Serenity, the stock-picker who back-maps the AI supply chain to find obscure upstream "chokepoints" — critical suppliers with few substitutes — spent Wednesday in a reflective, self-critical mood, acknowledging a sharp drawdown and laying out a revised framework for when markets actually pay him for being early.

Owning the losses. He jokingly rebranded his recent results "Diversified Losses," singling out his photonics-heavy book as the hardest hit. He conceded he had concentrated too much in optical names versus memory and other sectors without hedging. Among the casualties: AXT (AXTI), the indium-phosphide (InP) substrate maker he views as a near-monopoly material supplier for AI optics, which he said was punished mainly on dilution and float-expansion worries; Applied Optoelectronics (AAOI), the Made-in-America optical-transceiver maker, which he thinks simply got dragged down with the theme; and Soitec (SOI), the SOI-wafer supplier, where he disputes recent negative institutional reports. He stressed he can't give buy advice and that corrections are inevitable on the way up.

The "8-month rule." Serenity's main new idea was a timing heuristic: markets tend to price a theme in roughly eight months ahead — except for Space, Robotics and Quantum, which run hotter. By that yardstick he admits he was about four months too early on the furthest-out optical architecture. He mapped his photonics roadmap by stage: next-gen 1.6T pluggable transceivers are being priced now (Lumentum/LITE, Coherent/COHR, with AAOI and Jabil's 1.6T LRO product as later bloomers in 2027); CPO — co-packaged optics, putting the optical engine next to the switch die — "scale-out" names like POET arrive late 2026, while CPO "scale-up" sits in late 2027. He argued the selloff in some Asian CPO suppliers reflects markets disagreeing with how early he positioned.

Where he's repositioning. Serenity said he is reshaping the book around four themes — memory, neoclouds, photonics and robotics — leaning heavier where a catalyst is roughly eight months out, and trimming anything more than a year away. In memory, he reaffirmed the supercycle: Micron (MU), the lone US DRAM/HBM/NAND maker he frames as a structurally cheap re-rating candidate, sits "mid-bottleneck" as HBM3e and HBM4 ramp. In glass-core substrates — next-gen advanced packaging — he reiterated owning LPKF (LPK), the German laser-etching equipment near-monopoly, noting tool suppliers book revenue before the broader volume ramp. In neoclouds he highlighted Nebius (NBIS), his favored AI-cloud long, suggesting its $7–9B revenue run-rate target lands around Q4, with markets only now starting to price it. On robotics, he flagged Tesla's (TSLA) Optimus humanoid program as the anchor he plays through suppliers rather than the stock itself.

His clearest conviction tell: he singled out Sivers (SIVE) — his largest position, a tiny Swedish InP continuous-wave laser maker he calls the light-source chokepoint for CPO — as the kind of name he wants oversized precisely because it spans pluggables, NPO and every CPO architecture at once.

Separately, he noted Anthropic's accusation that Alibaba's (BABA) Qwen lab distilled its models via millions of fake exchanges, observing such practices seem widely known but unpenalized so far.

*This is derived commentary, not investment advice.*