Vantora raises $100M to build physical AI its clients own outright
Two stories about who owns the intelligence layer: a California venture studio that now builds startups its corporate backers can keep to themselves, and a Chinese agent stack being wired to domestic silicon.
Vantora, the startup builder formerly called UP.Labs, raised more than $100 million from Silversmith Capital Partners — its first outside capital after four years of building AI-native companies for customers including Porsche, Alaska Airlines and J.B. Hunt. Founder and CEO John Kuolt told TechCrunch that the money funds a change in the model as much as it funds more companies. Vantora still builds startups for corporate partners that invest in them and act as their first customers, but those partners can now fold the resulting company into their own business instead of watching it sold on the open market. The firm says it has launched 17 companies this way.
Kuolt calls the new arrangement a "proprietary M&A pipeline," and he is blunt about why it exists. "Imagine you're a Fortune 100 industrial company and you need to retrofit all of your hardware and machines for autonomy. You need to own that, it needs to be sovereign, and you can't rely on a third party to go do that for you," he said. "They're never going to let us go sell that to their competitors."
That constraint is what pushed Vantora toward physical AI — work that touches machines, maintenance schedules and logistics rather than chat windows. The firm says it previously killed ideas that mattered to a partner but were too sensitive to take to market, including one with the freight carrier J.B. Hunt, and that the proprietary model puts those back on the table. The trade is legible: a narrower market for each company, in exchange for the problems that the largest industrial buyers refuse to hand to a vendor who also serves their rivals. That is the same logic pulling banks and governments toward private deployments, and it is a harder sell for anyone selling intelligence as a shared service.
At Huawei Connect in Shanghai, 360 Group and Ascend AI said they have jointly built a solution that pairs 360's Agent Factory — its agent generation and orchestration layer — with Huawei's Ascend compute stack, aimed at long-horizon agent workloads. The two companies say Ascend was tuned for long-sequence inference and the frequent KV-cache reads and writes that agent loops generate, cutting first-token latency, and that 360's orchestration lets agents nest, form teams and share working memory. 360 says its internal beta ran more than 1,000 consecutive steps without interruption at better than 95 percent task success, and cut a multi-role collaboration task that used to take two hours down to 20 minutes, with DeepSeek, Qwen and GLM models running out of the box.
Those numbers come from a vendor's own beta, so discount them. The direction still matters: China's answer to long-running agents is vertical integration — models, orchestration and silicon from one stack — while the Western labs sell the harness and the compute as separate purchases. We covered an earlier sign of that split — iFlytek ships Spark X2.5, trained end-to-end on Chinese silicon.
What to watch: whether Vantora's partners actually exercise the option to keep what gets built, and whether "1,000 steps" survives contact with production traffic rather than a beta.
Would you buy an AI system you can't inspect — or would you rather own it outright? Tell us in the comments.
