Cloud storage options are being considered by more brands as Generative AI booms.
Cloud Storage Alternatives has Never Been More in Demand
For instance, this week, CoreWeave, a provider of GPU infrastructure initially established for cryptocurrency mining, secured an additional $1.1 billion in funding from Coatue, Fidelity, and Altimeter Capital, among others. The round raises the company’s post-money valuation to $19 billion and its total debt and equity financing to $5 billion—an extraordinary amount for a startup less than ten years old.
Not solely CoreWeave
Months after concluding a $320 million Series C round, Lambda Labs, which also provides a variety of cloud-hosted GPU instances, secured a “special purpose financing vehicle” of up to $500 million in early April. Voltage Park, a nonprofit organization supported by crypto magnate Jed McCaleb, announced in October last year that it would invest $500 million in GPU-powered data centers. Additionally, Together AI, a cloud GPU host researching generative AI, secured $106 million in March from a round led by Salesforce.
Why is there such a frenzy of investment and enthusiasm surrounding the alternative cloud space? As expected, the response would be generative AI.
The proliferation of generative AI has led to an increased need for hardware capable of efficiently operating and training generative AI models. GPUs are the logical architectural choice for training, fine-tuning, and operating models due to their thousands of cores that can execute the parallel linear algebra equations that comprise generative models.
GPU installation, however, is costly. As a result, most organizations and developers migrate to the cloud.
Prominent entities in the cloud computing industry—Amazon Web Services (AWS), Google Cloud, and Microsoft Azure—provide ample specialized hardware instances and GPUs specifically designed to handle generative AI workloads. However, for a minimum of some models and projects, alternative clouds may be more cost-effective while providing superior availability.
A popular model training and inference choice, the Nvidia A100 40GB is available for rental on CoreWeave for $2.39 per hour or $1,200 per month. The identical GPU costs $2,482 per month or $3.40 per hour on Azure and $3.67 per hour or $2,682 monthly on Google Cloud.
As generative AI workloads are typically executed on clusters of GPUs, cost deltas increase rapidly.
“Specific ‘GPU as a service’ cloud providers comprise a market segment that includes CoreWeave,” Sid Nag, vice president of cloud services and technologies at Gartner, explained to TechCrunch. “They have taken Nvidia GPUs and provided an alternate route to market and access to those GPUs in response to the hyperscalers’ high demand for GPUs.”
According to Nag, even some of the largest technology companies have begun to rely on alternative cloud providers due to computing capacity constraints.
Microsoft and CoreWeave reportedly signed a multimillion-dollar contract in June of last year to ensure that OpenAI, the developer of ChatGPT and a close Microsoft partner, had sufficient computing capacity to train its generative AI models. For leverage purposes, Nvidia, the primary supplier of CoreWeave’s processors, views this as a positive development; it has reportedly granted alternative cloud providers priority access to its GPUs.
Principal analyst at Forrester Lee Sustar believes that cloud vendors such as CoreWeave succeed partly because they lack the “baggage” of infrastructure that incumbent providers must contend with.
“Because the public cloud market is dominated by hyper scalers, which require substantial investments in infrastructure and a variety of services that generate minimal or no revenue, CoreWeave, and other challengers have the potential to thrive by concentrating on premium AI services without the overall financial burden of hypercar-level investments,” he explained.
Is this development, however, sustainable?
Sustar harbors skepticism. According to him, the expansion of alternative cloud providers will depend on their ability to consistently deliver GPUs online in significant quantities while maintaining competitively low prices.
Competing on price may become more challenging as market leaders such as Google, Microsoft, and AWS increase spending on specialized hardware to operate and train models. TPUs are provided by Google, Azure Maia, and Azure Cobalt, two custom processors recently unveiled by Microsoft, and Trainium, Inferentia, and Graviton are available from AWS.
Sustar stated, “Hypercalers will reduce their reliance on Nvidia by utilizing their custom silicon, whereas Nvidia will consider CoreWeave and other GPU-centric AI clouds.”
Additionally, none of the generative AI applications require GPUs to operate optimally, especially those that do not need them in a time-sensitive manner. Although CPUs can perform the necessary calculations, they could be more active than GPUs and custom hardware.
From a more existential standpoint, the generative AI bubble may explode, leaving providers with numerous GPUs but insufficient customers to demand them. However, according to Sustar and Nag, short-term prospects appear favorable, and both anticipate a consistent influx of fledgling clouds.
Sustar stated, “GPU-focused cloud startups will present formidable competition for incumbents, particularly among multi-cloud customers who can manage the complexities of risk, compliance, and security across multiple clouds.” “Customers of this nature are willing to test a new AI cloud if it possesses reputable leadership, substantial financial support, and GPUs that do not require waiting periods.”
