---

🧠 I. Large AI Models

Developers Share Early Benchmarks of Zhipu's New Model, Suspected to Approach Mythos Level

Benchmark tests circulating on X indicate that the model codenamed Ox Alpha is actually a new product from Zhipu AI's GLM series. Early scores are approaching Mythos level, leading industry speculation that the final name will be GLM-5. This is merely unilateral testing by developers; the official team has not confirmed it, and both naming and release timing remain uncertain. If the benchmark scores hold up, domestic large models will, for the first time, qualify to compete against top-tier overseas closed-source models, requiring a recalculation of pricing and inference cost structures, and rewriting Zhipu's own valuation story. Zhipu's strategy has always been parallel tracks of open source and commercialization. If the new model truly reaches the Mythos tier, the open-source ecosystem gains another competitive flagship option, significantly lowering migration costs for developers. For domestic model vendors, this rumor itself is a signal: the arms race among top players has reached the doorstep of overseas flagships, potentially accelerating the overall pace of model releases in the second half of the year.

• Anthropic Launches Claude Academy Educational Platform — Systematizes AI usage and learning methods, teaching users how to truly leverage models. Anthropic's judgment is direct: the gap between those who can use AI and those who cannot is widening, and education is becoming a new battlefield for model vendors to compete for users. Courses are deeply integrated with their Agent toolchain, allowing immediate application upon completion. The logic behind model vendors entering education is that proficiency drives stickiness; teaching users to use Agents brings them closer to commercial conversion than simply teaching them to chat. For developers, while entry barriers drop, the hidden downside is higher switching costs between models, as educational content itself becomes a layer of invisible customer lock-in.
• Research Observation: Creative Outputs from Large Models Are Converging — Generation results from different vendors are becoming increasingly similar, narrowing differentiation space, with homogenization possibly becoming the industry theme for the next phase. Competition among models is shifting from capability ceilings to cost and engineering. For content creators, choosing which model for creative production is becoming irrelevant. When creative outputs converge, true differentiation lies in who can integrate models into better workflows, further highlighting the value of toolchains.
• Google New Research: Integrating Movement Trajectories into Language Models — Uses mobility data to enhance models' understanding of location concepts, opening new ideas for map and local-life AI applications. The research proves that spatiotemporal information helps models better understand what a place truly means, directly improving location Q&A and navigation interaction experiences. For Google, feeding its accumulated mobility data into models constitutes a second layer of data moat beyond maps.

---

💻 II. AI Software

NVIDIA Releases Switchyard Model Routing, Dynamically Swapping Models Mid-Task

NVIDIA introduced the Switchyard router, capable of dynamically rearranging model combinations mid-task execution. Official self-tests show it compresses task costs to one-third of the original. Model routing turns selecting models by difficulty into a productizable capability, offering a new path for reducing inference-side costs and thickening the stickiness of NVIDIA's software stack. For developers, money-saving tricks that previously required proprietary middleware are now standard, out-of-the-box features. NVIDIA is already treating inference infrastructure like model gateways as a new growth point, moving from selling compute power to selling compute scheduling capabilities, expanding the imagination for its business model. The routing layer is the cash register of the model era; whoever makes this layer the default option controls the accounting rights for inference traffic. Switchyard targets not just its own models but cross-vendor scheduling, effectively holding the routing standards in hand, which poses pressure and serves as a reference for other inference service providers.

• Salesforce and ServiceNow Use Buybacks to Counter AI Panic — Software giants respond to the narrative of AI disrupting subscriptions with real money, calming market concerns about AI eroding software revenue. Both companies emphasize that AI is re-evaluating the value of their products rather than destroying subscription models. The scale of buybacks and stock price reactions have become barometers for gauging sentiment in the software sector. The pricing logic of AI panic is slowing revenue growth; buybacks can only support valuations, not save growth rates, so divergence within the software sector will continue.
• Top YouTube Creators Face Backlash for Taking AI Company Sponsorships — A batch of filmmaking creators lost subscribers after accepting sponsorships from AI firms, sparking debate over the boundary between AI marketing and content integrity. Commercial collaborations between creators and AI companies are becoming sensitive, with fans showing significantly lower tolerance for AI-sponsored content compared to regular ads. For AI firms, advertising channels are narrowing, requiring a reassessment of word-of-mouth marketing strategies.
• Microsoft Task Manager Adds AI Workload Monitoring — Visualizes local compute resource usage, marking the beginning of system-level tools tracking desktop AI infrastructure. For ordinary users, this provides the first intuitive answer to when AI is secretly running and how many resources it consumes. It also offers developers an entry point for troubleshooting performance issues in AI applications. With system-level monitoring added, transparency and observability for local inference will improve significantly.
• OpenAI Responds to Codex Quota Discrepancies — Heads of the Codex and ChatGPT teams publicly responded, stating that most affected users utilized resale/sharing channels like sub2api, triggering risk controls that caused quota anomalies. An official investigation is underway, reminding users to use legitimate subscription channels. Behind the risk control logic is the cleanup of the gray market for API reselling, meaning quota management for developer accounts will become increasingly strict.
• Frontier AI Rewrites Cybersecurity Economics — Cost structures on both offense and defense sides are changing simultaneously, redefining demand for security talent and tool forms. AI lowers attack barriers while greatly enhancing defensive automation levels. Reports suggest cybersecurity will shift from manpower-intensive to model-driven, with industry pricing models changing accordingly. Valuation logic for security vendors is switching from efficiency narratives to model narratives; whoever turns AI defense into a product will capture this revaluation wave.
• Linus Torvalds Uses AI to Debug Intel GPU Drivers — Kernel mailing list logs sparked heated discussion, signaling that AI-assisted kernel debugging is moving from demos to daily practice. If even kernel maintainers are using AI to find bugs, the next step for developer toolchains is clear. Static analysis, log troubleshooting, and patch generation will gradually be handed over to models, with humans only making final decisions. The Linux kernel is one of the toughest codebases to crack; if AI can help there, the demonstration effect on the developer community is significant.

