Do not use the notional value of Leopold’s put options. It is very misleading. Leopold Aschenbrenner’s Situational Awareness LP 13F: Common stock longs: $3.856B Call positions, underlying value: $1.362B Total long / call exposure: $5.218B The reported put value is mostly underlying notional, not the actual premium paid. Some estimate: he may have spent around $500M buying the puts. Very rough number, but directionally important. This looks less like a “massive bearish short” and more like a large downside hedge / protection overlay against his long AI infrastructure plays. Most likely, he was hedging for war / macro tail risk (seems reasonable hedging) and the puts got smoked after the massive semi rally in April. Now FinX is running with the fear narrative.

Kyung Kye-hyun, a standing adviser and former head of Samsung Electronics' Device Solutions division: "Chinese companies are aggressively expanding their production capacity (CAPA)," "The market could shift starting in the second half of next year or the first half of 2028, when memory supply will surge." There is actually nothing new here. The industry widely expected it and priced it in accordingly. The problem is that everyone could still be wrong about AI demand, just as they have been over the past few years.

Ben Reitzes of Melius Research raised PTs most on Micron Technology and SanDisk, citing long-term agreements and memory as a key AI bottleneck, with valuation multiples likely expanding toward HDD peers as AI demand stays strong.

The future of AI: The setup demonstrates a practical hybrid model: smaller, local models handle most execution cheaply and privately, while frontier models provide high-level direction only as needed, reducing costs and latency.

Samsung is reportedly developing next-gen HBM packaging for mobile on-device AI. According to ETNews, Samsung is working on “Multi Stacked FOWLP,” combining advanced copper-pillar stacking with fan-out wafer-level packaging. Today’s mobile LPDDR still uses copper wire bonding, which limits I/O to roughly 128–256 terminals and creates signal-loss, thermal, and power-efficiency bottlenecks. Samsung’s VCS technology improves this by stacking DRAM dies in a staircase structure and connecting them with copper pillars. The new approach appears to push that further, aiming to bring HBM-like bandwidth closer to mobile devices.

Interesting chart - NVIDIA's Rubin AI platform will drive massive LPDDR consumption in 2027 CPU-side memory has been an under-appreciated part of DRAM demand as I discussed in my DRAM model.

If AI, especially GPUs, is like nuclear weapons, then China should just take over Taiwan now to prevent the U.S. from getting them, since Dario wanted the U.S. to block China from getting them anyway. Either Xi or Dario is dead wrong here.

China will be just fine with its AI chip production. Dario may be worried that if China gets enough GPUs, it could beat OpenAI or Anthropic, so he advocates for export controls with AI fearmongering. But China is no longer relying on NVDA chips or CUDA. It is not hard to imagine that, in a few years, Chinese open or closed models will be used everywhere and may run best on Chinese AI stacks.