Chinese equities came under pressure as investors questioned whether AI hardware demand can sustain the sector’s rally and data pointed to a weaker domestic economy. Reports put the CSI 300 and Shanghai Composite near their lowest levels since July 2025, with Monday losses ranging from roughly 0.5% to 0.9% in the different accounts; Hong Kong’s Hang Seng rose. AI-linked hardware took heavier losses: the 5G Communication Index fell as much as 7.5%, STAR50 dropped as much as 5.2%, and optical-transceiver maker Zhongji Innolight fell as much as 12.3%. Slower October factory output and weak consumption added to concerns. Morgan Stanley favored individual-stock selection over broad exposure, while ViaBTC’s Jeff Ko said tighter U.S. export controls could accelerate investment in China’s domestic AI chips and computing infrastructure. Investors also looked to Nvidia’s earnings and Anthropic’s expected IPO as tests of AI-market sentiment.
Databricks published a guide and open code for running five tabular foundation models on its AI Runtime GPU service. The models make predictions from spreadsheet-style data using example rows rather than task-specific training, for uses such as credit-risk, customer-churn and demand prediction. Databricks says GPU inference costs more per prediction than CPU-based XGBoost and uses more memory as training data grows; it has not published latency, memory or dollar-cost figures that would let customers calculate cost per 100,000 predictions. TabICLv2, Mitra v2 and KumoTabular have commercially permissive terms, while TabFM and the latest TabPFN require separate agreements. Databricks says Nvidia H100 chips can run TabPFN 3 on tables of up to one million rows and 200 columns.
Epoch AI’s tracker recorded a sharp rise in disclosed software vulnerabilities at 21 major organizations: monthly high-severity disclosures climbed from about 590 to nearly 2,940 in roughly six months. Critical disclosures rose from around 90 to a peak near 625, then eased to about 450 in the latest reading. The increase followed the announcement of Anthropic’s Claude Mythos Preview, but Epoch says the data does not establish that AI caused it. The tracker measures disclosure dates, not discovery dates, and reporting practices vary. The figures may reflect faster identification of existing flaws, putting pressure on security teams to verify and patch them quickly.
European shares edged higher as investors awaited Nvidia’s earnings, due Tuesday, and pricing news for Anthropic’s expected IPO; London-listed miners fell after weak Chinese activity data. Bernstein estimated planned and under-construction data centers represent about $1.4 trillion of investment, despite political uncertainty and near-term supply constraints. Bloomberg reported that Anthropic’s IPO drew orders for nearly seven times the shares on offer. The company is targeting a $1.4 trillion valuation and could seek up to $90.4 billion including a separate private placement, with trading reportedly possible Thursday or Friday. The offering would exceed SpaceX’s $75 billion base offering; commentators also compared demand with Snowflake’s 2020 IPO, while warning that strong demand does not prevent post-listing declines.
In other AI developments, the FDA cleared Uptimify, software that analyzes CT angiography images to help physicians detect and follow possible brain aneurysms. CrowdStrike launched an AI Kill Switch in Falcon Guardian to tell connected AI gateways to block agents from the models guiding their actions. The accompanying security analysis warns that agents may restart, replicate or find alternate routes, so teams should test for lingering access and shadow agents and stop them at multiple control points, including endpoints. CrowdStrike’s George Kurtz said local open-weight models grew from about 25% to nearly 80% of token usage in a matter of months, a shift that can leave network-only controls blind to activity on devices.
A Microsoft-commissioned GigaOm comparison modeled costs and quality for Copilot, Claude and ChatGPT across 37 prompts and Microsoft productivity apps. For advanced users performing 140 monthly tasks, it estimated monthly costs of $96 for Copilot, $299 for Claude and $149 for ChatGPT, using October 2026 prices or assumptions where prices were unavailable. The study found broadly similar quality across tested workloads, with variation by task; Microsoft attributed Copilot’s lower modeled cost partly to included everyday use and usage-based charges for advanced agent workflows.
Menlo Ventures’ Anthropic investment exceeded $500 million and marked a shift in the firm’s risk approach. Wu Martin said the bet could have put Menlo at risk, credited partner Matt Murphy’s backing and noted support from limited partners. The firm now directs more than 90% of investments to AI, making early bets and larger investments in emerging winners. Menlo forecasts consumer AI spending rising from $12 billion–$14 billion in 2025 to $40 billion in 2026. Its survey found about two-thirds of U.S. consumers use AI, about one-quarter use it daily, and roughly half of AI users pay for a product, compared with about 3% last year.
