---

Today's main theme is AI moving from narrative to accounting. The enterprise software sector proved itself as the most direct monetization outlet for AI demand with two earnings reports; Salesforce and Okta surged 14% and 15% respectively after hours. On the domestic LLM front, MiniMax's H1 revenue grew 283% YoY, with losses narrowing simultaneously, proving commercialization is no longer just a slogan. Amidst the hype, there's cold water: the US data center boom is facing community backlash, with nearly half of planned projects potentially delayed, recalibrating the pace of compute expansion. The residual heat from US earnings week also transmitted to A-share compute chains; Cambricon stood firm above the 1000 yuan mark, while optical module leaders saw mixed movements. Below is the full version of 23 source articles.

---

🧠 I. AI Large Models

MiniMax H1 revenue grows 283% YoY, losses narrow 11% — Total operating revenue for H1 FY2026 was $117 million, approx. 785 million RMB, up 283.1% YoY. Net loss attributable to parent was $358 million, narrowing 11% YoY. Sales nearly quadrupled in China's AI race. Management disclosed during the earnings call that M3 Pro parameter scale is expected to increase to approx. 3T, and adaptation for domestic chips for M3 and H3 is underway, with large-scale domestic compute clusters coming online soon. Most domestic LLM companies are still burning cash for market share; MiniMax's revenue slope proves model capability can convert into real money. Market pricing for Chinese AI assets is shifting from comparing parameters to comparing revenue. This growth curve provides a new reference for peers in the primary market and puts the question of 'who can crack commercialization first' on the table.

• MINIMAX H1 earnings supplement details — Net loss attributable to parent was $358 million, narrowing 11% YoY. Management emphasized during the earnings call that reinforcement learning and long-range task training will determine the upper limit of model intelligence; next-gen products continue to bet on these two technical lines. Loss narrowing appearing simultaneously with revenue ramp-up indicates the first phase of burning cash for scale is entering an efficiency validation period. If this slope is maintained in H2, annual revenue has room to step up. For peers, MiniMax's loss curve is also a reference; the match between burn rate and revenue scale will be increasingly scrutinized by the primary market.
• OpenAI bans batch of Russia-linked ChatGPT accounts — Investigations show Moscow is using AI to generate influence content globally in bulk, turning ChatGPT into a political ghostwriter. OpenAI stated it identified and banned relevant accounts, with limited details disclosed. Generative AI drives the marginal cost of content production near zero, adding another empirical case to the entanglement of AI and information warfare. Platform governance pressure rises simultaneously. Generative AI compresses the marginal cost of content production to near zero, erasing personnel barriers in information battlefields. Whether more investigation details will be disclosed later is worth watching.
• European new open-source AI lab Sevren established — Focused on open-weight model technology R&D, adding another player to Europe's open-source ecosystem. Specific model paths and team composition are undisclosed; placeholder first, reveal later. Europe's intent to secure a seat in open weights is obvious. The debate between open-source and closed-source technical paths is taking on geopolitical colors. The supply side of open weights is multipolarizing, with players in the US, China, and Europe.

💻 II. AI Software

Salesforce beats earnings expectations, surges 14% after hours — Quarterly performance and guidance both exceeded Wall Street expectations; stock price surged up to 14% after hours. AI products like Agentforce drove growth, equity holdings in Anthropic contributed substantial investment income, and the company released Claudeforce, an AI entry point for sales scenarios, directly addressing concerns that SaaS would be replaced by AI. This earnings report shows AI hasn't eaten software but has become a new growth engine. For the software sector, guidance from top companies alleviated the disruption narrative from LLMs, allowing valuation logic to breathe easier. Optimistic expectations for AI monetization are spreading to traditional software assets; Agentforce penetration data will be a validation point for the next earnings report. For investors, the biggest signal from this earnings season is the authenticity of AI budgets, verified for the first time in software financial statements.

Okta beats earnings expectations, rises 15% after hours — Identity security company Okta beat Q2 expectations and raised full-year guidance; stock rose approx. 15% after hours. Management cited a surge in AI-driven attack methods, with enterprises accelerating identity security system upgrades. The double-edged sword effect of AI manifests in cybersecurity: generative AI lowers the cost of phishing and account attacks, making defensive businesses better off. CrowdStrike previously reported record quarterly numbers, corroborating the prosperity of the security sector in earnings. After AI threats became normalized, the rigidity of identity and endpoint security budgets is being repriced by the market, significantly improving visibility for security companies' performance. The bill for identity security is growing thicker; penetration rates for multi-factor authentication and passwordless login are in rapid ramp-up phases.

