A five-year-old Japanese startup just gave AMD something Nvidia has spent a decade making nearly impossible to take away: a real training workload, not a benchmark win.
Turing Inc., the Tokyo company racing to build fully autonomous vehicles, has added AMD Ventures as an investor and started running roughly 10% of its AI training on AMD chips, according to a report from Bloomberg. Until now, Turing had built its entire stack on Nvidia hardware, training and inference alike, since the company was founded in 2021.
Ten percent is not a lot. It's also more than most AI startups have managed to move off Nvidia in years of trying.
Turing was co-founded by Issei Yamamoto, the lead developer behind the shogi-playing AI Ponanza, and Shunsuke Aoki, an autonomous-driving researcher with a Ph.D. from Carnegie Mellon. The company has since built Heron, a vision-language model it says scales up to 70 billion parameters, and Terra, which it calls Japan's first generative world model for self-driving. None of that comes cheap. In November, Turing said it closed ¥15.27 billion in its Series A first round, about ¥9.77 billion in equity co-led by JIC Venture Growth Investments and Global Brain, plus ¥5.5 billion in syndicated loans arranged by Mizuho Bank. Denso, GMO Internet Group, Canon Marketing Japan and Yanmar Ventures all took stakes.
This week's news layers a new round on top of that one. AMD Ventures joined a $79 million extension to the Series A, which values Turing at roughly $600 million, according to people familiar with the matter cited by Bloomberg. AMD's chips showed up in the training pipeline at the same time, not just in the funding documents.
For AMD, that pairing is the harder win of the two. Nvidia doesn't just sell more AI chips than AMD does. It sells CUDA, the software libraries built around it, and a decade of developer habits that are expensive to unlearn. Turing's own executives told Bloomberg the switch came down to supply diversification and lower costs, not superior silicon. That's a modest claim, and a believable one. Nvidia's GPUs have been supply-constrained on and off for years, and a startup burning through compute to train driving models on real-world footage has every reason to hedge its bets rather than wait in line.
Turing's timeline explains why the hedge matters now. The company wants to put its full self-driving software into consumer vehicles and driverless robotaxis as early as 2028, two years ahead of the 2030 target Yamamoto described publicly back in 2024. Last year Turing ran a 30-minute autonomous test drive on the outskirts of Tokyo. It has since repeated the test in busier parts of the country, the kind of unglamorous, incremental mileage that autonomy programs actually run on, not the headline demo.
Most of Turing's training still runs on Nvidia H200 GPUs through GMO Internet's GMO GPU Cloud, which uses Nvidia's Spectrum-X networking in what the company says is its first deployment in Japan. AMD's 10% slice won't get Turing to market on its own. What it changes is how much leverage Nvidia holds over a customer that, until recently, had nowhere else to go.
Chip diversification has been an industry talking point for at least two years and a reality for almost nobody outside the largest cloud providers. Microsoft, Meta and Amazon have all bought AMD's MI300-series accelerators in volume, but those are trillion-dollar companies with the engineering headcount to run two GPU stacks side by side. Turing has a fraction of that headcount and one product to bet everything on. If a company that size can carve out a tenth of its training runs for AMD without an Nvidia supply crunch forcing its hand, that says AMD's ROCm software has closed enough of the gap to be usable by teams that can't afford to babysit two platforms.
Frankly, the figure that matters here isn't the $600 million valuation. It's the 10%. Nvidia has spent years fielding chip-for-chip comparisons while quietly counting on the fact that almost nobody outside a handful of hyperscalers actually switches providers. Turing moving a slice of its workload doesn't prove AMD has caught up on performance. It proves the cost of trying finally dropped low enough for a startup with everything riding on one product to bother at all.
Also read: UK startups raise $17 billion in H1 2026, their strongest half since 2022 • Kling AI's New Funding Priced It Below the Number Investors Floated in May • Nvidia's Next AI Rack Just Slipped a Year and Wall Street Noticed
