Welcome back to Level.UP, brought to you by UP.Labs.

This week, we’re covering this summer’s World Cup, which has become the largest live test of physical AI yet. Second, Japanese industrial companies are pouring capital into Western venture funds to get a front-row seat to the physical AI capabilities they can’t build themselves fast enough.

Plus: the tools we’re tracking, and hot takes on Mistral’s first robotics model and the humanoid-hands race.

Join us: On Monday, July 27, UP.Labs’ portfolio company Kairos is hosting a private Executive Roundtable on Manufacturing Flow and Operational Resilience. The session will feature insights from Gerald Johnson, an automotive industry veteran with over 40 years at General Motors, and Alston German, the former CIO of Federal-Mogul. They’ll discuss why scheduling often breaks down in manufacturing and explore where AI can provide real solutions. This is a senior, peer-level conversation designed to facilitate meaningful interactions, so it won’t be a typical webinar format.→ Request an invite

This year’s World Cup has been the most AI-saturated yet. The centerpiece is Lenovo’s Football AI Pro, a knowledge assistant that orchestrates multiple AI agents across more than 2,000 football-specific metrics and petabytes of tracking, performance, and historical data. 

Teams can query it in 15 languages, and it runs on players’ and coaches’ phones; a source close to FIFA says squads are leaning on it heavily after matches to break down performance. 

The tech extends well past the bench:

The company says the player avatars grew out of its robotics digital-twin work and the video analytics out of its own manufacturing quality-control lines. Together, the deployments mark Lenovo’s biggest bet yet.

Lenovo started with problems: controversial offside calls, coaching staff working from uneven data, and broadcasts running 40 seconds behind the play. Each one was concrete, each carried a fixed constraint, and each got its own build. 

Notably, the models doing the work are available to anyone. Lenovo rents the same frontier systems as its competitors. What cannot be rented is everything wrapped around the model. 

Football AI Pro runs on more than 2,000 domain-specific metrics and petabytes of proprietary tracking data. Engineers spent 18 months embedded in the sport to assemble it. The intelligence lives in that data architecture and that domain plumbing — slow to build, unglamorous, and effectively impossible to copy. 

For any operator sitting on decades of sensor logs, process telemetry, and field data, this is the lesson worth keeping. Data and orchestration continue to be the moat. 

Which brings us back to the pitch, and to a more useful question for operators watching from the sidelines: what does a build like this unlock for the rest of us?

Lenovo didn't win a hardware contract here. It found a new vertical hiding inside its own operations.

The player avatars came out of robotics digital-twin work. The video analytics came off its own manufacturing quality-control lines. Football AI Pro is, underneath, the same orchestration layer that runs factory floors, repackaged for the world's most-watched sporting event. The capability already existed; AI was the connective tissue that let Lenovo stitch it to a market it had never served.

Most large companies are sitting on the same raw material Lenovo had: decades of process telemetry, sensor logs, and field data that have quietly accrued value no one has bothered to price. Too often, that value is still sitting in server warehouses, doing nothing. 

AI changes this.

Stitch that history together with existing processes and a frontier model, and parts of the value chain that were previously closed — new revenue lines, new efficiencies, adjacent categories — suddenly open up.

Japanese industrial companies are moving capital into Western venture funds. And its most recent deals have been trending towards the physical layer.

These industrials are coming in as limited partners: the passive money behind a fund, not the managers who run it. And according to Venture Capital Journal, these corporations are using their fund stakes to integrate emerging technology into their operations, build relationships with founders, and set up future acquisitions or distribution deals as a scouting-and-access play.

Follow the strategic money. Right now, it is rotating toward the physical layer.

When an industrial company funds a venture firm to get an early look at robotics and industrial automation for early access, it is telling you where it expects applied AI value to concentrate. Chemical makers and auto suppliers are not only seeking venture returns. They’re also buying proximity to capabilities they cannot build fast enough on their own. 

For operators, the consequence is a change in who arrives at the table. Your next strategic investor, acquirer, or distribution partner may be a foreign industrial company using a venture fund as its front door. 

The capital comes with an operating agenda attached, and it is hunting the exact physical-AI capabilities you either already have or are building now. For a Head of BD, that reframes both fundraising and partnership strategy.

The move is to treat these corporate LPs as potential partners, not just check-writers. When a chemicals major or an auto supplier turns up as a limited partner in a fund circling your space, it is mapping which capabilities it needs and cannot produce internally. 

That map is worth studying. It shows you who might want to partner with you, buy from you, or distribute what you are building. The money is arriving early, and it is arriving with intent — which, for anyone building in the physical economy, is a door opening.

Ready to work smarter? Here are the tools we're tracking this week:

Mistral Ships A Robot Brain That Runs On A Single Camera. Mistral released Robostral Navigate, its first robotics model — an 8B system that lets a robot navigate complex, obstacle-filled spaces using a single RGB camera and a plain-language prompt, no LIDAR or multi-camera rig required. It's hardware-agnostic and trained entirely in simulation. The signal: physical AI is stripping out sensor cost, not adding it. And the economics of deploying robots on a factory floor shift the moment perception runs on a $30 camera instead of a $3,000 LIDAR stack. → Read more 

1X Unveils Tendon-Driven Humanoid Hand. 1X NEO’s new hand has 25 force-controlled degrees of freedom and tactile fingertips that sense shear. They’re built entirely in-house — tendons, motors, polymers, skin — with hundreds already off the line toward a 10,000-hand 2026 target. CEO Bernt Børnich’s framing is that robotics spent 70 years working around the hand problem, and the humanoid bet reverses it: the machine lives or dies at the fingertips. The bet: the dexterity race is really a manufacturing race. Hands are the most failure-prone, most expensive part of a humanoid; whoever industrializes them first sets the cost floor for everyone else. → Read more

The Nvidia-Escape Race Is On, From Both Sides. Anthropic is in early talks with Samsung to manufacture a custom AI chip. Meanwhile, DeepSeek is graduating V4 from preview to official release in mid-July, and Huawei has confirmed its Ascend AI cluster can run the model natively — cutting China's reliance on restricted Nvidia hardware. Notably, DeepSeek tuned V4 to work with agent tools, including Anthropic's own Claude Code. The takeaway: Nvidia holds roughly 74% of the AI chip market, and everyone is now engineering around that concentration. The compute layer is increasingly becoming the most contested asset. → Read more