Blackwell, Decoded
Foundations · Module 1

Why Blackwell?

The shift to reasoning AI5 min readModule 1 of 8

For years, better AI meant one thing: train a bigger model on more data. That recipe still works — but the frontier has moved. The newest models don’t just answer; they reason. And reasoning, it turns out, is expensive in a completely different way.

#The old playbook: bigger is smarter

The first decade of modern AI ran on a simple, reliable bet called pre-training scaling: make the model larger, feed it more data, and throw more compute at it, and it gets predictably smarter. This is what produced today’s billion- and trillion-parameter models. NVIDIA says the compute required for pre-training has grown 50 million-fold in five years.

But you can’t just keep making the base model bigger forever. Building genuinely useful systems now depends on two newer techniques that happen after the giant pre-training run.

#The new playbook: refine, then reason

Two newer “scaling laws” pick up where pre-training leaves off:

  • Post-training — teaching the model its job. After pre-training teaches a model the knowledge of the internet, post-training tunes it for specific tasks using domain-specific and synthetic data. Because synthetic data has no upper limit, this stage is hungry for compute.
  • Test-time scaling — letting the model think. Also called long thinking, this spends extra compute while answering. Instead of replying in a single pass, a reasoning model weighs multiple possibilities and refines its answer in real time.
1 · Pre-train Learn the world from internet-scale data 2 · Post-train Specialize for the job ~30× pre-train compute 3 · Reason Think before answering ~100× one inference pass Each stage adds intelligence — and demands more compute
The modern AI pipeline. Module 2 unpacks each of these three scaling laws in detail.
The compute problem in two numbers

Post-training can need 30× more compute than the original pre-training run to build a customized model. And long thinking can need 100× more compute than a single inference pass to crack a genuinely hard problem. Multiply those together across millions of users and you see why a new machine was needed.

#Why this needs new hardware

Reasoning AI doesn’t just want more compute — it wants that compute delivered with low latency, in real time, with extremely fast links between chips. That combination is exactly what the Blackwell architecture was purpose-built for: data-center-scale reasoning with up to 30× the energy efficiency of the previous-generation NVIDIA Hopper GPUs.

What “Blackwell” actually refers to

It’s a family, not a single product. This course covers the Blackwell and Blackwell Ultra GPUs, the GB300 NVL72 and GB200 NVL72 rack-scale systems, and the HGX B300 / B200 servers. Each is a different way to package the same underlying chip.

The payoff NVIDIA promises: any enterprise can deploy state-of-the-art LLMs with affordable economics, while unlocking a class of real-time, high-throughput AI that simply wasn’t possible before. The rest of this course shows how.

Continue →