Societe Generale Securities: What new AI narratives will the tech giants release in the near future?
The Zhitong Finance App learned that Societe Generale Securities released a research report saying that the recent narrative conveyed by the tech giant Earnings Call is: 1. Demand for computing power continues to exceed supply, and customers are already locking in long-term production capacity; LTA provides double certainty in quantity and price, increasing the visibility of long-term computing power demand; new technology is not being delayed; instead, it is being advanced and accelerated to provide a definitive total increase in production; insufficient supply of materials will drive the continuous expansion of upstream materials production.
2. Computing power leasing companies have been promoted to “new cloud factories”, and cloud business growth continues to accelerate; current AI returns can support higher computing power prices and drive continuous improvement in cloud factory ROI; customer advance payments are becoming an important source of capital expenditure.
3. The capital expenditure of major domestic Internet companies continues to accelerate, and AI ToB already has the ability to monetize.
Societe Generale Securities's main views are as follows:
Recently, leading US optical communications (Coherent, Lumentum), leading computing power leasing (CoreWeave (CRWV.US), Nebius (NBIS.US)), and domestic internet giant Tencent (00700), announced their latest results and held a conference call to release many new narratives in the AI industry, which has become an important support for the recent restoration of the global AI market.
Overall, the story conveyed by the tech giant Earnings Call recently was:
1. Demand for computing power continues to exceed supply, and customers are already locking in long-term production capacity; LTA provides double certainty in quantity and price, and the visibility of long-term computing power demand has increased; new technology has not been delayed; instead, it is being advanced and accelerated to provide a full increase in deterministic volume; short supply of materials will drive the continuous expansion of production of upstream materials.
2. Computing power leasing companies have been promoted to “new cloud factories”, and cloud business growth continues to accelerate; current AI returns can support higher computing power prices and drive continuous improvement in cloud factory ROI; customer advance payments are becoming an important source of capital expenditure.
3. The capital expenditure of major domestic Internet companies continues to accelerate, and AI ToB already has the ability to monetize.
Summarize the highlights of recent tech giants' performance and earnings conference calls for reference:
Coherent: Demand visibility is extremely high, CPO/NPO progress exceeds expectations, indium phosphide production capacity doubled ahead of schedule
1. Orders are full and demand visibility is extremely high: Demand for optical interconnection in AI data centers continues to explode, and demand shows no signs of abating. The fourth quarter reservation volume set another record. FY27 has been fully booked throughout the year, customers have begun to lock in CY28 production capacity, and LTA Changxie will extend to 2030 to provide both certainty in quantity and price; 800G continues to grow, and demand intensity for 1.6T has further increased compared to three months ago.
2. The CPO was not postponed; instead, it was accelerating and moving forward: the market was previously worried about CPO delays, but did not see any delays in CPO demand; instead, it accelerated and moved forward (pull-in). Ultra high power CW lasers have climbed the slope at the Texas plant, and Q2 FY27 (December quarter) began contributing revenue.
3. NPOs are also progressing at the same time: in the past 3-6 months, customer participation has increased significantly. Not only CPO, but NPO projects have also increased dramatically, and the market space for “Light” will increase significantly in the future. NPO and CPO are only different form factors. Since they provide a wide range of solutions, the amount of content is comparable, and they support any type according to customer preferences.
4. The scope of application of OCS has expanded significantly: initially it was mainly aimed at horizontal expansion, but now we can clearly see the adoption path of horizontal expansion and vertical expansion, making the previously predicted market size conservative; Q4 revenue has grown, and will continue to expand production capacity and drive faster revenue growth in the next few quarters, and will become a product line contributing considerable revenue.
5. Indium phosphide production capacity expansion: Previously, the only production bottleneck was indium phosphide, which will expand later. The company plans to double the internal 6-inch indium phosphide production capacity by the end of the year and achieve it one quarter ahead of schedule, and more than double by the end of 2027; the yield of 6-inch indium phosphide has already exceeded 3 inches, driving a significant improvement in gross margin.
Lumentum: Strong demand for 1.6T, increased CPO visibility, NPO provides all new volume, and accelerates the locking of indium phosphide supply
1. Strong demand for 1.6T optical modules: Strong deployment of AI clusters from leading Tier 1 hyperscale cloud customers is driving the rapid transition from 800G to 1.6T. The company has already started mass production of 1.6T optical modules, and the adoption rate is expected to accelerate in the next quarter and continue until 2027.
2. Increased visibility of CPO orders: Demand for ultra-high power laser chips is expected to begin in the second half of 2027, and large-scale deployment for leading customers is expected to begin in 2028. The company has received a purchase order for the first external light source (ELS) module, which will be delivered in the second half of 2027.
3. NPO is a completely additional incremental opportunity: NPO is not a replacement for CPO, but a new intermediate structure. The CPO plans of leading customers remain unchanged. Major existing major CPO customers are evaluating NPO solutions for specific new application scenarios, and NPOs are expected to enter the market from 2027 to 2028.
4. Long-term imbalance between EML supply and demand, acceleration of indium phosphide substrate supply: 100G/200G EML (electric absorption modulation laser) shipments reached a record high; EML supply and demand imbalance, maintaining a gap of more than 30%; production capacity continues to expand, and EML shipments are planned to increase 50% year-on-year in the December quarter of 2026; in order to ease bottlenecks, the company is stepping up the supply of more indium phosphide substrates.
