Autonomy's real bottleneck: Turning autonomous vehicle data into value

For more than a decade, the race toward autonomous driving has been sold as a contest of algorithms. Whoever trains the smartest perception model wins.

That story is incomplete, and it's quietly draining budgets. The harder problem under every self-driving program is autonomous driving data: capturing it, moving it, making sense of it and turning it into something a vehicle can actually learn from. Get that wrong, and the cleverest algorithm in the world has nothing useful to train on.

With competition intensifying across every major market and development cycles getting shorter, solving the data problem slowly has become a competitive risk in its own right. 

A modern test fleet doesn't generate gigabytes of data. It generates petabytes every week from cameras, lidar, radar and dozens of other sensors. Industry estimates suggest full autonomy will demand roughly a thousand times the computing power of today's advanced vehicles.

The cost isn't only storage. The bigger drain is human. Autonomous vehicle data processing involves teams of engineers spending their days converting files into usable formats, sorting recordings and hunting for the handful of moments that actually teach the system something (a near miss, an unusual pedestrian, weather the model has never seen).

When that work is manual, schedules slip and your most expensive specialists spend their time on logistics rather than engineering. For a finance leader, that looks like millions in R&D producing very little learning per dollar spent. 

There's an uncomfortable pattern here. A promising pilot proves the concept on a small slice of data. The team then tries to scale it across a global fleet, and the whole thing seizes up. The pipeline that worked for one car in one city buckles under the strain of thousands of vehicles across dozens of markets.

This isn't unique to automotive. MIT research shows that around 95% of AI pilots never deliver the results leaders expected, usually because the technology is rarely correctly connected to the people and processes around it. Autonomy is the most data-hungry version of that same trap. The lesson for boards is blunt: a successful demo tells you almost nothing about whether you can scale. 

DXC has spent years working to close this gap through DXC Robotic Drive, one of the largest petabyte-scale development platforms built specifically for autonomous and assisted driving, supporting everything from Level 2+ assistance to full Level 5 autonomy.

The outcomes matter more than the architecture. Aumovio, one of the world's largest automotive suppliers, uses DXC Robotic Drive for its data ingest service, the critical step of quickly getting vast sensor recordings off the vehicle and into a usable state.

In a separate program for a global carmaker, DXC built a system that analyzes data in the exact format the vehicle recorded it, removing a slow conversion step, then uses AI to automatically flag interesting encounters worth a closer look. The result is faster R&D and engineers who spend their time improving the driving software rather than stewarding files. 

The data challenge doesn't end when a vehicle ships. A connected car is a rolling data center, and software-defined vehicles keep changing through over-the-air (OTA) updates long after they leave the showroom. That turns a development problem into an operations problem, and the stakes rise because these systems can't fail. Those same streams also open new revenue, from predictive maintenance to subscription features, but only for automakers that can operate them reliably.

This is where DXC's newer thinking comes in. DXC OASIS, an agentic operations platform launched in 2026, sits atop an automaker's existing tools as a single layer, monitoring the entire estate, predicting problems before they become incidents and letting AI agents handle routine work while human experts make the judgment calls. For a connected fleet, that means fewer blind spots and a more rapid response when something drifts. It's backed by serious AI capability too, including a multi-year alliance with Anthropic to embed agentic AI into mission-critical systems. 

Tools alone don't fix the scaling problem. What separates programs that stall from programs that ship is a repeatable way of working. DXC's Xponential blueprint is built for that, with governance and security designed in from the start and a deliberate path from a small proof of value to enterprise scale. Applied elsewhere, the same approach cut service-desk tickets by 20% for 32,000 employees at Textron, a useful reminder that disciplined scaling produces measurable results rather than perpetual pilots. 

Stop treating autonomy as purely an algorithm investment and recognize connected vehicle data for the strategic powerhouse it truly is. Before greenlighting another round of model development, ask a simpler question: can we actually capture, organize and learn from fleet data at scale, affordably and safely? If the honest answer is no, that's where your focus and investment belong.

A convincing demo is easy to produce, but it tells you little. Lasting advantage is built by transforming masses of vehicle data into reliable, ever-growing intelligence. And that journey accelerates when you join forces with those who’ve already mastered the toughest challenges at scale.