Elon Musk's xAI self-developed inference chip exposed! Codename X1, TSMC 3-nanometer process, mass production next year

Elon Musk's xAI is reported to be developing its own inference chip!

Codenamed X1, it will use Taiwan Semiconductor's 3-nanometer process and is expected to achieve mass production in the third quarter of 2026 (with an initial production of 300,000 units).

In fact, as a rapidly rising AI startup, xAI's "chip shortage" issue has not been a recent development.

Just this July, The Wall Street Journal revealed that xAI would raise $12 billion to purchase NVIDIA chips.

And a day after the report, Musk made bold claims:

xAI's goal is to achieve 50 million units of H100 scale computing power within five years.

What does 50 million units mean? It's worth noting that the so-called "world's strongest AI cluster," the Memphis Supercluster, only has 100,000 H100 units.

In this context, it’s understandable that Musk is pursuing self-developed chips to fulfill his ambitious goal.

After all, buying chips isn't that simple; it requires a significant amount of capital, and even NVIDIA and AMD may not guarantee sufficient supply.

Moreover, looking around, xAI is not the first company aiming to "break through." Just from the partnerships with Broadcom, companies like Google, Meta, and OpenAI are all actively pursuing self-development.

Let's take a closer look—

Over a month ago, xAI was reported to be hiring

In fact, the news about xAI developing its own inference chip is not new; there have been clues for some time.

About a month ago, DCD Intelligence (a company focused on the data center market) revealed through previously undisclosed job postings that xAI was hiring for its chip development.

The job responsibilities included "co-designing the next generation of AI systems, covering everything from silicon to software compilers to models," as well as "designing and improving new hardware architectures to push the boundaries of computational efficiency," among others.

It also mentioned some job requirements, such as candidates needing to be familiar with accelerator-friendly hardware design languages (like Chisel, VHDL, and Verilog), and preferably having experience in "simulating training workloads on new AI hardware architectures."

But, the mention of "the relevant department will be led by xAI veteran Xiao Sun" seems to be unfulfilled at this point.

Because this individual just joined AMD in September.

However, who leads is a matter for later; what we can see is that xAI's intention to develop its own inference chip did not come out of nowhere.

Moreover, xAI recently announced a new move— it will establish an office in Seattle.

It has also released recruitment information for its image/video generation and GPU core teams, with annual salaries reaching up to $440,000 (approximately 3.14 million RMB).

Putting aside the already familiar high salaries, let's talk about the subtlety of the location choice.

Companies that have already established offices in Seattle include OpenAI, Anthropic, and others, so xAI's move is also seen by foreign media as directly competing with these rivals.

Highlighting the point, competition.

Indeed, the development of its own inference chip is undoubtedly aimed at allowing the newly established xAI, which has been around for just over two years, to further gain computational power in the competition.

Moreover, from a timing perspective, xAI doesn't have much time left.

Recently, the Financial Times reported that one of its competitors, OpenAI, is collaborating with Broadcom to develop an AI inference chip codenamed "XPU," which is expected to go into mass production in 2026, with Taiwan Semiconductor Manufacturing Company (TSMC) as the foundry. This chip will not be sold externally but will specifically meet OpenAI's internal training and inference needs.

With Broadcom involved, focusing on inference, and mass production set for 2026, how familiar this pattern is~ The strategies of these two companies can be described as a 1:1 replication.

And it's not just them; Google and Meta have also long joined the ranks of self-developed chips. In addition to Google already achieving results, Meta was just reported last month to be collaborating with Broadcom to create a custom ASIC chip.

Now it remains to be seen which of the other companies will move faster (doge).

Tesla is also advancing its own inference chip

In addition to xAI, Tesla, under Musk, is also developing its own inference chip.

Elon Musk's recent tweet has sparked heated discussions:

Had a great design review with the Tesla AI5 chip design team! This is going to be an epic chip.

The upcoming AI6 is expected to be the best AI chip to date.

He also added that for models with parameter counts below 250 billion, AI5 might be the best inference chip.

So far, the lowest silicon cost and the highest performance-to-power ratio.

In August, Tesla halted its chip design project Dojo, and the team was disbanded on the spot.

The former project leader, Peter Bannon, has started a new venture called DensityAI to develop chips, hardware, and software for driving cars and AI data centers in the automotive industry. According to insiders, about 20 former Dojo team members have already joined DensityAI.

Ahem, back to the point. According to Musk, the closure of the Dojo project was due to the unreasonable allocation of resources to two different AI chip designs—previously, Tesla had adopted a dual-track chip development strategy of "training + inference," with Dojo used for training and the FSD chip used for in-car inference.

Now, it's time to focus more.

Switching from two chip architectures to one means that all our silicon talent will focus on creating an incredible chip. Looking back now, it seems obvious.

It is reported that the AI5 and AI6 chips are expected to support Tesla's AI and autonomous driving training. The AI5 is a transitional or scenario-specific main chip, produced by Taiwan Semiconductor; the AI6 will be the "unified heart" of Tesla's future AI ecosystem, manufactured by Samsung Electronics.

It can be seen that whether it's Tesla or xAI, Musk's efforts in developing self-researched inference chips are now relatively clear.

The next point remains the same: speed, speed, and more speed.

I wonder how Musk, who once set the record for "building the world's largest supercomputing center in 19 days," will perform in this battle?

Author: Quantum Bit, Source: Quantum Bit, Original Title: "Musk's xAI Self-Developed Inference Chip Exposed! Code Name X1, 3nm Process by TSMC, Mass Production Next Year"

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at their own risk