In this note
- GB300 vs Vera Rubin NVL72 specs
- Is Vera Rubin cheaper per token than GB300?
- Vera Rubin availability and ramp risk
- What a 60-month term looks like when the next generation lands mid-term
- Will GB300 hold its resale value?
- Which to pick for your workload
- When waiting for Vera Rubin pays off
- Reserving instead of owning
Sign for GB300 now if you need capacity in the next two quarters, if your site is built for racks of about 135 kW, or if most of your work is BF16 training or batch inference. On that kind of work Vera Rubin (VR200) gains little per watt or per dollar. Wait for Vera Rubin if you serve interactive reasoning or agent workloads, or train in FP8 or FP4, and the project can slip two or three quarters.
The early public benchmarks put Vera Rubin at about 2 to 4 times GB300's throughput per megawatt on reasoning inference, for roughly 50% more ownership cost per GPU-hour. So it's cheaper per token. Over a 60-month term that gain can pay for a long wait, as long as the work can wait and your site can power 187 to 227 kW racks.
GB300 vs Vera Rubin NVL72 specs
Both are 72-GPU NVLink racks on the same Oberon rack design. GB300 NVL72 (Blackwell Ultra) has shipped in volume since late 2025. Vera Rubin NVL72 started revenue shipments in August 2026, and the first production cluster went up at a GPU cloud in September.
The figures below come from NVIDIA's October 2026 Vera Rubin datasheet (opens in a new tab) and the GB300 NVL72 page (opens in a new tab), all dense. GB300 rack power is the OEM's published TDP.
| Per rack | GB300 NVL72 | Vera Rubin NVL72 | Ratio | Per watt, Max-Q | Per watt, Max-P |
|---|---|---|---|---|---|
| NVFP4 dense | 1,080 PFLOPS | 2,520 PFLOPS | 2.3x | 1.7x | 1.4x |
| FP8 dense | 360 PFLOPS | 1,260 PFLOPS | 3.5x | 2.5x | 2.1x |
| BF16 dense | 180 PFLOPS | 288 PFLOPS | 1.6x | 1.15x | 0.95x |
| HBM bandwidth | 576 TB/s | 1,400 TB/s | 2.4x | 1.75x | 1.45x |
| NVLink bandwidth | 130 TB/s | 216 TB/s | 1.7x | 1.2x | 1.0x |
| Rack power | about 135 kW | 187 kW Max-Q, 227 kW Max-P | 1.4x to 1.7x | n/a | n/a |
The per-watt columns divide each ratio by the power ratio (187 or 227 kW against 135 kW). NVIDIA doesn't say which power profile its peak figures assume, so treat the two columns as a range.
BF16 is the row to look at if you train in it. Rubin doubled tensor core width for FP8 and FP4 only, as SemiAnalysis noted in February (opens in a new tab), so BF16 per watt is about where GB300 already is.
Is Vera Rubin cheaper per token than GB300?
Peak specs overstate some gains and understate others. The best public comparison so far is SemiAnalysis's July 2026 analysis (opens in a new tab) of DeepSeek R1 serving. It pairs early Vera Rubin results from a GPU cloud with SemiAnalysis's own July 2026 GB300 benchmarks and normalizes both to output tokens per all-in utility megawatt.
| Interactivity (tokens/s per user) | Rubin throughput per MW vs GB300 | Rubin cost per million output tokens, share of GB300's |
|---|---|---|
| about 100 | about 2x | about two thirds |
| about 200 to 250 | about 4x, the widest gap | about a third |
| 300 | 5.4x, with GB300 at the limit of what it can serve | about a fifth |
SemiAnalysis also puts Rubin's ownership cost per GPU-hour about 50% above GB300's, in a scenario where an operator owns the racks. The throughput lead is bigger than that at every point it tested, which is why the cost per token comes out lower across the board.
Treat these as early numbers. The Rubin results come from a pre-production rack at a GPU cloud, and SemiAnalysis says it hasn't independently verified them. They cover one MoE reasoning model in single-turn serving, so dense BF16 training will look more like the 1.6x peak ratio, at higher power. Rubin's software will also improve, as Blackwell's did. The 10x figure that GPU cloud highlighted is against GB200; against GB300 the gap is smaller.
Rack prices point the same way, less clearly. Morgan Stanley put a VR200 NVL72 at about US$7.8 million to hyperscalers and a GB300 NVL72 at about US$4 million, close to double (Tom's Hardware (opens in a new tab)). In March a Tom's Hardware source quoted US$5 million to 7 million for Vera Rubin and US$6 million to 6.5 million for GB300, which is about even. NVIDIA publishes neither.
