Musk, CEO of electric vehicle manufacturer Tesla, provided investors with the latest details on its artificial intelligence (AI) chip strategy, emphasizing that Tesla’s AI5 AI chips are produced at TSMC’s Arizona plant. Although Tesla is vigorously promoting customized AI chips, Musk has made it clear that it does not intend to replace NVIDIA (Nvidia).
CNBC reported that the AI5 chip is Tesla’s latest Autopilot product and specializes in processing self-driving function signals. Musk said at the third-quarter earnings conference that Tesla is pursuing overproduction of AI5 chips with dual suppliers Samsung and TSMC. After the number of AI5 chips exceeds the demand for vehicles and robots, the excess chips can be deployed to Tesla data centers at any time.
At the 2024 shareholder meeting, Musk announced that AI5 chips will be manufactured by TSMC and AI6 chips will be manufactured by Samsung. But he confirmed that the AI5 is co-produced at the U.S. facilities of both foundries. Samsung announced in July that it had secured a $16.5 billion contract, and Musk later confirmed that the customer was Tesla.
Musk revealed that the AI5 chip is suitable for half reticle, and the chip size is only half the full reticle AI chip of companies such as NVIDIA and AMD. Although Tesla has abandoned NVIDIA Drive chips since 2019 and uses its own processors. But in terms of data center and model training, NVIDIA's position is still unshakable.
Tesla will continue to use NVIDIA GPUs to train AI models. Tesla chips mixed with NVIDIA GPUs. Tesla also recently announced that the computing power is approximately equivalent to 81,000 NVIDIA H100 chips. Musk’s AI startup xAI is also already a major customer of NVIDIA. xAI is building a large-scale supercomputing center in Memphis, Tennessee, using high-end NVIDIA Grace Blackwell chips. Musk also expressed high praise for NVIDIA, saying that it has done an excellent job in meeting the almost impossible complex requirements.
Tesla chose to design the chip based on efficiency and cost considerations. Hyperscale cloud service providers like Google, Amazon and Microsoft are investing in developing AI chips, as some AI experts believe customized chips may be more cost-effective or faster for specific tasks.
Currently, Tesla chip design has a unique advantage because it only needs to meet one customer requirement. Musk noted that this makes design work extremely simple and allows the removal of a large number of complex components from the chip. After Tesla's design team removed the chip's traditional GPU, signal processor, and other common components, Musk predicted that the AI5 chip could achieve the best performance per dollar, possibly by more than ten times.