Learn / GPU hardware
GPUs for AI: the 2026 lineup, explained
From the Hopper H100 to Blackwell Ultra and rack-scale GB300, plus AMD's MI355X. What each part is for, how they compare on memory and generation, and how to choose the right one, then rent or reserve it in Europe.
The AI GPU lineup
The datacenter market runs on a handful of parts. NVIDIA spans two live generations in 2026, Hopper (H100, H200) and Blackwell (B200, B300, and the rack-scale GB300 NVL72), and AMD's Instinct MI355X is the credible alternative. They differ most on memory per GPU, the number that decides how much model and context fit before you have to shard across cards. The bars below are drawn to that number. Open a guide for the full spec sheet, or the live index for what it costs.
- NVIDIA Hopper
NVIDIA H100
80 GB HBM3The mature reference GPU. Widely available, fully tuned software, softened pricing. The default for anything that fits in 80 GB.
- NVIDIA Hopper refresh
NVIDIA H200
141 GB HBM3eSame Hopper compute, nearly double the memory. Buys longer context and bigger KV caches without moving to Blackwell.
- NVIDIA Blackwell
NVIDIA B200
192 GB HBM3eThe volume Blackwell part. Native FP4 and a large jump in throughput. The current-generation default for new training.
- NVIDIA Blackwell Ultra
NVIDIA B300
288 GB HBM3eMaximum memory per GPU. Frontier training and long-context serving that would otherwise spill across nodes.
- NVIDIA Blackwell Ultra rack
NVIDIA GB300 NVL72
288 GB per GPU72 B300 GPUs fused into one NVLink domain. A rack that behaves like a single very large accelerator.
- AMD CDNA 4
AMD MI355X
288 GB HBM3eThe credible second source. Matches B300 memory, runs the open ROCm stack, and typically prices below the NVIDIA part.
Two generations, one decision
Hopper is the incumbent. The H100 shipped in volume, so it is the part almost every cloud has in stock, with a software stack that has been tuned for years and a price that has softened as Blackwell arrived. The H200 is the same compute die with faster, larger HBM3e memory bolted on, which buys longer context and bigger batches without changing anything else about how you deploy.
Blackwell is the frontier. The B200 roughly doubles Hopper's useful throughput and adds native FP4, the low-precision format most new inference stacks now target, which is why it is the default choice for fresh training runs and high-volume serving. The B300, marketed as Blackwell Ultra, trades a little of that balance for the largest memory on any single GPU, 288 GB, and the GB300 NVL72 fuses 72 of those GPUs into one NVLink domain so a whole rack addresses memory as if it were a single accelerator. That is what makes trillion-parameter models tractable without a fabric of slow inter-node hops.
AMD sits alongside both. The MI355X matches the B300 on memory per GPU and runs the open ROCm stack with day-one support for the mainstream serving engines, so for memory-bound inference it is a genuine second source rather than a science project. It typically prices below the equivalent NVIDIA part, and the discipline that a real alternative imposes on the whole market is worth as much as the sticker saving.
How to choose
Start with one question: is your workload compute-bound or memory-bound? If you are compute-bound on current-size models, the newest silicon you can afford wins, and the B200 is the volume default. If you are memory-bound, long contexts, large KV caches, or very large mixture-of-experts models whose total weights set the floor, capacity per GPU matters more than peak FLOPS, and the B300 or MI355X earn their place by keeping the deployment on a single node instead of a fabric of many.
A useful shortcut is to size the model first. A dense model needs roughly one byte per parameter at FP8, half that at FP4, before you add the KV cache that grows with context and batch. Work out how many gigabytes that is, add headroom, and the memory number on each card above tells you how many GPUs you need. If one card holds the model, you have found your floor; if it does not, the question becomes how few cards you can spread it across, and that is exactly where the 288 GB parts change the arithmetic.
Then weigh the two practical constraints. Availability: Hopper shipped in volume, so H100 capacity is the easiest to secure on demand, while frontier Blackwell is tighter and mostly spoken for ahead of delivery, which is why reservation matters. Budget: reserved multi-year terms price well below on-demand, larger clusters come down per GPU, and a second source like AMD keeps the whole market honest. For most teams the answer is a mix, H100 for the bulk of production and Blackwell for the frontier, sized to what the workload actually needs rather than to the top of the spec sheet.
Era tracks live prices for every GPU above across the major clouds and neoclouds, and reserves multi-year capacity in Europe on air-cooled B300 today and liquid-cooled GB300 NVL72 from 2027. Tell us the workload and we will size it.
Reserve capacity in EuropeGPU guides
Deep dives on the individual accelerators, each with specs, comparisons, node-fit math, and live pricing:
Questions, answered
Which NVIDIA GPU is best for AI in 2026?
There is no single best GPU; there is a best GPU for a workload. The H100 is the mature, cost-effective reference for anything that fits in 80 GB. The B200 is the current-generation default for new training and FP4 inference. The B300 and rack-scale GB300 are for frontier training and the largest memory-bound models. Match the part to whether you are compute-bound or memory-bound, and to your budget and availability.
Should I still buy H100 capacity, or move to Blackwell?
If your models fit comfortably in 80 GB per GPU or a 640 GB node, the H100 is usually the better economics: it is widely available with softened pricing and a fully tuned software stack. Move to the B200 or B300 when memory capacity, context length, or model scale forces multi-node deployments on Hopper.
How do I choose between the B200 and the B300?
Same Blackwell generation. Choose the B200 when you are compute-bound on current-size models, where it is the volume part and the better value. Choose the B300 when you are memory-bound, big-batch serving, or running very large mixture-of-experts models that spill past a single B200 node, where its 288 GB per GPU consolidates the deployment.
Is AMD a real alternative to NVIDIA for AI?
For memory-bound inference, increasingly yes. The MI355X matches the B300 on memory per GPU, runs the open ROCm stack with day-one support for mainstream serving engines, and typically prices below the equivalent NVIDIA part. It is a genuine second source, with the caveat that you should validate your specific model and kernels on ROCm first.