16 September 2026
Heard In AI

How Waymo's own chip changes the robotaxi cost math

On Moonshots with Peter Diamandis, the panel walked through Waymo's sixth-generation driver: a purpose-built 5-nanometer chip, fewer but sharper cameras, and an autonomous-hardware estimate falling from $115,000 to $20,000. Peter had ridden in the new vehicle; Alex objected that the West is now white-labeling Chinese hardware, and argued Waymo is heading for full vertical integration.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

Peter Diamandis had ridden in the new car the day before. His verdict on its looks was mixed — "kind of not as sexy as the gold cybercab" — but the reason he wanted to talk about it was arithmetic.

For years, he said, the economics of Waymo against Tesla's Cybercab looked devastating. Elon Musk has said a Cybercab would sell at around $30,000 for the vehicle and its sensor hardware. Waymo's fifth-generation Jaguar, by the panel's accounting, came in at roughly $300,000: about $200,000 for the car and $100,000 for the full autonomous-driving hardware package. That is a tenfold disadvantage.

The number the panel put against it: autonomous-driving hardware falling from $115,000 to $20,000 on the sixth-generation system, a cut of more than 80%. Those are the panel's figures. Waymo's own published account of the new hardware does not quote prices.

What is actually in the trunk

The engineering claim behind the cost estimate is that Waymo now designs the computer that runs the car. In an August 20, 2026 engineering post, the company describes a purpose-built 5-nanometer ASIC — an application-specific integrated circuit, a chip built for one job rather than bought off the shelf. It takes in camera, lidar and radar streams and runs the sensor-fusion neural networks that merge them into a single picture of the road.

Diamandis reached for the headline figure: the chip processing all three data types in real time "at one quadrillion operations per second. We've gone past trillions. We're at quadrillions already. Quops." Waymo puts it as more than 1,000 TOPS of machine-learning performance from these chips, covering both front-end processing and the models themselves.

Waymo's own emphasis falls elsewhere. Its account stresses that driving decisions are computed onboard within milliseconds, and that the result comes from designing hardware and software together rather than from raw operation counts. The chips sit alongside conventional CPUs and GPUs, plug into the vehicle's liquid cooling, and are arranged as two independent processing engines so that one can take over if the other faults.

Fewer cameras, sharper ones

The sensing side of the sixth-generation driver was announced earlier in the year, in a February 12, 2026 post. The panel described the new configuration as 13 cameras, four lidar units and six radars, against 29 cameras and five lidars on the Jaguar generation — more than 40% fewer sensors, with Waymo saying performance improves rather than degrades.

The substitution that makes that possible is resolution. Waymo moved to 17-megapixel imagers, which let a smaller number of cameras keep a detailed view of the scene. Heaters, wipers and sprayers are built directly into the sensor pods, so the system can clear snow, ice and road grime instead of losing a sensor to the weather.

Diamandis recalled Musk's long-standing argument that if a human can drive with one eye, cameras alone should suffice. Waymo's stated reasoning for keeping the other two runs the other way: lidar supplies its own illumination and accurate distance measurements, including around opening car doors and vulnerable road users such as pedestrians and cyclists, while radar tracks motion across changing light and weather. The point of the combination is that perception survives when one type of sensor becomes less effective. Waymo attributes the lower cost partly to improving component economics across the industry and partly to moving processing into its own silicon.

The hardware figure is separate from the price of the car it goes into. The panel also described a purpose-built robotaxi minivan developed with a Chinese electric-vehicle maker at $75,000 per vehicle, against the roughly $200,000 Jaguar; the vehicle and manufacturer names are indistinct in the recording. Neither number is the cost of running a service.

"Careful what we wish for"

That sourcing is where the panel split. Alex began by "venting some pain here." Jaguar is owned by an Indian company, he noted, but built its cars largely in the UK. Now, he said, the headline about cheaper hardware conceals whose hardware: Google, Waymo and Alphabet are OEMing Chinese hardware to hit their targets. "Careful what we wish for," he said, with whoever suggests Google should simply fine-tune Chinese models. "I would rather see the West use a Western hardware stack rather than just white labeling Chinese hardware. That's somewhat disappointing."

His second reading of the same announcement was about control. Looking at the chip supply chain underneath, he said, it seems Waymo is moving away from Broadcom and vertically integrating: its own chips at the compute layer, OEMed vehicles at the hardware layer, and its own route to the passenger. It is "playing footsie with Uber for the moment for distribution," he said, but probably wants to own distribution in the long term — as he pointed out, when he takes a Waymo he opens Waymo's app, not an aggregator's.

He put a historical rhyme beside it. Tesla started with the high-end Roadster and pushed costs down until it hit autonomy, at which point the $25,000 Model 2 never launched. Below some threshold in car price, Alex argued, selling cheaper cars stops making sense and offering a hosted autonomy platform makes more. He expects Waymo eventually to drop third-party vendors and become "a Dell for Chinese hardware, or maybe American hardware," with its attention back on software.

Another voice on the panel took the pattern wider: the largest companies of earlier eras did different things — Exxon did oil, IBM did mainframes, GE did nuclear reactors and toasters — while today's biggest firms are each building AI chips, AI models and data centers, colliding into vertically integrated companies that own the whole stack. Robots, he added, are next, and TSMC sits at the center of it as "the lynch pin," waiting to be verticalized.

For Diamandis, the argument has already resolved into a habit. He uses Waymo regularly where he lives, he said, and it is much cheaper than Uber. Asked whether the vehicles will reach Europe, Emad Mostaque said Waymos are starting to appear in London, and that he thinks regulation there may turn out to be largely positive for them.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Tesla invites fleet buyers, with no price attached

Tesla opened an interest form for businesses that want to buy Cybercab fleets and build mobility hubs for its Robotaxi Network, without publishing prices, delivery dates or revenue-sharing terms. On Moonshots, Peter Diamandis said he had filled it out, and the panel turned the invitation into an argument about owning machines instead of working for wages.

6 min read

A superintelligence ban and a hands-off G20 land in the same week

On September 3, Senator Bernie Sanders and Representative Greg Casar announced legislation to permanently prohibit superintelligent AI and pause advanced development; a day earlier the White House reported unanimous G20 agreement on a non-binding, innovation-first framework. The Moonshots panel rejected the bill's single human-level threshold, then spent the rest of the segment arguing over what a credible middle position would be: universal chip logging, open weights, and a right to compute.

6 min read

Memory, not GPUs: the shortage that could redesign AI hardware

On Moonshots, Peter Diamandis reported back from meetings with SK hynix and Solidigm leadership with a claim that memory, not compute, now limits AI. The panel argued that a changed workload and a supplier industry scarred by past busts are pushing prices up faster than factories can respond — and that the fix may be new chip designs, including etching model weights into silicon, rather than simply paying more.

8 min read

OpenAI's Jalapeño chip has the Moonshots panel asking: could it sell compute to rivals?

On Moonshots EP #284, the panel read out the performance figures OpenAI published for Jalapeño, the inference chip it built with Broadcom, and argued that inference is moving off NVIDIA. One panelist went further, imagining an "OpenAI Compute" cloud that rents capacity to competitors — possibly even hosting an Anthropic model. Dave Blundin explained why CUDA no longer locks buyers in at inference time, and why NVIDIA's next defense is the networking between chips.

6 min read