NVIDIA GPU roadmap explained: from A100 to H200 and beyond

If you have spent any time provisioning AI infrastructure over the last few years, you have watched the ground shift under you more than once. A100. Then H100. Then H200. Now Blackwell and Rubin are showing up in every procurement conversation.

It is a lot to track. So here is the roadmap laid out plainly, generation by generation. The context in this will help you when you are the one deciding what to run your workload on.

Quick answer: NVIDIA's data center GPUs have moved through four major architectures in recent years: Ampere (A100), Hopper (H100 and H200), Blackwell (B200 and B300), and now Vera Rubin, arriving through the second half of 2026. Each generation brings more memory, faster interconnects, and lower-precision compute formats built specifically for AI workloads.

Why this roadmap matters more than a typical spec bump

In consumer hardware, a new generation usually means "faster." In data center AI hardware, a new generation usually means a workload that used to need four GPUs now needs two, or a model that used to require careful partitioning across chips now fits on one.

That difference changes budgets, power planning, and how long your current hardware stays useful. It is worth understanding, even if you never touch a GPU directly.

The generations at a glance

Now let's walk through why each of these actually mattered, not just what the spec sheet says.

A100: the chip that started the AI infrastructure boom

The A100 is the reason most of today's AI infrastructure exists in its current form. It was not built with chatbots and generative AI in mind. It was built for a mix of scientific computing and early deep learning, and it happened to be in the right place when the AI boom hit.

If you are still running A100 fleets, that is not necessarily a bad thing. For smaller models, batch inference, or validation runs before a production deployment, A100 capacity is often cheaper and still gets the job done. Teams frequently keep A100 around specifically for this reason: it is a low-cost way to test before committing to newer, pricier hardware.

H100: built for the transformer era

H100 is where NVIDIA stopped generalizing and started designing specifically around transformer-based models. It introduced FP8 precision through the Transformer Engine, which let training and inference run faster without a meaningful accuracy penalty for most workloads.

This is the chip that most large language model training happened on through 2023 and 2024, and it remains the default choice for a huge share of production inference today. If your workload is well understood and your team has already tuned around H100, there is rarely an urgent reason to move off it just because something newer exists.

H200: the upgrade people underestimated

H200 GPU gets described as a minor refresh, and that undersells it. The compute is the same as the H100. What changed is memory: H200 carries 141GB of HBM3e, close to double what H100 offers, with meaningfully higher bandwidth.

Here is why that matters in practice. Large language model inference is often memory-bound, not compute-bound. That means the bottleneck is how fast data moves in and out of memory, not how many calculations the chip can do per second. For those workloads, H200 delivers a real, measurable improvement, even though the core compute architecture did not change.

If your inference workload is currently split across more GPUs than feels necessary just to fit the model in memory, H200 is usually where that problem gets solved.

Blackwell (B200 and B300): rack-scale computing arrives

Blackwell (B200 and B300) is a bigger jump than H100 to H200. It is not just a faster chip. It is a shift toward thinking about the rack as the unit of compute, not the individual GPU.

A few things changed at once with Blackwell:

• More memory per GPU. Up to roughly 288GB depending on the variant, a substantial jump over H200.

• NVFP4 precision. A new 4-bit format that roughly doubles throughput over FP8 for workloads that can tolerate it, which covers a growing share of inference use cases.

• Rack-scale interconnects. Systems like the GB200 NVL72 link 72 GPUs together with extremely high bandwidth, so the rack behaves more like one giant accelerator than 72 separate ones.

That last point has a real consequence most teams do not plan for early enough: power and cooling. A fully loaded NVL72 rack draws well over 100 kilowatts and requires liquid cooling. If your facility was built around air-cooled racks, that is a planning conversation you need to have months before hardware shows up, not after.

What comes next: Vera Rubin and beyond

NVIDIA has publicly laid out its roadmap further ahead than it typically has in the past, which is itself a signal of how far in advance hyperscalers now lock in capacity.

Here is the short version:

• Vera Rubin (H2 2026): 288GB of HBM4 memory, a new NVLink generation with roughly double the per-GPU bandwidth of Blackwell, and rack-scale designs like the NVL144 that push well past what Blackwell racks deliver.

• Rubin Ultra (2027): A higher-bandwidth refinement, following the same pattern NVIDIA used going from B200 to B300.

• Feynman (2028): The next full architecture generation, still early in disclosure but already referenced in NVIDIA's public roadmap.

The pattern across every generation is consistent: more memory, faster interconnects, and precision formats tuned tighter to AI workloads instead of general-purpose compute.

What this actually means for your buying decisions

Here is where the theory turns into a practical decision, and where most teams either overspend or under-plan.

• Don't chase the newest chip by default: If your current generation handles your workload comfortably, the newest release is not automatically worth the switch. Match the generation to the actual bottleneck you have, whether that is memory, bandwidth, or raw compute.

• Check whether your bottleneck is memory or compute: This single question decides more than almost anything else. If you are memory-bound, jumping from H100 to H200 alone can solve real problems without a full architecture change.

• Plan facilities before you plan hardware: Power and cooling requirements have grown faster than most facility upgrade cycles. Confirm your infrastructure can support a generation before you commit a budget to it.

• Consider renting before you own: Given how fast this roadmap moves, renting capacity from a cloud provider avoids being locked into hardware that ages out of relevance in 18 months. This matters even more during periods when new hardware carries long lead times.

• Re-run your cost math every generation: A workload that needed four older GPUs might need two current-generation ones. The math changes enough each cycle that assumptions from a year ago are often already outdated.

The bottom line

NVIDIA's roadmap moves fast, but the underlying pattern is easy to follow once you see it: more memory, faster interconnects, and precision formats built for AI rather than general compute. Every generation from A100 to Rubin follows that same thread.

The teams that manage this well are not the ones chasing every new release. They are the ones that know exactly where their own bottleneck sits, memory, bandwidth, or compute, and pick the generation that actually solves it.