Nscale Drops $3.5B on Figure; Humanoid Chips Get Complicated
A data-center builder just made the biggest robotics bet of the year, Xpeng raised $900M, and engineers admit humanoid compute is messier than autonomous cars.
Nscale, a data-center infrastructure firm, signed a $3.5 billion deal with Figure on Tuesday. Details are thin—PYMNTS broke it, nobody elaborated—but the number alone makes it the largest publicly disclosed humanoid transaction of 2026. Figure's been shipping robots to BMW and piloting warehouse units; Nscale builds the racks and cooling that keep AI models running. Whether this is a supply deal, a joint venture, or Figure selling compute-as-a-service through Nscale's footprint is unclear. What's clear: someone with data-center money thinks humanoids are infrastructure, not science projects.
Meanwhile, the engineers building the silicon are calling timeout. Semiconductor Engineering published a deep-dive Wednesday on humanoid compute architecture, and the consensus is grim: it's more complex than autonomous vehicles. AVs run inference on structured sensor streams in predictable environments. Humanoids juggle vision, balance, tactile feedback, path planning, and object manipulation in real time, often with limbs that can break things. You can't just bolt an Orin in the torso and call it done. One unnamed chip architect told the outlet that current solutions are "overprovisioned and underoptimized," which is engineer for "we're winging it."
Xpeng's IRON humanoid platform raised over $900 million in recent funding rounds, per Trend Hunter. The Chinese EV maker has been telegraphing a manufacturing-focused bot for months; the capital suggests they're past the demo phase. Xiaomi showed up at IFA Berlin Thursday with its own humanoid and an in-house AI chip, according to Seoul Economic Daily. AMD is opening IFA tomorrow; Samsung bailed on the event, but humanoid robots are literally walking the catwalk. If the trade-show circuit is any indicator, 2026 is the year humanoids tried to go from vaporware to SKU.
UBTECH, the Shenzhen-based publicly traded humanoid maker, reported first-half revenue of 1.27 billion yuan—roughly $175 million—double last year's H1 figure. Shipments are already nearing 2025's full-year total, per Biggo Finance. The stock is "languishing," per AD HOC NEWS, despite the company's vertical-integration play into chips and consumer bots. Turns out doubling revenue in a category no one understands yet doesn't automatically juice the multiple.
On the infrastructure side, PwC dropped a forecast Tuesday that global AI capex will hit $31.6 trillion through 2050. Barron's, MarketWatch, and half a dozen outlets ran versions of the same line: that's more than the U.S. and U.K. spent building their railways and the internet combined. Fierce Network noted the caveat—it all hinges on power supply and politics. If the grid can't scale or permitting stays glacial, the number is fiction. If it does, data-center REITs and anyone selling backup generators just became a 24-year trade.
Nvidia's tying up with Hugging Face in what WION called a "huge open-source AI bet." Jensen Huang's playbook: give away the software, sell the shovels. PC GPU shipments were up 10.4 percent quarter-over-quarter and 1.1 percent year-over-year, per Jon Peddie Research, which is less "boom" and more "not collapsing." A Curzio Research piece argued Wall Street is watching AI stocks when it should be watching AI debt; a 24/7 Wall St. roundup flagged three "overlooked" data-center plays. Translation: the capex story is starting to fragment into who's paying for it and who's financing the people paying for it.
China's robotics sector is moving "deeper into the supply chain," learning to "feel" via tactile sensors, per Global Times. That's the state-run line; the subtext is Beijing wants domestic humanoid IP before the export bans tighten further. Xpeng's automotive-manufacturing push and UBTECH's consumer-bot vertical integration both point the same direction: China's trying to own the stack before the U.S. decides which layers to embargo.