新华财经上海12月28日电（记者高少华）2025年国产大模型DeepSeek引爆大模型的落地应用，国内算 力需求激增，但由于高端算力供应紧张、国产芯片百花齐放、智算中心扩容混建，由此形成大量异构算 力孤岛，给智算集群的资源利用效率和灵活性带来挑战。 记者从上海壁仞科技股份有限公司（以下简称"壁仞科技"）获悉，壁仞科技致力打造自主原创的高性能 图形处理器（GPU）软硬件体系，首代通用GPU产品基于原创训推一体芯片架构，实现中国首个四种及 以上异构芯片混训技术落地，有效突破了算力孤岛难题。 伴随国内算力需求爆发式增长，算力资源不足成为一大短板，业界由此逐步转向异构算力阶段，推动图 形处理器、神经网络处理器（NPU）、现场可编程门阵列（FPGA）等异构芯片成为算力供给的主力。 在海外供应链风险加剧、国产算力替代加速背景下，混合异构算力已成为企业人工智能基础设施建设的 新常态，但同时业界也普遍面临GPU利用率不足、异构资源管理困难等痛点。 针对异构算力问题，国内GPU及算力厂商正通过硬件兼容、软件框架优化、智能调度、国产化适配和能 效提升等举措，以此提升整体芯片算力，实现算力协同。 "壁仞科技目前有万卡集群整体解决方 ...

Core Insights - OpenAI has entered into a nearly $12 billion partnership with cloud service provider CoreWeave, aiming to diversify its computing capabilities away from Microsoft [2][3] - The five-year agreement will see CoreWeave provide computing power for training and running OpenAI's AI models, with OpenAI also acquiring $350 million worth of equity in CoreWeave at the time of its IPO [3][4] OpenAI's Strategic Moves - The collaboration with CoreWeave is part of OpenAI's strategy to reduce reliance on a single supplier, ensuring greater autonomy and flexibility in AI model training and operations [5] - OpenAI's CEO Sam Altman emphasizes the importance of infrastructure for developing cutting-edge AI models, while Microsoft CEO Satya Nadella has expressed concerns over "overbuilding" in computing capacity among large tech companies [5] CoreWeave's Background and Challenges - CoreWeave, established in 2017, initially focused on cryptocurrency mining before pivoting to provide cloud services using NVIDIA's high-performance GPUs for AI model training [4] - CoreWeave faced setbacks as its largest client, Microsoft, withdrew some commitments due to delivery issues, which impacted its revenue, with Microsoft accounting for 62% of CoreWeave's income last year [3][5]