Tesla’s AI5 FSD computer: 2,000+ TOPS, dual-sourced in the US, and a pivot away from Dojo

Tesla’s next-generation Full Self-Driving (FSD) computer, internally dubbed AI5 or HW5, is being framed as a step-change in in-car AI compute: 2,000–2,500 TOPS for low-precision inference, up to 40× performance over the current AI4 hardware in some workloads, roughly 9× more memory bandwidth, and a dual-sourced manufacturing strategy that spans TSMC in Arizona and Samsung in Texas. On paper, AI5 is powerful enough to make most previous “FSD computer” generations look like dev kits.

Underneath the performance claims, AI5 is also a signal of Tesla’s strategic shift. The company has shut down its Dojo supercomputer project and is leaning more heavily on external partners like Nvidia and AMD for training. AI5 is where Tesla is choosing to concentrate its custom-silicon effort: a high-volume, inference-focused chip that it controls from RTL to package and deploys in every car it sells.

What AI5 is supposed to deliver

Public information is a mix of Tesla commentary, leaks, and independent reporting, but several numbers are now reasonably consistent:

Raw compute: AI5 is expected to deliver around 2,000–2,500 TOPS (trillion operations per second) of low-precision AI throughput, roughly 5× the ~500 TOPS attributed to the current AI4 / HW4 platform.
Performance uplift: Elon Musk has talked about AI5 being up to 40× more performant than AI4 in some tasks, which implies not just more TOPS but better utilisation, memory bandwidth and architectural changes.
Memory and bandwidth: Leaks and analyses suggest around 9× more memory bandwidth and substantially larger on-package memory to feed heavier vision and planning networks.
Process technology: AI5 is planned to be manufactured on advanced nodes in the US, split between TSMC’s Arizona fabs and Samsung’s Taylor, Texas facility.
Deployment window: Volume production is generally indicated for 2026, with AI5 slated for future Tesla vehicles, robotaxis, and possibly datacenter-style inference boxes.

As with earlier Tesla FSD computers, AI5 is a heterogeneous SoC design: multiple AI accelerator blocks, CPU cores for general-purpose tasks, dedicated ISP and sensor-processing logic, and a lot of on-chip interconnect to tie everything together. The focus is squarely on inference, not training.

Why Tesla needs that much in-car compute

Tesla’s FSD stack leans heavily on vision: multiple camera feeds at relatively high resolution and frame rate, processed by deep neural networks for detection, segmentation, tracking and prediction. On top of that sits route and trajectory planning, plus all the conventional vehicle-control logic.

Several trends push compute demand upward:

End-to-end models: Tesla, like many in the industry, is moving towards larger “end-to-end” networks that take camera inputs and produce steering/throttle decisions more directly. These models tend to be bigger and require more FLOPs than collections of smaller, hand-engineered modules.
Higher-resolution perception: Better small-object detection, improved long-range visibility, and fewer corner cases all benefit from more pixels and more temporal context.
Redundancy and ensembles: Safety demands multiple networks cross-checking and adversarially testing each other, rather than a single brittle model.

AI5’s 2,000–2,500 TOPS budget gives Tesla headroom to push these directions without saturating the hardware, while still meeting tight latency budgets. The company cannot afford a 300 ms perception-to-action delay at highway speeds; shaving tens of milliseconds off end-to-end latency can translate into meters of stopping distance.

Comparisons: AI5 versus AI4 and consumer GPUs

For context, Tesla’s AI4 / HW4 FSD computer is often quoted at roughly 500 TOPS of low-precision inference throughput. AI5’s 2,000–2,500 TOPS target is a straightforward 4–5× jump at the headline level, with Musk and others pointing to additional efficiency gains that could yield up to 40× speed-ups in particular workloads.

Compared to high-end consumer GPUs, the numbers are not ridiculous. Some current enthusiast cards in the RTX 50-series reportedly achieve around 1,800–3,400 TOPS at INT8/FP8-class precisions, but at 360–575 W board power in a desktop environment. AI5 has to live in a car, where sustained power and heat budgets are much tighter and where functional-safety constraints are higher.

