Rafay chases one-hour GPU-cloud orchestration deployments

By Rebecca Stone

Canadian operator Telus moved about 20% of its token production to a Rafay Systems-powered “token factory” averaging five to six billion tokens an hour, providing a concrete throughput figure for AI orchestration. Rafay chief product officer Mohan Atreya supplied the figure.

Rafay’s orchestration and workflow-automation platform acts as the control plane between neocloud computing infrastructure and end-user applications. Firmus, Era4 and Sharon AI are among its named users. Its token-routing layer exposes AI models through one endpoint and switches among them automatically, helping providers keep expensive capacity busy.

Rafay CPO Mohan Atreya said: “Pretty much every GPU has been spoken for.” He was describing utilization at Rafay’s neocloud customers, not the entire market. No independent utilization data, customer count or measurement period was provided, so the statement remains a vendor-side indicator.

The wider backdrop is unsettled: Microsoft has reversed course on Nscale, Firmus has canceled its IPO and Elon Musk has characterized the neocloud market as a guppy-level death. Yet Atreya said neoclouds remain busy, protect margins and are trying to move beyond renting servers. Those goals make software, service quality and multitenancy more important than another bare GPU listing.

Can neoclouds sell more than raw compute?

Enterprise buyers now expect security, governance and service capabilities comparable to those they receive from hyperscalers, rather than server access alone, Atreya said. His prescription combines multitenancy support with marketplaces for:

• Bare-metal servers
• Kubernetes
• Token services

That shifts the proposition from hardware rental toward managed cloud operations, but Rafay disclosed no pricing, contract value or revenue contribution for these services.

Rafay also sells outside the neocloud market. Atreya said GPU-cloud customers were easier to convert because idle expensive processors threaten returns and purchasing decisions move quickly. Enterprise sales cycles run longer, although he compared large companies to small neoclouds because separate business units carry distinct complexity and regulatory duties. The same orchestration product applies to both, in his view.

What does utilization show about contracts and margins?

Atreya said those customers favor three-month commitments rather than annual prices. Shorter terms allow providers to raise prices when needed and fine-tune margins, mirroring flexible hyperscaler contracts. The approach supports margins, but it does not establish long-term price stability for GPU capacity.

Where does orchestration fit with Nvidia and hardware?

Rafay and Nvidia co-design offerings supplied to neoclouds, with Atreya describing their engineering relationship as tight. The company also works with five hardware partners:

• Dell
• Lenovo
• Hewlett Packard Enterprise
• Supermicro
• Cisco

Rafay sells through partner price books, bundling orchestration with infrastructure so customers receive the software earlier rather than discovering the need after procurement.

Sharon AI illustrates that route: the Australian neocloud began with a small Rafay deployment and then scaled quickly. Rafay gave no size, revenue or capacity figures, leaving the pace claim without a numerical benchmark.

Why does the Telus deployment matter?

Telus offers a different example because its internal teams are customer No. 1, Atreya said. Moving token production in-house shifts spending from an external supplier to internal operations; it does not show that a neocloud can resell the same service at scale.

He said the factory preserves the service-level agreement and quality an internal customer would expect from a hyperscaler or Anthropic. That operating equivalence matters to an enterprise building its own AI platform, but internal adoption does not establish a neocloud resale business.

Where else can the control plane sell?

Rafay sees demand among enterprises rebuilding AI data-center capacity instead of relying wholly on hyperscalers. Atreya cited companies in Turkey, Saudi Arabia and Dubai that have stepped back from hyperscalers, linking the change to attacks in the wider Middle East conflict. He did not identify those organizations or quantify their deployments.

He also pointed to classified US environments, data-sovereignty requirements and the need to run AI closer to home. These constraints can support localized orchestration demand, but deployments must still meet security, governance and compliance controls that Rafay did not detail.

Can deployment really shrink from weeks to an hour?

Rafay’s stated goal is to standardize deployment enough for a customer to move from delivered hardware to a working platform in roughly one hour instead of weeks. Atreya added: “The process, the practice, that’s where the IP is right now.” That claim concerns operating procedures as much as control-plane software.

Rafay did not provide a measured production implementation time or a comparison set for the one-hour target. Any move up the value chain must still reconcile rapid deployment with tenant isolation, governance and hyperscaler-like service levels.

Until Rafay supplies quantified production outcomes, the one-hour target and broader enterprise proposition remain plausible but largely vendor-argued. The next rollout cycle will test whether Nvidia- and hardware-partner channels can turn that case into repeatable production deployments without sacrificing the margin flexibility operators value in three-month commitments.