發生了什麼事
TechCrunch reports that Anthropic signed a roughly $45 billion agreement to rent AI computing from Nscale, using systems built around Nvidia’s Vera Rubin chips. The report says the six-year arrangement would draw capacity from Nscale’s flagship data center in West Virginia, with service expected to begin in late 2027. The deal has not been independently confirmed in the source by Anthropic or Nscale.
TechCrunch reports that Anthropic has agreed to rent about $45 billion worth of AI from Nscale, a British AI infrastructure company founded in 2024. The report attributes the information to a source familiar with the deal and says Bloomberg first reported it. Neither a public contract nor a direct confirmation from Anthropic or Nscale is included in the source, so the agreement’s final terms are not independently confirmed here. The report describes the arrangement as a six-year deal rather than a one-time equipment purchase, meaning the stated value appears to cover an extended supply of computing services.
According to TechCrunch, Nscale would provide the capacity through Nvidia’s Vera Rubin systems. The article describes Vera Rubin as a system that combines six different chips working together and characterizes it as Nvidia’s newest state-of-the-art chip system. TechCrunch says the computing capacity is expected to begin powering Anthropic’s services in late 2027 and that the infrastructure would come from Nscale’s flagship data center in West Virginia. The source does not specify the number of systems, the amount of electricity involved, the precise delivery schedule, or how the $45 billion value would be calculated.
TechCrunch places the reported agreement within a broader expansion of Anthropic’s computing arrangements. The article says Anthropic earlier signed a six-year, $10 billion deal with AI cloud startup Volta for capacity from a data center in Norway, and in July signed a $5 billion -related agreement with AMD. It also reports that Anthropic disclosed a larger relationship with Amazon involving an additional five gigawatts of compute, while expanding its relationship with Google and Broadcom. TechCrunch further says an earlier SpaceX arrangement was reportedly providing $1.25 billion worth of capacity each month. These surrounding figures and arrangements are reported by TechCrunch and have not been independently verified in the supplied source.
為什麼這很重要
The reported agreement would be a major expansion of Anthropic’s computing commitments and illustrates how frontier AI companies are securing infrastructure years ahead of expected demand. It also links model development and service availability to data-center construction, advanced chips, electricity, and the finances of relatively young infrastructure providers.
The reported Nscale agreement shows that the competition among leading AI companies is increasingly a competition for physical infrastructure. Computing capacity depends on advanced accelerators, data-center space, cooling, network connections, and electricity. If the agreement proceeds as described, Anthropic would be reserving access to a substantial future supply of those resources rather than relying only on short-term cloud availability. That could help the company train and operate larger or more heavily used systems, but the source does not establish what specific models or products the capacity will support.
The deal also highlights the financial and operational risks created by rapid AI infrastructure expansion. Nscale was founded only in 2024, according to TechCrunch, yet the reported commitment would make it responsible for delivering a very large, multiyear supply of . The source says Nscale has already reached agreements with companies including Microsoft, but it does not provide details about Nscale’s financing, construction progress, hardware orders, power contracts, or ability to meet Anthropic’s requirements. Those unknowns matter because a multiyear compute agreement can be valuable only if the facilities, chips, electricity, and networking arrive on schedule and operate reliably.
For the public, the infrastructure race may affect the availability and economics of AI services even when users never interact with a data center directly. More capacity could support broader access or more capable services, while the cost and scarcity of advanced computing could also reinforce the advantage of the largest companies. The reported concentration of supply among Nvidia, Amazon, Google, AMD, SpaceX, Volta, and specialized infrastructure firms suggests that AI development is becoming tied to a small group of hardware and cloud providers. That is an analytical implication of the reported arrangements, not a finding independently demonstrated by the article.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
What is a common training objective for an autoregressive language model?
接下來看什麼
The key questions are whether Anthropic or Nscale publicly confirm the agreement, how much computing capacity the $45 billion figure represents, and whether the West Virginia facility reaches operation on schedule. The deal’s effect on Anthropic’s model training, product availability, pricing, and dependence on outside infrastructure remains unknown.
The first verification step is a public response from Anthropic or Nscale, along with any contract, regulatory filing, financing disclosure, or facility announcement that clarifies the agreement. The source provides no confirmation from either company. It also does not say whether the $45 billion figure is a firm minimum commitment, a maximum potential value, or an estimate based on expected usage and future pricing. Those distinctions would materially change how the deal should be understood.
The reported late-2027 start date is another important checkpoint. Observers would need evidence that Nscale’s West Virginia data center has secured power, completed construction and equipment deployment, and can operate Nvidia’s Vera Rubin systems at the required scale. The source does not provide a detailed construction timetable, capacity figure, chip allocation, or contingency plan if delivery is delayed. It is therefore not possible from this report to determine when Anthropic users would see any practical effect.
Finally, watch for evidence about what the will actually be used for. TechCrunch says the capacity is expected to power Anthropic’s services, but does not identify a model launch, product expansion, training run, or customer commitment tied to the deal. Future disclosures could show whether Anthropic is using the infrastructure mainly for model training, for existing products, enterprise demand, or a combination of purposes. Until those details emerge, the central verified development is the reported multiyear infrastructure agreement—not a confirmed improvement in model capability, service reliability, or user access.