---

I. Large AI Models

Momenta's first semi-annual report since IPO: Revenue up 75.9%, adjusted net loss narrowed to RMB 14.097 million

H1 2026 revenue reached RMB 1.602 billion, a year-on-year increase of 75.9%, with a gross margin of 73.2%. After excluding non-operating factors such as fair value changes in preferred shares and share-based payments, the adjusted net loss narrowed to RMB 14.097 million, approaching break-even. More important than the growth itself is the signal: an autonomous driving company long labeled with "high R&D investment, high delivery costs" has proven for the first time with auditable financial data that its mass-production model can turn positive.

A 73.2% gross margin is close to software company levels, directly chipping away at the market consensus that "the Robotaxi story can only survive via financing." In Momenta's revenue structure, licensing fees and mass-production services each account for half, and economies of scale are beginning to cover R&D expenses. This serves as a template for algorithm companies moving toward a software P&L statement.

This creates valuation benchmarking pressure for already-listed smart driving supply chain companies like Horizon Robotics and RoboSense. The market is beginning to reprice autonomous driving companies based on unit economics rather than the number of design wins. High-R&D targets that are not yet profitable and rely on hardware deliveries to stack revenue will be scrutinized quarter by quarter under a magnifying glass. What Momenta secured is not just a break-even figure, but the pricing standard for the entire sector.

Tencent Hunyuan Hy4 preview queues immediately upon launch; compute expansion cannot keep up with concurrency

Released late at night on August 28, the Hy4 preview (770 billion total parameters, 49 billion active, 1 million token context, fully open-sourced under Apache 2.0) saw its WorkBuddy inference cluster overwhelmed by requests within just 3 days. Officials urgently expanded capacity and promised continuous dynamic resource allocation, but still did not rule out continued queuing during peak hours. This release without a press conference made headlines thanks to the queues.

Queues upon release are an extreme signal of the authenticity of demand for domestic open-source models. Free and open-source spending zero marketing dollars yet instantly maxing out high-end compute indicates that enterprise-level API call demand is far from satisfied. The other side of the coin is Tencent's insufficient total high-end compute and concentrated peak concurrency. Looking at this alongside Zhipu's announcement of the "Large Foundation" route on the same day, the conclusion is clear: the domestic model competition has officially entered the second half, where compute reserves are key. Only those holding GPUs dare to open-source and lower prices.

Direct beneficiaries are the domestic compute chain on the inference side and cloud service providers with their own compute pools. For small and medium-sized model companies relying on third-party compute, this is a warning bell on the cost side—in the era of free open-source, inference bills will only get more expensive. When raising funds, it is best to include GPU budgets in the business plan upfront.

• Zhipu discloses next-generation large model roadmap, confirming the Large Foundation direction. The goal is to support full-scale business calls on day one of release. Specific parameters are not disclosed for now, with subsequent work focusing on mid-training and post-training around the foundation. The core idea is to fully unleash foundation capabilities rather than stacking parameters to climb leaderboards. Model competition is shifting from leaderboard chasing to engineering delivery, with "paralysis upon launch" being treated as a product accident that needs to be eliminated in advance. Combined with its semi-annual report showing revenue growth but persistent losses, the Large Foundation route effectively bets the next round on compute reserves, making fundraising progress a critical variable.
• MiniMax submits its last interim report before issuing an AGI ultimatum; Zhipu reviews its interim results on the same day, and Moonshot AI is also paving the way for Kimi's IPO. In just a few days, the commercialization narratives of three independent large model companies are being compared on the same financial coordinate system. This is the first time domestically that there is a sample of pure AI companies available for horizontal comparison. API call volume and ARR quality are the sole basis for pricing. The scissors gap between fundraising valuations and real revenue will be exposed under interim reporting standards. MiniMax submitted first, thereby defining the anchor for market expectations for the entire sector.