• Open-source team replicates OpenAI Codex Micro for $40 — Developer built AgentPad13 from scratch, completing tasks like PCB routing for $40, verifying that open-source paths can also make agents cheap. The price gap between commercial closed-source products and community replicas is becoming the growth space for open-source agents, putting a straitjacket on closed-source vendors' pricing power. Such replication projects remind the market that the balance between closed-source pricing and open-source ecosystems can be broken at any time.
• Tsinghua University and HKUST propose real-time video editing framework LiveEdit — ECCV 2026 paper achieves 12.66 FPS real-time streaming editing with 4-step denoising, pushing video editing from offline cutting to online scenarios like live effects and video conferences. Breaking real-time constraints means the next battlefield for AIGC video tools is latency, not quality; end-to-end latency becomes a new competitive metric. For creators, once shoot-and-edit workflows run smoothly, the short video production chain will shorten further.
• Dataset of AI agent security incidents open-sourced — 1000 incidents classified, covering jailbreaks, data breaches, permission abuse, etc., transforming security risks from anecdotal rumors into quantifiable research subjects. Before large-scale deployment of agents, the sample library of security incidents is the industry's experience bank. Security research finally has a public material pool, giving improvements on both offense and defense sides a reference. Jailbreaks, data breaches, and permission abuse are three high-frequency incident types; protection priorities can be front-loaded accordingly.

🤖 III. Humanoid Robots

• Open-source humanoid robot football league launched — Unitree G1 battles daily in realistic physics simulations, with frontier AI models acting as club managers. Match results and replays are fully public. Testing embodied intelligence decision-making through competitive formats is more intuitive than any benchmark test. 'Sports events' for humanoid robots are becoming algorithm testing grounds, accumulating massive adversarial data for the open-source community. Adversarial environments are the fastest soil for algorithm iteration; the value of such public leagues will accumulate over time.

🚗 IV. Autonomous Driving

• BYD Executive: Cars are getting bigger, small-size models will disappear — Executive VP He Zhiqi stated that the proportion of cars with width ≥1.9 meters rose from 10% in 2018 to 40% in Jan-Jul this year. In a few years, cars with wheelbases under 2.8 meters will vanish. Behind the size race is the engineering logic of batteries, chassis, and cabin space. Survival space for small cars is being compressed, serving as a wind vane for automaker product planning; the micro-car market will accelerate shrinking. Size growth raises per-unit value and battery usage, benefiting the supply chain, but pressure on battery swap and charging infrastructure increases simultaneously.
• HarmonyOS Intelligent Mobility Shangjie Z7 Allroad model first exposed — The Shangjie brand plans to launch an Allroad new car belonging to the Z7 series, expected to officially appear in H2 this year. HarmonyOS Intelligent Mobility continues to complete its model matrix beyond sedans and SUVs; Allroad positioning targets multi-scenario mobility. As product lines widen, HarmonyOS Intelligent Mobility's multi-brand matrix strategy is forming (multiple product lines co-built by Huawei and automakers advancing in parallel), and the installed base of Huawei's intelligent driving solutions is expanding. Shangjie previously laid out sedans and SUVs; Allroad fills the crossover mobility gap.

🌐 V. Physical AI

• AI hearing aid glasses startup Legato secures $12 million funding — Approx. 50 million adults in the US suffer from hearing loss, but only about 20% seek treatment. This company aims to lower the barrier with AI hearing aid glasses, embedding voice enhancement and real-time subtitles into glasses. Hearing aids are transitioning from medical devices to consumer electronics, with AI audio processing becoming the new product watershed. Hearing aid penetration has been stuck between 'visibly conspicuous' and 'expensive'; AI glasses aim to untie both knots simultaneously.

📈 VI. Macro & Market Data

US data center boom faces community backlash — New projects are slowing and being reshaped, with increasing opposition from local governments and communities. Construction cycles and site selection logic are changing. Electricity, water usage, and noise have become core community disputes; some projects are forced to relocate or shrink. Project developers are starting to include community communication in preliminary planning. Compute expansion bottlenecks are shifting from chips to electricity and public opinion. Data center site selection is becoming a social issue, not just a technical one, introducing new variables to the delivery rhythm of the compute supply chain.

• Kimmeridge: At most half of planned US data centers will be delayed — Power supply and approval bottlenecks are the biggest constraints on compute expansion; funding and site selection are queuing again. Existing projects that secured power early are actually more valuable. Differentiation within compute assets has begun; secondary market related stocks need to re-examine project landing timelines. For compute leasing and cloud vendors, grid connection time of projects under construction is the starting point of the revenue curve; delays directly rewrite long-term valuations.
• Ohio to welcome 'World's Largest Data Center' — Joint bet by OpenAI, Nvidia, and Japanese capital. Project located in Appalachian foothills; local residents have mixed hopes and concerns. Employment and taxes are hopes; electricity and water resources are concerns. Whether large compute projects can pass the community gate is a key variable for landing rhythm. Similar cases will become more common; community relations are becoming an implicit approval item for data center site selection.
• US candidates sign data center and AI safety pledges — With new elections approaching, multiple candidates included data center siting and AI safety in campaign pledges. Infrastructure footprint, power consumption, and job displacement are becoming ballot issues, increasing uncertainty in compute policy. Industry lobbying will also become more active.
• Huawei pitches building AI data center to Egypt — Huawei proposed constructing for Egypt