A share hardware and storage names sold off as if AI's build out had peaked. A Jefferies cloud survey, rising DRAM prices, and sourced insider pushback point the other way.
When Meta moved to monetize part of its AI compute capacity by selling it to outside buyers, China's A-share tech complex fell hard the same day. The 创业板指 dropped 5.71% and the 科创综指 lost 5.64%, with semiconductor, AI-hardware, and memory-chip names leading the slide. The market's read went bearish: Meta selling compute, in that frame, meant AI's infrastructure build-out had finally outrun demand.
Industry insiders pushed back in 证券时报 coverage, and they pushed back with a specific reframing. Selling AI compute to third parties is not a sign the capex cycle has peaked. It is a sign the AI infrastructure business model is starting to mature: capacity that once sat idle can now be productized and turned into revenue. The frame is sourced to those insiders, not the publication's editorial line.
Two demand-side data points back the insider reading. A Jefferies enterprise cloud survey found 95% of respondents expect 2026 cloud budgets to grow year on year, with Microsoft's Azure widening its lead to a 27-point margin over the next preferred cloud supplier. That is not the demand profile of an AI infrastructure market that has run out of buyers.
Hardware pricing tells the same story. Beijing Junzheng, a Chinese memory designer, said DRAM customers in China raised prices again in the second quarter, with overseas customers beginning to adjust. The company expects third-quarter prices to keep rising, with fourth-quarter supply still tight and further upside possible. NOR Flash also rose in the first and second quarters, though by less than DRAM. Tight supply and rising contract prices are the data shape of a cycle still short of capacity, not one that has overshot.
The original 36Kr evening digest that surfaced the Meta story also carried a second discrete event worth flagging: China's securities regulator cleared Unitree Technology, a humanoid-robotics firm, for registration on the STAR Market, the Shanghai exchange's tech-focused board for growth-stage companies. The two stories did not move markets together, but both sit inside the same AI-and-frontier-hardware complex that the selloff compressed.
What to watch next is not whether Meta's compute sale converts commercially. That is a business-execution question. The harder question is whether upcoming capex prints and DRAM contract pricing keep walking away from the stock reaction. If they do, the insider reading stops being a counter-narrative and becomes the basic interpretation.