J.P. Morgan Asset Management strategist David Kelly urged investors to diversify AI holdings amid volatility and technological uncertainty, saying technology stocks have left many U.S. investors overweight and speculative shares could fall sharply. He recommends rebalancing toward value stocks and developed international equities. Kelly expects AI to lift productivity but says benefits will not reach every company. He also anticipates a cooler economy, lower tariffs, limited fiscal stimulus amid congressional gridlock and easing year-over-year inflation as gasoline prices decline.
Key Takeaways
- Chinese stocks fell to or near their weakest levels since July 2025 as AI hardware shares sold off and October factory output and consumption weakened.
- Nvidia earnings and Anthropic’s anticipated IPO became key near-term tests for investor confidence in the AI trade.
- Databricks says its tabular foundation models cost more per prediction than CPU-based XGBoost; it has not provided figures for latency, memory or cost per 100,000 predictions.
- Databricks lists TabICLv2, Mitra v2 and KumoTabular as commercially permissive; TabFM and the newest TabPFN require separate agreements. Nvidia H100s can run TabPFN 3 on tables up to one million rows and 200 columns.
- Epoch AI found monthly high-severity vulnerability disclosures at 21 major organizations rose from about 590 to nearly 2,940, but says the data does not prove AI caused the increase.
- Anthropic’s planned IPO reportedly attracted orders for nearly seven times the shares offered; the company targets a $1.4 trillion valuation and may seek up to $90.4 billion including a private placement.
- The FDA cleared Uptimify to help physicians detect and assess possible brain aneurysms in CT angiography images.
- CrowdStrike’s AI Kill Switch blocks agents through connected gateways, but the article argues endpoint controls and checks for restarted or shadow agents are also necessary.
- A Microsoft-commissioned comparison modeled advanced-user monthly costs at $96 for Copilot, $299 for Claude and $149 for ChatGPT, with broadly similar quality across tested workloads.
- Menlo Ventures has invested more than $500 million in Anthropic and now directs over 90% of its investments to AI; J.P. Morgan strategist David Kelly recommends diversifying AI-heavy portfolios.
China stocks slip as AI hardware shares and weak data weigh
Chinese stocks fell on Monday as investors questioned how long the artificial intelligence rally could last and fresh signs of economic weakness added to concerns. The CSI300 Index dropped 0.6% by the lunch break, while the Shanghai Composite Index fell 0.4%; Hong Kong’s Hang Seng rose 0.8%. The 5G Communication Index fell more than 4%, and the STAR50 Index lost nearly 3%. Zhongji Innolight, which makes optical transceivers for AI data centers, fell 7.7%. October factory production slowed and consumption stayed weak, while the CSI300 had fallen 14% so far this year.
Weak data and AI hardware losses drag Chinese stocks lower
Chinese stocks fell on Monday as worries about the artificial intelligence rally and signs of weakness in the domestic economy weighed on investors. The CSI300 Index fell 0.6% by the lunch break and the Shanghai Composite Index dropped 0.4%, while Hong Kong’s Hang Seng rose 0.8%. The 5G Communication Index lost more than 4%, the STAR50 Index fell nearly 3%, and Zhongji Innolight dropped 7.7%. October factory production slowed and consumption remained weak. Morgan Stanley analysts said China’s deflation measures were incremental and favored picking individual stocks over broad market exposure; the CSI300 was down 14% for the year.
China stocks hover near 16-month low as AI hardware losses deepen
Chinese stocks ended lower on Monday and stayed near their lowest level since July 2025. Hardware shares fell as investors questioned the strength of the artificial intelligence rally. Fresh signs of weakness in China’s domestic economy added to the negative mood.
China stocks close lower as AI hardware selloff deepens
Chinese stocks ended lower on Monday, near their lowest level since July 2025, as investors worried about the artificial intelligence rally. The CSI 300 closed at 3,979.80 points, down 0.91% from the previous session. Losses in AI hardware shares had accelerated over the past few days. The report also lists a 5-day change of -14.04% for the index.
China stocks near 16-month low as AI hardware shares plunge
Chinese stocks ended lower on Monday, near their lowest level since July 2025, as AI hardware shares fell and concerns about the artificial intelligence rally grew. The CSI300 and Shanghai Composite Index each closed down 0.9%, while Hong Kong’s Hang Seng rose 0.5%. The 5G Communication Index fell 7.5%, the STAR50 dropped 5.2%, and Zhongji Innolight lost 12.3%. Jeff Ko of ViaBTC said tighter US export controls could speed investment in China’s domestic AI chips and computing infrastructure. October factory production slowed and consumption remained weak, and the CSI300 was down 14% for the year.
