Short-term funds take profits, Bitcoin inflow to Binance hits new high since February.

Insight Beating AI News Flash: AI Physics Lab Physical Superintelligence (PSI) announced that using models including GPT-6 Astra and Claude, it has made new progress on 5 long-standing unsolved quantum information problems. For example, researchers proved that in specific quantum fields, quantum entanglement persists even when two regions are far apart, but once the distance exceeds a certain threshold, nonlocal correlations can no longer be detected via the CHSH Bell inequality test. Another study demonstrated that a certain type of quantum operation, when performed consecutively enough times, will definitely destroy entanglement. During the research, physicists were responsible for guiding and verifying, while models like GPT and Claude participated in the derivation. The team also had AI Agents in Codex and Claude Code write Lean formal proofs, with computers checking the validity of the reasoning. All five results have released their papers, proof codes, and verification materials. However, not all five results fully resolved the original problems; some conclusions only apply under specific conditions, and several issues remain unsolved.

According to data from the BIT Stock Platform, U.S. pre-market crypto-related stocks showed mixed performance, with MSTR down 0.75%, ASST (Strive) down 1.56%, CRCL down 1.24%, COIN down 0.69%, BMNR down 1.24%, SBET up 0.10%, HOOD up 0.16%, and PURR down 0.88%.

Beating AI News Flash: The shortage list spurred by the AI boom has expanded beyond GPUs and memory to circuit boards. When planning their 2027 supply chains, HP and Dell have found that even the once-overlooked printed circuit boards (PCBs) could face shortages. In September this year, Amazon invested around NT$1.59 billion to take a stake in Taiwanese PCB maker Kinwong Electronics, shifting from a client to a shareholder. AI servers require PCBs with more layers and lower signal loss, while chip packaging substrates are growing larger, driving a sharp surge in demand for key materials such as specialty glass fiber cloth and copper-clad laminates. Suppliers are prioritizing higher-margin AI orders, squeezing demand for general-purpose servers. Lead times for PCBs for some general-purpose servers have reached nearly a year. However, production capacity for key materials will struggle to meet demand in the short term. TrendForce estimates that Japan’s Nitto Denko controls around 90% of the global T-glass specialty glass fiber cloth market, with new capacity not coming online until mid-2027 at the earliest. The chairman of Taiwanese leading substrate manufacturer Unimicron Technology warned on October 20 that some key materials could face a roughly 20% shortage in 2027, with lead times for some chip packaging substrate production equipment stretching as long as 30 months.

Beating AI Express: Local time October 26, the U.S. Department of Justice filed an appeal notice with the Ninth Circuit Court of Appeals, seeking to overturn a California court’s August ruling and reinstate the full federal agency ban on Anthropic. The same day, Anthropic publicly submitted its IPO prospectus. The case has two parallel tracks: In September, the Washington Court of Appeals already allowed the Pentagon to exclude Claude from military systems and related defense procurement citing "supply chain risks". However, the California court blocked a broader ban: it not only ordered all federal agencies to cease using Claude, but also prohibited defense contractors from engaging in non-military commercial partnerships with Anthropic. The judge ruled the government lacked sufficient basis to prove the relevant risks, and the ban was in fact retaliation for Anthropic’s public criticism of the government. Anthropic’s prospectus shows government agency contract revenue accounts for only around 1% of its total revenue, but the ban has impacted a larger commercial market. Evidence cited by the California court indicates three government contractors have terminated or been ordered to terminate their contracts with Anthropic. Separately, three deals totaling over $180 million that were near completion fell through. Some clients have requested to add unilateral termination clauses, while others have directly switched to competitors. Anthropic previously estimated to the court that if the broadest ban were fully implemented, its 2026 revenue could decline by billions of dollars.

SK Hynix will release its third-quarter 2026 financial results after the U.S. market closes on Tuesday, alongside a new shareholder return plan. Market focus is not only on high-bandwidth memory (HBM) demand and earnings performance, but also whether the company will announce a large-scale special dividend. Other storage manufacturers including Samsung Electronics and Japan’s Kioxia will also report their results in succession. AI infrastructure demand continues to drive an uptick in the storage chip cycle. KB Securities analysts estimate that over 70% of the memory chips South Korea’s two leading storage chip makers plan to produce next year are already covered by five-year supply agreements signed with major tech firms. However, as the market begins to worry about returns on AI capital expenditure and financing costs, investors are paying closer attention to order quality, future demand sustainability, and production expansion pace. On the shareholder return front, South Korean media previously reported that SK Hynix is considering a special dividend of around 16 trillion won (approximately $12 billion); another report noted that the total new shareholder return package could reach up to 60 trillion won (about $45 billion). None of the above figures have been officially confirmed by the company. Earlier, SK Hynix announced it would repurchase and cancel 40 trillion won in treasury stock, and plans to allocate more than 50% of its free cash flow from 2025 to 2027 to shareholder returns. This earnings report and future demand guidance will serve as a key basis for the market to judge whether the AI storage chip supercycle can continue.

Insight Beating AI Flash News: Cloudflare has disclosed an AI data breach. A user isolation flaw in its Workers AI platform’s GLM-5.3 model led to 155 responses mixing content previously submitted by other customers—meaning messages you send to the AI could appear directly in strangers’ outputs. The issue traces to GPU cache management: Cloudflare set a 1.05-million-token context window for the model, but the GPU cache handling inputs only holds ~680,000 tokens. Long requests that should have triggered an error instead ran via underlying inference engine SGLang, letting the model read residual cache from other users. All 155 affected responses came from long requests exceeding cache capacity. Cloudflare admitted it knew the cache was insufficient pre-launch, but tests only checked for normal interface responses, not verifying content, so missed the data leakage. Joint troubleshooting uncovered another cache flaw: when a user cancels a mid-process request, the system may allocate cache to another user early, while original data is incomplete, then overwrite the new user’s content. The SGLang team reproduced the issue with two H200 GPUs: the AI, normally counting from 1, suddenly started outputting prompts from another request. Such errors leave no abnormal logs. Cloudflare fixed the actual leak on September 9, confirmed no other Workers AI models were affected, and found no evidence of malicious exploitation. The SGLang team has since patched related cache management defects, recommending relevant deployments upgrade to v0.5.22 or newer.