Trump waives sanctions on Russian diesel imports, three oil tankers arriving in the U.S. next week spark legal controversy.

According to Taiwanese media reports citing supply chain sources, TSMC is considering raising prices for its advanced packaging services by 10% to 20%, with the increase primarily targeting its CoWoS packaging technology used in high-end AI chips. If the price hike takes effect, chipmakers including NVIDIA and AMD will face upward pressure on production costs. CoWoS, which integrates chips and high-bandwidth memory (HBM) into a single package, is a critical process for high-end AI accelerator chips. As demand expands beyond GPUs to server CPUs and custom AI chips (ASICs), advanced packaging capacity remains tight. The report projects TSMC’s monthly CoWoS production capacity will exceed 130,000 units in 2026, with an expected further rise to over 200,000 units by 2027. Even with ongoing capacity expansion, demand growth is likely to continue driving packaging price increases. Should costs rise, manufacturers heavily reliant on CoWoS such as NVIDIA could face margin pressure, while clients like Apple may also be indirectly impacted by the higher supply chain costs. However, the relevant price hike plans are still under consideration.

Mizuho Securities cut the price target for USDC stablecoin issuer Circle Internet Group (CRCL) from $45 to $43 in its latest research report, while maintaining an "Underperform" rating. The bank warned that AI personal assistant engines could pose "significant risks" to Circle’s mid-term revenue. The report noted that with the widespread adoption of AI agents, USDC’s market capitalization could face downward pressure of $4 billion to $9 billion, while Circle’s reserve revenue could see a mid-term decline of 10% to 20%. Mizuho’s simulations showed that AI agents including Muse, when answering the question “Where should funds be held,” tend to recommend users shift money from USDC to higher-yielding fiat alternatives protected by FDIC insurance, such as savings accounts offered by SoFi. Even if users stick to the crypto space, AI agents “appear to prefer other platforms” over holding USDC directly within Circle’s ecosystem. Mizuho further outlined a faster AI adoption scenario: roughly $70 billion could flow to fiat products, and another $20 billion to higher-yielding stablecoins, which would reduce USDC’s market cap by up to $90 billion and cause an 18% drop in Circle’s reserve revenue. The bank also cut its 2027 financial forecasts: it lowered the expected average circulation of USDC from $108 billion to $101 billion, reserve revenue expectations from $43 billion to $41 billion (a roughly 5% decline), and adjusted EBITDA expectations from $635 million to $603 million. The market reacted negatively. According to data from the BIT stock platform, Circle’s stock closed 7% lower at $79.84 in Monday trading, but Mizuho’s new price target implies roughly 46% additional downside.

Morgan Stanley’s latest research report points out that the "Three A" forces—AI agents, rising demand for air conditioners, and industrial reshoring—are collectively driving global power demand into a super cycle. The bank has raised its global power consumption growth forecast for 2025–2030 by 10%, projecting global electricity use will hit 36.3 trillion kilowatt-hours (kWh) by 2030, roughly 15% higher than market consensus. Around half of the power demand driven by AI agents, Asian industrial expansion, and new air conditioners and heat pumps has not yet been priced into related companies’ valuations. Morgan Stanley notes power demand growth is no longer driven solely by AI data centers. Global industrial power use is expected to rise from 11.1 trillion kWh in 2025 to 13.1 trillion kWh in 2030; global cooling power demand is forecast to grow from 2.9 trillion kWh to 4.1 trillion kWh over the same period, with a 7% compound annual growth rate (CAGR), outpacing the overall global power demand growth rate. Regionally, Asia is set to contribute around 64% of global incremental power demand between 2025 and 2030. The U.S. power supply gap is also severe: Morgan Stanley projects a 43 gigawatt (GW) power shortage for U.S. data centers by 2028, with some demand potentially spilling over to Asia, driving local power infrastructure investment. On the investment strategy front, Morgan Stanley believes pricing power in the power market is shifting from equipment suppliers to power generators, grid operators, and natural gas infrastructure owners. The bank is bullish on stocks including Bloom Energy, Solaris Energy, and China Yangtze Power, while adding Korea Electric Power Corporation (KEPCO) and Tata Power to its avoid list.

