Tesla Renames AI Chips to SI and Promises an 80-Fold Performance Gain With SI6

Tesla has once again repackaged its custom hardware—and made plenty of noise about it. The accelerators previously known as AI chips are now called SI. The company links the name to “superintelligence,” while its updated roadmap promises an 80-fold increase in compute performance for the sixth-generation chips, SI6.

What exactly has been announced

The announcement contains few facts, and they are worth stating plainly. First, the accelerators have been given the new SI name. Second, SI6 is expected to deliver an 80-fold increase in compute performance. Third, the Dojo 3 system will be used not only in terrestrial infrastructure but also in space.

What the available description does not say is which chip provides the baseline for the “80-fold” figure, which numerical formats were used to measure performance, at what power consumption, or when SI6 will actually launch. Without those details, the number remains a marketing claim. “80 times faster” could refer to an improvement on a single die, in one rack, or across an entire cluster—and those are three very different stories.

How to interpret figures like these

Consider a simple example. Suppose you are training a model and it takes 80 days on your current cluster. If the 80-fold improvement refers to the peak arithmetic throughput of a single chip, that does not mean training will finish in one day. You may run into memory-bandwidth limits, interconnect bottlenecks between chips, software-stack maturity issues, and reliability constraints: hardware fails in large clusters, and data must be supplied quickly enough. Real-world gains are almost always substantially below the peak figure.

That is why every roadmap should face the same questions. What about memory? How does the network between chips scale? How many watts are consumed per unit of work? Are there a capable compiler and debugging tools? The “times faster” multiplier answers only the first of these questions.

Why the rebranding matters

The SI name is primarily about positioning. The word “superintelligence” taps into the rhetoric currently driving the market: companies are selling not merely compute, but a path toward systems more capable than humans. For an engineer, the name changes nothing about the architecture. For investors and recruitment, it does: it makes the case for custom silicon easier to explain.

The space angle is intriguing in its own right. Dojo 3 in space is an ambitious idea, immediately raising practical questions about cooling without convection, radiation, power supply, communication latency, and maintenance. The available description does not say exactly what Tesla intends to compute in orbit—training, inference, or on-site data processing.

What practitioners should do

There is no immediate action for developers and product teams that do not buy Tesla chips: SI6 remains a plan, not a product that can be rented or ordered. Still, there are three useful takeaways.

First, do not build someone else’s roadmap into your budgets. Plan around available hardware and keep your code portable so that changing accelerators does not require rewriting the entire stack.

Second, if you are selecting a compute provider, demand comparable metrics: tokens per second per watt and the cost of training a specific model, rather than multipliers with no comparison baseline.

Third, watch whether Tesla opens access to its compute capacity and software. The available description does not make this clear, so it is too early to discuss any impact on the AI infrastructure market.

Limitations

This entire article is based on a brief summary of the announcement. The figures come from the company itself, there are no independent measurements, and Tesla has a history of ambitious timelines that later slipped. The sensible position, therefore, is to note the signal that the company continues to invest in its own accelerators, then wait for specifications, benchmarks, and firm dates. Only then will it become clear where the marketing ends and the engineering begins.