1. US Four Major Cloud Providers Post Strong Q1 Earnings, Capex Shifts to AI Inference
In late April, four major North American cloud providers (Google, Microsoft, Amazon, Meta) reported quarterly earnings far exceeding expectations, with capital expenditure shifting from model training to inference applications. Amazon's custom chip business has surpassed $20 billion in annualized revenue, ranking among the top three data center chip providers globally. Its Trainium series chips are in high demand, with over $225 billion in revenue commitments already secured. Google and Microsoft also reported record cloud revenue and capex plans. These robust results dispel widespread concerns about an AI bubble and confirm that inference computing demand is becoming a new growth engine.
2. NVIDIA Showcases Next-Gen Chips and Ecosystem at GTC 2026, Aiming to Drastically Cut Computing Costs
At its recent GTC conference, NVIDIA announced several key advancements, focusing on significantly reducing AI computing costs. The next-generation Vera Rubin chip is scheduled for mass production, reportedly cutting the number of chips needed to run the same AI model to one-fourth, potentially lowering inference costs by 90%. The conference also demonstrated technology integration following the acquisition of Groq, with its LPU chips working alongside GPUs to reduce latency in AI agent applications. Additionally, NVIDIA launched the open-source platform NemoClaw to help enterprises deploy AI employees.
3. Microsoft Expects GPU and Storage Shortages to Last All Year, Hardware Price Hikes to Raise Capex by ~$25 Billion
During its quarterly earnings call, Microsoft acknowledged that rising prices for GPU and other hardware components would increase its full-year 2026 capital expenditure by approximately $25 billion. The company expects cloud computing supply constraints to “at least run through 2026,” and even with continuous acceleration of deliveries, the constrained supply situation is unlikely to ease. Nevertheless, Microsoft emphasized that its AI business margin is healthier than during its cloud transition years, a strength the company believes is underestimated by the market. It also announced that GitHub Copilot now has over 20 million paid enterprise seats, with its business model shifting toward usage‑ and value‑based pricing.
4. Chinese GPU Leader Cambricon Posts Record Q1 Results, Quarterly Revenue Exceeds RMB 20 Billion for First Time
Chinese GPU maker Cambricon released its 2026 first-quarter financial results, achieving operating revenue of RMB 28.85 billion, a year-on-year increase of 159.56%, marking the first time quarterly revenue surpassed RMB 20 billion. Net profit attributable to shareholders reached RMB 10.13 billion, up 185.04% YoY. The significant growth was mainly driven by sustained AI computing demand, with its cloud product line contributing the majority of revenue. Cambricon achieved Day‑0 adaptation upon the release of DeepSeek V4, demonstrating its rapid response capability within the domestic AI chip ecosystem. Following the earnings release, Cambricon's market value briefly overtook that of Kweichow Moutai, becoming the most valuable company on the A‑share market.
5. Moore Threads Achieves First Quarterly Profit, M-Xi Narrowed Losses Significantly
Performance divergence is evident among domestic GPU makers. In Q1, Moore Threads reported revenue of RMB 738 million, up 155% YoY, with net profit attributable to shareholders of RMB 29.35 million, marking its first quarterly profit since inception. However, non‑recurring net profit remained negative at a loss of RMB 54.28 million, with profitability mainly reliant on government subsidies, indicating its core business has yet to achieve a sustainable profit model. M‑Xi expects Q1 revenue of RMB 562 million, up 75.37% YoY, with losses narrowing sharply by 57.49%. The top five customers accounted for 91% of Moore Threads’ sales, highlighting a customer concentration risk.
6. DeepSeek V4 Released, For the First Time Listing Huawei Ascend Alongside NVIDIA GPU in Hardware Validation List
On April 24, DeepSeek officially released its V4 model, and in its technical report, for the first time listed Huawei Ascend NPU alongside NVIDIA GPU as verified hardware platforms, marking a “historic rendezvous” between major domestic models and domestic computing power. Official data anticipates that a super node equipped with 8,192 Ascend 950DT chips will be available in Q4 2026, delivering 8 EFLOPS of FP8 compute power. The Ascend 950 super node has already begun providing critical computing support for running the DeepSeek‑V4 model and is expected, after mass production in the second half of the year, to drive down API call prices significantly.
7. GPU Rental Market Prices Surge Nearly 40%, Domestic Cloud Providers Raise Prices
Driven by surging AI inference demand and ongoing supply tightness, the global GPU spot rental market is experiencing a high‑prosperity cycle of “volume and price increases.” Data shows that the one‑year rental price for an NVIDIA H100 GPU has risen nearly 40% from its low six months ago. The most advanced Blackwell series chips now rent for $4.08 per hour. On the supply side, Blackwell chip capacity has been booked through the second half of the year. Facing cost pressure, domestic cloud providers such as Tencent Cloud and Alibaba Cloud have announced price increases, with some products seeing hikes as high as 154%, signaling continued supply tightness.
8. Intel and AMD Jointly Release ACE Technology, Promoting x86 AI Computing Standard
Facing the AI wave, former rivals Intel and AMD have joined forces. They jointly published a white paper on AI Computing Extension (ACE) technology, proposing it as a “standard matrix acceleration architecture for x86 chips,” aiming to dramatically upgrade AI computing power of the x86 architecture by boosting matrix multiplication performance. Intel noted that the deployment ratio of CPU to GPU in AI workloads is tightening from 1:8 toward 1:4 or lower. This move seeks to strengthen the x86 ecosystem’s competitiveness in the AI era, lower adaptation barriers for developers, and differentiate from NVIDIA’s CUDA ecosystem. ACE instructions have already gained support from mainstream frameworks such as PyTorch and TensorFlow.
9. Demand for Huawei Ascend 950 AI Chips Surges Following DeepSeek V4 Release, Major Domestic Tech Firms in Talks
According to a Reuters report citing sources, following DeepSeek’s release of its V4 AI model running on Huawei Ascend chips, market demand for the upcoming mass‑produced Huawei Ascend 950 AI chips has surged dramatically. Leading domestic internet firms, including ByteDance, Tencent, and Alibaba, have entered discussions with Huawei regarding orders for the new chip. Industry analysts note that while the Ascend 950 may still lag behind NVIDIA’s more advanced H200 chip in performance, export restrictions barring the latter from the Chinese market present a rare market opportunity for Huawei.
10. Largest Domestic Scientific AI Computing Cluster Put into Operation in Zhengzhou, Equipped with 60,000 Domestic AI Accelerator Chips
The Zhengzhou core node of the National Supercomputing Internet has officially been put into operation. This is currently the largest scientific AI computing cluster in China, equipped with 60,000 domestic AI accelerator chips and delivering peak computing power of 10 EFLOPS. The cluster adopts a fully self‑developed technology stack, using high‑speed interconnection, immersion phase‑change liquid cooling, and other technologies to solve core challenges in ten‑thousand‑card clusters such as interconnection, heat dissipation, and reliability. It introduces a “Super Scientific Computing Agent,” allowing researchers to dispatch computing resources via natural language commands, aiming to lower the barrier to AI computing power for universities and small‑to‑medium enterprises.