Tesla AI5
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,075 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,075 words
Add missing citations, update stale details, or suggest a clearer explanation.
Tesla AI5 is a custom artificial intelligence inference chip developed by Tesla, positioned as the successor to the company's AI4 (Hardware 4) self-driving computer. Also referred to informally as Hardware 5 (HW5), the chip is designed primarily to run Full Self-Driving (FSD), Tesla's advanced driver-assistance and autonomy software, and is also intended for the company's Optimus humanoid robot and for large clusters used for AI inference and training. Tesla chief executive Elon Musk announced on April 15, 2026 that AI5 had reached tape-out, the point at which a finalized design is sent to a foundry for fabrication. As of mid-2026 the chip is a pre-production part: it has been taped out and demonstrated as an early silicon sample, but it is not yet in mass production, which Tesla has targeted for 2027.
AI5 is the latest in a line of bespoke automotive AI processors that Tesla designs in-house and has fabricated by outside foundries. Unlike a general-purpose accelerator, the chip is tailored to the specific neural-network workloads of Tesla's autonomy stack. Musk has summarized the design philosophy by saying the team knows "what the chip needs to do, and just as importantly, we know what the chip doesn't need to do," a reference to stripping out functions that a single-customer, single-workload part does not require.[1]
The chip is notable for several reasons: the large performance gains Tesla claims over the previous generation, a dual-foundry manufacturing arrangement split between TSMC and Samsung, and its role in a broader strategic shift in which Tesla wound down its in-house Dojo training supercomputer and consolidated around a single family of inference-and-training chips. Because the part has not yet shipped in volume, much of the public information consists of forward-looking claims by Musk and reporting on Tesla's manufacturing agreements rather than independent measurements.
Tesla has iterated through several generations of self-driving hardware. Early Autopilot systems used Nvidia hardware before the company moved to its own designs.
The third-generation computer, Hardware 3 (HW3), introduced Tesla's first fully in-house FSD chip in 2019. It was followed by Hardware 4 (HW4), also marketed as AI4, which shipped in vehicles beginning in 2023 and uses a redundant dual system-on-chip (SoC) board. According to Musk, AI4 is manufactured by Samsung, and Tesla also uses AI4 silicon for training work in its data centers alongside Nvidia accelerators.[1] Musk has stated that AI4 already achieves a safety level above the human baseline for supervised driving, which Tesla has used to argue that an AI5 retrofit is not an urgent requirement for existing vehicles.
In parallel with its automotive chips, Tesla pursued Dojo, a custom supercomputer program intended for training the neural networks behind FSD. In August 2025 Musk confirmed that Tesla was shutting Dojo down and disbanding its team, describing the program's Dojo 2 effort as "an evolutionary dead end."[2][3] He framed the decision as a consolidation: "It doesn't make sense for Tesla to divide its resources and scale two quite different AI chip designs," adding that the AI5, AI6 and later chips would be "excellent for inference and at least pretty good for training," and that a supercomputer could instead be built by placing many such chips on a board.[3] The Dojo wind-down redirected attention to AI5 and its planned successor, AI6.
Tesla announced the AI5 tape-out via a post by Musk on X on April 15, 2026, which read: "Congrats to the @Tesla_AI chip design team on taping out AI5! AI6, Dojo3 & other exciting chips in work."[4] Alongside the announcement, Musk shared an image of a first silicon sample and predicted the part would become one of the most-produced AI chips ever.[5] Tape-out marks the final design milestone before fabrication; it does not mean the chip is in production.
The headline performance figures are forward-looking claims made by Musk, principally on Tesla's third-quarter 2025 earnings call on October 22, 2025, and on subsequent calls and posts. Musk's most-quoted claim is that AI5 is "up to 40x faster" than AI4 in certain tasks or by certain metrics, though he did not specify which workloads produce that figure.[6][1] Other Tesla statements describe the gain in more conservative terms, including that AI5 is roughly an order of magnitude (about 10x) more powerful than AI4 overall, and that a single AI5 die delivers roughly five times the useful compute of the dual-SoC AI4 board it replaces.[4][7] The figures below are the values most commonly attributed to Musk and to Tesla; independent measurements are not yet available.
| Attribute | AI4 (Hardware 4) | AI5 (reported / claimed) | Notes |
|---|---|---|---|
| Configuration | Redundant dual SoC | Single die | Move away from dual-chip redundancy[7] |
| Headline performance | Baseline | "Up to 40x" faster in certain metrics | Musk's framing; metric unspecified[6][1] |
| Useful compute | Baseline (dual board) | ~5x a dual-SoC AI4 board | Per Tesla statements[7] |
| Overall power | Baseline | ~10x (order of magnitude) | Musk's more conservative framing[4] |
| On-chip / system memory | ~16 GB | ~9x AI4 (reported ~144 GB) | Reported figure; not officially confirmed[7] |
| Power efficiency | Baseline | ~3x | Tesla claim[8] |
| Memory devices | n/a | GDDR packages from SK hynix | Reported from sample image[5] |
The memory configuration reported from the demonstration sample uses standard graphics memory (GDDR) packages supplied by SK Hynix on an organic substrate, an approach consistent with a high-volume, cost-sensitive automotive part rather than the exotic packaging used in data-center accelerators.[5] These hardware-level details are inferences drawn from the published sample image and have not been formally specified by Tesla.
