA has no new consumer GPU generation coming in 2026 at all; the next GeForce architecture is not expected until late 2027, and AMD's next gaming architecture is on a similar horizon.
But the same forces shipping Rubin are emptying the gaming market. Our November 2026 GPU price guide documented the consumer side in detail: NVIDIA quietly pulled the RTX 5080 from its US online store, AMD is reportedly preparing another price increase on the GPU-plus-memory kits it sells to board partners, and street prices sit far above launch MSRPs. Fresh November 1 data from a Japanese market survey found NVIDIA and AMD card prices up 18% to 45% in a single month — an RTX 5090 climbing from the equivalent of $6,150 to $8,100 — with analysts blaming a buying spree on high-VRAM cards driven by on-premise AI tools, including a new open-source project that sent buyers hunting for large-memory GPUs through October.
The connection is direct: every HBM stack that goes into a Rubin or Helios rack is memory that does not go into a GDDR6 kit for a gaming card, and every wafer TSMC can print is fought over by data-center orders with data-center margins. Gamers are not competing with other gamers for GPUs anymore — they are competing with the AI build-out. That is the market Rubin defines, and it is not changing in 2026. For the full consumer picture, see our October 2026 graphics card price guide alongside the November update.
Cloud GPU Prices Right Now (November 2, 2026)
Rental prices are the reality check on every spec sheet. The figures below are live Spheron marketplace rates as of November 2, 2026 — they move with availability, so treat them as a snapshot, not a quote:
| GPU | Rate (Nov 2, 2026) | Notes |
| H100 | $2.64/hr on-demand | Widely available; the cost leader, ~2 years past cloud launch |
| H200 | from $3.35/hr spot | Premium over H100 justified when 141 GB fits your model on one GPU |
| B200 | $5.25/hr spot (~$10.20/hr on-demand) | Available at multiple providers; spot supply improved through 2026 |
| B300 | $12.36/hr on-demand (~$7.24/hr spot) | 288 GB capacity; same 8 TB/s bandwidth as B200 |
| Vera Rubin | Reservation only (est. $8–12/hr spot) | Not yet on the spot market; pre-orders open, broader availability 2027 |
Two things stand out. First, H100 pricing has compressed hard — $2.64/hr on-demand makes Hopper the undisputed value tier for anything it can still run well. Second, the B200's spot price has fallen far enough that the generational upgrade is no longer a luxury purchase for production inference. Rubin's estimated $8–12/hr spot range, when it arrives, would price it as a premium tier — which is exactly what a 50-PFLOPS-FP4 card should be.
Which GPU Should You Rent? The Decision Guide
Match the hardware to the workload, not to the hype cycle:
Fine-tuning 7B–13B models: H100 at $2.64/hr on-demand is the cost leader. B200 spot offers roughly 2x the training throughput — choose based on whether wall-clock time or cost per experiment matters more to your team.
Fine-tuning 70B models: H200 (141 GB fits a 70B model in FP16 plus adapters on one GPU) or B200. Single-GPU fine-tuning avoids tensor-parallelism overhead and simplifies checkpointing.
Inference, 7B–30B models, low concurrency: H100 remains viable. At low batch sizes these workloads are memory-bandwidth-bound, and the gap between H100 and B200 narrows compared to the price gap.
Inference, 70B models, production serving: B200 or B300 — roughly 4x the H100 throughput in independent benchmarks, with native FP4 optimization available on Blackwell for quantization-friendly models.
Inference, 200B+ models or long context: B300's 288 GB handles 200B models in FP8 on one GPU. For 400B+ at FP8, multi-GPU B300 is required until Vera Rubin ships.
Planning a new deployment for 2027: Vera Rubin is the target architecture, but Blackwell is the bridge — for production workloads you need running now, B200 or B300 beat waiting. Rubin capacity will arrive first inside the big clouds' managed instances, and the spot market follows later.
7 Mistakes to Avoid With the Rubin Generation
1. Planning against the launch-day specs. The 22 TB/s and 3.6 TB/s figures are superseded. Use the revised shipping specs — 19.2 TB/s HBM4, 3.0 TB/s NVLink 6 — for capacity planning.
