Key narratives emerging from recent tech giant earnings calls include: computing power demand consistently outstripping supply, with customers securing long-term capacity; LTAs providing dual certainty in volume and pricing, improving visibility into long-term computing needs; new technologies accelerating rather than being delayed, creating fresh incremental demand; and component shortages driving expansion in upstream materials.
Additionally, computing power rental companies are evolving into "new cloud providers," with cloud business growth accelerating. Current AI returns can support higher computing power prices, improving cloud provider ROI. Customer prepayments are becoming a significant source of capital expenditure. Chinese internet giants are also accelerating their capital expenditure, with AI-to-business models already demonstrating monetization potential.
Leading optical communications firms Coherent and Lumentum, computing power rental leaders CoreWeave, Inc. and NEBIUS, and Chinese internet giant Tencent have recently released their latest results and held earnings calls, unveiling several new AI industry narratives. These have been a key support for the recent global AI market recovery.
A summary of the core content from these recent tech giant earnings calls is as follows:
Coherent: High Demand Visibility, CPO/NPO Progress Exceeds Expectations, Indium Phosphide Capacity Doubled Ahead of Schedule
Coherent reported order backlogs, extremely high demand visibility, and no signs of demand attenuation for AI data center optical interconnects. Q4 bookings hit a record, with FY27 fully booked and customers locking CY28 capacity. LTAs extending to 2030 provide dual certainty in volume and pricing. 800G continues to grow, and 1.6T demand intensity is higher than three months ago.
Contrary to market concerns, CPO is not delayed but accelerating. High-power CW lasers are ramping in the Texas plant, contributing revenue from Q2 FY27. NPO is also progressing, with significantly increased customer engagement over the past 3-6 months. Both CPO and NPO projects are increasing, significantly expanding the addressable market for optics. The application scope for OCS has expanded significantly from initial horizontal scaling to include vertical scaling, making previous market size estimates conservative. The company plans to expand capacity and drive faster revenue growth in coming quarters. The only previous capacity bottleneck was indium phosphide, with plans to double internal 6-inch indium phosphide capacity by year-end, one quarter ahead of schedule, and more than double it again by the end of 2027. Yields for 6-inch indium phosphide now exceed those for 3-inch, driving significant gross margin improvement.
Lumentum: Strong 1.6T Demand, Improved CPO Visibility, NPO Provides Full Increment, Accelerating Indium Phosphide Supply Lock-in
Lumentum cited strong 1.6T optical module demand driven by AI cluster deployments at top-tier Tier 1 hyperscale cloud customers. The company has started production of 1.6T optical modules, with adoption expected to accelerate next quarter and continue through 2027. CPO order visibility has improved, with demand for ultra-high-power laser chips expected to ramp in H2 2027 and large-scale deployments by leading customers anticipated in 2028. The company has received its first purchase order for an external light source (ELS) module for delivery in H2 2027. NPO is a completely additive incremental opportunity, not a replacement for CPO. Leading customers' CPO plans remain unchanged, and existing top CPO customers are evaluating NPO for specific new applications, with NPO expected to enter the market between 2027 and 2028. EML supply-demand imbalance is long-term, with a gap remaining over 30%. The company is expanding capacity, planning a 50% year-over-year increase in EML shipments in the December 2026 quarter. To alleviate bottlenecks, the company is actively securing more indium phosphide substrate supply.
CoreWeave, Inc.: Computing Power in Short Supply, Prices Rising Across the Board, Business Expanding to AI Inference, Depreciation Concerns Overblown, Customer Diversification Accelerating
CoreWeave, Inc. reported Q2 revenue up 112% year-over-year, with a backlog of $104 billion, and an additional $25 billion in customer commitments added in early Q3. AI computing power demand continues to outstrip supply, with every new GPU attracting multiple customers. Recent capacity is essentially sold out. The company implemented a roughly 25% price increase across its entire product line in July, with Blackwell and Vera Rubin SKUs achieving record pricing and margins. Older architectures like Ampere and Hopper have not seen significant price declines. New contracts contribute 5-10 percentage points higher gross margins, as customers see AI returns sufficient to support higher computing power prices.