That difference in environment shapes the design trade-offs. AI5 is optimised for deterministic low-latency inference across a fixed set of workloads and sensors, not for arbitrary gaming or general-purpose compute. It can sacrifice some of the flexibility of a GPU in exchange for better efficiency and easier validation.

Dual-sourcing: Samsung Texas and TSMC Arizona

One of the more unusual aspects of AI5 is Tesla’s plan to manufacture it at both Samsung and TSMC in the United States. Musk has publicly praised Samsung’s Taylor, Texas facility as having “slightly more advanced equipment” than TSMC’s Arizona fab, while stressing that both will participate in AI5 production.

The rationale is straightforward:

Supply-chain resilience: Relying on a single foundry for a core product is risky. Dual-sourcing reduces the chance that yield issues, geopolitical shocks or local disruptions knock out AI5 supply.
Negotiating power: Having two foundries competing for volume gives Tesla leverage on pricing and capacity.
Political optics: Building AI5 in US fabs meshes with US industrial policy and “onshoring” narratives, which can matter when regulators and policymakers scrutinise autonomy and AI deployments.

There are engineering costs to this approach. Designing one chip that can be built on two different advanced nodes, with different design rules, libraries and process quirks, is non-trivial. Tesla will need to ensure that AI5 silicon from Samsung and from TSMC is functionally equivalent across temperature, voltage and ageing, and that any subtle differences are well-understood.

Dojo is dead, long live AI5

In mid-2025, Tesla effectively pulled the plug on its in-house Dojo training supercomputer, disbanding the team, reassigning engineers and allowing key leaders to leave. Reports and company commentary indicate that Tesla will lean more heavily on external partners—Nvidia and AMD on the compute side, Samsung and TSMC on manufacturing—for training, while keeping its custom-silicon focus on inference chips like AI5 and the planned AI6.

Several factors likely fed into that decision:

Scale and capital requirements: Competing with Nvidia in training silicon demands enormous, sustained investment in architecture, software, packaging and process integration. For a single OEM with a still-narrow AI product line (FSD, robotaxis, robots), the ROI is uncertain.
Rapid GPU evolution: Nvidia’s cadence from A100 to H100 to B100/GB200 has been aggressive. Dojo would need to match or beat that just to maintain parity, not leadership.
Opportunity cost: Every dollar and engineering hour spent on a Dojo D2 or D3 is a dollar not spent on in-car hardware, software and fleet-scale deployment issues that only Tesla faces.

By stepping away from Dojo and concentrating on AI5 and successor chips, Tesla is narrowing its bets: own the inference boxes where it has a unique problem and can deploy at volume, outsource most of the training horsepower to specialists who can amortise their R&D across the whole industry.

How AI5 fits into Tesla’s broader AI stack

In Tesla’s ideal workflow, the stack looks something like this:

Data from millions of cars is uploaded, processed, and used to train large vision and planning models on third-party accelerators (primarily Nvidia, potentially AMD in the future).
New models are compiled and quantized for inference on AI5, with tooling to ensure that behavioral regressions are detected and that performance targets are met.
Over-the-air updates push these models to vehicles equipped with AI5 (and to older hardware where possible, with tailored versions).
AI5 runs the models in real time in the vehicle, with enough overhead to handle worst-case scenarios, sensor noise, and redundancy checks.

AI5 therefore, sits at a critical pivot point: it has to be powerful enough to run what the training side produces, deterministic enough to satisfy safety and certification requirements, and efficient enough not to blow the car’s power and thermal budgets. It also has to be stable over the long term; Tesla can’t afford to ship a chip that needs frequent silicon respins once hundreds of thousands of cars are on the road.

Safety, redundancy and functional safety

More TOPS does not automatically mean safer autonomy. Regulators and safety engineers will look at how Tesla uses AI5’s additional compute to improve robustness, not just capability. Key questions include:

Redundant pathways: Does AI5 host redundant networks or logic paths that can cross-check each other’s outputs and detect anomalies?
Graceful degradation: If parts of the chip fail or overheat, can the system degrade to a “safe” mode (e.g., reduced autonomy or manual-only) without sudden failure?
Determinism and verification: Are the AI accelerators designed to be sufficiently deterministic for safety analysis, or do they have non-deterministic behaviors that complicate validation?