---

🤖 III. Humanoid Robots

• Computer and Phone Repair Technicians Mass Migrating to Robot Repair — Traditional 3C repair businesses are shrinking, driving technicians toward the robot sector. Repair livestreams are becoming new traffic entry points, and zero-experience training businesses are growing simultaneously. As robot ownership rises, after-sales repair is emerging as an undervalued industrial chain, with a window for skill transfer for repair talent opening up. Unlike phone repair, robot repair commands higher average transaction values and more stable repeat purchases, attracting the first wave of pioneers. In the early stages of the industry, what's missing isn't machines, but people who can get hands-on and fix them. The linkage model between training and repair is establishing a template. These roles have low educational requirements but high demands for manual dexterity and learning speed, perfectly matching the profile of veteran repair technicians, resulting in lower-than-expected transition costs.

---

🚗 IV. Autonomous Driving

• US Department of Energy Labs Review Chinese LiDAR — Labs under the US Department of Energy have launched an investigation to assess potential safety risks after large-scale deployment of Chinese LiDAR. If issues are found, it could impact exports of domestic LiDAR to the US and the pace of smart driving solutions going global. Previously, Chinese LiDAR manufacturers have repeatedly addressed safety doubts. This technical review led by DOE labs differs in nature from political bans, but outcomes remain to be seen. Against the backdrop of normalized geopolitical reviews, export strategies for domestic smart driving supply chains need to reserve more margin for compliance. The scope of review expanding from complete units to core sensors indicates regulatory focus is sinking to the component level, requiring exporting companies to re-sort supply chain disclosure boundaries.

---

🌐 V. Physical AI

• Google Releases Biomarker Discovery Framework — A multi-agent system filters candidate biomarkers from wearable sensor data, bringing AI-assisted medical discovery to personal devices. Moving lab research processes to watch and band data represents the most life-relevant sample of AI for Science implementation. This framework uses multi-agent division of labor to handle data cleaning, feature mining, and priority sorting, shortening the distance from sensor signals to clinical hypotheses. The value of wearable data has been consistently overestimated yet underestimated; what's missing is the middle layer translating signals into medical questions, which Google is precisely filling.

---

📈 VI. Macro & Market Data

JPMorgan: Alibaba Cloud Margins Systematically Undervalued

JPMorgan's latest report points out that Alibaba Cloud's current 12% profit margin is undervalued. Capital expenditure has ramped up quickly over past quarters, with massive GPUs and data centers just coming online, still in a 60% utilization ramp-up phase, yielding only ~6% ROIC for the first year of batches. Extrapolating via stacked vintage models, mature-state ROIC should approach 20%, suggesting the market's valuation based on current status may be systematically low. The report's logic is that ramp-up period financial statements miss potential returns from assets already online; once utilization peaks, profit elasticity will release concentratedly. For Alibaba, if this logic is accepted by the market, the revaluation of cloud business will directly lift the overall valuation center, providing a new reference for valuing domestic cloud vendors. The report also notes that Alibaba Cloud's revenue structure is tilting toward AI services, whose gross margins and growth rates differ from traditional IaaS, warranting separate analysis.

Amazon's 7.65GW Texas Power Plant Approved, Potentially Largest Carbon Emission Source in US

Amazon's planned gas-fired power plant for its AI data center in Texas has been approved, with a total installed capacity of 7.65GW composed of 35 gas turbines. Annual emission authorization is approximately 33 million tons of greenhouse gases, potentially making it the largest single carbon emission source in the US. Self-built AI power plants are becoming the new normal, bringing conflicts between compute expansion, grid load, and carbon neutrality goals to the forefront. On one side, exploding electricity demand from data centers; on the other, grid construction cycles can't keep up. Cloud vendors building their own power plants is shifting from isolated cases to industry convention. When electricity becomes a hard constraint on compute expansion, whoever locks in power first locks in the next round of data center orders. For environmental groups and surrounding communities, this authorization implies a commitment to continuous emissions for decades, with legal battles over plant siting just beginning.

• Concerns Over US Capital Scarcity Rising — High debt levels collide head-on with the AI spending frenzy, making fiscal pressure explicit. Markets are beginning to reprice the idea that compute must pay interest, challenging the narrative of AI capital expenditure in a high-interest-rate environment. Analysis suggests that if the central bank rate floor continues to rise, tech giants' hundreds-of-billions-dollar compute investment plans will face refinancing cost pressures. The overlap of the AI cycle and the debt cycle is the macro line to watch most closely in the second half of the year.

-