China stocks hit 16 month low as AI hardware shares slide
China’s CSI300 and Shanghai Composite indexes fell on Monday to their lowest levels since July 2025. The CSI300 closed at 4,263.65, down 0.8%, while the Shanghai Composite closed at 2,514.51, down 0.5%. Investors worried that demand for AI hardware may not sustain the rally, and signs of weakness in China’s economy added to the concern. The CSI300 has fallen 12.4% this year, while the Shanghai Composite has lost 14.1%.
AI hardware selloff and weak data weigh on China stocks
Chinese stocks fell on Monday as investors questioned whether demand for AI hardware could support the sector’s recent rally. At midday, the CSI300 was down 0.6% and the Shanghai Composite was down 0.4%, while Hong Kong’s Hang Seng rose 0.8%. The 5G Communication Index fell more than 4% to a one year low, and the STAR50 dropped nearly 3%; Zhongji Innolight lost 7.7%. October figures pointed to slower factory output and weak consumer spending, while financial and consumer staples shares gained. Investors were also looking ahead to Nvidia’s earnings and Anthropic’s IPO this week as tests for the AI rally.
Databricks weighs GPU costs for AI models that predict from tables
Databricks says GPU inference for tabular AI models costs more per prediction than running XGBoost on a CPU. The article compares models including TabICLv2, with about 28 million parameters under a BSD 3 Clause license, and TabFM, with about 400 million parameters under a noncommercial license. KumoTabular and Mitra v2 list commercially permissive terms, while TabPFNv3.5 does not. TabICL compresses each row into four tokens, allowing test rows to attend only to training rows, and TabFM and KumoTabular use a similar approach. Databricks has not published latency, memory, or dollar cost figures, leaving buyers without a clear cost per 100,000 predictions.
Databricks offers five AI models for predictions from table data
Databricks published a guide and open code for running five tabular foundation models on its AI Runtime GPU service. These models predict from spreadsheet style data without task specific training, instead learning from a company’s example rows when making predictions. They can help with tasks such as predicting credit risk, customer churn, and demand, but Databricks says they cost more per prediction than XGBoost on CPUs and use more memory as training rows increase. TabICLv2, Mitra v2, and KumoTabular can be used commercially, while TabFM and the newest TabPFN require a separate deal with their makers. Databricks says Nvidia H100 chips can run TabPFN 3 on tables with up to 1 million rows and 200 columns; AI Runtime has been in preview since March.
High severity software flaws rise about fivefold in six months
Epoch AI data shows that monthly high severity vulnerabilities disclosed by 21 major software organizations rose from about 590 to nearly 2,940 in roughly six months. Critical severity disclosures also climbed, from around 90 to a peak of about 625, before falling to roughly 450 in the latest reading. The increase followed the announcement of Anthropic’s Claude Mythos Preview, but Epoch cautions that the data does not prove AI caused the rise. The tracker counts when vulnerabilities are disclosed, not when they are found, and reporting practices differ across organizations. The surge may reflect faster discovery of existing flaws, which raises pressure on security teams to validate and patch them quickly.
European stocks edge higher as investors await Nvidia earnings
European stocks mostly edged higher, while London-listed miners fell after weak Chinese activity data pointed to pressure on industrial production and domestic demand. Investors focused on turmoil around Nvidia leadership and the company’s earnings report, due Tuesday, as well as Anthropic’s expected IPO pricing. Bernstein estimated planned and under-construction data centers represent about $1.4 trillion in investment, despite political uncertainty and near-term supply limits. U.S. stock futures fell after weakness in Asian AI stocks, while the euro remained near a 19-month low against the dollar. Federal Reserve Chair Kevin Warsh was set to speak Monday at a central bank conference on productivity and jobs.
Anthropic IPO draws orders for seven times the shares on offer
Bloomberg reported that Anthropic’s planned IPO has drawn orders for nearly seven times the shares available. The AI company is targeting a $1.4 trillion valuation and, with a separate private placement, may seek to raise as much as $90.4 billion. Reports said its shares could begin trading Thursday or Friday. The deal would be larger than SpaceX’s $75 billion base offering, and demand is being compared with Snowflake’s 2020 IPO, whose stock rose 112% on its first trading day. Some market strategists warned that major IPOs can lose value after launch, even when demand is strong.