Tesla has historically taken a more software-first approach to safety than some traditional automakers, who lean heavily on ISO 26262-style functional-safety processes and redundant hardware paths. AI5 will need to be evaluated in that light: does its architecture make it easier or harder to build a safety case that satisfies regulators in multiple markets?

Upgrade paths and hardware fragmentation

Another practical issue is how AI5 fits into Tesla’s fleet, which already spans multiple generations of FSD hardware (HW2, 2.5, 3, 4/AI4). Musk has previously claimed that earlier FSD computers would be “all the hardware you need” for autonomy, but the introduction of AI4, and now AI5, complicates that narrative.

Key questions for owners and regulators include:

Will AI5 be offered as an upgrade to existing HW3 or HW4 owners who paid for FSD, and at what cost?
How will Tesla handle software compatibility across multiple hardware generations without creating safety discrepancies between older and newer cars?
Could regulators insist that certain autonomy levels only be permitted on AI5-equipped vehicles?

From a technical standpoint, Tesla can ship trimmed-down models for older hardware and full-fat models for AI5, but that increases testing and validation complexity. From a commercial standpoint, AI5 could become a lever to encourage owners to buy new cars rather than expecting free or cheap hardware upgrades.

AI5 beyond cars: robots and datacenters

Tesla has hinted that AI5 will not be limited to vehicles. Musk has referenced using future AI chips, including AI5 and AI6, in robots (such as the Optimus humanoid) and even in data centers for certain inference workloads.

That opens up several possibilities:

Common platform: Using the same AI chip in cars and robots simplifies software development and allows Tesla to leverage volume for better manufacturing economics.
Edge inference nodes: AI5-based boards could be packaged as inference appliances for warehouse robots or other industrial systems, not just Teslas.
Supplementing data-center GPUs: For certain neural workloads, AI5-based boards might offer attractive perf-per-watt, though they are unlikely to supplant Nvidia or AMD in general-purpose training.

Whether Tesla seriously productizes AI5 outside its own ecosystem remains to be seen. Historically, it has talked about licensing FSD and other tech to other automakers but delivered little in terms of external productization.

How AI5 fits into the wider automotive silicon landscape

Tesla is not the only automaker with custom AI silicon ambitions. Nvidia’s Drive platforms provide off-the-shelf, high-performance options for OEMs that don’t want to design their own chips. Mobileye, Qualcomm, and others have automotive SoC lines that integrate AI acceleration, IS,P and connectivity.

AI5’s differentiator is vertical integration: Tesla controls the chip, the software stack, the vehicle hardware, and the data pipeline. That allows it to make trade-offs a traditional Tier 1 supplier cannot, but it also means it carries all the risk. If AI5 underperforms, runs into yield or reliability issues, or fails to unlock meaningful progress in autonomy, there is no obvious fallback except more reliance on external silicon.

For other OEMs, AI5 is more of a benchmark than a template. Few have the scale, risk tolerance or in-house software capability to justify their own automotive AI chips; Nvidia, Mobileye and Qualcomm remain their logical partners.

Editor’s take

AI5 is, in many ways, the chip Tesla was always going to end up building: a big, US-fabbed inference engine that gives it headroom for the next generation of FSD and robotics workloads. The dual-sourcing from Samsung and TSMC is a savvy hedge in a world where supply chains and politics matter as much as clocks and cores. The shutdown of Dojo and pivot towards external training hardware is equally pragmatic; Tesla doesn’t need to win the data-center arms race to ship cars.

The open questions are about execution. Can Tesla get AI5 into cars on its stated timeline, with consistent behavior across two fabs? Can it turn 2,000+ TOPS into fewer crashes, fewer disengagements, and regulatory approvals rather than just more impressive demos? And can it manage the expectations of owners who bought into FSD on earlier hardware now that the company is openly touting a far more capable chip?

Until those questions are answered on real roads, AI5 will remain what it is today: an ambitious, promising piece of silicon at the center of Tesla’s new AI strategy—but not yet proof that strategy will work.