1,016 TOPS per car: Pony.ai's Plan B for life after Nvidia
1,016 TOPS per car: Pony.ai's Plan B for life after Nvidia Pony.ai's robotaxis run on four Nvidia Orin X chips each — 1,016 TOPS per car — and the company plans to more than double its fleet to 3,500 vehicles by the en
1,016 TOPS per car: Pony.ai's Plan B for life after Nvidia
Pony.ai's robotaxis run on four Nvidia Orin X chips each — 1,016 TOPS per car — and the company plans to more than double its fleet to 3,500 vehicles by the end of 2026. Its CFO just told TechCrunch the industry's Plan B for losing Nvidia is already years old, and it still isn't the cheap option.
What was actually said
The quote comes from TechCrunch Mobility, published November 15, 2026, from an interview at Web Summit in Lisbon with Leo Wang, CFO of Pony.ai. The context matters: Web Summit opened with co-founder Paddy Cosgrave saying the superstars of the future are coming from China, not California — and days later the news spread through the conference that Nvidia founder and CEO Jensen Huang had resigned after admitting to an affair with an alleged spy.
Wang was asked what happens if the fallout leads to even stricter export controls on Nvidia chips shipped to China. His answer, in full, is worth quoting because of how un-dramatic it is:
"If you look at the whole situation, this has been developing for already a couple of years. Everyone, I would say, is also practicing kind of like a Plan B for a few years, so at least if you think about that, I would say there are alternatives. Although cost-effectiveness-wise, probably it's not the best choice yet. But still, it's an alternative way that already exists and which is more mature than a few years ago."
Two things in that answer. First, the trigger is hypothetical — "if the fallout results in even stricter export controls." No new restriction has been announced. Second, Wang is describing a hedge that already exists, not a switch he is about to flip. The news here is a CFO saying the quiet part out loud, three days after the CEO of his industry's dominant supplier resigned.
The exposure, counted
Pony.ai's dependence on Nvidia is not abstract, and it sits in two places: training compute in the data center, and inference compute in every car.
On the car side, the Gen-7 robotaxi runs four Nvidia DRIVE Orin X chips, 1,016 TOPS of deep-learning inference per vehicle. The fleet stood at 1,446 vehicles in March 2026, and Pony.ai raised its 2026 target from 3,000 to 3,500 vehicles. That is 2,054 more cars — a 142% increase — and at the current configuration, roughly 8,200 more Orin X chips just to keep the fleet growing.
The next generation makes the dependence sharper, not weaker. Pony.ai's next autonomous driving controller migrates to Nvidia's DRIVE Hyperion platform with dual DRIVE Thor chips, delivering 4,000 FP4 TFLOPS — roughly twice the floating-point throughput of the Orin setup, at about $3,500 per Thor chip. There is no domestic drop-in for that named anywhere in Wang's remarks. The Plan B he describes is about training compute, not the car.
The economics explain why the fleet matters enough to plan around. Pony.ai said its Gen-7 vehicles reached unit-economics breakeven in Shenzhen, at about $47 of net revenue per vehicle per day on 23 orders a day. A 3,500-car fleet at that rate is roughly $164,500 a day in net revenue — if the cars keep running. Every car that can't be built or can't be supplied with chips is a revenue line that doesn't exist.
The domestic alternative, measured
So what is Plan B actually made of? Wang didn't name a chip. The honest way to size it is against the one domestic accelerator with public numbers: Huawei's Ascend 910C.
- Raw compute: the 910C delivers 800 FP16 TFLOPS. Nvidia's H200 delivers 1,979 FP8 TFLOPS. Different precisions, so the comparison is already apples-to-oranges — and even the generous reading puts Huawei's best at roughly 40% of Nvidia's previous-generation flagship on headline throughput.
- The export-control metric: on total processing performance (TPP, the US Commerce Department's metric), the 910C scores 12,032 against the H200's 15,840 — 76% of the way there. Against the downgraded H20 that Nvidia was allowed to sell in China, the picture flips: Huawei's older 910B scores 5,120 TPP against the H20's 2,368, and Cambricon's Siyuan 590 scores 4,493. On paper, the domestic chips already beat the chip Nvidia was allowed to ship.
