Blackwell: The revolutionary GPU transforming AI — Faster and more energy-efficient. What’s next?

Summary

The Superchip Revolution: How Blackwell Made AI Affordable and Powerful for Everyone

What is NVIDIA Blackwell?

NVIDIA just released a new type of computer chip called Blackwell. A chip is like the brain of a computer. Blackwell is much more powerful than the old chips.

It is made for artificial intelligence, which is computer programs that can learn and think. Companies are using Blackwell to run AI programs faster and cheaper than before.

Think of it like this: the old chips were like a regular car. The new Blackwell is like a high-performance race car. It goes much faster and uses much less fuel.

How Much Faster is Blackwell?

Blackwell is much faster than the old chips. Here are the numbers:

The old chips were called Hopper. A Hopper chip could do 10 petaFLOPS of work. A petaFLOP is a very big number—it is 1 quadrillion calculations. Blackwell can do 20 petaFLOPS. That is double the power.

But here is the amazing part: when you put multiple Blackwell chips together, they work 30 times faster than the old systems. Think about it like this: if an old system with the old chips took 1 hour to answer a question using AI, Blackwell could answer the same question in 2 minutes.

The electricity bill also goes down. Blackwell uses 25 times less electricity to do the same job. And it costs 25 times less money to buy and run.

Real Examples of How Companies Use Blackwell

Microsoft, Amazon, and Google all announced that they will use Blackwell in their cloud services. They are building gigantic data centers full of Blackwell chips. Here is why:

Amazon said they can now run more AI programs at the same time using the same building and electricity. It is like fitting 40 times more cars into the same parking lot.

Microsoft is building custom supercomputers with Blackwell. These are computers with thousands of Blackwell chips all connected together, all doing work at the same time. Google and Oracle are doing the same thing.

In Malaysia, a company started an AI gym where anyone can rent Blackwell computers by the hour. Just like you rent a gym membership to use the gym, now companies can rent AI computers. This is new. Before Blackwell, you had to buy your own expensive GPU or go without. Now, you can rent what you need.

Why Does Blackwell Exist?

The problem with the old chips was a simple thing called memory bandwidth. Imagine you have a worker in a factory. The worker is very fast at making things. But if you only deliver materials slowly, the worker sits around waiting. That is what happened with the old chips. The chip could do calculations very fast, but the data (the materials) moved slowly to the chip.

Blackwell solved this problem. It has a much wider pipe for data to flow. The memory can send information 2.4 times faster. Think of it like going from a normal garden hose to a fire hose. More water flows faster.

Blackwell also has much more memory. The old chips had 80 gigabytes of memory. Blackwell has 288 gigabytes. That is like going from a cup to a bucket.

A New Thing Called FP4

Blackwell has a new trick called FP4. FP means floating point, which is a way to store numbers in a computer. FP8 uses 8 bits (a bit is a 1 or 0) to store a number. FP4 uses only 4 bits.

Think of it like this: if FP8 is storing a number with lots of details, FP4 stores the same number with fewer details. You do not need all the details to run an AI program. FP4 lets you do the same work with half the memory and twice the speed.

This is a big deal because it means you can train bigger AI models using less memory.

What About Cost?

Blackwell costs a lot of money. If you rent a Blackwell chip for three years, it costs between 3 and 4 dollars per hour. That might not sound like much, but for a whole company using many chips, it adds up.

But here is the thing: because Blackwell is so fast and uses so little electricity, companies save much more money in the long run. If Blackwell costs twice as much to buy, but it does the work 30 times faster and uses 25 times less electricity, you come out way ahead.

A company that buys 8 old chips might generate 300 million AI responses per second. The same company buying 4 Blackwell chips can generate billions of responses. That is huge.

The Problem With Power

Blackwell needs a lot of electricity. One Blackwell chip uses 1,200 watts. That is like running 12 refrigerators at once.

A room full of Blackwell chips needs between 60,000 and 120,000 watts. That is enough electricity to power 200 houses.

This is starting to cause problems. Data centers in Virginia where many AI computers are located are hitting electrical limits. The power companies cannot deliver electricity fast enough. It is not that they cannot make the electricity, but they cannot move it to where the computers are.

Some estimates say that by 2030, AI data centers will use 945 terawatt-hours of electricity per year. That is like the entire electricity use of Japan. This is a huge challenge.

The Competition

NVIDIA is winning the competition. The company controls 92 percent of the GPU market. That means 9 out of 10 computer chips sold for AI are made by NVIDIA.

AMD makes a chip called MI300. It is good, but not as good as Blackwell. AMD only has 7 percent of the market. Intel just started making GPUs and has 1 percent of the market.

Why does NVIDIA win so much? Because NVIDIA has something called CUDA. CUDA is like the language that programmers use to write programs for NVIDIA chips. CUDA has been around for 20 years. Everyone knows CUDA. It is much harder to learn AMD or Intel chips. That is why NVIDIA wins.

What Comes Next?

NVIDIA has announced new chips coming in 2026 and 2027.

The new chips will be called Vera Rubin and Rubin Ultra. Each new chip will be better than the one before.

The new Rubin chip uses 40 percent less electricity. That is amazing because it means the electricity problem might get better.

Companies are talking about AI agents.

An AI agent is a program that can think for itself, plan things, and solve problems without a human telling it what to do every step.

Blackwell is perfect for AI agents because they need lots of computing power to think through many possibilities.

Conclusion

Blackwell is here, and it changes everything about AI. It is much faster, much cheaper, and much more efficient than before. For the first time, many companies can afford to do serious AI work.

But Blackwell creates new problems.

It uses lots of electricity. It requires expensive infrastructure. And NVIDIA controls almost all of the market, which means one company decides the future of AI.

The next few years will show if Blackwell lives up to all the hype. If it does, AI will become as normal in business as email is today. But if the electricity problem gets worse, we might have to slow down.

Either way, Blackwell is a big deal.