技术档案

AI训练与推理算力AI training and inference compute

训练与运行AI的算力

每条都摘自下方编号的来源原文；只写事实，不写预测或建议。

技术简介

  • CPU独自计算的数据科学流水线受串行计算方式制约。3
    These operations are bottlenecked by the serial nature of CPU-only computing, which is compounded when scaling out for large processes.
  • Anthropic购买的算力用于训练和运行Claude，使用Trainium、TPU及GPU等AI硬件。6
    We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them.
  • AI训练和推理需要高密度计算、先进网络、优化存储及管理复杂分布式系统的软件。10
    Both training and inference require high-density compute, advanced networking, optimized storage, and software capable of managing complex distributed systems.
  • NVIDIA 加速计算平台以 GPU 为基础，用于神经网络训练和推理等并行负载。11
    At the foundation of the NVIDIA accelerated computing platform are our GPUs, which excel at parallel workloads such as the training and inferencing of neural networks.
  • 模型推理算力用于让AI模型理解并响应新的数据和请求。12
    Model inference workloads are composed of tasks (and their associated computing power) necessary for AI models to interpret and respond to brand-new data and requests.
  • AI工作负载是用于完成AI、机器学习和深度学习任务的计算过程、应用和实时计算资源集合。12
    AI workloads are collections of individual computer processes, applications and real-time computational resources used to complete tasks specific to artificial intelligence, machine learning and deep learning systems.
  • 模型训练工作负载用于教AI框架识别模式，以作出准确预测。12
    Model training workloads are used to teach AI frameworks to identify patterns in order to make accurate predictions.

已有应用

  • Anthropic使用AWS Trainium、Google TPU、NVIDIA GPU和AMD GPU组成多芯片算力架构。7
    We have achieved this through our research breakthroughs and our multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, enabling us to orchestrate workloads to optimize for price and performance.
  • Meta的Muse使用AMD EPYC处理器提供计算。8
    AMD said that its Helios racks were running at Meta and its EPYC processors were powering Meta's Muse.
  • Cognition 在 CoreWeave 上运行 Devin 的训练、强化学习和生产推理。5
    Cognition, the applied AI lab behind Devin, runs training, reinforcement learning and production inference for Devin on CoreWeave.
  • Anthropic与亚马逊合作推出Rainier集群，并已使用逾百万颗Trainium2训练和运行Claude。9
    Together we launched Project Rainier, one of the largest compute clusters in the world, and we currently use over one million Trainium2 chips to train and serve Claude.
  • Anthropic用AWS Trainium、Google TPU和NVIDIA GPU训练与运行Claude。6
    We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them.

为何此时

  • 2026年4月Anthropic称消费端快速增长已影响免费及付费用户高峰时段的可靠性和性能。9
    Growth at this pace places an inevitable strain on our infrastructure; our unprecedented consumer growth, in particular, has impacted reliability and performance for free, Pro, Max, and Team users, especially during peak hours.

主要参与者

  • 英伟达提供DGX、HGX等加速计算硬件系统及软件栈，硬件和软件亦经合作伙伴提供。3
    These hardware systems and software are available from NVIDIA and our partners.
  • Anthropic的多芯片算力架构使用AWS Trainium、谷歌TPU、英伟达GPU和AMD GPU。4
    We have achieved this through our research breakthroughs and our multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, enabling us to orchestrate workloads to optimize for price and performance.
  • CoreWeave 云平台已上线 NVIDIA Vera Rubin NVL72，Cognition 是首个运行生产负载的客户。5
    CoreWeave Inc. (Nasdaq: CRWV), The Essential Cloud for AI™, today announced the availability of NVIDIA Vera Rubin NVL72 on CoreWeave, with Cognition as the first customer anywhere running production workloads on the system.
  • 博通按与谷歌签署的协议开发和供应后续世代定制TPU，并供应AI机柜网络等部件。2
    Broadcom Inc. (“Broadcom”) and Google LLC (“Google”) have entered into a Long Term Agreement for Broadcom to develop and supply custom Tensor Processing Units (“TPUs”) for Google’s future generations of TPUs and a Supply Assurance Agreement for Broadcom to supply networking and other components to be used in Google’s next-generation AI racks through up to 2031.
  • 亚马逊是Anthropic的主要云服务商和训练合作伙伴。6
    Amazon remains our primary cloud provider and training partner, and we continue to work closely with AWS on Project Rainier.

未定事项

  • 2026年10月26日披露的SpaceX GPU租赁承诺仍有交割条件，预计年底前交割、2027年上半年随设备交付启动租期。1
    Entry into the lease remains subject to certain specified closing conditions and is expected to close prior to year-end 2026, with lease schedules expected to commence upon delivery of GPUs and related equipment in the first half of 2027.
  • 博通披露，Anthropic使用扩展TPU算力取决于其持续商业成功。2
    The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic’s continued commercial success.

相关卡片

来源

  1. apollo.com · SpaceX Secures $40 Billion in the Largest GPU Financing Ever · 2026-10-26
  2. SEC · 8-K · 2026-04-06
  3. nvidia.com · Accelerated Computing for Enterprise IT | NVIDIA
  4. SEC · Anthropic, PBC - S-1 (Q2) · 2026-10-26
  5. CoreWeave · CoreWeave Delivers NVIDIA Vera Rubin NVL72 Performance at Production Scale, Starting With Cognition · 2026-09-30
  6. anthropic.com · Anthropic expands Google and Broadcom compute deal · 2026-04-06
  7. SEC · Anthropic, PBC - S-1/A-1 · 2026-11-02
  8. CNBC · Nvidia, AMD to invest $4 billion in Meta as tech giants deepen AI partnership · 2026-10-27
  9. anthropic.com · Anthropic and Amazon expand compute collaboration · 2026-04-20
  10. SEC · crwv-20251231 · 2026-03-02
  11. SEC · nvda-20260125 · 2026-02-25
  12. ibm.com · What are AI Workloads? | IBM · 2025-05-09