Vultr Claims AI Infrastructure 50-90% Cheaper Than AWS

The pitch from Vultr is simple and aggressive: the same Nvidia hardware that runs on AWS costs 50-90% less on their infrastructure. That’s a large enough gap that, if accurate, it changes how platform teams should be doing AI compute cost modeling.

The claim isn’t coming from a startup making noise. Vultr is a $3.5 billion cloud platform (valued December 2024, $333 million in equity funding led by LuminArx Capital Management and AMD Ventures) with 32 data center regions and more than 80 million cloud servers deployed. Their recently announced partnership with SUSE Rancher and a $1 billion+ Ohio AI cluster — planned around 24,000 AMD MI355X chips — signals this is a sustained infrastructure bet, not a marketing cycle.

What the Numbers Actually Say

The headline comparison holds up under scrutiny, but requires context. Google Cloud H100 pricing sits at $14.19 per GPU per hour at list rates. Neocloud providers including Vultr are running equivalent H100 configurations at approximately $2.49 per hour — a 470% price differential at spot pricing. AWS cut H100 prices by 44% in mid-2025 and still sits at multiples of neocloud rates on equivalent hardware.

The gap narrows in enterprise contracts. Hyperscalers negotiate aggressively on committed use discounts, and their bundled pricing includes services — global redundancy, compliance certifications, managed networking — that neoclouds price separately or don’t offer at the same maturity level. The honest comparison for an enterprise platform team isn’t list price vs. list price; it’s total cost of ownership across compute, data transfer, compliance, and operational overhead.

That said, the core problem Vultr is solving is real. According to the company’s positioning, the failure mode they see repeatedly is enterprises approving AI workloads based on pilot costs, then discovering that production scale on hyperscalers runs “multiples of what they budgeted.” The hyperscaler pricing model — layered, negotiated, contract-dependent — is structurally difficult to forecast. Vultr’s counter-positioning is transparent, pay-as-you-go rates on specific hardware configurations with no long-term contract requirement.

Where Vultr Makes Sense

For AI training runs, batch inference, and compute-intensive workloads where the requirement is GPU access at known cost, Vultr (and the neocloud category broadly) deserves to be in the evaluation. AMD MI300X starts at $1.85 per GPU per hour; MI325X at $2.00 per hour. These are meaningful options for organizations running regular training jobs that don’t require the full hyperscaler ecosystem.

The limitation is ecosystem depth. AWS Bedrock, Azure AI Foundry, and Google Vertex AI bundle model management, monitoring, fine-tuning pipelines, and compliance tooling into their compute pricing. Teams that have built against those managed services can’t easily port to a lower-cost compute tier without rebuilding the surrounding infrastructure. The neocloud value proposition is strongest for teams running their own MLOps stack — not teams that have invested in hyperscaler-managed AI services.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.