[scuttleslops] Semiconductors and AI Infrastructure: AAPL, CRWV, NBIS, HPE, DELL, Analog Chips
The Circuit: Apple’s Siri AI, Nebius Inflection, China Trip Takeaways; June 15, 2026
The most important Apple insight is that the speakers view Siri’s opportunity not as competing directly with Gemini or ChatGPT for broad web search, but as turning AI into a productized, system-level feature. Bajarin says Apple’s strength is “integration at the systems level,” where the user can ask practical questions like finding a report in messages or pulling flight information from email. Goldberg frames this as “making a product out of AI,” arguing that most consumers do not care whether something is labeled AI; they care whether the device can do something useful that it could not do before. Their shared point is that Apple’s AI moat, if it exists, will come from owning the device, the operating system hooks, the private personal index, and the user experience rather than having the world’s best general-purpose model.
A related strategic point is Apple’s attempt to clarify how much of the new Siri is “Apple” versus Google/Gemini. Bajarin says Apple went out of its way to emphasize that its on-device models, broad world knowledge, and contextual Siri work are not simply white-labeled Gemini, even if the cloud model appears to use some base Gemini technology that Apple fine-tuned or adapted. Goldberg’s interpretation is that this is mostly brand protection: Google is not perceived as privacy-friendly, while Apple’s AI pitch depends on trust. The speakers also highlight Apple’s use of Nvidia and Intel for confidential compute, which Bajarin views as notable because Apple historically preferred to control silicon and had long-running bad blood with Nvidia. The implication is that Apple may use Nvidia infrastructure for now, but Goldberg expects Apple eventually to build its own inference accelerator if the scale and economics justify it.
The Nebius discussion centers on whether NeoClouds are real long-term cloud businesses or merely temporary arbitrage vehicles around GPU scarcity and power access. Bajarin comes away more positive on Nebius, saying it looks “closest-ish to a hyperscaler” among the NeoClouds because it has more of a software stack than peers and a deeper technical bench. Goldberg adds that the core question for all NeoClouds is whether they are “real companies” or just “landlords or short-term arbitrage in electricity and GPU availability,” and he says Nebius and CoreWeave are among the few that look capable of building sustainable businesses. Nebius’ Yandex heritage matters here: Goldberg notes that Yandex was a legitimate tier-two hyperscaler with strong engineering talent, giving Nebius more credibility than a pure GPU leasing startup.
The growth opportunity for Nebius is not just training clusters but broader AI infrastructure, including software workloads, dedicated CPU racks for inference, and eventually physical AI/robotics infrastructure. Bajarin says he pressed Nebius on whether it was hearing demand for CPUs, and management indicated it is preparing to scale both Nvidia Vera and Intel CPU racks. He also highlights Nebius’ claim that it can design more efficient racks than standard OEM machines through better networking, wiring, power handling, and latency optimization. This is important competitively because, in his view, AI infrastructure companies need to prove they are technical operators, not merely financiers of GPUs and power contracts. The risk, however, is still that most enterprise customers will remain multi-cloud, with some workloads staying on hyperscalers and only certain AI workloads going to NeoClouds.
Goldberg’s China trip produces a blunt takeaway: U.S. export-control policy, especially the Entity List, is not working as intended. He says the Entity List once created real fear among Chinese companies, but now “nobody’s afraid of it in the same way” because companies such as YMTC have survived and, in his telling, are now “thriving.” The most striking example is a Chinese company on the Entity List that still sources roughly 90% of its critical components from major U.S. technology suppliers, which leads Goldberg to ask what the Entity List is actually accomplishing if U.S. companies can still sell into restricted entities. He also notes that enforcement capacity is weak, claiming there are only five BIS inspectors in China and that they are highly constrained. The risk for U.S. policy is that restrictions may accelerate domestic substitution in China without actually stopping the flow of key components.
InPractise Interview with Former Group CFO, Nebius Group; June 9, 2026
The speaker frames Nebius less as a hastily assembled AI-infrastructure play and more as the surviving AI core of the old Yandex: after the forced Russia disentanglement, the company kept Nebius, Toloka, AVRide, TripleTen, and a large remaining stake in ClickHouse, while selling the Russian and gray-zone assets. The strategic origin story matters because management wants investors to see Nebius not as a GPU reseller, but as an engineering-heavy company that inherited a culture of building search, cloud, fintech, self-driving, and AI systems in-house.
They contrast that background with CoreWeave’s crypto-mining heritage, arguing that Nebius has a more “full-stack” builder mentality and therefore more control over costs, supply chain, deadlines, and customer experience. One of the most important early strategic choices was the clean break from Russia: the company hired a Big Four firm to verify that it had “no Russian customers, no Russian suppliers, nothing,” and Nvidia apparently required that proof before investing.