The computing power rental business is expanding from traditional GPU rental to the AI inference market. Managed inference ARR surged from $1 million to over $100 million within months of launch, with a year-end target of at least $250 million. Depreciation concerns are overblown; the company successfully signed a contract for a 2020-era A100 GPU through 2029 at full pricing. Computing power shortages have spread to older cards. While cutting-edge training tasks will continue to chase newer GPUs, inference, evaluation, reinforcement learning, and enterprise tasks do not always require the latest chips. Old card value does not drop to zero upon new card introduction; it shifts from cutting-edge training to inference and cost-sensitive tasks. Customer diversification is accelerating, expanding from large model companies and hyperscalers to industrial, financial, life sciences, quantitative, government, and traditional enterprise sectors, indicating AI adoption is becoming industry-wide.
NEBIUS: Extremely Strong Computing Power Rental Demand, Continuously Improving ROI and Rising Prices, Customer Prepayments Becoming a Source of Capital Expenditure
NEBIUS reported a near 4x quarter-over-quarter increase in total new contract value in Q2, with new customer contracts growing over 9x. Demand carries strong pricing power; under current contract terms, the company could completely sell out its entire planned 2027 capacity. ARR reached $3 billion by end of June, up from $1.9 billion at end of March. Computing power prices are being continuously raised; the company's first Blackwell capacity auction yielded a price 15% higher than the previous highest transaction and 20% above the Blackwell sales pipeline quote. Customer prepayments are becoming a source of capital expenditure, with approximately 70% of contracts signed this quarter involving customer prepayments, covering 50-60% of related CapEx. The company aims to increase this ratio further. The expected project payback period has shortened from the previously anticipated 2-3 years to 1 year and 10 months. The next-generation Rubin architecture is already undergoing laboratory validation, with initial results meeting expectations. Management noted the technical difficulty of migrating from Grace Blackwell to Rubin is less than the previous migration to Grace Blackwell. Deployment is planned from late 2026 or early 2027, continuing through 2027.
Tencent: Chinese Internet Giants Accelerating CapEx, AI-to-Business Already Monetizable, Relying on Product Strength to Recover Investment
Tencent's Q2 capital expenditure of 52.78 billion RMB significantly exceeded market expectations (65% higher than the 32.1 billion estimate), representing a 176% year-over-year and 65% quarter-over-quarter increase. The company has clearly stated it will continue to increase computing power procurement in Q3, with full-year capital expenditure expectations further revised upward. AI computing power demand from Chinese internet giants remains in an accelerating phase. Tencent is advancing its AI strategy through a three-layer architecture: the intelligence layer (the Hy3 large model, prioritizing cost-effectiveness and in-ecosystem scenarios over general model benchmarks), the application layer (WorkBuddy for AI office tools and CodeBuddy for AI programming, focusing on B-end productivity), and the infrastructure layer (proactively increasing computing power procurement to convert model and application usage into actual revenue, with the WeChat ecosystem's internal AI agent as a core differentiator).
Regarding AI returns, given the current supply-demand dynamics and pricing in the computing power rental market, Tencent could immediately rent its purchased computing power to third parties, covering depreciation and achieving decent returns. However, the company has chosen a different path, investing most new computing power into its own research models and proprietary AI applications, prioritizing leading model capabilities and market-leading applications to generate stronger long-term economic returns, such as selling tokens through WorkBuddy. WorkBuddy is already the leading enterprise AI agent in China by DAU, demonstrating strong user willingness to pay for subscriptions and tokens. China's AI-to-business sector already possesses commercial monetization capabilities. Due to lower token production costs in China, the token business can already achieve positive gross margins at current prices. WorkBuddy's paid user gross margin and Model-as-a-Service (MaaS) gross margin are now comparable to Tencent Cloud's overall gross margin. WorkBuddy is essentially a new platform, envisioned as a very flexible Agentic AI workspace. Its target users include not only large enterprise employees but also regular office workers, freelancers, entrepreneurs, and single-person companies. The platform will eventually host many models, skills, and developers.
Comments