Google could reportedly invest $100M in AI cloud operator Fluidstack – News

Cloud startup Fluidstack Ltd. is reportedly in talks with Google LLC about a $100 million investment.

The Wall Street Journal today cited sources as saying that the deal could value Fluidstack at $7.5 billion. According to Forbes, the low-profile startup previously raised about $25 million in equity funding over two rounds. It has also secured a $10 billion line of credit to finance data center construction initiatives.

Fluidstack operates a cloud platform built to power artificial intelligence training and inference workloads. One of its main selling points is that it enables companies to quickly spin up large graphics card clusters. One early customer, venture-backed AI coding startup Poolside Inc., provisioned a cluster with more than 2,500 GPUs in 48 hours.

Fluidstack’s website indicates that the most advanced chip in its catalog is Nvidia Corp.’s GB200, which combines two Blackwell B200 graphics cards with one central processing unit. The company also offers several standalone GPUs.

Customers can manage their Fluidstack-hosted infrastructure using Kubernetes or Slurm. Both tools automate the task of allocating hardware resources to AI models. Kubernetes can power both training and inference workloads, while Slurm is mainly geared towards the former use case.

Fluidstack’s platform abstracts away much of the manual work involved in using the two open-source tools. A centralized dashboard enables administrators to provision servers, release updates and perform other maintenance tasks in one place.

The other core software component of Fluidstack’s platform is an observability tool called Lighthouse. It alerts administrators when a GPU uses too much power or generates excessive heat. According to Fluidstack, it also detects technical issues that affect other hardware components. For example, Lighthouse can spot malfunctions in the PCIe link that connect a GPU to its host server.

Google’s potential $100 million investment is reportedly part of an effort to boost the adoption of its tensor processing unit artificial intelligence chips. The newest TPU in the series, Ironwood, made its debut last April. It can carry out 4,614 trillion computations per second and provides double the performance per watt of Google’s previous-generation silicon. 

According to the Journal’s sources, the search giant is “increasing its financial support to a network of data-center partners” in a bid to grow TPU adoption. The report didn’t name the partners. It’s possible they include Hut 8 Corp. and TeraWulf Inc., two data center builders that recently inked infrastructure construction contracts with Fluidstack. Google is backstopping the contracts.

Separately, Fluidstack is collaborating with Anthropic PBC on a $50 billion initiative to build AI data centers in the U.S. The large language model developer expects the first facilities to come online this year.

Image: Google