Vera Rubin availability and ramp risk
A generation that has just started shipping carries risks a mature one doesn't.
| Factor | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| Volume shipments since | late 2025 | August 2026 |
| Orders | Shipping for about a year | Purchase orders from every major hyperscaler, AI cloud and system OEM, per NVIDIA |
| Spec stability | Settled | HBM4 target of 22 TB/s per GPU; the October datasheet works out to about 19 TB/s |
| Software | Mature | Blackwell kernels run; Rubin tuning still in progress |
| Site | About 135 kW per rack, about 10% of heat to air | 187 to 227 kW per rack, 100% liquid |
| Price trend | More than 15% rise reported for systems shipping from early 2027 | Same |
NVIDIA guided about US$20 billion of Vera Rubin revenue for its fiscal third quarter, which Tom's Hardware worked out (opens in a new tab) at roughly 2,550 to 3,975 racks. It's the fastest ramp NVIDIA has had, and most of those racks go to the biggest buyers. If you're outside that group, plan on delivery in 2027 and ask any provider for a date in the contract instead of a target.
Memory is the other risk. By Morgan Stanley's estimate it's about a quarter of a Vera Rubin rack's cost. In August, Bloomberg reported that NVIDIA had told its largest customers server prices would rise by more than 15% in many cases (opens in a new tab), on both Grace Blackwell and Vera Rubin systems shipping from early 2027. Waiting doesn't get you out of that.
What a 60-month term looks like when the next generation lands mid-term
NVIDIA now ships a new data center generation about once a year. A 60-month term that starts in 2027 ends in 2032, and at least two newer generations will ship before then.
| Generation | Status as of late October 2026 |
|---|---|
| Blackwell Ultra (GB300 NVL72) | Volume since late 2025 |
| Rubin (Vera Rubin NVL72) | Revenue shipments since August 2026 |
| Rubin Ultra (NVL72, Kyber NVL144, NVL576) | Planned for 2027; the Kyber rack reportedly slips to 2028 (opens in a new tab), which NVIDIA disputes |
| Feynman | Announced as the generation after Rubin Ultra, on the Kyber rack |
The market rate for a GPU-hour falls under any fixed contract. Each generation lowers the cost of a token, so an older GPU-hour is worth a little less every year while your contracted rate stays put. That's as true of Vera Rubin as of GB300. Vera Rubin just starts the clock about a year later.
What you run in years 4 and 5 matters more than the launch benchmark. A rack that serves a frontier model in year 1 is often running smaller or distilled models, batch jobs and evals by year 5, and the real question is whether that work is still worth the rate.
And the wait doesn't end. Rubin Ultra is planned for 2027, inside the first year of a 2027 Vera Rubin deployment. Picking a generation decides when you start; it won't keep you current.
The defenses are structural. Split the commitment across generations (GB300 now for what you need in 2027, Vera Rubin for later), and prefer contracts with a known exit point and a way to roll part of the commitment into the next generation. Our post on GPU depreciation and useful life covers how owners account for this.
Will GB300 hold its resale value?
If you own the racks, residual value at the end of the term is part of the price. So far older NVIDIA GPUs have held more value than the yearly cadence would suggest, at least while capacity stays short.
One listed GPU cloud has depreciated its GPUs over six years since 2023. In November 2025 its CEO said a batch of H100s coming off contract was rebooked at 95% of the original price, CNBC reported (opens in a new tab). Amazon moved the other way in 2025, cutting the useful life of part of its server fleet from six years to five because of the pace of AI hardware, and Microsoft gives a range of two to six years for computer equipment.
The rental market tells a similar story. On Silicon Data's H100 index (opens in a new tab), H100 rental rates rose about 6% through October 2026, four years after the part first shipped.
Rack-scale systems complicate this. A GB300 NVL72 rack can only go to a buyer with a liquid-cooled site that can feed about 135 kW per rack. That's a much smaller pool than for air-cooled HGX servers, and there's no long resale history for these racks yet. At 187 to 227 kW, a Vera Rubin rack narrows the pool again.
Which to pick for your workload
| Your situation | Lean toward |
|---|---|
| Capacity needed within six months | GB300 |
| Mostly BF16 training, batch or low-interactivity inference | GB300 |
| Site designed for racks of about 135 kW, no power upgrade planned | GB300 |
| Interactive reasoning or agent serving above about 100 tokens per second per user | Vera Rubin |
| FP8 or FP4 training of large MoE models | Vera Rubin |
| Fixed megawatts and a site that can take 227 kW racks | Vera Rubin |
| Long term, work that can wait two or three quarters | Vera Rubin, or GB300 now plus Vera Rubin later |
When waiting for Vera Rubin pays off
Here's one way to frame the wait for a finance memo. Say Vera Rubin does k times as much useful work per dollar as GB300 and arrives d months later. Over a window of W months it catches up on total work when d equals W times (1 minus 1/k).
