Tesla CEO Elon Musk stated at the shareholders' meeting that, to meet the rapidly growing demand for AI chips, Tesla is considering building its own "TeraFab," a facility larger than TSMC's "Gigafactory." However, chip manufacturing involves extremely high technical barriers and massive investments. Nvidia CEO Jensen Huang cautioned that advanced manufacturing processes cannot be replicated simply by throwing money at the problem; the difficulty is considerable.

Musk indicated that to meet the company's massive semiconductor demand in the field of artificial intelligence (AI), it may directly invest in building its own chip manufacturing business, requiring the establishment of a chip manufacturing base called "TeraFab."

He compared it to TSMC's "Gigafab," which has a monthly capacity of over 100,000 wafers, emphasizing that the new factory will be "much larger."

Currently, TSMC refers to its facilities with a monthly capacity of 30,000 to 100,000 wafers as "Megafabs," while those exceeding 100,000 wafers are designated as "Gigafabs."

If Tesla completes TeraFab, its monthly capacity will far exceed 100,000 wafers, far surpassing current mainstream wafer foundries, and could even place it among the world's largest chip manufacturers.

For reference, TSMC's Fab 21 in Arizona, with a total investment of $165 billion, is expected to become a Gigafab-level campus in the future, but Musk stated that Tesla's plans will be even larger.

However, regarding Musk's vision, Nvidia CEO Jensen Huang responded at a recent TSMC event, pointing out that the complexity of chip manufacturing is often underestimated.

He stated bluntly, "Building advanced chip manufacturing capabilities is extremely difficult. Besides the facilities themselves, the engineering technology, scientific research, and process experience accumulated by TSMC are all highly challenging."

Tesla's chip demand is rising rapidly

As a company with both AI supercomputers and significant automotive computing needs, Tesla purchases a large number of Nvidia GPUs and, after the discontinuation of Project Dojo, is pushing forward with the development of its own AI5 processor for use in autonomous vehicles, robotics, and data centers.

To ensure a stable supply, Tesla currently employs a dual-source foundry model with TSMC and Samsung. Musk also stated that Intel (INTC-US) may be a potential partner, but no agreement has been signed yet.

Musk emphasized that as Tesla's AI applications continue to expand, external supply will be insufficient to meet demand, thus necessitating the consideration of becoming a vertically integrated manufacturer (IDM) like TSMC and Samsung.

Musk stated, "Even if we project the best-case scenario of supplier chip production, future chip supply will still be insufficient. Therefore, we may have to build a TeraFab; this is imperative."

Building your own chip factory presents numerous challenges

Mastering advanced chip manufacturing requires an investment of far more capital and technology than most people imagine. At the current industry level, a wafer fab capable of mass-producing cutting-edge processes with a monthly capacity of approximately 20,000 wafers requires an investment of tens of billions of dollars, not including subsequent process development and mass production tuning costs.

As an example, the Japanese chip startup Rapidus is attempting to challenge this hurdle.

The company plans to establish 2nm process mass production capabilities within the next few years and estimates a commercially viable fab to be completed by 2027, with total expenditures reaching approximately 5 trillion yen (about US$32 billion).

While this goal demonstrates strong ambition, in today's highly mature global semiconductor race, the success of a new player directly entering the most advanced nodes remains highly uncertain.

Analysis points out that the development process of advanced processes is itself a lengthy and highly interdisciplinary challenge. From process route development and material and transistor architecture design to extensive TCAD simulations verifying electrical, stress, and leakage behavior, any deviation in any step can force the entire process to be restarted. Therefore, the initial "startup" phase alone often takes several years.

While Rapidus has obtained a license for IBM's 2nm GAA transistor architecture and enjoys some technical collaborations with imec in Belgium and CEA-Leti in France, the transistor architecture is only the starting point of the entire R&D chain.

Next, the engineering team must design and fine-tune thousands of process steps, including front-end transistor forming (FEOL), mid-stage contact layer (MOL), and back-end metal interconnect (BEOL). Processes such as deposition, etching, lithography, and annealing all require atomic-level precision.

Each step involves numerous parameters, requiring extremely deep engineering experience and repeated testing to ensure mass production reliability and yield.

Even after the aforementioned process flow has been completed, it's still necessary to establish a PDK (Process Design Kit), SPICE models, and standard cell libraries to ensure chip design teams can actually use the process for circuit design.

At this point, the factory must simultaneously adjust the production line equipment settings to maintain stable output in a real mass production environment. This, too, is a process that cannot be accelerated simply by throwing money at it.

Ultimately, the real test lies in "yield." Whether a new entrant can achieve a profitable, high yield for advanced processes in a short period is crucial to its market position. This often requires experienced engineering teams to be stationed at the factory for extended periods, making repeated adjustments, and enduring countless failures.

As for whether Rapidus can deliver results in 2027, the industry generally adopts a wait-and-see attitude. The outcome remains to be seen.

Source: Content from Juheng.com

Reference link https://news.cnyes.com/news/id/6226812

View more at EASELINK