Allbirds Didn't Need GPUs. They Had Something Better.

Allbirds just sold its shoe business for $39M, raised $50M, and rebranded as NewBird AI, a GPU rental company aiming to serve AI developers. The stock jumped nearly 600% in a single day.

Makes sense on a balance sheet. Doesn't make sense as an AI pivot.

Allbirds walked away from something rare: material science deployed at consumer scale, with proprietary inputs, real-world performance data, and millions of customers. Instead, it's entering one of the most crowded layers of the AI stack, where differentiation exists but rarely compounds the way data driven systems do.

A real AI pivot would have used what they actually had: a direct line to the customer. By incentivizing trade-ins for worn out pairs, every returned shoe becomes a data point. Computer vision maps material wear, foam compression, and fiber degradation against geography and use patterns, capturing the real world data that lab tests can't easily replicate. Nike knows how to market. Allbirds could have known how molecules behave in the wild. A technical moat instead of a cultural one.

Observe → learn → design → deploy.

The product itself becomes the learning system. Whoever closes that cycle builds a data advantage no one can buy.

That's the consumer version. The industrial version is bigger and harder.

Wind turbine coatings degrade after years offshore. Aircraft composites develop invisible microcracks. Pipeline corrosion shows up where no model predicted. These systems fail in expensive ways, and they're still designed using lab tests that can't replicate real conditions.

Meanwhile, the data already exists: maintenance logs, inspection reports, sensor readings, decades of failure events across deployed assets. But it’s fragmented, unstructured and rarely fed back into design.

That's starting to change.

The Two Loops of Physical AI

The Lab Loop is automated discovery: AI designs experiments, robots run them, and results continuously retrain the system.

A new generation of early-stage autonomous labs is turning this into reality. Periodic Labs is searching for new superconductors and magnets that could reshape computing. Radical AI is running high-throughput experimentation, testing dozens of new materials per day for extreme environments like hypersonic flight. CuspAI uses AI instead of a lab to design new materials based on properties like strength, conductivity and heat resistance.

The World Loop is learning from materials already deployed in the field. It’s a slower, more messy process, but real. A robot generating 10,000 materials is powerful. A turbine blade cracking after eight years offshore is ground truth.

That ground truth is finally being captured by a few early-stage companies. BrightAI is aggregating operational data across power grids, water systems, HVAC, and manufacturing, predicting failures, dispatching crews, and feeding outcomes back into the system so each cycle improves. Archetype AI is building a foundation model for the physical world trained directly on sensor data from deployed assets.

Both loops are critical. The Lab gives you velocity. The World gives you defensibility.

What Changed

The unlock is that foundation models can now learn directly from raw sensor data and generalize across physical scenarios they weren't explicitly taught. On the materials side, models can now train on multiple types of data at once: what a material is made of, how it behaves under stress, how it conducts energy, what experts have written about it. They no longer look at each property in isolation. The standard approach in materials science until recently could only handle one dimension at a time.

Sensors have been cheap for years. Telemetry has been collected for longer. What's new is the ability to learn from that mess across modalities and feed it back into design.

A decade ago you could pull data off a turbine but couldn't use it to design the next one. Now you can, and incumbents aren't set up for it. The teams that own the field data don't design materials. The teams that design materials don't see the field data. No one inside a 50-year-old industrial company is incentivized to build a feedback loop that compounds over years.

The Thesis

When real-world data shapes how products get designed, the system that captures and learns from that data becomes the control point.

That's the bet: software-first, data-defensible, sitting between deployed reality and the next generation of design.

In ten years, every materials company that matters will be a software company underneath. The ones that aren't will get acquired by the ones that are.

Allbirds had a shot at this. They had the customer line, the materials, and the brand to pull worn-out pairs back as data. Wrong margins, wrong capital base, probably never going to build it. But someone in industrials will.

If thats you, I would love to connect.

-Kiswana