FDA clears Uptimify AI software to help detect brain aneurysms
The U.S. Food and Drug Administration has cleared Uptimify, an AI software tool designed to help physicians spot possible brain aneurysms on CT angiography images. The software analyzes the images and supports detection and follow-up assessments. Radiologists and other physicians can now use it in clinical settings. Its purpose is to help improve the speed and accuracy of brain aneurysm detection.
CrowdStrike’s AI kill switch needs more than one control point
CrowdStrike launched its AI Kill Switch in Falcon Guardian to direct connected AI gateways to block agents from accessing the models that guide their actions. The article warns that rogue agents may restart, create new copies, or use other routes, so one control point may not be enough to stop them. CrowdStrike Chief Product Officer AJ Shipley said agents must be stopped at every available control point, including on the endpoint where they run. Network-only tools may miss local open-weight models, which Kurtz said rose from about 25 percent to almost 80 percent of token usage in a matter of months. The article urges security teams to test whether agents return after termination, check lingering access, and look for shadow agents and local models.
Microsoft study finds Copilot had the lowest modeled AI costs
Microsoft commissioned GigaOm to compare the cost and quality of Copilot, Claude, and ChatGPT for everyday and advanced AI users. The study tested 37 prompts across Copilot Chat and Cowork and Copilot in Word, Excel, and PowerPoint, modeling 100 monthly tasks for everyday users and 140 for advanced users. For advanced users, it estimated monthly costs of $96 for Copilot, $299 for Claude, and $149 for ChatGPT, using October 2026 prices or assumed prices where list prices were not available. The research said quality was broadly similar across the products for the tested workloads, though results varied by task. Microsoft said Copilot’s included everyday use and usage-based charges for advanced agent workflows helped explain its lower modeled total cost.
Menlo’s Anthropic investment shifts its approach to venture risk
Menlo Ventures invested more than $500 million in Anthropic, a bet Wu Martin said could have put the firm at risk. She credited partner Matt Murphy for backing the investment and said limited partners helped fund it. Menlo now puts more than 90% of its investments into AI, combining early bets on teams with larger investments in emerging winners. Menlo’s Consumer AI Report forecasts global consumer AI spending will rise from $12 billion to $14 billion in 2025 to $40 billion in 2026. Wu Martin said about two thirds of surveyed U.S. consumers use AI, about one quarter use it daily, and about half of AI users pay for a product, compared with about 3% last year.
JPMorgan strategist urges investors to diversify AI holdings
David Kelly, chief global strategist at J.P. Morgan Asset Management, urged investors to diversify their AI holdings as the sector faces more volatility and technological uncertainty. He said technology stocks have left many U.S. investors overweight and warned that speculative stocks could fall sharply. Kelly recommends rebalancing toward value stocks and developed international equities. He believes AI will change the world and expects it to boost productivity, but says not every company will benefit. He also expects a cooler economy, lower tariffs and limited fiscal stimulus due to congressional gridlock, as well as easing year-over-year CPI as gasoline prices decline.
Sources
- China’s stocks fall as AI hardware selloff, weak data weigh
- China Stocks Fall as AI Hardware Selloff and Weak Data Weigh
- China stocks end near 16-month low as AI hardware selloff deepens
- China stocks end near 16-month low as AI hardware selloff deepens
- China stocks end near 16-month low as AI hardware selloff deepens
- China stocks end near 16-month low as AI hardware selloff deepens
- China stocks fall as AI hardware selloff, weak data weigh
- Databricks bets H100s fix tabular AI cost
- Databricks pitches AI that predicts from tables without training
- Critical Software Vulnerabilities (CVEs) Have Jumped 5x In The Last 6 Months, Shows Epoch AI Data
- European Midday Briefing : Stocks Mixed Ahead of Pivotal Week for AI Trade
- Anthropic IPO Reportedly Oversubscribed 7 Times – But Will The AI Euphoria Survive The Stock Market Debut?
- FDA clears new AI software for brain aneurysm detection
- CrowdStrike’s AI Kill Switch Shows Why It’s Going to Take More Than a Single Off Button to Stop a Rogue AI Agent
- What AI at work really costs
- Menlo's $500M Anthropic bet reset venture risk
- Diversify AI holdings as tech trade enters volatility and uncertainty, says JPMorgan’s David Kelly
Comments
Please log in to post a comment.