- The volume gap: in 2024, Nvidia sold more than one million H20 chips in China. Huawei sold about 200,000 AI chips. That is a 5-to-1 ratio in Nvidia's favor, in the year Chinese hyperscalers were supposedly already diversifying.
- The system-level claim: Huawei says its CloudMatrix384 system — 384 Ascend 910C chips lashed together — delivers 3 to 4 times the per-card inference performance of an H20 in specified scenarios, and a 2025 paper measured DeepSeek-R1 inference at 6,688 tokens per second per NPU on prefill. That is Huawei's own claim about its own system, not an independent benchmark, and it is a claim about a rack, not a chip.
And then the tax Wang actually named: cost-effectiveness. Domestic chips exist and are "more mature than a few years ago," but they are "probably not the best choice yet" on cost. The reason everyone in the industry knows but rarely quantifies is software: there is no domestic equivalent of CUDA with the same maturity, and porting a training pipeline from CUDA to Huawei's CANN is engineering work measured in quarters, not a purchase order. The chips are the cheap part of Plan B. The migration is the expensive part.
The other half of Wang's answer is the part that costs nothing: use mid-tier compute more efficiently. "Demand is already far greater than supply," he said, calling it one of the pain points Pony.ai struggles with. That is true before any new export control. The Plan B is not just a hardware swap; it is rationing — running training on mid-tier hardware instead of high-end, because the high-end hardware was already scarce.
The same week, from the other end
This is the third Nvidia story I have written this week, and the three of them are the same supply chain viewed from different ends. On Thursday, the Blackwell fingerprint report showed about 12,000 banned Blackwell GPUs were already inside China, identified by their traces in the open weights of six Chinese frontier models. On Friday, the Wall Street Journal reported roughly $530 billion in Nvidia supply commitments and guarantees — Nvidia, in effect, funding its own demand.
Wang's interview is the third leg: the customer side, saying out loud that the industry has been rehearsing life without Nvidia for years. Put together, the picture is not "China is cut off" or "China has replaced Nvidia." It is both at once: banned hardware already in the country, domestic hardware maturing but still second-best on cost, and every major player keeping two supply chains warm because neither one is reliable.
That is why the most interesting number in Wang's interview is not a chip spec. It is "a few years." Plan B has been practiced for a few years, and it is still not the best choice. A backup plan that takes years to become the cheap option is not a backup plan — it is a second business, running in parallel, at a permanent discount to the first. Pony.ai's 3,500-car target for the end of 2026 will be built on Nvidia chips if they can get them. The domestic alternative is real, it is maturing, and Wang just told you it is not ready to carry the fleet alone.
Sources: TechCrunch, "TechCrunch Mobility: The Nvidia backup plan for China's self-driving industry" (November 15, 2026); Pony.ai blog, "Pony.ai and Toyota Launch Fully Driverless Commercial Services with Gen-7 bZ4X Robotaxi" (November 16, 2026 China time — around 1,000 bZ4X robotaxis planned); Pony.ai fleet figures (1,446 vehicles March 2026, 3,500 target for end of 2026, Gen-7 breakeven at ~$47/day and 23 orders/day, dual DRIVE Thor at 4,000 FP4 TFLOPS) via company disclosures as reported; Ascend 910C/H200/TPP figures via Huawei Connect 2025 disclosures, the Institute for Progress report, and July 2026 Bernstein analysis as reported; 2024 China chip volumes (Nvidia >1M H20, Huawei ~200,000 AI chips) via MERICS as reported; CloudMatrix384 inference figures via Huawei Cloud claims and the 2025 CloudMatrix384 paper. The Blackwell fingerprint (~12,000 GPUs) and $530B guarantees figures are from my own earlier posts this week. Wang did not name specific domestic chips, and no new export control has been announced — the trigger in his remarks is explicitly hypothetical.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.