A weekend of commentary from Serenity — the stock-picker known for front-running the AI buildout by buying its obscure upstream suppliers — centered less on new positions than on defending the durability of the AI capex cycle itself.

Oracle's resale data as a depreciation counterargument. Serenity flagged a detail from Oracle's (ORCL) recent earnings that he said he had missed: the company reportedly resold renewing GPU capacity at roughly a 20% premium over prior contracts, much of it on equipment four years or older. Oracle is a legacy database giant that has pivoted hard into leasing AI datacenter capacity, much of it tied to OpenAI's Stargate project. Serenity treats it as a demand datapoint rather than an investment — he has it on his "avoid" list over its debt load — but here he read the resale premium as evidence against the "rapid GPU depreciation" thesis (associated with bear Michael Burry) that aging chips lose value fast. He framed it as a positive for the neocloud names: his favored long Nebius (NBIS), a Netherlands-based AI cloud he backs for its Nvidia funding, and even Iris Energy (IREN), the Australian miner-turned-cloud he is otherwise firmly bearish on for its heavy dilution. Cheaper-than-feared depreciation, he argued, is a broad win for Nvidia (NVDA) and the compute-leasing ecosystem.

A wry portrait of finance social media. In a lighter post, Serenity contrasted the modest holdings of tech workers he knows (Tesla, Alphabet, SpaceX) with the leveraged bets of his own online audience — outsized memory options, exotic freight futures riding AXT's (AXTI) enormous run, and long positions in Swedish continuous-wave laser makers picked because they echo his own photonics-bottleneck logic. AXT is his core conviction name, the near-duopoly supplier of indium phosphide substrates that feed the lasers powering optical interconnects in AI datacenters; CW lasers are the light source for co-packaged optics, the technology of moving the optical engine next to the switch or GPU die. He seemed half-proud, half-bemused at having helped seed such speculation. He also nodded again to SpaceX (SPCX), whose blockbuster debut he has framed as a defining mega-IPO, and referenced the long-Palantir (PLTR) trade popular in that crowd.

Dismissing the AI-slowdown narrative. His most substantive thread argued that no frontier lab, nor the US or China, will meaningfully slow AI development, and that most proposed "regulation" amounts to regulatory capture aimed at rivals. His evidence was demand scarcity: memory procurement from Samsung stretching to 2031, and the 70–100%-plus supply-constrained growth guidance from Nvidia and Broadcom (AVGO), the custom-ASIC and networking bellwether he uses to validate his photonics longs. He urged followers to ignore commentators calling for a crash in names from Nvidia to TSMC (TSM) to Micron (MU) — the latter a centerpiece of his memory-supercycle thesis — arguing such fear-mongering only shakes out retail. He closed by needling Anthropic for pairing calls to limit training compute with a rapid push toward its own IPO.

*This is derived commentary, not investment advice.*