据知情人士透露，OpenAI打算从Alphabet旗下的Google购买AI芯片，这是这家ChatGPT开发商扩建其AI基础设施庞大布局中的最新一环。该知情人士表示，任何交易的最终条款目前尚不得而知。OpenAI和Google则均拒绝置评。
随着企业寻求更便宜且适合运行人工智能模型的芯片，Google的张量处理单元(TPU)作为英伟达市场领先的AI芯片的替代方案，越来越受欢迎。Google今年开始直接向部分客户销售TPU。
此举也体现了OpenAI多管齐下、急迫提升自身AI算力的战略意图。
该公司此前已宣布，计划在今年晚些时候开始启用自研芯片，以支持其AI模型的运行。这意味着其加入了这场由多家企业掀起的“自研芯片潮”，共同向英伟达在算力领域的主导地位发起挑战。
值得一提的是，OpenAI与博通合作开发的内部芯片Jalapeno，就是在Google资深员工的协助下研发的。负责OpenAI内部芯片团队的Richard Ho此前曾参与GoogleTPU的研发。
Google在十多年前就开始研发其TPU。起初，这些TPU被用于加速该公司的网络搜索引擎并提高效率。但这些芯片很快被证明在处理AI工作方面极具价值，并在近年来成为英伟达GPU的替代方案。
英伟达算力霸权遭遇新挑战
此次采购计划将加深OpenAI与Alphabet旗下Google之间的联系——尽管这两家公司在人工智能产品和搜索领域存在直接竞争关系。
而长期以来，OpenAI与英伟达的关系一直颇为紧密。从初代GPT模型的训练到ChatGPT席卷全球，OpenAI几乎所有的核心模型演进，都建立在数以万计的英伟达GPU集群之上。黄仁勋曾亲自将全球第一台DGX AI超级计算机交付给OpenAI，而OpenAI的爆火也直接将英伟达推上了全球市值之巅。
今年8月，英伟达CEO黄仁勋曾表示，OpenAI已承诺到2030年大规模部署英伟达AI基础设施，现有及计划中的部署规模约为12吉瓦，并有机会扩大至约16吉瓦。
然而，这并未阻止OpenAI寻求多元化供应链。面对日益高昂的推理成本、动辄数百万颗芯片的庞大算力缺口，以及过度依赖单一供应商带来的议价劣势与断供风险，OpenAI显然也不愿将全部身家性命系于英伟达一身。
无独有偶。甲骨文公司本周二也宣布，将首次在其云平台上向客户提供基于Alphabet旗下GoogleTPU芯片的算力。
对于这一布局，甲骨文云基础设施总裁Mahesh Thiagarajan在声明中表示，不同的AI工作负载有不同的算力需求，这种选择让客户能够灵活地为每项AI工作负载挑选在性能、吞吐量和经济性上最合适的基础设施。
Google高管Amin Vahdat在宣布与甲骨文交易的声明中表示，“下一代人工智能将需要多样化的算力。”
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to sources, Open AI intends to purchase an AI chip from Google, under the banner of Alphabet, the latest in the large layout of the ChatGPT developer to expand its AI infrastructure. According to the source, the final terms of any transaction are not yet known. Open AI and Google refused to comment.
As enterprises seek cheaper chips that are suitable for the operation of artificial intelligence models, Google's tension processing unit (TPU) is becoming more and more popular as an alternative to the AI chip, which is the leading market in Britain. Google started selling TPU directly to some clients this year.
This also reflects Open AI's strategic and strategic intent to urgently improve its own budget.
The company had previously announced plans to launch the self-study chip later in the year to support the operation of its AI model. This means that it joins this “self-research chip surge” launched by a number of companies to challenge Britain's dominance in the field of computing.
It is worth mentioning that the internal chip Jalapeno developed by Open AI in cooperation with Chase was developed with the assistance of senior Google staff. Richard Ho, who is responsible for the Open AI internal chip team, was previously involved in the development of GoogleTPU.
Google began to develop its TPU more than a decade ago. Initially, the TPUs were used to speed up the company ' s network search engine and to improve efficiency. However, these chips quickly proved extremely valuable in dealing with AI work and have in recent years become an alternative to GPU in Weidar.
Yin Weidar is facing new challenges.
This procurement plan will deepen the connection between Open AI and Google under Alphabet - Despite the direct competition between the two companies in the area of artificial intelligence products and search.
For a long time, Open AI has been very close to Yen Wida. From the training of the first-generation GPT model to ChatGPT's global roll, almost all of Open AI's core model evolution is based on tens of thousands of British Wyda GPU clusters. Wong In-hoon personally delivered the world’s first DGX AI supercomputer to Open AI, and Open AI’s fire directly pushed Britain to the top of global market value.
In August of this year, the CEO Hoang In-hoon of the United Kingdom of Great Britain and Northern Ireland stated that Open AI had committed itself to a large-scale deployment of the British Virgin Islands infrastructure by 2030, with an existing and planned deployment of approximately 12 Giva and an opportunity to expand to about 16 Giva.
However, this has not prevented Open AI from seeking a diversified supply chain. Open AI is also clearly reluctant to pin its entire life on Britain in the face of the increasing cost of reasoning, the huge arithmetical gap in moving millions of chips, and the risk of bargaining disadvantage and severance from overdependence on a single supplier.
No one. Oracle also announced this Tuesday that for the first time on its cloud platform, it would provide its clients with the calculus based on GoogleTPU chips under Alphabet.
For this layout, Mahesh Thiagarajan, President of Oracle Cloud Infrastructure, stated in his statement that the different AI jobs contained different calculus requirements, and that this option gave clients the flexibility to select the most appropriate infrastructure for each AI job load in terms of performance, throughput and economy.
In a statement announcing dealings with Oracle, Google executive Amin Vahdat stated that “the next generation of artificial intelligence will require a diversified calculus”.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.