Are AI software-only startups losing competitive edge?

When NVIDIA showed off Figure’s humanoid robot during Jensen Huang’s recent keynote, one viewer posted a short prompt on X, in reference to the excavator in Figure’s demo:

"@ahmedshubber25 Curious to know your first reaction when you see the Caterpillar move"

https://x.com/ahmedshubber25/status/1983533889210708034

Ahmed Shubber, founder of construction startup Lumina, replied: “We own our entire hardware stack. The rest rely on OEMs (Original Equipment Manufacturers) that will eventually lock them out.” His statement points to something that is increasingly evident: that control of autonomy is moving back to the companies that build the machines.

The story played out in self-driving cars. Cruise, Argo, and Aurora tried to retrofit autonomy onto existing vehicles, while Tesla built its cars, sensors, and software as one system. The companies that depended on others for hardware struggled to scale and are now cutting back or shutting down. Tesla, which owns the full stack, continues to ship and refine its systems.

That pattern is repeating in construction and mining. Startups like Built Robotics and SafeAI add AI layers to equipment made by others. But OEMs such as Caterpillar, Komatsu, and John Deere are developing their own autonomy systems. Once those mature, they have no reason to leave room for outside software.

Caterpillar’s next phase

Caterpillar is the clearest example of this shift. It just reported revenue of $17.6 billion, beating Wall Street estimates of $16.77 billion, with adjusted earnings per share of $4.95 versus expectations of $4.52.

Its Energy & Transportation unit, responsible for power systems, engines, and some mining equipment, rose 17% to about $7.2 billion and now contributes 40% of total revenue. Executives linked the growth directly to data-center construction and AI energy demand. The same business line also builds autonomous haul trucks and drilling systems used in mining.

Caterpillar has already logged over one million tons hauled autonomously at Luck Stone’s Bull Run Quarry and runs fleets of 32 Cat 794 trucks at the Quellaveco mine in Peru. These are production systems, not trials.

Caterpillar also announced it would acquire RPMGlobal, an Australian mining software company, for about A$1.12 billion (~$728 million). The deal strengthens its in-house digital and autonomy capabilities and reduces reliance on external providers.

The company is using rising profits from energy and data-center demand to invest in autonomy and electrification, closing the gap between hardware and AI.

Why hardware control matters

At the compute level, autonomy is becoming standardized. NVIDIA’s Jetson Thor and Isaac stack give companies off-the-shelf hardware and models to run vision, planning, and control. As NVIDIA CEO Jensen Huang said, “The next wave of AI is robotics.” When compute and base models are the same for everyone, the edge comes from owning the machine and the data that train it.

Companies that make their own machines can gather field data, retrain faster, and improve reliability. Retrofit startups depend on access to someone else’s hardware and telemetry. When that access closes, so does their business model.

In another tweet, Shubber quoted an investor who said: “It’s easier to tell LPs you backed an ex-Waymo executive than a founder from a garage.” It explains why many autonomy startups keep getting funded despite fragile positions. Investors follow safe resumes and familiar narratives, not necessarily strong economics.

That bias leads to crowded funding of software-only autonomy firms that depend on OEMs’ goodwill. When the OEMs take the autonomy stack in-house, those startups lose their supply base and their path to market.

Like Caterpillar, other major manufacturers are already moving in this direction too.

• John Deere integrated its Blue River Technology acquisition into “See & Spray” systems for precision agriculture. • Komatsu partners with NVIDIA on its “Smart Construction” platform. All three control both machine design and operational data. They can roll out updates, gather results, and feed them back into new models, something retrofit startups can’t do at scale.

The new wave of robotics companies (Figure, Bedrock, Agility, Covariant) also work this way. They design hardware and software together, creating feedback loops that improve performance over time. Owning the full loop, that is, machine, data, and model, is becoming the only stable structure in autonomy.

The autonomy market is consolidating around companies that own their machines. As NVIDIA standardizes AI infrastructure and OEMs integrate autonomy directly into their products, the middle layer of retrofit and middleware startups is shrinking.