Xai’s Grok 3: All graphics processing units, none of the breakfasts

At the end of February, Elon launched his latest style. Of course, it was “the best in the world”.

Is it really the smartest artificial intelligence on earth?

As usual, the musk brought the noise train. But there was not much objective data at launch. Xai short Blog post He stated that he was still in the experimental version and the models were active exercises.

Some criteria that appear GROK 3 have shown in the future. However, they did not give access to the application programming interface. It is important because independent standards use it for evaluation.

Therefore, Elon claims that Grok 3 is “wrongly smart” and seizes everything else. But the only ways to verify are chatting with them yourself or looking at their standards.

And those criteria? Take a look:

Do you see that area lighter on the right? This is Boost Grok who got more Calculating (Test Time Calculate) to get more consistent answers. It is not exactly a fair battle.

You may know that artificial intelligence models often give slightly different answers every time – sometimes better, sometimes worse. Most of the criteria ignore this contrast, and the first response evaluation (Passing@1). It is simpler and coincides with how artificial intelligence already uses – we expect a good answer in the first attempt.

But GROK results were displayed using Cons@64. Meaning, I got 64 attempts for each question and chose the most common answer. Then, XAI compares that reinforced result versus dozens of Pass@1 of competitors.

So on the one hand, they claim to be a model of the next generation. On the other hand, they use very cheap tricks.

To be fair, in such a competitive field, all laboratories destroy the rules. They are cherry sewing criteria or exclude stronger models of comparisons-but they are rarely flagrantly.

Well, aside criteria. What do experienced users actually say? General consensus is:

The model is huge but did not make penetrations. Halosa is still tilted towards excessive long responses.

In terms of performance, Grok 3 lands somewhere near Openai Top, and perhaps a little better than Deepseek and Google’s Stuff at the time of the release.

However, two months later, the new Gemini 2.5, Claude 3.7 and GPT-4O arrived. We also finally got a partial API access to Grok 3 and its small copy. Unfortunately, the mini version only received the API thinking.

So today we know that it is costly and certainly not the best.

But hold, there is still more for the story.

The model is interesting and deserves to be considered. You have to hand them over to them, Elon and Kay jumped to the market quickly, becoming a major player in record time.

1 – The device

The big story here?

In 2024, Xai built a huge account block. We speak 100000 Nvidia H100 graphics processing units and run in just 4 months. Then they doubled this to 200,000 cards in another 3 months.

Nvidia CEO, Jensen Huang, Mentioned This usually takes about 4 years.

This was a huge engineering work. This time, there is no funny work – it’s the largest data center in the world. No one else was able to connect many graphics processing units in one place.

These groups are usually multiple regular data centers linked to costly Infiniband cables. During training, these centers need to switch a lot of data constantly. If the connection is slow, the expensive graphics processing units are sitting in lethargy, which is bad news.

The typical data center may contain 10,000 to 20,000 graphics processing units, and absorbs 20-30 megawatts of power. For exampleMicrosoft (Openai) runs the 100K graphics processing units in Arizona, and runs 128 kg.

Seeing the HH -shaped buildings? This is the position of standard identification data next to each other.

Energy needs for upper groups have exploded up to 10x since 2022. We are now talking about about 150 megawatts per group. This is like running a small city. This creates a huge load on regional energy networks. In some places, it is actually cheaper to generate energy from serving it because there are not enough energy lines.

Therefore, Elon brings this market behind. And … is “something Elon”. I hate his tweets everything you want, a man knows how to build factories like anyone else.

He bought an old electricity factory in Memphis and decided to create one giant data center instead of a network like anyone else.

As expected, power has become a problem.

The factory had only 7 megawatts from the local network – with only 4000 graphics processing units. The local tool, the Tennessee Valley Authority, promised another 50 MW, but not until August. A subtitle is still built of 150 megawatts of Xai, and not ready until the end of the year.

But waiting is not musk style.

Dylan Patel (from half solutions) I was monitored Through satellite images that just brought Elon 14 huge diesel generators from Voltagrid. A drug addict of up to 4 mobile mobile stations and operated the data center. Literally transporting the truck in electricity.

Patel stated that they may have bought 30 % of the entire American market for these generators (although I could not find anything on that).

Incalously, the data center also uses liquid cooling. Google has really done this really widely before. This is a big problem because the next generation of NVIDIA, Blackwell B200s requires liquid cooling. Anyone else will have to amend the current data centers.

You can check the first few minutes of this video to see what appears inside. I got a muffled laugh at the extent of the man to pressure the gray boxes and cables:

It is seriously great engineering – just look at cable management.

No one has done this huge work in a very short time.

2 – More devices!

Eleon says by Summer 2025, they will have a group of GPU 300K with Blackwell B200 chips. Looking at the Musk habit of exaggeration, let’s say it is somewhere between 200-400 thousand new chips by the end of 2025. B200 is about 2.2 times better than H100 for typical training (based on November 2024 estimates).

