GB200 vs GB300 NVL72: What Actually Changed, and Is It Worth Waiting For?
From the outside, the NVIDIA GB200 NVL72 and GB300 NVL72 are almost impossible to tell apart. Both are liquid-cooled racks with 72 GPUs and 36 NVIDIA Grace CPUs joined into one NVLink domain. Both ask the same things of your data centre. The difference is inside the GPU, and it matters a great deal for some workloads and hardly at all for others.
The unpopular summary: GB300 is not better at everything. It is better at the work most AI buyers are moving towards. If your work is different, the older rack may be the smarter purchase.
GB200 vs GB300 NVL72 at a glance
| What | GB200 NVL72 | GB300 NVL72 |
|---|---|---|
| 🧠 GPU | 72 × Blackwell | 72 × Blackwell Ultra |
| 💾 GPU memory in the rack | Up to 13.5 TB |
About 20.7 TB (72 × 288 GB) |
| 🚀 AI performance (NVIDIA's claim) | Baseline |
About 1.5× |
| ⚡ Rack power (HPE) | 132 kW |
132 to 140 kW |
| 💧 Cooling split (HPE) | 115 kW liquid, 17 kW air | About 90% liquid, 10% air |
| 🏗️ Rack and facility needs | Effectively the same: liquid loop, CDU, high-density power | |
Sources: HPE GB200 NVL72 product page, HPE GB300 NVL72 product page, NVIDIA developer blog on Blackwell Ultra. Bars are scaled to the larger value in each row.
What changed
1. Memory is the headline
Each Blackwell Ultra GPU carries 288 GB of memory, up from roughly 186 GB. Across 72 GPUs that is about half as much again. Memory decides how large a model fits, how long a conversation can run, and how many users one rack can serve at once. For reasoning models and long-context inference, it is the number that matters most.
2. Low-precision AI maths got faster
NVIDIA positions GB300 NVL72 at about 1.5 times the AI performance of GB200 NVL72. Independent testing suggests the real gap depends heavily on the model. SemiAnalysis's InferenceX benchmark, for example, reported up to 2.83 times the throughput per GPU on one very large model in May 2026, because that model was limited by memory on the older rack. On smaller models the gap narrows.
Source: SemiAnalysis InferenceX
3. The facility bill barely moved
This is the part buyers under-appreciate. HPE lists the GB200 rack at 132 kW and the GB300 rack at 132 to 140 kW. If your data hall can host one, it can very nearly host the other. The expensive part of going rack-scale, the liquid loop and the power feed, is a one-time decision that covers both generations.
The part nobody advertises
Blackwell Ultra was tuned for AI inference. According to HPC architect Glenn K. Lockwood's published notes, the B300 GPU's double-precision (FP64) performance is dramatically lower than the B200's. Double precision is what traditional scientific simulation uses: weather, fluid dynamics, molecular modelling.
So for a research group running classic simulation codes alongside AI, the "old" rack may be the faster one for half its workload. That is worth confirming with NVIDIA's own figures for your codes before ordering.
Source: Glenn K. Lockwood, B300 notes
Who should pick which
| Your situation | Lean towards |
|---|---|
| 💬 Serving very large or long-context models to many users | ✅ GB300 |
| 🤖 Agentic and reasoning workloads that are memory-bound | ✅ GB300 |
| 🌦️ Mixed AI and double-precision simulation | 🤔 Test both; GB200 may win on simulation |
| 💰 Models that already fit comfortably, budget is tight | ✅ GB200, if price and lead time are better |
| ⏳ Order would not land for a year or more | 👀 Look at Vera Rubin NVL72 timing first |
Is it worth waiting for?
If you are choosing today between the two, the facility work is the same, so the decision is about workload and price. Buyers whose models are limited by memory will feel the GB300 difference immediately. Buyers whose models already fit will mostly be paying for headroom.
And if you are tempted to wait for the next generation instead, remember that every rack generation has a successor already announced. We covered that trade-off in AI Hardware Depreciation: How Fast Does It Lose Value, and When Should You Upgrade?




