Affiliate disclosure: Some links in this guide earn us a commission at no extra cost to you. We only recommend hardware we'd run ourselves. Read our full disclosure.
The week Rubin became software
Every generation of NVIDIA accelerators gets its real launch when the software arrives, not when the silicon does. For Vera Rubin, that moment happened on October 9, 2026: the vLLM project confirmed it runs on Vera Rubin NVL72 today, with daily container builds and day-0 support for the models that matter most on the agentic frontier — DeepSeek, Kimi (Moonshot AI), GLM (Z.ai) and MiniMax. Engineers from NVIDIA, Inferact and Red Hat had been bringing vLLM up on Rubin hardware since GTC, and the message from the preview post is unusual in server-AI circles: bring-up finished ahead of volume ramp.
The reason is architectural continuity. Rubin is not a from-scratch design — it builds on Blackwell's architecture family, so vLLM's mature Blackwell kernel stack carries over rather than starting from zero. On top of that base, vLLM picked up Rubin-tuned attention, GEMM and MoE kernels through FlashInfer 0.7.0, including a MiniMax Sparse Attention prefill kernel re-tuned for Rubin. And the cleverest piece is locality-aware MoE: using CUDA 13.4's new locality domains, the weights of a mixture-of-experts model get split so that each SM reads its expert weights only from the HBM stack physically nearest to it — turning Rubin's raw bandwidth into usable throughput instead of leaving it on the datasheet.
This is the CUDA moat working exactly as designed, in public. The software ecosystem shows up on day zero, in open source, and the "when will the tooling exist?" question that used to dominate every new accelerator launch — a question Intel's Crescent Island and the custom inference ASICs still have to answer — is simply not a Rubin question. We looked at that contest in our Vera Rubin vs Crescent Island inference shakeout, and the software gap is where it is being decided.
What the early numbers say — and how to read them
The spec deltas NVIDIA is quoting for Vera Rubin NVL72 versus GB200 NVL72: 5x the NVFP4 inference FLOPS, about 2.4x the HBM bandwidth (the jump from HBM3e to HBM4), and 1.7x the bidirectional NVLink bandwidth on the sixth-generation fabric. There is also a quieter change that matters more than it sounds: Rubin boosted exponential throughput — 2x in FP32 and 4x in BF16/FP16 versus GB200 — because softmax is the operation that has quietly throttled attention as everything else got faster.
Then the early benchmark claims: 7.8x per-GPU throughput versus GB200 NVL72 on AgentX at matched interactivity, and up to 3.7x higher VLM throughput versus GB300 NVL72 in MLPerf. Read those carefully — both names are doing work. AgentX is an agentic serving workload, precisely the decode-heavy, long-context regime where memory bandwidth and locality domain tricks pay off most, so it flatters Rubin's strengths the way any debut benchmark does. "Per GPU at matched interactivity" is not the same as a rack-to-rack end-to-end claim, and the vLLM team itself frames these results as early, with optimizations still landing.
Our read: treat 7.8x as the ceiling-shaped direction of travel for agentic inference on Rubin, and expect conservative deployments to see less. What is not in dispute is the structural point — the gap per dollar of rack will be decided in inference throughput, not training FLOPS, and Rubin's spec sheet leads with inference this time. For the competition side of that argument, Intel's 480GB Crescent Island is betting on memory capacity over FLOPS in the same fight.
The supply story is genuinely split
Now the part that should worry anyone waiting for AI hardware prices to cool. Three data points landed within ten days of each other, and they do not agree at first glance. On October 6, KeyBanc reported NVIDIA had cut 2026 Rubin production to roughly 1.5 million GPUs from about 2 million planned — a supply-chain rationing move, which we unpacked separately in our Rubin production-cut reality check. Yet on October 5, India's AM Intelligence announced two further binding orders for 20,000 Rubin GPUs, on top of 9,000 committed in August in Hyderabad — roughly 29,000 GPUs total, adding around 70 MW of capacity across India and Malaysia, with delivery scheduled into Q2 2027. Demand for the product being "rationed" is literally being pre-ordered a year out.
The third piece is the memory teardown that explains the cut. On October 10, Counterpoint Research argued that most Rubin products will ship with 12-layer HBM4 packages rather than the 8-layer configuration that had been rumoured. More layers per stack means more DRAM dies, more yield pressure, and a higher bill of materials per GPU — which both tightens HBM4 supply industry-wide and makes Rubin cheaper to cut in volume than to discount.
