Google put the model in the car: a Pixel on the CAN bus is the deployment shape nobody budgeted

Google Cloud mounted a Pixel 10 Pro inside a Formula E GEN4 car at the 2026 Goodwood Festival of Speed, connected it to the car's CAN bus, and ran an agent on the handset that spoke to driver Dan Ticktum through his earpiece ^[1][5][10]. The racing is the packaging; the substance is an inference topology that very few enterprise AI plans currently fund.

According to Forbes, which interviewed both companies at the event, the split worked like this: Google's Gemma model ran on the phone and analysed vehicle telemetry locally, while agent-to-agent communication could reach Gemini outside the car for context such as timing feeds or broadcast data ^[8][9]. John Abel, a managing director in Google Cloud's Office of the CTO, described it as "running a little micro agent on the actual phone," calibrated for the questions the driver cares about: where he is losing time, how his traction is, how the balance is from left to right ^[6][7]. Abel said the system could, "within a matter of a second, give him immediate feedback in his earpiece" ^[10]. Formula E CTO Dan Cherowbrier said the aim was to get the insights closer to the driver ^[11].

The physical conditions are what make the choice non-optional. The GEN4 produces up to 600kW, roughly 815hp, with permanent all-wheel drive and 0-100 kph in about 1.8 seconds ^[2], an average of roughly 1.6 g ^[2]. The car goes up the hill at more than 150 mph ^[4], which is about 67 metres per second ^[1]; a one-second advisory loop is spent while the car covers 67 metres of narrow tarmac. Forbes notes the hill is hot and dusty and that connectivity there cannot be assumed ^[12]. Round-tripping to a data centre, which is how most generative AI is delivered today ^[16], is not a design option in that envelope.

For operators, the interesting cost is organisational. A workload split between a small model on an endpoint and a large one in a region is two build pipelines, two update cadences, two security boundaries, and a routing decision about which side handles which task that somebody has to own and audit. Abel's own non-racing example was manufacturing: cameras and models embedded in production equipment that spot a displacement or fault and alert an operator immediately ^[13]. That is capital in endpoints and integration work at the machine, not incremental spend on tokens, and it lands on plant and field engineering teams rather than the central platform group.

Two caveats belong on the record. The reported latency is "a matter of a second" ^[10] with no published measurement, model size or accuracy figure, and nothing in the account claims the agent made the car faster; Ticktum's 42.46-second run is described as one of the quickest ever recorded on the hillclimb ^[14] without attribution to the phone. Google Cloud and Formula E have also been working together on race strategy and broadcast insights already ^[15], so the promotional incentive is not hidden.

What to watch: whether the on-device and cloud agent split shows up as a shipped product with a documented routing policy rather than a festival demonstration; whether it appears in competitive race weekends, where the rules on driver aids bite differently; and whether the industrial version Abel described arrives with an answer for who patches a fleet of inference endpoints bolted to moving machinery.