According to South Korea’s Maeil Economic Daily, major Chinese cloud service providers including Alibaba, Tencent, and ByteDance are in negotiations with Samsung Electronics and SK Hynix for long-term server DRAM supply agreements. Some clients reportedly aim to extend the typical 3–5 year long-term supply contracts to 2035, signing ultra-long 10-year deals. The companies seek to lock in products including server DDR and low-power LPDDR. As demand for AI servers continues to surge and memory supply tightens, cloud providers are rushing to secure memory capacity in advance, even willing to pay higher prices to meet the supply needs of their AI business expansion. Earlier, large U.S. tech firms had already locked in memory resources via long-term supply agreements; the move by Chinese cloud providers reflects that competition in AI infrastructure is further extending into the server memory supply chain.

Beating AI Express (Dongcha) reports that Baidu has released and open-sourced Unlimited-OCR 2, a model with approximately 3 billion parameters, of which only around 500 million are activated during inference. The first version, launched in June, could parse around 40 consecutive document pages, but as page count grew, generating each token required attention to visual information from all pages, leading to progressively slower decoding. The second version introduces the SR-SWA sparse attention mechanism: instead of accessing visual data from all pages each time, the model focuses on the current page, selects a small amount of relevant visual information, and continues recognition in combination with recently generated text. In 100-page tests, single-step decoding is roughly 5.6 times faster than the previous R-SWA attention scheme and about 7 times faster than traditional full attention, with significant memory savings. In 100-page document trials, chunked decoding and cross-page cache reuse cut the peak KV cache memory for single-path inference from 47GB to 1.77GB. In another 16-path concurrent test, generation throughput jumped from 360 tokens per second to 1315 tokens per second. Baidu also built MarathonDocBench, a dataset of 100 documents ranging from 103 to 300 pages, to test long-document parsing capabilities. Unlimited-OCR 2 scored 77.06, outperforming Gemini 3.8 Flash (28.61) and Qwen 3.8 Max (22.93) in Baidu’s in-house benchmark tests. For extremely long documents, the system can use windowed processing and reuse part of previous window caches to further reduce memory overhead, though too small a window may harm cross-page content coherence. The paper also notes the new mechanism cannot fully retain a document’s global information, with overall structures like multi-level headings still prone to misidentification.

Shenwan Hongyuan released a macro research report stating that U.S. AI investment growth is likely to peak around the fourth quarter of this year, entering a decline phase in 2027. The report notes that roughly 74% of capital expenditure in the global AI industry chain flows to data centers. As of the second quarter of 2026, hyperscale cloud providers account for 66% of total capital expenditure in the AI stack, making them the primary spenders. Per Bloomberg consensus forecasts, the five major U.S. cloud providers—Amazon, Google, Meta, Oracle, Microsoft—will see their combined capital expenditure exceed $800 billion in 2026, with a year-on-year growth rate of around 94%, accounting for approximately 2.5% of U.S. GDP. Their capital expenditure will continue to expand in 2027 and 2028, but growth rates are projected to slow to 40% and 10% respectively. On the allocation front, around 74% of global AI industry chain capital expenditure goes to data centers, with about 55% of that amount invested in servers and GPUs. Based on project-level data calculations, Shenwan Hongyuan estimates the average data center construction cycle is roughly 20 months, with an average shelving rate of around 33%; annual investment is primarily determined by existing plans and ongoing projects. Under three shelving rate scenarios (20%, 33%, 45%), all calculations point to the peak of U.S. AI investment growth likely in the fourth quarter of 2026—about one quarter later than the market consensus on the five major cloud providers’ capital expenditure growth rate—making a shift to a growth decline phase in 2027 a high-probability event.