CoreWeave 签署 A100 云合同，将 Nvidia（英伟达）2020 年的芯片延用至 2029 年

Why it matters

A 2029 A100 contract gives cloud operators evidence that older AI GPUs can outlive their first leases, reducing a key risk behind debt-funded data centers.

Jensen Huang (@JensenHuang) 在 8 月 13 日表示，Nvidia 的 A100 机群可以在 2029 年之前保持“mission-capable”（任务可用）状态，这一说法放大了 CoreWeave 的一项合同，该合同挑战了 AI 加速器在几年内就成为被搁置资产的假设。

“Nvidia 的强大 A100 机群从 2020 年到 2029 年都能执行任务，”Nvidia 的联合创始人兼首席执行官在一篇 X 帖子 中写道。Huang 将此归功于 CUDA，为开发者和 Nvidia 工程师提供了一个共同的软件平台，以便在 Ampere、Hopper 和 Blackwell 架构上构建的系统在其整个使用寿命中进行升级。

直接的证据来自 CoreWeave。在其 8 月 11 日的财报电话会议 中，这家 AI 云服务提供商表示已签署了一份延伸至 2029 年的 A100 合同，且“价格具有吸引力”。CoreWeave 将这笔交易描述为旧有 Nvidia 基础设施在原客户合同到期后仍能产生收入的一个例子。

A contract tests GPU depreciation assumptions

Nvidia introduced the A100 on May 14th, 2020 as the first data-center GPU based on its Ampere architecture. A contract running into 2029 would put some of that infrastructure to work as much as nine years after the product's introduction, spanning several generations of faster Nvidia hardware.

CoreWeave has a direct financial reason to emphasize that lifespan. Chief Financial Officer Nitin Agrawal told investors that a typical five-year customer contract requires heavy upfront capital spending, financed through debt, customer prepayments and corporate capital. CoreWeave expects the initial contract to repay the asset-level debt attached to a cluster. Any renewal or resale after that point can add revenue without the original financing burden.

That argument matters because CoreWeave's expansion remains capital intensive. CoreWeave reported $2.575 billion in second-quarter revenue, alongside a $626 million net loss and $640 million in net interest expense. Capital expenditures reached $9.4 billion during the quarter as CoreWeave continued adding data-center capacity.

Longer GPU revenue lives could improve the returns on that spending. Shorter lives would leave CoreWeave replacing expensive hardware while carrying the financing and infrastructure costs associated with its expansion. CoreWeave said its earlier Ampere and Hopper fleets remain largely sold out and that prices for older generations are holding near or above levels seen a year earlier.

The 2029 contract supplies one concrete data point. It does not establish that every A100 cluster will remain fully utilized or economically competitive for the same period. Workload requirements, electricity costs, customer pricing and the performance gains delivered by newer hardware will determine which older systems continue earning revenue.

CUDA is Nvidia's durability argument

Huang's endorsement turns CoreWeave's contract into a broader defense of Nvidia's full-stack model. Nvidia sells accelerators on a rapid release schedule while CUDA gives developers a stable programming layer across generations.

Nvidia's current CUDA architecture matrix lists toolkit and driver support for Ampere as ongoing. The documentation also says the CUDA driver API is backward compatible, allowing newer drivers to run applications compiled with older CUDA toolkits. Nvidia's framework support matrix continues to list the A100 beside Hopper and Blackwell systems in supported configurations.

Software support cannot erase the performance difference between a 2020 A100 and newer accelerators. It can keep the older hardware available for workloads that do not require the fastest training or inference performance. CoreWeave told investors that customers are using prior-generation infrastructure across a wider range of AI workloads, while its newest clusters serve the most demanding jobs.

Nvidia benefits from the same asset-life thesis

Nvidia also has a financial stake in CoreWeave's success. On January 26th, Nvidia invested $2 billion in CoreWeave Class A shares at $87.20 each. The two businesses agreed to deepen their relationship, including deployments of multiple Nvidia generations and the use of Nvidia's balance sheet to help CoreWeave secure land, power and data-center buildings.

Huang's post supports both sides of that relationship. CoreWeave gains validation for the useful life of the assets underpinning its debt-funded cloud. Nvidia reinforces the case that customers are buying a computing platform whose software can preserve older hardware even as Nvidia introduces faster chips.

The A100 contract shows how that model can work in practice. New architectures command the workloads that need maximum performance. Older GPUs can move to inference, enterprise deployments and other jobs where availability and price matter as much as benchmark leadership. Keeping both generations productive allows Nvidia's annual hardware cadence to coexist with infrastructure designed to operate for most of a decade.