2. Treating the 30x/10x/35x claims as interchangeable. They use different metrics and different baselines (GB300 vs GB200, different models, different pairings). Compare like with like.
3. Assuming FP4 is free performance. FP4 needs quantization-friendly models and software support. Hopper cannot do it at all; Blackwell and Rubin need the right stack. Budget the engineering time.
4. Counting Google's 960,000 GPUs as existing capacity. It is an intended build-out plan across multiple sites, not hardware in racks today.
5. Expecting 14-day data centers on day one. DSX POD's compressed timeline applies to prepared sites, and partner shipments start in early 2027.
6. Waiting for Rubin when you need GPUs now. B200/B300 are available today at much-improved prices; Rubin spot capacity is a 2027 story.
7. Ignoring AMD. The MI455X's 432 GB of HBM4 is a genuine memory-capacity advantage for memory-bound workloads, and Helios rack commitments from OpenAI, Anthropic, Microsoft, Oracle, Meta, and xAI are real. A one-vendor plan is a risk.
Frequently Asked Questions
Is NVIDIA Rubin actually shipping?
Yes. Vera Rubin NVL72 systems are in full production and shipping, with the first systems generally available at hyperscalers in fall 2026 — on the schedule NVIDIA held all year. Broader availability, including spot cloud markets, is expected through 2027.
What specs did NVIDIA revise?
Per-GPU HBM4 bandwidth went from 22 to 19.2 TB/s, NVLink 6 from 3.6 to 3.0 TB/s, while scale-out networking rose from 0.4 to 0.45 TB/s. At NVL72 level: HBM4 bandwidth 1,580 → 1,400 TB/s, NVLink 260 → 216 TB/s, networking 28.8 → 32.4 TB/s. Capacity (288 GB per GPU, 20.7 TB per rack) and compute (50 PFLOPS FP4 inference) are unchanged.
Why did the specs change?
NVIDIA says the earlier figures were preliminary architectural targets and the revised figures reflect what systems sustain in qualified, volume production. The numbers now describe shipping hardware rather than engineering goals.
Should I wait for Rubin or rent Blackwell now?
For production workloads you need running today, rent Blackwell — B200 spot at around $5.25/hr is the current sweet spot for serious inference. Reserve Rubin capacity if your roadmap includes 400B+ models and your timeline extends into 2027.
How does Rubin compare to AMD's MI400?
Rubin leads on FP4 compute (50 vs 40 PFLOPS) and per-GPU bandwidth efficiency; AMD's MI455X leads on memory capacity (432 GB vs 288 GB) and slightly on bandwidth (23.3 vs 19.2 TB/s). They are different trade-offs: NVIDIA optimized the compute-bandwidth balance, AMD optimized memory capacity. Both ship in rack-scale systems (NVL72 vs Helios).
Will there be a GeForce Rubin gaming GPU?
No. Rubin is data-center only, and NVIDIA has no new consumer GPU generation in 2026. The next GeForce architecture is not expected until late 2027. Gamers should read our November 2026 GPU price guide instead.
What is the Vera Rubin Superchip?
It is the base compute unit pairing a Rubin GPU with NVIDIA's Vera CPU. Revised specs: 38.5 TB/s HBM4 bandwidth, 6 TB/s NVLink, 0.9 TB/s networking. Seventy-two Superchips make an NVL72 rack.
What happened to Rubin Ultra?
Rubin Ultra — the larger NVL144/NVL576 system on the new Kyber rack — slipped from 2027 to 2028 over circuit-board manufacturing issues, per industry reporting. The base NVL72 is unaffected.
How much does it cost to rent these GPUs?
As of November 2, 2026 (Spheron marketplace): H100 $2.64/hr on-demand, H200 from $3.35/hr spot, B200 $5.25/hr spot, B300 $12.36/hr on-demand. Vera Rubin is reservation-only with an estimated $8–12/hr spot range. Prices move with availability.
Does the spec revision mean Rubin is a disappointment?
No — it means the launch numbers were targets and the shipping numbers are guarantees. The 13% bandwidth revision changes cluster-planning math for memory-bound workloads, but the compute figures, the 288 GB of HBM4, and the architectural leaps over Blackwell are all intact. Datasheets that get corrected when hardware ships are healthier than datasheets that never get corrected.
