With 200,000 GPUs, Musk staged “AI Revenge”

Wen| Silicon-based laboratory kiki

“An AI that pursues the ultimate truth.”

At noon on February 18, Beijing time, Musk and xAI released their latest flagship model Grok-3 series and the latest chat robot Grok. With Musk’s powerful aura, although xAI is a latecomer in this AI competition, But every move it has attracted much attention.

In a press conference that lasted less than an hour and was watched by millions of people, xAI demonstrated Grok-3 ‘s impressive modeling capabilities, from Musk’s strongest data clustering to direct comparison of evaluation data, seemingly responding to his previous comments on Grok-3 ‘s evaluation of “the strongest AI on the surface.”

Musk and xAI release their latest flagship model Source: xAI

However, according to former OpenAI co-founder Anderj Karpathy,”Grok-3+Thinking” feels similar to OpenAl’s strongest model, o1-pro, but achieving the same ability requires a prerequisite: Only six months have passed between Grok-2 and Grok-3. ldquo; The timetable for reaching the most advanced fields is unprecedented. Anderj Karpathy said.

The release of Grok-3 is in line with Musk’s view of competition. He is always accustomed to using the fastest timetable to push teams to complete innovation. This is also the story he is good at telling. Using huge computing power (200,000 GPUs, Grok-3 uses hundreds of times more computing power than DeepSeek-V3), a small-scale team (xAI was originally established with only 12 people), cannot be a person at a table.

1. How did Grok 3 perform as the “strongest AI on the surface”?

In the live broadcast, the xAI team described Grok-2 as a “toy”, which of course was to highlight the power of Grok-3.

Grok-3 released by xAI is a model series that includes inference models and mini models.

In terms of model capabilities, Grok-3 has made new breakthroughs in many aspects such as reasoning, mathematics, code, and mathematics. Grok-3 ‘s reasoning models Grok-3 Reasoning and Grok-3 mini Reasoning scored 96 points on AIME and 85 points on GPQA, outperforming o3 mini, DeepSeek-R1, etc.

Source of Grok-3 ‘s performance in mathematics, science and code: xAI

In AIME 2025 ‘s latest mathematical benchmark, Grok-3 Reasoning surpassed the best version of the o3-mini, the o3-mini high.

Grok-3 ‘s performance in AIME 2025 ‘s latest mathematics benchmark source: xAI

On the large model arena LMSYS, the early version of Grok-3 (chocolate) ranked first in the overall list and was the first model to score more than 1400 points. Especially in the “coding” category, Grok-3 surpassed top inference models such as o1 and Gemini-thinking.

Grok-3 ranks first in LMSYS. Source: lmarena.ai

xAI demonstrated Grok-3 ‘s reasoning and creative programming capabilities live, such as allowing Grok-3 to generate a 3D animation code from Earth launch, landing on Mars, and returning to Earth, as well as upgraded versions of Tetris Mini games involving reasoning capabilities.

Generating code maps using Grok source: xAI

These model core capabilities have also been upgraded by Musk into the new Grok application, integrating three models: DeepSearch, Think and Big Brain in the form of Agents, providing high-level capabilities such as programming and mathematics for user search scenarios. DeepSearch can network and scan X to analyze information, provide queries and summaries, while Big Brain can do more and more careful distributed inference programming.

Grok has three model sources: DeepSearch, Think and Big Brain: xAI

In addition, Musk also revealed that new functions such as subsequent voice interactions and multimodal interactions will be launched. xAI will also establish an AI game community. Musk has revealed at this moment that he will open an AI game studio.

Grok-3 will not be open to all users immediately. Premium+ subscribers for X will be unlocked first. The membership service “SuperGrok” will also be launched on Grok’s independent APP for US$30 per month or US$300 per year.

Member service “SuperGrok” Source: xAI

Anderj Karpathy, former OpenAI co-founder who obtained early rights to use Grok-3 earlier today, said that Grok-3 is one of the most advanced thinking models, with performance comparable to o1-pro, and we need practical and realistic evaluation to observe it. rdquo; He gave an example. He uploaded a GPT-2 paper through the Grok-3 Think model and asked him a bunch of simple search questions and asked him to estimate the number of training flops needed to train GPT-2. This tested the model’s ability to find, mathematics and knowledge. According to his test results, GPT-4o failed to complete this task, and o1pro also failed, but Grok-3 with Thinking solved this problem well.

Source of Anderj Karpathy’s review: @Anderj Karpathy

In the past, Grok also left the impression of a more humorous and interesting AI. xAI also emphasized this point in the live broadcast. However, according to Anderj Karpathy’s test, the model’s sense of humor did not seem to improve significantly, and it was also too sensitive to “complex ethical issues.”

Objectively speaking, as a latecomer, xAI launched Grok-3 in less than a year, once again confirming Musk’s advantage of “vigorously creating miracles”. However, evaluating the actual capabilities and implementation of the model still depends on subsequent product functions.

2. Musk’s AI chips

In the competition for global models, Musk hopes that xAI will follow a typical “late mover, first-come” route.

