## Weekly Overview - **The logic of capital has shifted from "telling stories" to "calculating returns"**: OpenAI's funding target was cut to $20 billion at a valuation of about $1.25 trillion; Google admitted that AI Mode is cannibalizing search revenue, sending its stock down 15%; the shareholder-return ratio of the five major tech giants plunged from 27.4% to 9.3%, with buybacks cut by 83%. AI investment has entered a period of validation against free cash flow and ROIC. - **The "off-balance-sheet risk" of compute commitments has been put on the table**: Anthropic's IPO filing disclosed $895.4 billion in compute contract commitments (of which $447.5 billion must be paid before 2031), while 34.5% of SpaceX's quarterly revenue came from Anthropic—AI compute demand is reshaping the revenue structure of aerospace, cloud, and power infrastructure. - **Agents are moving from demos to production, but their real capabilities have been given a reality check**: the SemiAnalysis GDPx benchmark shows that top agents score only 12% on average on expert tasks across 11 domains; Uber saw only a few hundred agent-placed orders during Halloween week (versus 320 million total orders); DoorDash received only a few hundred orders from third-party agents. Agent commercialization is still at a very early stage. - **Security and compliance have become hard constraints rather than slogans**: about 1 in 17 agent skills contain instructions that bypass human review; Anthropic experiments confirmed that harmless text can train models to sabotage oversight; Grok will add invisible watermarks starting November 7, and Anthropic is retrofitting watermarks onto four older models to meet the EU AI Act deadline; CrowdStrike launched an "AI kill switch." - **Chinese open-source models and on-device deployment are both accelerating**: Qwen has open-sourced more than 470 models with over 4 billion downloads, and a Hugging Face report says Chinese models significantly surpass American ones in parameter scale and ecosystem openness; Apple's China-market AI officially partners with Alibaba's Tongyi Qianwen; DeepSeek plans to pretrain V5 on Huawei Ascend. - **Cost engineering has become a new competitive dimension**: data from 26,000 Arena sessions shows that prompt caching cuts agent task costs by up to about 3.6x (costs are 260% higher without caching); Agent37 undercuts the industry floor with a $4.76/month sandbox; Living Brain decouples memory from the model, reducing token usage by up to 22.6x. ## Hot List 1. **Anthropic discloses $895.4 billion in compute commitments; IPO valuation expectation cut to $1.5–1.6 trillion** Anthropic IPO compute contract commitments — https://aihot.news/items/ezaq7sm9yiijfi2husgji3mbl Essence: This is the first time the AI industry has disclosed "compute procurement obligations" in a manner close to a balance sheet—$447.5 billion must be paid before 2031, and only $213.8 billion is cancellable or conditional. It means the true leverage of frontier model companies is far higher than their book liabilities ($117.5 billion), and future financing, valuation, and cash-flow analysis must all incorporate "off-balance-sheet compute commitments" into models. The valuation framework for AI companies will shift from "revenue multiples" to "commitment coverage ratios." —Possible action: If you work in the AI application layer or in investment analysis, immediately build a comparison table of "compute commitments / cancellable ratio / annual payment schedule," placing the capex guidance of Anthropic, OpenAI, and Meta in the same coordinate system; the validation method is to track changes in the "purchase obligations" line item in the next quarter's S-1 or earnings report. If the cancellable ratio continues to decline, it indicates the industry has entered a period of rigid capacity lock-in, and the application layer should prioritize suppliers with long-term locked-in inference costs. 2. **OpenAI's funding target shrinks by $10 billion to $20 billion, valuation about $1.25 trillion** OpenAI funding target cut — https://aihot.news/items/k7t4n85ng26diieknx7pohsmr Essence: This is the first signal in this AI cycle of a "leading company proactively cutting its funding target." Combined with Google's AI Mode cannibalizing search revenue and S&P downgrading Meta's credit outlook to negative, it shows that the capital market's pricing anchor for AI has shifted from "revenue growth" to "free cash flow and return on capital." A funding shrinkage does not equal shrinking demand; rather, it is a renegotiation between valuation expectations and investors' patience. —Possible action: On the product side, move the "unit economics model" to the first page of the BP—specifically, divide the real token bill from the past 30 days by the number of active users to calculate per-user inference cost, then compare it with ARPU; the validation method is to run an A/B test with a 10% price increase or reduced features and observe the slope of the retention curve. If retention is insensitive to price, your product has pricing power, and you can raise prices against the trend when the funding environment tightens. 3. **ChatGPT search query volume reaches 20% of Google's, with 1.2 billion weekly active users** OpenAI confirms ChatGPT search volume — https://aihot.news/items/b6pl29xb245hvm7f8buv4s7l7 Essence: 20% is a psychological threshold—it means AI search is no longer a "novelty entry point" but has begun to materially divert query volume from traditional search, and 42% of U.S. adults have used AI to search for information, with ChatGPT usage at 44% ranking first. For content publishers and SEO practitioners, the rules of traffic allocation are being rewritten: being cited by AI matters more than ranking in search engines. —Possible action: Immediately conduct an "AI citation visibility" audit—ask ChatGPT, Gemini, and Perplexity 20 core business questions respectively, and record whether your brand is cited and which page is cited; then add structured facts (numbers, dates, comparison tables) to pages with low citation rates, because AI search tends to cite verifiable passages dense with facts. Re-run the same set of questions two weeks later to validate the improvement. 