Einride will test Nvidia’s Alpamayo reasoning AI on trucks, road trials in 2027

Swedish autonomous freight company Einride said on October 21 that it will evaluate Nvidia’s Alpamayo reasoning model inside its Einride Driver system, as Reuters reported. A live road evaluation is planned for the second half of 2027.

Shadow mode first, no control of the truck

According to Einride’s announcement, Alpamayo is Nvidia’s Vision Language Action (VLA) model. It reads the driving scene, proposes how the vehicle should move and explains its reasoning. Einride says it intends to be the first trucking company to deploy the model on Nvidia’s Hyperion platform.

The company plans to post-train and distill Alpamayo on its own operational data, adapting it to its vehicles and routes. The first stage is shadow mode: the model runs live, but it does not steer, brake or accelerate. Its outputs are compared with the decisions of Einride’s driving planner and with a human driver’s performance, while the existing safety systems keep full control. Einride says the model will sit inside a safety architecture that retains authority over the vehicle.

The stated goal is commercial: shorter development cycles and a lower cost of extending the Einride Driver to new vehicle types and routes.

An expansion of a deal made a month ago

The announcement builds on the relationship the two companies announced on September 21. In that deal, Einride is building the next generation of its driver on Hyperion, Nvidia’s Level 4 reference architecture, integrating Nvidia’s Halos safety system and using Nvidia Cosmos and Blackwell infrastructure to train and validate its models. As electrive reported at the time, the collaboration covers autonomous electric trucks. Coverage of that deal cited an Einride projection of 1,500 to 2,000 vehicles by 2028, with about 80% of its customer demand suitable for automation in the medium term. Those are company estimates, not independent figures.

Henrik Green, Einride’s CTO, said in the release: “Our advantage has always been running autonomous freight in the real world, for real customers.” Rishi Dhall, Nvidia’s vice president of automotive, said: “Physical AI is opening new possibilities for autonomous freight.”

Why a reasoning model changes the trucking stack

The technical significance lies in what kind of component Alpamayo would be. Conventional autonomous-driving stacks split the job into perception, prediction and planning modules, each tuned separately. A vision-language-action model takes camera input and outputs both a proposed maneuver and a plain-language rationale, which gives engineers something to inspect when it disagrees with the production planner. That is exactly what shadow mode is built to exploit: every disagreement with the planner or the human driver becomes a labeled case to study, with no risk on the road.

The word “distill” in Einride’s plan matters too. A large reasoning model is generally too heavy to run in a truck cab, so the company says it will post-train and distill the model on its own data. The release does not say what size the onboard model will be or what hardware it will need, and that is a gap worth watching.

Einride has some operational credibility to bring to the data side. Coverage of the September deal noted that days earlier the company and Lidl had deployed Germany’s first cab-less Level 4 autonomous truck. The new announcement still establishes little about performance. It gives no benchmark, no safety result and no criteria for what shadow-mode agreement would justify putting Alpamayo in control. The live evaluation is nearly a year away, so this is a research and validation plan, not a product launch.

What the announcement does not settle

Einride is listed on Nasdaq under ENRD. One market-data page showed the stock at $2.38, down 4.8% after the news, which suggests investors did not treat the plan as an immediate catalyst; that is an observation of the price, not an explanation of it. The release gives no financial terms, no commercial agreement for Alpamayo licensing and no date for driverless operation using the model.

For readers following Nvidia’s push into physical AI, the point is the pairing: Nvidia is supplying not only compute and simulation tools to a freight operator, but now a reasoning model meant to be tuned on the operator’s own routes. Whether that approach beats module-based planners on heavy trucks will only be answered by the 2027 road data, which Einride has not promised to publish.