Musk is even planned to build a 2.2 GB custom power plant. This is the power of more than a medium -sized city consumes.

It is not alone – all the big players do something similar:

• Meta is building two gas plants in Louisiana.
• Openai/Microsoft prepares something similar in Texas.
• Amazon and Google also build Gigawatt databases.

Why not nuclear? I got the power, but building a nuclear factory takes a long time. Not only can one run next to your data center per year. Wind and solar farms in addition to promising batteries, but they take a long time to publish the required scale.

As a result, both Microsoft and Meta had already had to back down from their green and green promises. They broke their backs to raise Molosh to heaven!

3 – Grok 3 huge

Therefore, Elon built this huge expensive box. Now what?

It is estimated that Grok 2 was trained on about 20 thousand H100, while Grok 3 used more than 100,000. For context, GPT-4 was trained for 90-100 days on approximately 25 thousand A100 chips, with approximately H100 2.25X faster.

When doing mathematics, Grok 2 got twice the computing power that you threw compared to GPT-4. Grok 3 has received five times more than GROK 2. Gewing Gemini 2.0 is likely to use a similar amount of devices (100,000 TPUV6 chips), but the same model is likely to be smaller.

Basically, the total Check Calculation For Grok 3 is an arrangement in size (10 times!) Higher than its closest competitor. Unfortunately, we do not have general data for GPT-4.5 or Gemini 2.5.

So they poured crazy quantities of resources in building this huge group, and the resulting model is … on an equal footing with job occupants. Certainly, the league championships are not better.

XAI’s experience appears to be still backward from Openai, Google or Anthropic. They are mainly brutal to the top layer. There are no magic tricks, only: “If the brute force does not solve your problem, you do not use it enough.”

But there is hunting with this approach. The era of artificial intelligence Estimates Over the past decade, algorithm improvements have formed about a third of progress in the capabilities of the model. The other two -thirds came from just throwing more devices and data in larger models.

The brute force has worked in favor of GROK 3 this time, but the costs will increase significantly with the provision of less and less improvement. Xai needs to catch up next to the algorithm. The good news is that they are now looking at him as they pay the borders, so it is possible that it is easier to attract higher talents.

4 – What is good in Grook?

1. It is completely free (perhaps even the full version).

Without the narrow limits of anthropoor, the interruption of Deepsic, or paid Openai levels.

Even with the decline in all new models in the past two months, Grok still keeps itself near the top of Chatbot Arena Total board.

We also have independent standards Epochai:

and Livebench:

2. Deep thinking and research mode

Again in February, the feature of deep free search was often puzzled. Now, Google and Openai offer some in a basic layer – did Grok pay them?

This mode automatically analyzes 30-100 links (you may do more Google) in minutes and collect a detailed (mixed) summary that you only need to disagree and check the facts. It is an easier way to search for anything than scratch. I found that the Grok version is working faster than others, so I started using it when I need to search for something. Like, when buying new headphones.

3. Integration with x

This may be the advantage of the killer: semantic search is not only for keywords, but for what you mean. You can also ask her to summarize posts on a topic to track trends. Or to find recent publications from a specific user.

Twitter is the closest to the actual time information platform, so this is great. But so far Grok has lags backward, and draws data from the last two days instead.

4. Unarmed things

To get the great end, put 18+. It is easy to help break the protection without great effort. You can get it … well, whatever you want, from FLRETY sounds to questionable recipes. Examples of sound mode are particularly wild.

Listen to the end, it’s honey!

Ironically, Grook itself does not seem to carry musk (or Trump) in a great appreciation. When this appeared, the XAI tried a repair – hardliner in a hard -line GROK criticism of ELON. When this exploded, he blamed a former Openai employee for “not installing the company’s culture.” Super Range.

The real issue is that Grok’s views are just a reflection of their training data (i.e. the Internet), not some deliberate bias. Trying to correct these observations without tampering with the entire form is difficult.

5 – Should you care about her experience?

Certainly try it, but as a second pilot for you.

TLDR:

• The cost of a method of training is more than the models of competitors.

• Nevertheless, the performance is equal with the best.

• It is very fast and free (currently).

• Deep search mode is really useful – try it if you don’t do it.

• More vulnerable to hallucinations and jumping to conclusions very quickly.

• The responses are usually well organized but often feel bloated.

• Farid access to Twitter data.

Xai has proven that it is able to build a world -class infrastructure at an unprecedented speed. But in actual artificial intelligence capabilities, they mainly buy their way to the top with a transparent account force.

This adds another powerful player who presses Openai, Google and Noteropic, pushing the artificial intelligence industry towards commodities. Competition is competing and the uniqueness of upper models fades.

I enjoyed this? Give voting or subscribing to my news message. I will appreciate that!