The table applies that formula, so it's an estimate. The values of k are our assumptions, from roughly flat (BF16 training) up to the cost-per-token gains SemiAnalysis measured on interactive inference.
| Work per dollar, Rubin vs GB300 (k) | Break-even delay, 36-month window | Break-even delay, 60-month window |
|---|---|---|
| 1.1 (BF16 training, batch) | about 3 months | about 5 months |
| 1.5 | about 12 months | about 20 months |
| 2.0 | about 18 months | about 30 months |
| 3.0 (interactive reasoning) | about 24 months | about 40 months |
The formula assumes you have the demand to use the extra throughput, and a Vera Rubin delivery date you can rely on, which in 2027 is the hardest input to pin down. It also gives no weight to having capacity now. For a product launch or a research deadline, that can outweigh everything in the table.
Reserving instead of owning
Lightyard reserves GB300 NVL72 as whole racks today. Vera Rubin NVL72 is on the waitlist, with first allocation, and can't be reserved yet. Either way it's a single-tenant cluster in the region you choose, run to an SLA, with failed parts replaced as part of the service. The contract includes an exit option, on terms set out in your proposal. Part of the commitment can roll into a newer GPU generation when it ships, so a GB300 commitment doesn't have to outlast the point where the next generation makes sense. See pricing, contract terms, or talk to us.
Questions this note answers
It depends on the precision and the workload. On NVIDIA's peak figures, FP8 compute per watt is about 2 to 2.5 times GB300's and BF16 per watt is roughly flat. On measured DeepSeek R1 serving, SemiAnalysis found about 2 times GB300's throughput per megawatt at moderate interactivity and about 4 times near 200 tokens per second per user.
For interactive reasoning inference and FP8 or FP4 training, yes, once you can get it and your site can power it. For BF16 training or batch inference the gain per watt is small, and a GB300 rack you can have sooner usually wins.
On early benchmarks, yes. SemiAnalysis estimates a Vera Rubin GPU-hour costs about 50% more to own than a GB300 one, but the throughput lead is bigger, so cost per token came out at about two thirds of GB300's at 100 tokens per second per user and about a third at 200 to 250.
Its market rate will fall over time, as every generation's does, but older GPUs have kept earning while supply is short. In late 2025 a listed GPU cloud said H100s coming off contract were rebooked at 95% of their original price, and H100 rental rates were still rising in late 2026.
Yes. According to SemiAnalysis, Rubin runs Blackwell kernels from the main libraries (CUTLASS, DeepGEMM, FlashMLA), which makes bring-up easier. Getting peak performance still takes Rubin-specific tuning.
Sources
- NVIDIA Vera Rubin datasheet (October 2026) (opens in a new tab)
- NVIDIA GB300 NVL72 product page (opens in a new tab)
- OEM product guide: GB300 NVL72 rack-scale AI (opens in a new tab)
- SemiAnalysis: Vera Rubin NVL72 vs GB200 NVL72, inference TCO (July 23, 2026) (opens in a new tab)
- SemiAnalysis: Vera Rubin, Extreme Co-Design (February 25, 2026) (opens in a new tab)
- NVIDIA Technical Blog: NVIDIA Vera Rubin POD (March 16, 2026) (opens in a new tab)
- NVIDIA Blog: first production Vera Rubin cluster (September 30, 2026) (opens in a new tab)
- Tom's Hardware: Large-scale GB300 shipments to start in September 2025 (July 18, 2025) (opens in a new tab)
- Tom's Hardware: Nvidia expects US$20 billion of Vera Rubin sales in Q3 FY2027 (August 27, 2026) (opens in a new tab)
- Tom's Hardware: Price of Vera Rubin NVL72 racks (March 24, 2026) (opens in a new tab)
- Tom's Hardware: Morgan Stanley VR200 NVL72 cost estimate (May 21, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia reportedly warns of 15% price hikes on AI servers (August 23, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia's Kyber rack for Rubin Ultra reportedly slips to 2028 (July 6, 2026) (opens in a new tab)
- CNBC: How long before a GPU depreciates? (November 14, 2025) (opens in a new tab)
- Silicon Data: H100 Rental Price Index (SDH100RT) (opens in a new tab)