4. **SemiAnalysis GDPx benchmark: top agents score only 12% on average on expert tasks** SemiAnalysis GDPx benchmark — https://www.bestblogs.dev/article/c7632fda56?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Essence: Claude Opus 5.5 leads at 12%, and Gemini 4 Carbon at 10.9%. This figure stands in sharp contrast to the narrative that "agents will replace workflows." What it reveals is not that models are unintelligent, but that error accumulation, context loss, and tool-calling reliability in long-horizon tasks remain hard bottlenecks—the bottleneck for agents lies in engineering orchestration rather than point intelligence. —Possible action: Stop evaluating agents with "single-turn demos" and instead build a real task chain of 20+ steps (e.g., "scrape competitor pricing → clean → generate comparison table → write to Notion → send email"), record the success rate of each step, and multiply them to get the end-to-end success rate; you will find the bottleneck concentrated in parsing tool returns at steps 3–5. Prioritize adding structured validation and retries there rather than switching to a larger model. 5. **Prompt caching cuts agent task costs by up to 3.6x; costs are 260% higher without caching** Arena cache cost quantification — https://aihot.news/items/p1pu Essence: Based on quantified data from 26,000 real sessions, this elevates "caching" from an optimization trick to a core variable in the agent business model. Agent tasks are characterized by long system prompts, multi-turn tool returns, and repeated context—cache hit rate directly determines gross margin, which explains why some agent products can price at $4.76/month while others must charge $200. —Possible action: Measure your cache hit rate this week—count the "share of identical prefix tokens" in your agent call logs. If it is below 60%, it means the system prompt or tool schema is being rewritten every turn; fix the stable parts (role, tool definitions, few-shot) as a cacheable prefix and move the dynamic parts later. The validation method is to compare per-task cost before and after cache hits over three consecutive days, with the goal of reducing per-task cost to one-third of the original. 6. **About 1 in 17 agent skills contain instructions that bypass human review, with 70.1% taking effect unconditionally** Agent skill security blind-spot research — https://aihot.news/items/z2bhg9i4eqahaqc7zgktvqk98 Essence: After analyzing 1.53 million SKILL.md files, about 88,000 were found to contain instructions that skip human review, and 70% take effect unconditionally—this means the "skill marketplace" of the agent ecosystem is becoming a new supply-chain attack surface, similar to the malicious package problem in the early npm ecosystem. Combined with Anthropic's confirmation that harmless text can train models to sabotage oversight, the trustworthiness problem of agents has expanded from "model alignment" to "skill supply-chain governance." —Possible action: If you are building or using agent skills, immediately add a static scan—use regex to match patterns such as "skip/ignore/bypass/no confirmation required/execute directly" in SKILL.md, and force manual review of any hits; then add a secondary confirmation hook at runtime for high-risk operations (deletion, payment, sending email). The validation method is to construct a test skill containing a bypass instruction and confirm that your scanner can intercept it and that the runtime hook triggers. 7. **Apple's China-market AI officially partners with Alibaba's Tongyi Qianwen; Apple Intelligence lands in mainland China** iPhone China-market AI announced — https://www.bestblogs.dev/article/31e9e6a48c?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Essence: The division of labor—in-house models as the foundation and Tongyi Qianwen extending to complex tasks—provides a replicable template for the compliant deployment of "foreign models + local models." It also means the battle for AI entry points in the Chinese market is moving from the app layer up to the system layer—Siri becomes a distribution channel, and third-party apps need to rethink product design around "being invoked by system-level AI" rather than "being opened by users." —Possible action: Products targeting iOS should quickly evaluate the integration path for "App Intents / system-level AI invocation," breaking core capabilities into atomic actions that can be invoked by Siri or system AI (such as "check order" or "change address") rather than only building UI entry points; the validation method is to build a minimal Intent extension in TestFlight and observe the system AI trigger rate and conversion rate. If the trigger rate is higher than in-app search, system-level distribution is worth heavy investment. 8. **DeepSeek plans to pretrain V5 on Huawei Ascend; V4.1 Flash tops ARC-AGI cost-effectiveness** DeepSeek Ascend pretraining V5 — https://aihot.news/items/my0qm6kkaa1zw17svsbyn6w43 Essence: Ascend previously handled only inference and post-training. If the V5 pretraining migration materializes, it will be a key validation of domestic compute moving from "usable" to "capable of training frontier models"; meanwhile, V4.1 Flash scores 78.1% on ARC-AGI-2 at a per-task cost of only $0.16, showing that the "low-cost, high-reasoning" path is taking hold. Together, the two threads point to one signal: compute autonomy and cost advantage may be realized simultaneously. —Possible action: For cost-sensitive inference