Same computer for everyone. The software drives.

Five driverless cars line up at Imola on the evening of 5 September 2026. It is the first time the Abu Dhabi Autonomous Racing League has raced outside Abu Dhabi, and it runs at 18:30 as part of the ACI Racing Weekend.

The number worth carrying into that grandstand is not the lap time. It is the distance these cars have travelled since testing began, and where that distance came from.

The experiment underneath the race

A2RL is built on a constraint that is rare anywhere in artificial intelligence. Every team gets the same car and the same computer.

The chassis is a modified Dallara SF23, the Super Formula car, redesignated EAV24 for autonomous use. The cockpit comes out and a steer-by-wire system, a sensor array and a compute stack go in where the driver sat. Radar, lidar, cameras and satellite positioning replace a pair of eyes.

The organiser supplies that stack to every entrant, along with base software already capable of driving the car on its own. What each team owns is the code it writes on top.

That makes the racing a controlled comparison. When one car finishes ahead of another, the reason is not a bigger budget for silicon. It is somebody's perception, planning or control code, running against everyone else's on identical parts, on the same afternoon, on the same tarmac.

Very little in this field is measured that cleanly.

What is actually in the car

Two components do the thinking. An AMD EPYC 7313P processor with sixteen cores, and an NVIDIA RTX 6000 Ada graphics card.

Both are described in the technical literature from the 2025 season, including the paper the winning Munich team published on its own approach and the dataset paper A2RL released this summer.

The RTX 6000 Ada is worth pausing on. It carries 48 GB of memory with error correction, moves data at 960 GB per second, and draws a maximum of 300 watts under an active fan. It is a workstation card. One of them.

That is the whole of it, and all of it sits in the car. A car at racing speed, working out continuously where to put itself relative to four other cars doing the same, runs on a single graphics card and a sixteen-core processor.

What changed

When testing started, the cars lapped about three and a half minutes slower than a human reference.

In April 2024 the first race at Yas Marina was a difficult afternoon. Coverage from the day describes cars stopping on track and spinning, and a race that had to be restarted.

At an exhibition in Japan later that year, the former Formula 1 driver Daniil Kvyat was around ten seconds a lap quicker than the machine.

In November 2025 the two met again at Yas Marina. Kvyat set a best lap of 57,569 seconds. The Munich car, called HAILEY, set 59,154. The gap was 1,585 seconds.

Then the six-car race ran, and on the final lap the same Munich car recorded 58,183 seconds. Against Kvyat's best from earlier that day, that is a gap of about six tenths of a second.

One caveat belongs with those numbers. The ten-second figure and the 1,585-second figure were measured at different places: the first at the 2024 exhibition in Japan, the second on a shortened layout at Yas Marina. They describe a real and fast improvement, and they are not two readings from the same instrument. Anyone quoting them as a clean sequence is skipping that.

What is comparable is the last pair. Kvyat's 57,569 and the Munich car's 58,183 were set at the same event on the same day.

Why this is a compute story

Physical AI is the part of the field where a model has to perceive a moving world, decide, and act on the decision while the world keeps moving. It is what sits inside a delivery robot, a warehouse forklift and a car that brakes for something the driver has not seen yet.

The received view is that this is waiting on hardware. Not enough compute at the edge, not enough throughput inside a power budget, come back when the next generation ships.

A2RL says something different, and it says it under unusually clean conditions. The compute is set by the organiser and is the same for every team. Nobody in this series wins by buying a bigger processor, and the cars have gone from three and a half minutes behind a professional driver to six tenths of a second behind him.

That is the encouraging part, and it is worth being precise about why. It means the compute already available is enough to get close to human performance in a real-time control problem that leaves almost no room for error. Vibration from a combustion engine, a stiff chassis, tyre wear, changing grip, four other vehicles making their own decisions, and no opportunity for the software to think it over.

It also means the constraint that moves fastest is the one currently doing the work. Silicon arrives on a multi-year cadence. Software ships on a Tuesday.

The same pattern, one industry over

This is not the first time we have run into it. In the round of MLPerf Inference results published in April, software optimisation delivered up to 2,77 times more throughput on hardware that had not changed, and NVIDIA reported per-GPU throughput on one model rising roughly 1,3 times in about six months from software alone.

Data centre inference and a race car are different problems with the same shape. In both, a chip bought today gets faster without being replaced, and the size of that gain depends on how many people are working on the code.

For anyone deciding where to put money into physical AI, that is a useful thing to know about the shape of the curve.

What to watch at Imola

The five teams are TUM, Kinetiz, Unimore, Constructor and Polimove. Entry is free, through the ACI Racing Weekend.

Watch how the cars behave when the plan breaks. Autonomy on a clear track is a solved demonstration. Autonomy when a car ahead brakes early, when grip changes, when a sensor returns something ambiguous, is the part that decides whether any of this reaches a public road.

Watch the spread between the leaders. Identical hardware means a small spread is a statement about how much of the available performance the best teams are now extracting.

And watch what Imola itself does to the software. Both A2RL races so far have been run at Yas Marina, a modern circuit built flat on reclaimed land. Imola is older, narrower and hilly. A stack that transfers well to it is a stack that has learned something general.

That is the question worth asking of any physical AI system, on a track or off one. The question is how much of the tuned performance survives contact with somewhere the system has never been.

Sources

Event details, timing and the team list come from A2RL's own page for the Imola race. The compute and vehicle configuration come from two papers on the 2025 season: the Munich team's account of its head-to-head approach, arXiv 2602.08571, and the A2RL Vmax dataset paper, arXiv 2607.17813. Specifications for the RTX 6000 Ada come from NVIDIA. Lap times and the 2025 result come from ASPIRE's announcement of 17 November 2025 and from contemporaneous coverage of the event. The MLPerf figures come from the MLCommons results published on 1 April 2026 and from NVIDIA's submission summary for the same round.

Where two figures were measured in different places, this article says so.