$Tesla (TSLA.US)$ Investment Notes Elon Musk today disclose...

Elon Musk today disclosed more details about Tesla's I5 chip, the company’s fifth-generation self-developed AI training chip.

It is 40 times faster than NVIDIA chips, with computing power increased by 8 times, memory enhanced by 9 times, and bandwidth improved by 5 times.

The code path has been shortened to approximately five steps, and the cost per inference is ten times cheaper compared to NVIDIA.

Energy efficiency per watt has tripled. The chip will be manufactured by Samsung using Taiwan Semiconductor’s lithography technology, with all production taking place in Texas and Arizona by 2026, Musk added.

Taiwan Semiconductor and Samsung will produce slightly different versions due to variations in how they translate the design into its physical form.

Samples and a small number of products will be available in 2026, but mass production can only commence in 2027. I6 will utilize the same factory, with an expected performance improvement of approximately twofold, aiming for rapid follow-up.

I5 aims to achieve mass production by mid-2028. I6 will require a different factory due to its higher risk. What do these figures reveal?

The first layer is the de-NVIDIA-ization process. Tesla has been using NVIDIA’s GPUs to train FSD and Optimus, which is costly.

The unit price of NVIDIA’s H100 ranges from $30,000 to $40,000, and a training cluster requires tens of thousands of GPUs.

Tesla’s self-developed chips reduce inference costs by tenfold compared to NVIDIA, implying that training costs could decrease by an order of magnitude.

This is not just about cost savings; it's about strategic independence. NVIDIA faces supply shortages, long delivery cycles, and significant price fluctuations.

After developing its own chips, Tesla is no longer constrained by the supply chain.

The second layer is vertical integration. Tesla designs its own chips, with Samsung and Taiwan Semiconductor manufacturing them at factories in Texas and Arizona. The entire supply chain is based in the United States.

Not relying on overseas resources aligns with Trump’s America First policy and mitigates geopolitical risks.

Musk mentioned that I7 requires a different factory due to higher risk, implying it could involve an entirely new process technology or architecture.

This may involve a three-nanometer or even more advanced process, or potentially a completely new chip design concept.

The third layer serves as the foundational support for strategy; behind the updates to the FSD version lies extensive data training. Tesla has over 6 million vehicles on the road, generating billions of miles of driving data daily for training, which demands immense computing power.

If the I5 chip truly delivers 40 times faster performance, Tesla can iterate FSD versions more rapidly, accelerating the release pace of versions 15, 16, and 17.

Technological progress will accelerate, and the same applies to Optimus humanoid robots, which need to learn to walk, grasp, and identify objects.

These tasks require extensive training, and the I5 chip can enhance Tesla’s learning speed by several times.

The fourth layer involves translating cost advantages into competitive advantages. A tenfold reduction in per-inference costs means Tesla can offer FSD subscriptions at a lower price or achieve higher gross margins at the same price point.

At $99 per month for the FSD subscription, a tenfold reduction in costs would significantly increase gross margins, transforming FSD from a conceptual feature into a genuine profit engine.