---

II. AI Software

• Clipto valued at $250 million, a three-year-old company using AI for natural language retrieval in TB-level video libraries. Clients cover media and brand content teams. Generative content makes production easier but searching harder; as enterprises accumulate more assets, Adobe and Apple are eyeing this space. Media asset management has become a new necessity. This track has produced its first independent unicorn, indicating that "organizing data after AI" is becoming a standalone business.
• Blue Voice raises $6 million. David Lawrence, a Harvard Law School dropout who left due to campus shooting incidents and police compliance controversies, started the company to build a "Police Version of Harvey," focusing on law enforcement documentation and compliance review. Vertical legal AI is accelerating segmentation by profession. From law firms to policing, insurance claims, and HR, the pricing power of legal AI is sinking from general tools to job-specific scenarios. Job knowledge barriers are the new moat.
• Datadog reveals saving over $1 million monthly by optimizing AI usage. Methods include routing different tasks to models of different sizes, caching high-frequency requests, and cutting low-value calls. Engineering teams are putting the FinOps knife to AI bills for the first time. Internal enterprises are starting to view AI spend management as a profit source. This is the first major corporate proof of the shift in the AI cost-side narrative.
• Newell states AI cut marketing costs by 80%. The consumer goods giant follows the cost-cutting narrative. Note that the 80% figure likely refers to the marketing material production phase rather than total marketing expenses. However, for brand CFOs, the collapse of outsourced content production budgets is already impacting the P&L. AI's impact on traditional industries is appearing systematically in earnings calls.
• India's voice AI price war drops to 0.15 cents per minute. Local industry veterans publicly warn that the track is sliding into a race to the bottom. India is the fastest-growing market globally in terms of call volume, but also the thinnest-margin market. For Chinese voice AI companies preparing to expand into Southeast Asia, this is a ready-made cautionary tale. Scaling without pricing power is just turning losses into an assembly line.

---

III. Humanoid Robots

• Luming explores scalable implementation solutions for embodied intelligence. Leiphone dissects its path from demos to production line delivery: first lock onto a single industrial scenario to go deep, then replicate horizontally. This year, the financing logic for embodied intelligence has changed. Investors no longer pay for backflips. Companies that can clearly articulate delivery paths and single-unit economic models are starting to secure pricing for the next round.
• Perceptron AI releases Isaac 0.5, a 36 billion parameter open-weight embodied foundation model, focusing on general policy transfer for robot manipulation. Following large language models, the embodied track is skipping the closed-source stage and jumping straight into open-source ecosystem battles. The fight for standard-setting authority has begun early.

---

IV. Autonomous Driving

• Saudi HUMAIN partners with Applied Intuition to develop autonomous trucks. Sovereign AI funds are betting their export strategy on commercial heavy truck scenarios. Middle Eastern capital is starting to look for a second curve for autonomous driving beyond passenger vehicles: trunk logistics routes are fixed, shipper willingness to pay is clear, and policy resistance is much smaller than for Robotaxis. Additionally, Applied Intuition's simulation platform helps HUMAIN skip building its own test tracks. This partnership also shows that HUMAIN's ambition goes beyond selling compute; it wants to extend sovereign cloud to the vertical application layer.

---

V. Physical AI & Compute Infrastructure

• Don't just watch NVIDIA; utility stocks have independent upside logic. With PJM grid electricity auctions approaching, analysts believe the US midterm elections could serve as a clearing event. After the elections, developers and utilities in both red and blue states may be more willing to announce data center deals. The next catalyst for the AI power chain is on the grid side, not the chip side. Moreover, utility stock valuations do not include the crowded premium seen in NVIDIA-style stocks.
• Huawei's profit decline widens. Rising storage prices and supply tightness drag down profitability. Even with H1 revenue growth, margins were not maintained. Cost pressures in the hardware chain are transmitting downstream along the industry chain: DRAM/NAND contract prices continue to rise, and the storage supercycle is squeezing terminal manufacturer gross margins. Smartphones, servers, and automotive electronics are all affected. Conversely, this is a window of simultaneous volume and price increases for domestic storage chain companies.
• Advanced packaging is the true AI moat. Independent analysis points out that market attention is overly focused on GPU designers. The scarcity of CoWoS-like advanced packaging capacity faces revaluation. There are increasingly more competitors in the design phase, but production lines capable of achieving yield targets when packaging multiple dies together are rare worldwide. The compute bottleneck is shifting from chips down to the packaging stage. Foundry and OSAT players' influence and pricing power in the industry chain are rising. This is a layer in the AI hardware rally that has not been fully priced in.

