〔Overnight Brief〕AI Infrastructure Shifts: CoreWeave Warns on Nvidia Dependence, Cerebras Cloud Surges, and DeepSeek V4 Pro Debuts

CoreWeave Flags Nvidia Concentration Risk

CoreWeave, in its latest quarterly filing, warned that if market demand shifts and it must abandon Nvidia chips, it would face significant time and capital pressure to secure replacements, potentially affecting its ability to deliver products to customers ^[1]. This is the first time the company has formally listed 'de-Nvidia-ization' as a business risk ^[1]. Despite the warning, CoreWeave's order backlog reached $104B by the end of Q2, and shares jumped 19% to $107.73 on Wednesday ^[1]. Nvidia holds about 10% of CoreWeave and supplies all its customers' GPU compute, creating a highly concentrated supply structure ^[1].

Cerebras Q2: Cloud Soars, Hardware Slips

Cerebras reported Q2 GAAP revenue of $180.1M, up 74% YoY, but hardware revenue declined 23% to $54.1M ^[2][3]. Core revenue (excluding certain items) grew 103% to $209.9M, driven by cloud and other services revenue of $127.7M, up 287% ^[2][3]. The company's CFO noted 'the demand for fast inference is unlimited,' attributing growth to OpenAI's expanded deployment and other cloud customers ^[2]. Management raised full-year guidance and reiterated a 2027 target of '3x+ growth,' while the stock fell about 17% after hours ^[2]. CEO Andrew Feldman expects manufacturing capacity to increase 10x in 2026 and expand further in 2027 ^[4][5]. For Q3, Cerebras expects revenue of $214-216M, above the $212M consensus ^[6].

DeepSeek V4 Pro and Grok 4.6 Challenge Frontier Models

DeepSeek officially launched V4 Pro (version DeepSeek-V4-Pro-0813) on August 12, featuring a 1M-token context window and 384K max output, with enhanced agent capabilities including Responses API and Codex integration ^[7][8]. In benchmarks, it scored 87.9 on Terminal Bench, close to Fable 5's 88.0, and outperformed Fable 5 on CyberGym and AutomationBench ^[8]. Pricing is 3 yuan per million input tokens (cache miss), 6 yuan per million output tokens, with cache hit at 0.025 yuan ^[7][8]. Meanwhile, SpaceXAI released Grok 4.6, claiming it matches or exceeds GPT-5.6 Sol Max and Fable 5 on several tests, with API pricing at $2 per million input and $6 per million output, about half the cost of competitors ^[9].

Supply Chain Investments: Liquid Cooling, MLCC, and Trade Facilitation

AI server liquid cooling investments are deepening into components, as companies like 强瑞技术 (Qiangrui Technology) plan a private placement up to 1.05 billion yuan for precision liquid cooling parts, and 飞荣达 reported rising orders and prices ^[10][11]. The industry is shifting from technology route competition to production consistency, automation, and overseas certification ^[10][11]. Separately, MLCC demand is rising with AI servers, with multiple stocks gaining over 300% year-to-date ^[12]. The U.S. announced a pilot platform in Panama to accelerate AI-related trade among allies, backed by up to $50M in foreign assistance ^[13].

Other Notable Developments

Apple is reportedly in talks with publishers for multi-year content licensing deals worth up to several hundred million dollars to feed AI-powered Siri ^[14]. 滴普科技 (Dipu Technology) agreed to place 14.3 million new H-shares at HK$35 each, raising about HK$500 million ^[15]. 宇树科技 (Unitree) saw strong institutional participation in its IPO, with A-class investors (including mutual funds and insurance) allocated 85.36% of the online portion ^[16]. Cisco's CEO expects AI infrastructure revenue to grow to $7.5 billion in fiscal 2027 ^[17]. Google executives held a company-wide meeting to address AI department restructuring concerns ^[18]. Palo Alto Networks will run OpenAI cyber models inside customer networks ^[19]. Nvidia's biggest risk may come from its customer base, per analyst commentary ^[20].