Put together, the picture is not a demand collapse — it is allocation plus memory-cost physics. The buyers with binding orders and power capacity (the AMI-class customers) are locked in through 2027; everyone else competes for the allocation that remains, and the memory inside every GPU is the scarce input. That same dynamic is what has kept retail DDR5 prices climbing all year — our RAM budget 2026 guide walks the line items, and the DDR5 tracker has watched kits swing between roughly $448 and $466 this week alone.
Why inference-first silicon changes the 2027 buy math
Zoom out and Vera Rubin's lineup tells you where NVIDIA thinks the money is: the launch platform spans five purpose-built rack-scale systems for agentic AI — the NVL72 flagship, a Vera CPU rack, and LPX/SPX-class inference, networking and storage variants — with Blueprint-grade co-design across the rack. That is a portfolio built for serving agents at cloud scale, and it makes the benchmark scoreboard we watch (token throughput per dollar, not training peaks) the one that actually moves purchases.
For self-hosters, the honest answer about Rubin itself is simple: it is not a homelab product. NVL72 is a megawatt-class buyer's system — AMI's ~29,000-GPU commitment is the buying unit here, not your rack. But the trickle-down clock starts with every generation ramp: as hyperscalers refresh from GB200/GB300 toward Rubin through 2027, prior-generation cards flood the secondhand market, and that is where the local LLM hardware tiers math gets interesting — the used RTX 3090 tier already remains the best token-per-dollar play, and Rubin's ramp adds future supply to exactly that market segment.
Want the trickle-down play today? The used 24GB tier still punches hardest per dollar:
The second-order effect is more interesting long-term: the kernels that make Rubin fast — FlashInfer's attention, GEMM and MoE work plus locality-domain scheduling — land in open source first and then propagate back down the product ladder. Day-0 ecosystem maturity at the top of the market is eventually what makes prior-generation accelerators better than they were at launch, because the same inference stack gets tuned on them.
What to do right now
Three practical moves for October 2026. If you build inference servers: keep your vendor planning independent of Rubin headlines — rack-scale availability stays allocation-bound into 2027, so capacity decisions should price in waiting lists, not datasheets. If you are a homelab builder: Rubin's rack-scale fabric is a reminder that the fat-pipe era has reached everyone — our 10GbE NIC guide for Proxmox covers the cards that stop latency-sensitive VM traffic from bottlenecking long before you ever see rack-scale gear.
And if you are buying memory or storage this quarter: the Rubin/HBM4 story says the input costs that drove retail DDR5 up all year are not reversing — Counterpoint's 12-layer finding and the datacenter DRAM pull both point to sticky prices into 2027. Buy when your kit list needs it and a genuine dip shows up (the DDR5 price tracker is the daily pulse), rather than waiting for a crash the supply data does not support. For the security-minded, remember our earlier point in the GPU firmware supply-chain analysis: as more accelerator generations flow through the market, firmware provenance is the part nobody budgets for.
Memory stays tight while HBM4 hogs the wafers — price your build accordingly:
FAQ
What is NVIDIA Vera Rubin NVL72?
It is NVIDIA's agentic-inference rack-scale system: 72 Rubin GPUs in one NVLink 6 domain paired with Vera CPUs. NVIDIA specs it at 5x the NVFP4 inference FLOPS and about 2.4x the HBM bandwidth of GB200 NVL72, aimed at serving large MoE models like DeepSeek, Kimi, GLM and MiniMax.
Can you run Vera Rubin in a homelab?
No. Rubin NVL72 is a rack-scale product sold into megawatt-class data centers. The realistic homelab impact is trickle-down — as hyperscalers refresh, prior-generation GPUs enter the secondhand market, and the used RTX 3090/4090 tier is still the best token-per-dollar play for local inference.
Are the 7.8x Rubin performance claims trustworthy?
Credible early numbers with clear context: the vLLM team's 7.8x figure is per-GPU throughput versus GB200 NVL72 on the AgentX agentic workload at matched interactivity — not a spec-sheet claim, and agentic serving is exactly where memory-bandwidth and locality gains show up most. Treat it as a direction, not a guarantee for every workload.
Does the Rubin ramp mean DDR5 price relief?
The opposite, near-term. Rubin's HBM4 stacks (Counterpoint says most products will use 12-layer packages, not the rumoured 8) plus datacenter DRAM demand keep memory supply tight and prices sticky into 2027. Watch contract pricing, not headlines, before timing a desktop RAM buy.