---

VI. Macro & Market Data

BoE Governor Bailey warns next-gen AI models may trigger disorderly adjustment in global financial markets

Bailey publicly stated that threats posed by frontier AI models are deepening and could cause a disorderly correction in financial markets. Almost simultaneously, Reuters reported that central bank governors are collectively discussing the impact of AI agents on financial markets. AI enters the central bank agenda for the first time as a source of financial stability risk, rather than as a strategic industry.

This is a key turning point in regulatory narrative. Previously, AI was a supported entity in policy discussions; now it is being treated as a systemic risk exposure. Once "AI bubble + AI stability risk" becomes a joint statement by central banks, allocation discipline for sovereign funds regarding the AI sector will tighten. Volatility amplifiers like agent trading may also attract targeted regulatory rules.

Valuation-wise, this directly points to the risk premium of high-valuation AI leaders like NVIDIA and Palantir, which have seen massive YTD gains. On the flip side, AI safety, model auditing, and trade regulation tech companies may become hedge beneficiaries of policy. Government procurement budgets in this niche are worth tracking.

• China publicly criticizes Anthropic and sets preconditions for key China-US AI dialogue, placing export controls and model access disputes directly on the bilateral agenda. Model games enter the realm of regulatory diplomacy. Compliance costs for top labs are becoming geopolitically localized. Valuation discounts for China-related business will reflect in any subsequent fundraising pricing.
• Morgan Stanley says AI adoption finally reflects in corporate fundamentals. Its head of US thematic research judges that direct contributions from AI can now be stripped out of corporate earnings. Skepticism about the "productivity paradox" is receding in stages. The weight of this statement lies in coming from the biggest AI bull sell-side firm. Stock picking logic for the second half shifts to "who is actually making money using AI." Narrative premium gives way to delivery premium. Pure concept stocks unrelated to the application layer will be abandoned by capital first.
• Australian Treasury warns commercial adoption is too slow, risking missed AI economic dividends. It lists insufficient enterprise AI penetration as a national economic risk. This mirrors the Bank of England's financial stability warning: the same technology sees Southern Hemisphere governments calculating the opportunity cost of missing dividends, while Northern Hemisphere governments calculate the cost of systemic risk. Both sides have valid points and interests. Global regulators have not yet formed a unified coordinate system for AI, so cross-border compliance arbitrage windows will persist for some time.
• LG Energy Solution signs lithium supply agreement with Smackover in Arkansas. Another win for North American battery supply chain localization. Lithium prices remain in the bottom range, so locking in upstream resources now costs far less than in the previous cycle. The local content requirements of the US Inflation Reduction Act give "Made in USA Lithium" an inherent policy premium. Resource lock-in cards remain effective during price volatility cycles. This is another supply security approach besides Tesla's self-produced 4680 cells.
• Bolt founder Breslow raises up to $27 million in bridge funding for self-rescue, conditioned on pay-to-play: shareholders who don't participate will be diluted out. The cash survival game for the ride-hailing unicorn enters its final stage. The size of the bridge funding is a drop in the bucket relative to its burn rate. Willingness to accept such terms is itself a danger signal. The next round will either be a discounted acquisition or liquidation.

Closing Quotes (8/31): A-share AI chain showed a rotation towards strong domestic compute while optical modules took a breather. Cambricon +6.26% closed at RMB 1112.18, hitting a new historical high. Hygon Information +3.43%, Kingsoft Office +3.46%, Foxconn Industrial Internet +1.25%, and Innolight +0.72% closed higher. Zhongji InnoLight -0.75% saw short-term consolidation. Capital diffused from overseas chains to domestic autonomous chains and the AI application layer. Overnight US AI leaders diverged significantly. NVIDIA +1.48% (added $3.5 billion stake in MediaTek to solidify chip alliance), Tesla +5.51% (FSD pothole avoidance feature launching soon). Meta -0.98%, Alphabet -2.09%, Microsoft -1.22%, and Amazon -2.50% all closed lower. Capital rebalanced along the line of strong compute hardware vs. weak software platforms.

---