Goldman Quintuples Humanoid Forecast; Oracle Spends $28.5B on AI

Goldman Sachs just raised its 2035 humanoid robot forecast from roughly 1.3 million units to 6.5 million—a 5× jump that either reflects rapidly improving economics or a rapidly improving PowerPoint. The revision landed Monday amid a flurry of hardware announcements and a fresh wave of capex numbers that suggest the infrastructure spend is now large enough to have its own restructuring line item. Oracle disclosed it's spending $28.5 billion on AI capex this year while simultaneously booking a $2.8 billion restructuring charge, a combination that makes the term 'investment cycle' feel quaint.

XPeng announced its Turing AI chip—previously confined to electric vehicles—is now being deployed in its humanoid robots. The move mirrors Tesla's Silicon-to-steel playbook but with a tighter go-to-market: XPeng already has the IRON humanoid in production and is positioning the chip as a vertical integration play rather than a research moonshot. MediaTek also entered the agentic AI chip conversation with its Dimensity 9600 Pro, signaling that the race to power on-device reasoning is now a mobile-to-humanoid continuum.

The capex arms race is no longer theoretical. SpaceX reportedly spent $15.8 billion on AI infrastructure in a single quarter, with significant allocation toward orbital data center trials—a bet that only makes sense if terrestrial power and cooling constraints are now considered unsolvable at current latitudes. Amazon and the other cloud hyperscalers posted mixed earnings reactions despite massive AI spend, and data-center adjacencies took a beating: Eaton dropped 7.57 percent and Amphenol fell 5.67 percent on fears that the AI build-out is either peaking or about to hit a demand wall.

The term 'islanded data centers' appeared in infrastructure trade press this week, describing facilities being built faster than grid interconnects can be approved. It's a polite way of saying some of the largest compute projects in history are being designed to run off-site or off-grid because permitting is slower than NVIDIA's shipping schedule. Meanwhile, cloud gaming services are reportedly scaling back capacity as AI training workloads crowd out consumer GPU time, a shift that makes the resource crunch legible to anyone who's tried to queue a render lately.

The humanoid deployment narrative is also getting more specific. One outlet framed the 'first mass market' as night-shift labor in roles America can't staff—warehouse, fulfillment, light assembly—where the cost of a robot is now being compared not to a day-rate human but to an empty shift. That's a different ROI conversation than the one happening in demo videos. Separately, the Wall Street Journal reported on a defense contractor pursuing 'robot soldiers,' a reminder that the same actuation and autonomy stack can point in several directions depending on who's writing the check.

Motley Fool and others published lists of 'AI robotics stocks worth owning over Tesla,' a framing that suggests the market now believes the humanoid story has enough named players to support relative-value calls. Whether those calls are right depends on production cadence, margin structure, and who actually has volume contracts—all of which remain harder to verify than a Goldman model revision. The revision itself is notable not because 6.5 million is the right number, but because a major bank is now willing to anchor expectations that high in public.

Oracle's $28.5 billion capex figure and $2.8 billion restructuring charge in the same breath is the cleanest signal yet that AI infrastructure spend has crossed into 'balance sheet event' territory. The company is simultaneously building and cutting, a posture that suggests even the winners of the AI build-out are trying to figure out what the next twelve months of demand actually looks like. For now, the spending continues—but the stocks that make the shovels are no longer moving in lockstep with the narrative.