Coreweave: Computing power is in short supply and prices are rising across the board; business expansion into the AI inference market; no need to worry too much about depreciation; customer diversification is accelerated
1. The shortage of computing power continues, and the computing power leasing business is in short supply: Q2 revenue increased 112% year over year, revenue backlog (backlog) reached 104 billion US dollars, and customer commitments of more than 25 billion US dollars were added at the beginning of Q3. Demand for AI computing power continues to exceed supply. Every newly launched GPU has multiple customers competing for it, and production capacity has actually been sold out recently.
2. Product price increases across the board, and AI returns can support higher computing power prices: The company raised prices for the full range of products by about 25% in July, the prices and profit margins of Blackwell and Vera Rubin SKUs reached new highs, and the prices of old architectures such as Ampere and Hopper did not drop significantly; the new contract contributed to an increase in profit margins by 5-10 percentage points, and the AI return seen by customers was enough to support higher computing power prices.
3. The computing power leasing business is expanding from traditional GPU leasing to the AI inference market: ARR, the managed inference business, surged from $1 million to over 100 million US dollars in just a few months after launch, and is expected to reach at least 250 million US dollars by the end of the year.
4. The value of the old card has not returned to zero, and there is no need to worry too much about depreciation: the company has successfully signed a 2020 A100 GPU with a full pricing contract until 2029, and the shortage of computing power has spread to old cards; cutting-edge training tasks will continue to chase next-generation GPUs, but reasoning, evaluation, reinforcement learning, and enterprise tasks do not necessarily require the most advanced chips. Once the new card is launched, the value of the old card will not quickly return to zero; it only shifts from cutting-edge training to reasoning and cost-sensitive tasks.
5. Accelerating customer diversification: Customers are expanding from big model companies and hyperscalers to industrial, financial, life sciences, quantitative, government, and traditional enterprises, and AI is being adopted throughout the industry.
Nebius: Demand for computing power leasing is extremely strong, ROI continues to improve, computing power prices continue to rise, and customer advance payments are becoming a source of capital expenditure
1. Demand for computing power leasing is extremely strong, and ROI continues to improve: the total value of new contracts signed by the company in the second quarter was nearly 4 times month-on-month, and new customer contracts increased more than 9 times; demand has strong pricing power. According to current contract terms, the company can immediately sell out all planned production capacity for 2027; ARR reached 3 billion US dollars at the end of June, a significant increase from 1.9 billion US dollars at the end of March.
2. The price of computing power continues to rise: The transaction price of the company's first Blackwell production capacity auction was 15% higher than the highest transaction price in the past, and 20% higher than the price in the Blackwell sales pipeline.
3. Customer prepayments are becoming a source of capital expenditure, and the payback period has been drastically shortened: customers have begun to make advance payments for AI infrastructure, and about 70% of the contracts signed this quarter are customer advance payments, which can cover 50-60% of the relevant capex. The company hopes to continue to increase this ratio in the future. The expected payback period of the project was shortened from the 2-3 years previously anticipated to 1 year and 10 months.
4. Rubin has begun experimental verification: Management said that the next-generation architecture Rubin has been verified in the laboratory, and the initial results are in line with expectations. The technical difficulty of migrating from Grace Blackwell to Rubin is less difficult than the previous generation to Grace Blackwell. Deployment is scheduled to begin in late 2026 or early 2027 and continue throughout 2027.
Tencent: Capital expenditure of major domestic Internet companies continues to accelerate, AI ToB already has monetization capabilities, and relies on product power to recover investment
1. Capital expenditure greatly exceeded expectations, clearly increasing computing power procurement in the future: 26Q2 capital expenditure of 52.78 billion yuan surpassed market expectations (65% higher than the estimated 32.1 billion yuan), +176% year-on-year and +65% month-on-month. It is clear that Q3 will continue to increase procurement of computing power, and capital expenditure expectations for the whole year will be further raised. Demand for AI computing power from major domestic Internet companies is still accelerating.
2. Three-tier architecture promotes AI Tencent: ① Intelligent layer: Hy3, the big model, focuses on high cost performance, does not blindly pursue common model scores, and prioritizes implementation scenarios; ② Application layer: WorkBuddy (AI office), CodeBuddy (AI programming), focusing on B-side productivity tools; ③ Infrastructure layer: actively increasing computing power procurement to turn model and application usage into actual revenue. AI agents within the WeChat ecosystem are a core differentiating advantage.
3. Regarding AI returns: Considering the current supply, demand and pricing in the computing power rental market, the company can actually rent the purchased computing power to a third party immediately, immediately cover depreciation and get a good return — just like many emerging cloud vendors are doing. However, the company chose a different path: investing most of the added computing power into self-developed models and its own AI applications, first bringing the model to the leading level, putting the application first in the market, and then exchanging long-term economic returns through stronger intelligence, such as selling tokens through WorkBuddy.
4. About WorkBuddy: In terms of DAU, it is already the number one enterprise AI agent in China, and users have shown a strong willingness to pay for subscriptions and token purchases — China's TOB AI business already has commercial monetization capabilities; thanks to lower domestic token production costs, even at current prices, the token business can achieve positive gross profit, and the gross profit of WorkBuddy paying users and the gross profit of Model-as-a-Service (MaaS) is now comparable to Tencent Cloud's overall gross profit level; WorkBuddy is essentially a new platform, and the company wants it to be a very flexible Agentic AI workspace. The target users are not only employees of large enterprises, but also ordinary office workers, freelancers, entrepreneurs, and one-person companies. In the future, WorkBuddy will have many models + many skills + many developers.
Risk Alerts
Only public data collation, not research opinions or investment suggestions