Cloud AI Market Update: NVIDIA’s Cloud Strategy, Hyperscalers' ASICs, and DeepSeek

NVIDIA-approved cloud services, such as DGX Cloud, have long been available through hyperscale providers like Google Cloud, Microsoft Azure, and Oracle Cloud. This is in addition to the plethora of NVIDIA Graphics Processing Unit (GPU) instances available through the same hyperscaler group, as well as smaller providers like the “neo-cloud”—or GPU-as-a-Service—challengers exemplified by CoreWeave, Lambda Labs, and Crusoe. As is typical for a chip vendor, NVIDIA has allowed the cloud ecosystem to provide services powered by its compute technology and CUDA developer environment, with the inner circle of hyperscalers providing the deeply integrated and fully managed DGX Cloud supercomputing platform.

NVIDIA has previously leased its own data center capacity for its own computing needs, which include research, testing, and hosting—much like most corporations. However, a significant bump in capacity could signal a change from the above strategy by enabling the provision of DGX Cloud and AI Enterprise services—as well as “simpler” GPU instances—via its own infrastructure, outside of any hyperscaler and neo-cloud providers’ services. However, NVIDIA will not abandon its existing business model and reliable revenue stream, especially given the vast, unprecedented and increasing Capital Expenditure (CAPEX) by hyperscalers throughout 2024 on Artificial Intelligence (AI) compute, much of which lands in NVIDIA’s coffers and will not drastically change in the medium term, despite recent news about the relatively “compute-cheap” model released by Chinese DeepSeek.

An increased focus on cloud by NVIDIA might appear to some as a strategy to undercut cloud players, which are the biggest GPU consumers, by providing gold-plated in-house AI cloud services with hardware supplied at cost, but this would be counterproductive in the short and medium term. Zooming out, a clearer picture emerges when considering cloud providers’ familiar focus on general compute: by spreading their expertise broadly to accommodate a plethora of customers and workloads, from the fine-tuning of domain-specific Large Language Models (LLMs) to basic database and storage management, they are able to address a much larger market. A consequence of this is lower overall hardware utilization compared to what can be achieved in niche or High-Performance Computing (HPC) use cases with highly optimized setups. However, the AI and HPC needs of customers in healthcare or industrial simulation are more narrow and less addressable by the broader hyperscaler business model, and a gap emerges that NVIDIA could fill with its verticalized solutions and AI Enterprise platform—coupled with highly-optimized hardware dedicated to these workloads.

Another interpretation of such a move could be a hedge against the progress of hyperscalers’ own efforts in developing AI chipsets for use in their clouds, as well as “first-party” workloads for their own AI model development, which has also manifested in the recent financial successes of custom silicon designers Marvell and Broadcom. Examples of these chips include AWS’s Inferentia, Google’s Tensor Processing Unit (TPU), and Microsoft’s Maia. By building its expertise and offering it in narrow accelerated compute domains, customers are tempted to move toward NVIDIA’s platform and away from public clouds. Thus, by providing more specialized capacity today in areas where NVIDIA’s hardware and software excel, the moat around CUDA deepens, protecting against the commoditization of cloud AI chips—a process that is already well under way.

Whatever the motivation behind an alleged jump in cloud capacity at NVIDIA, ABI Research believes that specialized, verticalized offerings on top of AI hardware serve both short- and longer-term interests. By coupling AI development—especially in fields with narrow requirements like healthcare for drug discovery research—with NVIDIA’s software value add, the company will ensure that demand does not dry up, even if the market is flooded with energy-efficient hyperscaler Application-Specific Integrated Circuits (ASICs).

NVIDIA’s main data center competitors, Intel and AMD, are still behind, to varying degrees. AMD’s Instinct platform did ramp up in 2024, but did not make a significant dent in NVIDIA’s revenue, despite the much-lauded Total Cost of Ownership (TCO) benefits. AMD’s acquisition of ZT Systems should help with its large-scale AI systems proposition, presumably enabling the expansion beyond the 8-GPU Universal Baseboard (UBB) form factor designs that currently make up the most performant trays that its Original Equipment Manufacturer (OEM) and Original Design Manufacturer (ODM) partners sell. Intel’s Gaudi platform did not take the market by storm, and will likely be subsumed by the upcoming Falcon Shores GPU incorporating Gaudi IP with Intel’s existing GPU architectures. Nonetheless, it is the software value add where NVIDIA really excels, and where a significant chunk of its Research and Development (R&D) budged is consumed—in both verticalized solutions, as well as library optimizations that squeeze more performance out of its GPUs. These are areas where competitors should take heed.

The news of the performance of the less compute-intensive Chinese DeepSeek model has had a seismic effect on the earnings expectations of NVIDIA and other data center players. However, training such models is but one of many AI workloads around today, each of which is linked to specific server form factors and data center infrastructure. ABI Research expects the workloads of tomorrow to require different hardware optimizations, which include fine-tuning, and the training and inferencing of agentic AI. These important considerations form the basis of ABI Research’s upcoming AI data center forecasts.