Dangerous divergence intensifies in US stocks: Median S&P 500 constituents are down 15% from their peaks, with AI giants propping up the index’s performance.

According to Bitget’s official data, the sixth batch of rewards for the platform’s "Peer Program" has been fully distributed, with total rewards amounting to 6,182,523 USDT and a total of 4,515,913 recipients. Additionally, after users visit the event page to claim their rewards, they will receive an extra chance to enter the 8th anniversary surprise blind box draw. The event will run until October 26.

Ethereum co-founder Vitalik Buterin published a post noting that roughly half of Bitcoin mining costs come from capital expenditure on mining hardware, with the other half attributed to electricity consumption. By contrast, AI inference costs are almost entirely dominated by computing hardware capital expenditure, a difference that reflects the distinct cost structures of the two computing businesses. Bitcoin miners must cover ongoing electricity bills, with profit margins determined by mining machine efficiency, Bitcoin prices, and changes in network difficulty; AI inference, meanwhile, relies far more on investments in expensive hardware such as GPUs, as well as device utilization rates and computing efficiency. The industrial boundaries between the two sectors are also shifting. As demand for AI computing power grows, some Bitcoin mining firms are repurposing existing power resources and data center facilities for AI and high-performance computing (HPC) services, aiming to transition from pure mining to computing infrastructure services. Reuters previously reported that some mining companies are leveraging power capacity originally built for mining to enter the AI data center market. However, this transition does not eliminate cost pressures: mining businesses are more directly exposed to fluctuations in electricity prices and Bitcoin prices, while AI computing operations face risks including GPU procurement, depreciation, and underutilization. Both sectors compete for power and computing resources, but their ultimate profitability depends on how effectively these resources are converted into revenue.

Australian AI anti-fraud startup Apate.ai—derived from research at Sydney’s Macquarie University and founded in 2024—has recently gained attention. The company deploys between thousands and roughly 350,000 AI agents, which pose as ordinary victims of varying ages and accents. These agents handle scam calls around the clock, respond to phishing text messages, click malicious links, and infiltrate scammers’ group chats. Their core mission is to feign ignorance to waste scammers’ time, extracting intelligence such as bank card numbers, crypto wallets, and fake websites from fraudsters, then package this data for banks and telecom firms to use for blocking operations. Apate.ai states it has completed over 2.5 million conversations, cumulatively occupying scammers for approximately 17 years of work time, and extracted more than 250,000 actionable intelligence items. Its clients include Australia’s Commonwealth Bank, TPG Telecom, and others, with some institutions reporting a notable decline in fraud losses. In August 2026, Apate.ai secured an $8.15 million seed round led by Lobby Capital, bringing its total funding to nearly $10 million. The startup is now pursuing global expansion, including establishing a London office and strengthening its footprint in the United States.

Beating AI Express News: Anthropic has announced performance optimization results for its latest Claude model. From August to October this year, the time to first token (TTFT) for Claude Code local sessions dropped from 4.07 seconds to 1.33 seconds, a roughly 3x speedup. One key optimization is using gzip compression for conversation data sent to the model each round, cutting network transmission overhead from repeated uploads of long conversations. For Claude Code cloud sessions, TTFT fell from 16.99 seconds to 6 seconds, a roughly 2.8x speedup. Beyond data compression, Anthropic also introduced a pre-launch mechanism: as soon as a user begins inputting a task, the system preps the remote code execution environment, removing the need to wait for command submission to initiate it. The wait time for Claude web chat and Cowork dropped from 3.36 seconds to 2.17 seconds, a roughly 1.5x speedup. Claude API’s time to first token fell from 1.53 seconds to 1.01 seconds. Notably, the first three sets of data exclude model computation time, while the API figure includes it.

According to data from Crisil and Coalition Greenwich, global banks are projected to rake in $8.5 billion in investment banking fees from AI-related transactions this year, a sharp surge from $3.5 billion last year. The revenue covers M&A advisory fees and underwriting fees for bond and equity financing. However, as data center financing expands and interest rates climb, Wall Street has grown more cautious in evaluating risks of AI investments. Goldman Sachs CEO David Solomon noted that future capital and computing power demands for the AI ecosystem remain uncertain; Morgan Stanley emphasized it will stay highly selective about risks taken and capital deployed, avoiding transactions that relax credit standards. The bond market has seen a divergence: JPMorgan Chase data shows investment-grade bond issuance cooled in September during the third quarter, while speculative-grade financing stayed weak throughout the period. From the end of June to October 13, the 10-year U.S. Treasury yield rose from 4.47% to 5.21%, and the worst yield on high-yield bonds climbed from 7.16% to 8.30%, pushing financing costs higher. The AI financing boom still brings hefty fees for banks, but issues like high interest rates, data center returns, and borrowers' debt-servicing capacity are prompting Wall Street to tighten risk controls while chasing revenue.

The AI boom has driven semiconductor companies to repeatedly hit record profits, but the market is starting to worry the industry may repeat its cyclical downturn. Cited data shows global semiconductor sales growth in the first half of 2026 reached its highest level since the 1980s. TSMC has broken its profit record for six consecutive quarters, while Samsung Electronics has also posted four straight quarterly profit highs. However, the Philadelphia Semiconductor Index has fallen nearly 14% since October 6, prompting investors to re-examine the sustainability of the AI supercycle. Unlike past cycles, this semiconductor expansion features higher supplier concentration and stronger corporate cash flow. TSMC’s capital expenditure this year is expected to hit around $62 billion, and could rise to $76 billion next year, with its expansion relying mainly on operating cash flow rather than large-scale borrowing. Advanced manufacturing processes and memory chip markets are concentrated among a small number of vendors, boosting suppliers’ pricing power. More notably, chipmakers are shifting some demand risks to customers via long-term contracts. NVIDIA has committed over $250 billion to secure chip supply and related components through January 2029; Micron, meanwhile, has locked in part of its business through "take-or-pay" contracts, with a committed value of $32 billion through 2030. These contracts require customers to pay as agreed even if they do not take delivery. This means even if the AI investment boom cools, chip manufacturers may still have a buffer from advance payments and long-term orders, but customers will face greater procurement and financial risks. Still, if demand plummets and capital expenditure shrinks, chip prices, corporate profits, and high-valuation stocks could still take a hit. The semiconductor industry’s cyclical pattern has not disappeared — it is just that the cost of the next downturn may be shared by both suppliers and customers.