Tesla has announced that its next-generation AI5 chip will offer up to 40 times the performance of its AI4 predecessor, as revealed during the company’s third-quarter 2025 earnings call.
The AI5 chip was first showcased at Tesla’s 2024 shareholder meeting, and CEO Elon Musk described it as “an amazing design,” attributing the performance gains to the company’s integrated hardware and software knowledge. “By some metrics, the AI5 chip will be 40x better than the AI4 chip,” Musk stated. The architectural improvements over the AI4 include eight times more raw compute power, nine times more memory, and five times better memory bandwidth.
Tesla stated that the 40x performance increase was achieved by removing processing bottlenecks found in the AI4 chip. Specifically, SoftMax operations that required 40 emulation steps on the older chip will now run natively in just a few steps on the AI5. The new chip also incorporates native support for mixed-precision models and features sparse tensor operations optimized for AI workloads.
The decision to use both Samsung and TSMC for AI5 production marks a change from a previous strategy where TSMC was designated for the AI5 and Samsung for a future AI6 chip. Musk explained that this dual-foundry approach is meant to secure the supply chain. “Our explicit goal is to have an oversupply of AI5 chips,” he said. Excess chips not installed in vehicles or the Optimus robot will be used to power the company’s data center operations.
While developing its own silicon, Tesla will maintain its partnership with Nvidia. Musk clarified that the company does not intend to replace Nvidia as a hardware supplier for its data centers. Instead, Tesla will use its AI5 chips “in conjunction with” Nvidia’s systems. The company currently operates data centers with computing power equivalent to 81,000 Nvidia H100 chips.
According to Musk, the AI5 chip’s design eliminates legacy components such as traditional GPUs and image signal processors, effectively becoming “a GPU” itself. He predicted that this streamlined architecture will deliver “the best performance per watt, maybe by a factor of two or three, and the best performance per dollar for AI, maybe by a factor of 10.”
