Recent #AI Infrastructure news in the semiconductor industry
➀ China's AI data center expansion in 2023-2024 has led to significant underutilization of billions of dollars in infrastructure;
➁ Rushed projects resulted in poor planning and technical irrelevance;
➂ Market shift towards inference workloads has reduced demand for high-end GPUs, leading to a decrease in rental prices.
➀ SoftBank announced a significant acquisition, purchasing US server CPU company Ampere Computing for $6.5 billion in cash, expected to be completed by the second half of 2025;
➁ Ampere, founded by former Intel executives in 2017, specializes in designing chips for Arm architecture servers and has early customers like Microsoft, Google, ByteDance, and Tencent;
➂ Despite facing financial challenges with three years of cumulative losses of nearly 14 billion RMB, Ampere's technical strength and market potential attracted SoftBank's attention.
➀ DeepSeek disclosed its profit margin for the first time, revealing a theoretical daily income of $562,027 and a cost profit margin of 545%;
➁ The disclosure sparked a debate between You Yang and Yuan Jinhui, founders of Luochen Technology and Silicon-Based Flow, respectively;
➂ The debate revolves around the profitability of AI infrastructure companies and the efficiency of DeepSeek's model architecture.
OpenAI's release of GPT-4.5 has been delayed due to a shortage of GPUs. CEO Sam Altman mentioned on X that the rollout will be staggered and will involve adding tens of thousands of GPUs next week. This shortage is prompting OpenAI to develop its own AI silicon in partnership with Broadcom.
The high cost of GPT-4.5, at $75 per million input tokens and $150 per million output tokens, is a concern, but Altman emphasizes its unique intelligence.
NVIDIA's GPUs are in high demand, with Blackwell GPUs sold out until October. The expansion of AI infrastructure, including massive AI supercomputers and data centers, is driving this demand.
➀ Lumotive, a specialist in programmable optical semiconductors, has raised $45 million in a Series B funding round.
➁ The funding will be used to accelerate growth in global market expansion, data center AI infrastructure, and aerospace and defense applications.
➂ Lumotive's LCM technology enables ultra-reliable optical circuit switches for high-performance optical switching solutions.
➀ IDC预测,到2027年,公司将在AI相关的基础设施、平台、软件和服务上投入超过300亿美元,以支持他们在高度个性化的客户体验方面的竞争能力。
➁ IDC的Abhishek Kumar指出,营销领导者必须记住,技术本身不是差异化因素,而是差异化的推动者。
➂ 他们需要与IT、数据、数字团队等紧密合作,建立必要的AI融合营销基础设施和工具,以实时访问准确的客户数据档案,并部署更快、更精准、更有效的活动。
➀ The US expands its blockade on Chinese semiconductor companies, including 140 new entities on the Entity List.
➁ China's four major industry associations voice opposition, suggesting cautious procurement of US chips.
➂ Huawei, the only domestic company offering a full range of self-developed AI infrastructure, outlines its strategy to counter the blockade.
➃ Huawei's 'super-node + cluster' architecture enhances training efficiency and reduces failures.
➄ Huawei Cloud's new CloudMatrix architecture to launch in 2024, providing high-efficiency and reliable AI computing services.
➅ China's industries are increasingly adopting AI, with Huawei Cloud supporting numerous large models.
➀ The AI industry's annual revenue exceeds $600 billion, which is enough to pay for data centers and AI infrastructure such as GPU cards. However, the spending on GPU infrastructure is still controversial.
➁ The rise and fall of the H100 GPU market, from its initial leasing price of $4.70/hour to the current price of less than $2/hour, indicates a potential bubble in the GPU market.
➂ The author suggests that users should rent rather than buy computing power when needed, as the market is oversupplied with H100 computing power.