Xpeng spends $500M/year on AI training to beat Tesla FSD

Xpeng's head of autonomous driving Xianming Liu told Electrek the company spends roughly 300 million RMB (~$41 million) per month, close to $500 million annually, on AI model training alone, and claims VLA 2.0 has reached parity with Tesla FSD v13 with v14-level performance expected before the end of summer 2026. Liu disclosed Xpeng's training infrastructure has achieved a 4,360% gain in single-job training efficiency over 12 months with GPU utilization rising from 40% to 90%, and that VLA 2.0 consumes over four trillion tokens per model iteration.

Why it matters

$500 million a year on training alone, against a ~200,000-unit delivery base and $6.5 billion cash, is the number to anchor on. That is closer to a frontier-lab training budget than an automaker R&D line item, and it tells you Xpeng is willing to burn margin per car to keep VLA 2.0 on the scaling curve with Tesla. The Volkswagen licensing deal and the rumored VW plant acquisition in Europe are the monetization hedge: if Xpeng can amortize that training spend across VW's volume starting 2027, the unit economics rerate from "expensive car company" to "AD supplier with a captive hardware channel."

The architectural disclosure matters for anyone building VLA stacks. Liu is explicitly removing language tokens as an intermediate representation while keeping them as input, and folding the world model into the same foundation model rather than running it as a separate predictor. Two billion visual tokens per second compressed to ~20 control tokens, with no language detour in the middle, is a direct shot at the LLM-as-planner orthodoxy that dominates current humanoid and AV research. If VLA 2.0 holds up at v14 parity by August, the "language is poison" framing stops being a hot take and starts being a recruiting tool.

Background reading

Related signals

← Read other signals from this week

Friday issue, by email. Free. No spam.

Background reading

Related signals

Subscribe