The Bottom Line
NVIDIA Rubin is real, shipping, and slightly less superhuman on paper than the launch slides suggested — which is exactly what shipping hardware looks like. The revised 19.2 TB/s of HBM4 and 3.0 TB/s of NVLink 6 are still generational leaps over Blackwell, the 50 PFLOPS of FP4 compute is untouched, and the first NVL72 racks are landing at hyperscalers on schedule. Google's 960,000-GPU A5X plan and the DSX POD modular data-center push show where NVIDIA thinks this goes: AI infrastructure as a manufactured product, not a construction project.
The honest caveats all point the same way. The multipliers need their baselines stated. The 14-day data center needs a prepared site and a 2027 calendar. The 960,000 GPUs are a plan, not an inventory. And AMD's Helios rack — with more memory per GPU and gigawatt-scale customer commitments — means Rubin enters a real fight, not a coronation. For buyers, the playbook is unchanged: plan against shipping specs, rent Blackwell for what you need today, and reserve Rubin for the 400B-parameter workloads of 2027.
Sources
Key facts in this article are drawn from the following publications, retrieved November 1–2, 2026. Figures are quoted as reported; rumours and estimates are labeled in the text.
NVIDIA quietly revises shipping specs for Rubin GPU as first production hardware arrives — ServeTheHome report via DEV Community, October 31, 2026: spec revisions spotted in September datasheet update; first Vera Rubin NVL72 systems generally available at hyperscalers.
Vera Rubin spec changes: NVIDIA's revised figures — NeoTeo, November 1, 2026: NVIDIA's explanation via Wccftech (preliminary targets vs qualified volume production); four-configuration comparison table; 30x/10x/35x claim breakdown; May 31 production ramp.
Nvidia's Build-Out Gets Faster — the Bill Still Arrives on Schedule — ad-hoc-news, October 26, 2026: Google Cloud A5X instances (GA November 2026, up to 960,000 GPUs planned); DSX POD modular data centers (6 months to 14 days at prepared sites, partner shipments early 2027); DSX AI Factory power/water targets.
NVIDIA Rubin vs Blackwell vs Hopper: Generation-to-Generation Comparison (2026) — Spheron, November 2026: full per-GPU spec table (H100/H200/B200/B300/Vera Rubin); workload decision guide; live rental pricing as of November 2, 2026.
Vera Rubin Release Date: What Changes From GB300 to Rubin — Spheron, November 2026: fall 2026 first-cloud-cohort window; Rubin Ultra/Kyber slip to 2028; CUDA 13.4 preview support; Rubin GPU aliases.
AMD MI400, MI450 and Helios Guide: Status and Specs (2026) — Aquanode, October 2026: MI455X specs (432 GB HBM4, 23.3 TB/s, 40 PFLOPS FP4); Helios rack (72 GPUs, 31 TB HBM4, 2.9 exaFLOPS FP4); July 23, 2026 launch; volume deployments expected 2H 2026.
AMD Instinct MI400 Launches as Helios Rack, Not a Windows GPU — WindowsForum, August 2026: launch timing correction; Helios in full production per Reuters; availability caveats.
AMD Instinct: From MI300X to Helios — Timeline — WhatLedTo, updated October 25, 2026: gigawatt-scale customer commitments; $6.7B data-center revenue; Helios systems running at early customers.
A shortage of NVIDIA and AMD graphics cards has led to a price increase of up to 45% in a month — GameGPU, November 1, 2026: Japanese market survey (Aile) — prices up 18–45% in one month; on-premise AI buying spree.
AI reality check triggers $2 trillion tech stock rout — NationScope, November 2, 2026: October's $2T Nasdaq-100 selloff; Alphabet -12.6%, Nvidia -4.2%, AMD -10.9%, Intel -28.6%.
Related Reading on MyThrift Hub
GPU Prices in November 2026: AMD's Second Price Hike, the Vanishing RTX 5080, and Prices 77% Above MSRP — the consumer side of this story: what the AI build-out is doing to gaming GPU prices.
Graphics Card Prices in October 2026: Why GPUs Cost So Much — the October baseline: GDDR6 costs, the RTX 60 delay, and the buy/wait framework.
Comments
Post a Comment