Musk's wild plan for idle Teslas

What if your car started moonlighting as an AI server while parked in your driveway?

During Tesla's Q3 2025 earnings call, Elon Musk floated a wild idea: tapping into parked vehicles to build a giant distributed inference fleet. In Musk's words, if the cars are sitting around and get bored, their onboard computers could be put to work running AI workloads.

Musk went on to throw out some staggering math. If Tesla eventually scales to a fleet of 100 million vehicles, with each offering around one kilowatt of inference capacity, that adds up to 100 gigawatts of distributed compute. Best of all for Tesla, each car already comes with its own power conversion and liquid cooling baked in, which Musk pitched to investors as a massive untapped asset.

Overpowered chips looking for a side gig

Musk arrived at this idea while discussing the jump in Tesla's onboard computing power. He noted that while the current AI4 hardware (also known as HW4) is formidable, the upcoming AI5 chip is testing at up to 40 times the performance of AI4. That jump is intended to power unsupervised Full Self-Driving that Musk claims will be ten times safer than a human driver.

With that much muscle under the hood, Musk openly wondered if these chips might end up packing too much intelligence for just driving. Since personal cars spend the vast majority of their lives parked, turning idle silicon into a distributed compute engine seemed like a natural leap to him.

On the business side, Musk claimed Tesla could reach an annualized production rate of three million vehicles within 24 months. The biggest volume driver will be the Cybercab, which is slated to enter production in Q2 of next year. According to Musk, the true killer app for autonomous driving is simply letting passengers bury their faces in their phones without touching the wheel.

Bandwidth, memory, and angry owners

Musk's comments immediately drew comparisons to legacy distributed projects like SETI@home and Folding@home. But taking that concept commercial with customer-owned vehicles raised instant pushback in online hardware communities.

Technical critics were quick to point out major bottlenecks. Teardowns of current HW4 units reveal just 16GB of RAM and 256GB of storage. Running modern, memory-heavy AI models on that footprint is severely constrained.

Then there is connectivity. High-performance data centers rely on ultra-fast internal interconnects to act as a single cluster. Relying on cellular or satellite links from millions of parked cars introduces massive latency. As some commenters pointed out, high latency means this setup might only handle isolated, asynchronous tasks rather than true cluster computing.

Car owners had different concerns altogether. Putting extra computational load on a parked EV means drawing wall power, accelerating battery cycles, and exposing silicon to prolonged thermal stress. Without explicit user consent and clear financial compensation, many owners won't tolerate their personal cars being used as compute nodes. Some skeptics even joked that Tesla might just bury compute-sharing clauses deep inside future purchase agreements.

Why it matters

Musk's distributed inference vision hits directly at the biggest bottleneck facing artificial intelligence right now: data centers are running out of power, grid connections, and cooling infrastructure. Offloading inference workloads to customer driveways, where power converters and thermal systems are already installed and paid for, is a clever way to bypass data center CapEx entirely.

However, we think the economics for actual owners make little sense today. Cars are expensive personal assets, and owners are hyper-aware of battery degradation, electricity bills, and wear on electronics. Unless Tesla builds an air-tight incentive model that pays drivers significantly more than the cost of wall power and hardware wear, user adoption will face massive resistance.

From an engineering standpoint, this looks more like an aspirational moonshot than an actionable near-term roadmap. Severe memory constraints on current vehicle boards and erratic wireless upload speeds mean this fleet cannot replace dense hyperscaler clusters. It might work for minor, asynchronous background tasks, but turning 100 million cars into a cohesive 100-gigawatt AI supercomputer remains a distant fantasy.