Before the release of Grok-3, xAI conducted three major iterations of the flagship models of the Grok series in the past two years. In terms of model capabilities, the Grok series has demonstrated good performance in reasoning, reading understanding, mathematics, science, coding, etc. In terms of model lightweight and multimodal direction, xAI has also released the first multimodal models Grok-1.5V and Grok-2mini, constantly enriching its model family.

In terms of product interaction forms and business models, Musk has also been optimizing Grok. In terms of product interaction forms, on the one hand, he updates the interface, functions and product components. For example, combining X’s real-time insights with web search to launch a new citation function to improve the accuracy of answers. In January this year, unlike being embedded inside X, xAI also announced the launch of a separate iOS application and launched new content components around sports, finance and other scenarios to enhance the user content experience.

In terms of business models, through free and open APIs, the threshold for model use is lowered. At the end of last year, xAI announced that the Grok-2 model is free and open to X platform users (of course, there are usage restrictions), and simultaneously launched a public beta version of the enterprise API. A query by the “Silicon-Based Research Laboratory” found that xAI currently provides two model calls: Grok-2-1212 and Grok-2-vision-1212. Taking Grok-2-1212 as an example, its API pricing is US$2.00/million inputs and US$10/million outputs. At the same time, xAI also launched a data sharing plan to provide participating teams with free API points of US$150 per month.

xAI API pricing map source: xAI

During the live broadcast, xAI said that the Grok-3 model will appear in xAI’s enterprise API together with the DeepSearch function. Regarding open source issues that people are concerned about, when Grok-3 is mature and stable, Grok-2 may be open source within a few months.

In today’s era of big model competition and rising valuations, those who firmly believe that Grok and Musk can break through believe in Grok’s own unique advantages, namely data, cards, money and “anti-OpenAI stories.”

First, there is data. Grok is highly bound to X and is a closed-loop content ecosystem, high-quality data and stable scenarios, which is inherently advantageous in itself. Musk has repeatedly emphasized that by synthesizing data, Grok circumvents legal challenges related to data privacy and intellectual property that plague other AI models, while ensuring user data privacy.

Secondly, there is card, which is computing power. Tesla and xAI have stockpiled a large number of Nvidia H100 series chips. Musk previously questioned the organization’s ranking of “Meta as the world’s most hoarded H100 GPUs”, pointing out that “if the calculation is correct, Tesla should be the second place, xAI will be the third place.” He also spent 122 days transforming a home appliance factory into a super computing power cluster that collects 100,000 H100 chips. Even Nvidia founder Huang Renxun couldn’t help but sigh: To complete it in such a short period of time is simply a superhuman achievement.

During the live demonstration, Musk also took the lead in demonstrating his strongest data cluster before the release of Grok-3. The team said that they encountered many problems in February this year, such as cooling and energy consumption, and also wasted a lot of computing power. But in the end, Grok-3 was launched in six months with 10 times the computing power resources of Grok-2.

Musk’s data cluster source: xAI

xAI also does not seem to be short of money. According to Bloomberg, xAI is raising US$10 billion at a valuation of US$75 billion. Existing investors such as Sequoia Capital, Andreessen Horowitz and Valor Equity Partners all participated in the negotiations.

Moreover, there are “people”. Core members of the xAI team have worked in companies such as Google DeepMind, Tesla, OpenAI and Microsoft.

The two Chinese who appeared in the live broadcast are Jimmy Ba and Yuhuai Wu Photo source: xAI

Finally, Musk’s “anti-OpenAI” story has also gained many fans. Musk’s view of artificial intelligence has always been anti-OpenAI. He frequently emphasizes the differences in xAI with political neutrality and security.

3. Can Grok’s ambitions succeed?

However, behind Musk’s ambitious plan, Grok is also facing “internal and external troubles.”

First of all, as far as Grok itself is concerned, on the one hand, on the B-side and enterprise API calls, Grok currently does not have complete capabilities to serve enterprises. However, we have seen that in serving enterprises, coding and other capabilities are used to attract enterprise-level customers, OpenAI and Anthropic are making faster progress. The information previously reported that Anthropic’s annualized revenue from customers who use its models for software development and code generation increased 10 times. On the other hand, on the C-side, the current integration between Grok and X is not enough, especially for voice, video and other functions that have not yet been launched.

The confusion caused by some industry insiders is that now that OpenAI has integrated voice, video and other functions, and China companies including Doubao are also doing a good job. What new ideas can Grok and X jointly make in this regard? ldquo; If it was just a TTS (text-to-speech technology), it would not change. (Although during the live broadcast, xAI said it would not be TTS).

Secondly, externally, under the impact of China model companies such as DeepSeek, more intense competition has begun. OpenAI has released GPT-4.5 and GPT-5 roadmaps, and Anthropic has also announced that it will launch the Claude 4 series.

To a certain extent, the exploration and experimentation in AI reflect Musk’s consistent “view of competition” with a surprising timetable that pushes the team forward regardless, thereby approaching his expected goals. This has been verified in Tesla and SpaceX.

On social media, most people express excitement about Musk’s attempt, which also happened during the wave set off by DeepSeek. One artificial intelligence entrepreneur wrote: The new LLM competition is already popular, who will win in a week?& rdquo;