SpaceX's first earnings report as a public company disclosed AI segment revenue of $2.56B, up 247% year-over-year, against $15.8B in AI capital spending within an $18.4B total capex quarter — more than double revenue for the whole company. On the same call, Elon Musk said Nvidia is now the exclusive chip supplier for Starmind AI1, a satellite compute payload built on Nvidia's Vera Rubin architecture, extending the hyperscaler capex-now-monetize-later playbook into orbit as a disclosed, audited line item rather than a press release. The stock fell as much as 8% in after-hours trading despite the beat, on investor concern about the pace of that spending. (CNBC)
Private Companies
Anthropicother
Anthropic confirmed for the first time that it is building an in-house silicon team to design custom chips for Claude, per a company spokesperson and an active job listing. Claude's annualized revenue nearing $30B this year is what makes owning chip design economical rather than renting Nvidia, Google TPU, and Amazon capacity outright, the same logic that pushed OpenAI to ship its own inference chip with Broadcom last month. (Business Insider)
Public Markets
AMD's Q2 revenue rose 50% year-over-year to $11.5B, with Data Center revenue up 107% to $6.7B, but shares fell as much as 8% in after-hours trading despite the beat, as investors weighed Q3 guidance against Nvidia's continued grip on GPU-cluster demand. (CNBC)
Astera Labs posted record Q2 revenue of $392.4M, up 104% year-over-year, and guided Q3 to $540M–$560M on Scorpio X-Series fabric-switch demand — shares added another 5% after hours, a direct read on PCIe/CXL interconnect spend tracking the GPU buildout. (GlobeNewswire)
Emerging
Efficient inference: University of Michigan researchers fabricated a 22nm test chip for AdaMX, a numerics-and-silicon co-design that varies low-bit precision per weight/activation block instead of using one fixed 4-bit format. It recovers 82–83% of the accuracy MXFP4 loses versus full precision, beats Nvidia's newer NVFP4 format by roughly 30–40%, and adds only about 1% system energy overhead — evidence that inference-cost reduction is still available through numerics and architecture, not just the next process node. (arXiv)
Grid policy: Texas governor Greg Abbott ordered the state's grid operator and utility commission to audit all 1,800+ queued data-center interconnection requests — representing more than 474GW of demand, up from 233GW in January — before approving any new grid connections, halting ERCOT's review process pending disclosure of tax incentives, water use, and power plans. It's the first instance of a major state actually freezing new data-center buildout rather than just studying it, a preview of the fight every other interconnection-queue-strained grid operator will eventually face. (Texas Tribune)