Note: The X account is not owned by the original Enron entity. Users should be aware of the risk.

Insight Beating AI News Flash: After the launch of GLM-5.3 Flash, Zhipu AI confirmed a previously undisclosed detail: all inference computing power for Ox Alpha during its anonymous testing phase came from domestically produced Chinese AI chips. Ox Alpha processed 23.2 trillion tokens in just six full calendar days after going live on OpenRouter, more than twice the throughput of DeepSeek-V4-Flash over the same period. OpenCode previously even claimed its backend can provide a daily free quota of 100 trillion tokens. Zhipu AI noted that it has optimized end-to-end inference performance on the same domestic Chinese hardware to three times its original level, and its current hardware efficiency and per-token cost are nearly on par with mainstream NVIDIA GPUs. SemiAnalysis specifically singled out this fact. Earlier, when the public saw the daily 100 trillion token supply capacity, they wondered if such a scale could only be supported by top-tier labs and NVIDIA GPUs. It turns out that all operations here were entirely powered by domestically produced Chinese chips from start to finish.

Beating AI Express News: Despite being the traditional off-season for copper consumption from July to August, the scrap copper market in Taizhou, Zhejiang has seen exceptional activity. Prices of some high-quality scrap copper have exceeded 90,000 yuan per ton, even hitting 100,000 yuan per ton. Market demand has surged two to three times compared to earlier periods, with some goods pre-booked by clients before even being processed. Recycled copper enterprises are also operating at full capacity. Zhejiang Judong Co., Ltd. plans to raise its recycled copper production capacity to 100,000 tons this year, and is currently operating at almost full capacity, with orders booked for about a month ahead. Data shows that the copper spot price index of the Shanghai Nonferrous Metals Exchange rose by 31% in the first half of this year; from February to August, COMEX and LME copper futures rose by 52% and 58% respectively. Industry insiders believe that the core driver of this round of copper price hikes remains supply-demand mismatch: shrinking overseas copper mine output and rising raw material costs have slowed supply growth, while AI computing centers, power grids, and new energy infrastructure continue to generate new copper demand. As the traditional peak consumption season arrives from September to October, whether the supply-demand gap will further widen remains a key point of attention. However, analysts caution that the rapid rise in copper prices has already exerted significant pressure on downstream processing enterprises, with demand not transmitting smoothly, meaning the upward momentum of this round may have entered its mid-to-late stage. In the medium and long term, the progress of global copper mine resumption and the pace of AI demand realization will be the two main lines determining copper price trends.

According to a Wall Street Journal (WSJ) report, as stablecoins accelerate their penetration into payment and financial systems, traditional banks that once actively lobbied to restrict stablecoin development are starting to shift their stance. Some large banks have begun researching or advancing their own stablecoin initiatives to counter competitive pressure from crypto firms and non-bank giants. The report notes that JPMorgan Chase recently evaluated the possibility of issuing its own stablecoin, though no formal product plans have been finalized yet. Meanwhile, more than a dozen financial institutions, including Bank of America, Wells Fargo, and Santander, are advancing a global stablecoin project, which will initially focus on a U.S. dollar stablecoin before expanding to the euro and other G7 currencies, primarily serving commercial clients. Additionally, the BankChain Alliance—comprising 39 state banking associations and approximately 3,000 banks—plans to launch a bank-owned and governed blockchain platform in the first half of 2027, supporting both tokenized deposits and stablecoins. Banks previously leaned toward developing tokenized deposits, as these retain the regulatory, accounting, and credit risk frameworks of traditional bank deposits and keep funds within the banking system. However, as non-bank institutions including Visa, BlackRock, and Google have all entered the stablecoin market, the banking sector has grown concerned that stablecoins could erode traditional deposit and payment businesses, and has increasingly come to view issuing stablecoins as a potential defensive strategy.

Research indicates the tokenized stock market is experiencing rapid growth. Year-to-date in 2026, trading volume of tokenized stock perpetual contracts has surged from roughly $16 billion in 2025 to over $590 billion, while spot trading volume has topped $88 billion. Yet beneath this market expansion, the underlying legal and equity framework of tokenized stocks has emerged as a key concern for investors. Three main models currently dominate the market: issuer-backed, custodian, and synthetic. Issuer-backed tokens directly represent stock ownership, entitling holders to rights such as voting, dividends, and corporate actions; custodian models deliver corresponding economic rights via securities intermediaries; synthetic tokens, by contrast, are essentially contractual claims against third parties that do not confer direct ownership of the underlying stock, exposing holders to risks including counterparty, tracking, and corporate action transmission. The report notes that as regulatory frameworks gradually take shape and blockchain settlement infrastructure matures, demand for tokenized stocks will keep rising. However, investors must prioritize distinguishing whether their holdings represent actual equity or synthetic exposure built on it—a distinction that will directly determine their entitlements, risks, and legal protections.

Nick Timiraos, a Wall Street Journal reporter known as the "Fed's mouthpiece," analyzed that Federal Reserve Chair Kevin Warsh will deliver his first major speech since taking office at Jackson Hole this week, with the market’s focus centered on how he explains why U.S. inflation remains above target and how this assessment will shape future interest rate paths. The core division within the Fed currently lies in whether inflation is primarily driven by one-off external shocks such as tariffs and war-induced energy price hikes, or if U.S. economic demand is still overheating. If it is the former, inflation could fall as the shocks fade; if the latter, it would mean the Fed may need to tighten policy further. Timiraos believes Warsh’s speech needs to clarify his assessment framework and the conditions that could lead to a shift in his interest rate stance. The article also notes that Warsh’s communication strategy of reducing forward guidance since taking office is being tested by the market. If he judges that previous rate cuts were too aggressive and the economy is still overheating, he may need to push for a reversal of those cuts; if inflation is mainly driven by external shocks, the need for further policy tightening will diminish. The market will also closely watch whether he proposes an inflation analysis framework adapted to the new environment of deglobalization, AI-related capital expenditures, and geopolitical conflicts.

Insight Beating AI News Flash: Zhipu AI’s newly open-sourced GLM-5.3 Flash is directly competing with DeepSeek-V4-Flash. Both models prioritize low cost, coding, and agent capabilities, but based on currently available overlapping evaluations, Zhipu’s model outperforms its rival. On Terminal-Bench 2.1, GLM-5.3 Flash scored 84.3 points, while DeepSeek-V4-Flash notched 82.7 points; on DeepSWE, the scores stand at 63.4 vs. 54.4, marking a 9-point lead for GLM. The former natively supports image and video input. Pricing is also more competitive: after GLM-5.3-Flash’s limited-time 50% discount, its per-million-token rates are $0.075 for regular input and $0.25 for output. Even at DeepSeek V4 Flash’s cheapest off-peak rate, it charges $0.22 per input million tokens and $0.66 per output million tokens. GLM’s input price is nearly three times cheaper, and its output price is 2.6 times lower. After the two-week promotion ends, GLM will revert to $0.15 per input million tokens and $0.5 per output million tokens—still lower than DeepSeek’s off-peak price. DeepSeek is currently only cheaper on cache-hit pricing.