On November 10, 2025, Musk's Tesla has entered an intensive iteration stage in the development of AI chips, and its self-developed chip roadmap and supercomputing system layout are reshaping the technological landscape of autonomous driving and robotics. The following is a deep analysis based on the latest developments:
1、 AI5 Chip: Architecture Innovation and Countdown to Mass Production
1. Technological breakthroughs and performance parameters
Tesla's latest AI5 chip has completed design review, marking the official entry of the chip from the research and development stage to the production preparation stage. According to Musk's disclosure, AI5 has achieved significant improvements in multiple key indicators:
- Computing power and energy efficiency: The original computing power reaches 2000-2500 TOPS (trillions of operations per second), which is 5 times that of the current HW4 chip, and the inference speed in some scenarios is 40 times faster than AI4. The energy efficiency ratio (performance per watt) is three times higher than similar chips from Nvidia, and the performance per dollar is 10 times higher
- Architecture optimization: Adopting TSMC's 3nm N3P process and Samsung's 3nm process for joint manufacturing, the half mask design is achieved by removing redundant modules such as traditional GPUs and image signal processors (ISPs), significantly improving computational density and energy efficiency. The memory capacity of AI4 has skyrocketed from 16GB to 144GB, the bandwidth has increased by 5 times, and it can support giant neural networks with parameter scale expanded by 10 times
- 2. Mass production plan and supply chain strategy
- The AI5 chip is expected to release samples and conduct small-scale trial production in 2026, and achieve large-scale production in 2027. Tesla adopts a "dual supplier+overproduction" strategy:
-
- Division of OEM: TSMC is responsible for the early production of factories in Taiwan, China and Arizona, while Samsung is responsible for the subsequent production of factories in Texas. The chip versions produced by the two contract factories may differ slightly due to process differences, but performance consistency is ensured through software optimization
- Capacity layout: Tesla aims to produce "excess" AI5 chips, which will be deployed to data centers in addition to being used in cars and robots, to enhance computing power redundancy and reduce dependence on Nvidia GPUs
- 3. Application scenarios and strategic significance
- AI5 chip will become the core computing engine of Tesla's intelligent ecosystem:
- Autonomous driving: supports more complex end-to-end deep learning algorithms, capable of processing over 1 billion frames of camera data per second, significantly improving the FSD system's ability to recognize edge cases. It is expected that the upgraded Model Y model equipped with AI5 will be the first to be launched in 2026
- In the field of robotics: As the "brain" of Optimus humanoid robots, AI5 can support their understanding of complex environments, complete fine operations, and run GPT-4 level large models. Musk expects that AI5's low-cost advantage (with a single chip cost of approximately $5000-6000) will drive Optimus to achieve its $20000 commercial target
- Data Center: In collaboration with Dojo Supercomputing, AI5 can be used to train FSD models, predict protein folding, and other tasks. Its energy efficiency is 8 times higher than traditional GPU clusters, and power costs are reduced by 70%
- 2、 Dojo Supercomputing: The Computing Revolution from Autonomous Driving to General AI
- 1. Production and Performance Breakthrough of Dojo2 Chip
- Tesla's second-generation Dojo supercomputer chip has started mass production and is expected to be deployed by the end of 2025. Its core innovations include:
- Architecture Design: Adopting TSMC InFO SW packaging technology, a single training module consists of 25 D2 chips, providing 1.8 EFLOPS computing power. The performance is 10 times higher than the first generation Dojo, and the cost is only 1/5 of traditional GPU clusters
- Communication technology: Our self-developed TIA (Tesla Interconnect Architecture) achieves nanosecond level latency, enabling tens of thousands of chips to work together like a supercomputer, completely eliminating the memory walls and bandwidth bottlenecks of traditional data centers
- Energy efficiency advantage: When running GPT-4 level model training, Dojo's energy efficiency is 8 times higher than traditional GPU clusters, and the hourly training cost is reduced by 60%
- 2. Application extension and ecological openness
- The application of Dojo supercomputing has extended from single autonomous driving training to multiple fields:
-
- General AI Service: Tesla announces the opening of Dojo cloud service, allowing enterprises to rent its computing power for tasks such as large model training and video generation. OpenAI has migrated some training workloads to Dojo, reducing training time by 65%
- Robot development: Dojo is being used to train Optimus' physical interaction skills, enabling it to quickly learn complex tasks such as moving objects and operating tools
- Creative Industry: Supports real-time generation of 4K video content, multiple film studios have started using Dojo for special effects rendering, reducing costs by 80%
- 3、 Follow up roadmap: Continuous iteration from AI5 to AI7
- 1. Technical planning for AI6 and AI7
- AI6 chip: planned for mass production in mid-2028, with performance doubled compared to AI5, and will continue to be manufactured by TSMC and Samsung. Its design goal is to become the "best inference chip for 500 billion parameter models" and further optimize the energy efficiency ratio
- AI7 chip: Due to higher research and development complexity, new foundries (such as Intel) will be introduced, adopting a new process architecture, and is expected to be launched around 2030
- 2. The ultimate goal of supercomputing systems
- Tesla plans to launch the Dojo3 supercomputer in 2026, integrating more chips through TSMC's SoW-X packaging technology to increase computing power by 40 times compared to Dojo2, with a goal of reaching the ExaFLOPS level. This will provide unlimited computing power support for cutting-edge fields such as fully autonomous driving and General Artificial Intelligence (AGI)
- 4、 Industry impact and competitive landscape
- Tesla's self-developed chip strategy is reshaping the AI hardware market:
-
- Challenge to Nvidia: Tesla gradually reduces its dependence on Nvidia GPUs through AI5 and Dojo systems. Although it will still purchase its products for training, the advantages of self-developed chips in energy efficiency and cost have become a threat
- Leading the industry trend: Tesla's "end-to-end+dedicated chip" model has been imitated by many car companies, such as Volkswagen and Toyota, accelerating the development of their own autonomous driving chips, promoting the industry's transformation from general-purpose GPUs to customized AI chips
- Supply chain transformation: TSMC and Samsung engage in fierce competition for Tesla orders, leading to rapid release of 3nm process capacity and indirectly driving global semiconductor technology progress
- 5、 Summary and Prospect
- Musk's AI chip development has formed a closed-loop ecosystem of "car chips+supercomputer systems+robot applications". The mass production of AI5 chips and the implementation of Dojo2 supercomputing not only provide a solid hardware foundation for Tesla's FSD and Optimus projects, but also mark its comprehensive transformation from an automotive company to an AI technology giant. With the successive launches of AI6, AI7, and Dojo3, Tesla is expected to achieve its ultimate goal of "processing camera data from all vehicles worldwide every second" by 2030, completely changing the way humans interact with intelligent machines
Global
Hot News
沪公网安备31011702889749号
