Why Automakers Chose Different Compute Platforms for Autonomy

Some automakers buy their autonomy compute off the shelf. Tesla built its own instead.

NVIDIA's DRIVE platform and Mobileye's EyeQ chips are the two most widely licensed options in the industry, used across a broad range of automakers and suppliers who'd rather integrate a proven compute and perception platform than design silicon from scratch. Licensing gets a manufacturer to market faster and spreads the enormous cost of chip development across every customer using the same platform. For a traditional automaker without a large in-house chip-design team, that's often the only realistic route to a competitive perception stack on a normal production timeline.

Building in-house is the other route, and it's a much larger upfront bet. Tesla is the most prominent example of a company that chose to develop its own autonomy compute silicon rather than license anything from an outside supplier. The appeal is control: a company that owns its own compute stack can iterate on perception and planning software without waiting on a third-party vendor's roadmap, or negotiating for access to change something at the hardware level. That control comes at a price beyond the chip design itself, though. An in-house effort also has to build and maintain the tooling and the software stack a licensed platform's vendor would otherwise supply as part of the deal.

The build-versus-buy choice is also not as binary in practice as "license a platform" or "build from scratch" suggests. A company can license the underlying compute silicon while still writing its own perception and planning software on top of it, which is a common middle path: the hardware comes from an outside supplier, but the part of the stack that actually decides what the vehicle does in a given situation stays proprietary. That split matters because the compute chip and the software running on it are separable decisions, even though they're often bundled together commercially by the same supplier for convenience.

What a licensing deal typically includes beyond the silicon itself is also easy to undervalue from outside the industry. A mature compute platform usually comes with years of accumulated tooling: compilers and drivers tuned to the hardware, pre-trained perception models a customer can start from rather than build from zero, and a support relationship with a supplier that's already solved problems a newer in-house effort hasn't encountered yet. Reproducing that tooling ecosystem, not just the chip's raw processing power, is a large part of what makes an in-house effort slow and expensive to stand up, and it's a cost that doesn't show up on a simple chip-versus-chip comparison.

The choice also isn't easily reversed once a company has committed. Software written against one compute platform's specific architecture, memory model, instruction set, driver interfaces, doesn't port over cleanly to a different platform, which means switching suppliers later means re-validating a substantial share of the perception and planning stack rather than simply swapping a part. That lock-in cuts against a licensing customer's negotiating position over time even as it protects the supplier's business, and it's part of why a platform decision made early in a program tends to stay in place for years afterward, regardless of whether a competing platform later becomes more attractive on paper.

Neither path is obviously correct, and the choice says something about how a company treats the compute layer strategically rather than just as a parts decision. Buying gets speed and shared cost. Building gets control and, eventually, differentiation, assuming the in-house effort actually catches up to what a specialist supplier has spent years refining. That build-versus-buy split is also a large part of why Mobileye and NVIDIA both sit in the D20 index as compute and perception suppliers to the rest of the industry, rather than as vehicle operators themselves.