Tesla has also drawn comparisons to merchant silicon. Musk has claimed that AI5 offers roughly three times the performance per watt of Nvidia's Blackwell generation at a small fraction of the cost, and Tesla has framed the chip as running inference far more cheaply than buying comparable Nvidia hardware.[8] As with the AI4 comparisons, these are company claims rather than benchmarked results.
A distinctive feature of AI5 is that Tesla plans to have it built by two foundries rather than one. Reporting around the April 2026 tape-out consistently described the chip as dual-sourced between TSMC and Samsung, with production split between TSMC's fab in Arizona and Samsung's plant in Taylor, Texas.[9][10] In thanking the foundries, Musk credited "Taiwan Semi" and Samsung for their support in bringing the chip to production.[5] Dual sourcing is generally understood as a way to add capacity and reduce supply-chain risk for a part Tesla expects to manufacture in very large numbers.
The manufacturing arrangement evolved over time. When Tesla and Samsung announced a roughly $16.5 billion chip supply contract in July 2025, that agreement was reported as focused on the AI6 chip, with Samsung's new Texas fab dedicated to it, while AI5 was described at the time as a TSMC part.[11] Musk said the deal's "actual output is likely to be several times higher" than the headline value and that Samsung "does manufacture our AI4 computer and does a great job."[11] He has also remarked that Samsung's equipment was, in his view, "slightly more advanced" in some respects.[1] By the time of the AI5 tape-out in April 2026, reporting described AI5 production itself as split between the two foundries.
Tesla and SpaceX additionally announced a large semiconductor project, sometimes referred to as a "Terafab," in Texas during 2026, reflecting Musk's stated ambition to scale chip output substantially beyond near-term needs.[2]
AI5 is designed to serve more than one product line, which is part of why Tesla consolidated its chip roadmap:
Tesla has indicated that the earliest AI5 volume will be directed to Optimus and to its supercomputer clusters, with consumer-vehicle deployment following later.[7] Because AI4 already meets Tesla's stated safety threshold for supervised driving, the company has not announced a retrofit program to upgrade existing AI4 vehicles to AI5.
As of mid-2026, AI5 is a pre-production chip. The publicly reported milestones are as follows:
| Date | Milestone |
|---|---|
| June 2024 | Musk says AI5 will be in vehicles in the second half of 2025[4] |
| July 2025 | Musk says the AI5 design is "finished"; Samsung supply deal announced[4][11] |
| August 2025 | Tesla shuts down Dojo and pivots to AI5/AI6[2][3] |
| October 22, 2025 | On the Q3 2025 earnings call, Musk claims AI5 is "up to 40x" faster than AI4[6] |
| November 2025 | Tesla pushes AI5 volume production toward mid-2027[4] |
| April 15, 2026 | AI5 taped out; first silicon sample shown[4][5] |
| 2027 (targeted) | Planned start of high-volume mass production[9][12] |
Electrek characterized the April 2026 tape-out as arriving "nearly two years behind" the original 2024 timeline.[4] Reporting indicated that high-volume production at Samsung's Texas plant is expected in the second half of 2027, roughly a year after tape-out, and that Tesla's first robotaxi, the Cybercab, would launch on the existing AI4 hardware rather than AI5.[12] Until fabricated parts are validated and produced at scale, the chip's real-world performance remains unverified.
AI5 is significant primarily for what it represents about Tesla's strategy rather than for shipping results, since it has not yet reached volume production. The chip anchors a consolidated roadmap in which Tesla abandoned a dedicated training supercomputer (Dojo) in favor of a single inference-first chip family that can be scaled into clusters when needed. It also deepens Tesla's reliance on outside foundries while spreading risk across two of them, and it ties Tesla's autonomy and robotics ambitions to the same piece of silicon.
The performance claims, if borne out, would represent a large generational jump for an automotive AI part. They remain claims, however, and several of Tesla's own AI5 timelines slipped substantially before tape-out, which has tempered expectations about when the chip will reach customers.
AI5 sits within a wider industry trend of large technology companies designing custom AI silicon rather than relying solely on merchant accelerators. Tesla's effort is unusual in spanning vehicles, a humanoid robot, and data-center clusters with a single chip family. The next part in that family, AI6, is planned to be manufactured by Samsung and was described by Musk as the eventual successor to Dojo; a Dojo 3 concept and additional chips were also said to be in development at the time of the AI5 announcement.[3][4] AI6 is sometimes discussed as a more capable, all-in-one design intended to scale from FSD to Optimus to training, while AI5 is the nearer-term inference workhorse.