GPU warehouses are easy to admire and brutally hard to operate like a real business. Rafay Systems landing on Deloitte's 2026 Technology Fast 500 for a third straight year is another signal that the AI infrastructure scramble has moved well beyond simply finding chips. Somebody has to turn all that expensive compute into services customers can provision, use, measure and pay for. That less glamorous software layer is where Rafay says most of its growth is coming from.
The company is pitching itself to neoclouds, telecommunications providers, enterprises and sovereign cloud operators building what it calls AI Factories. The core argument is sensible: buying GPUs may be the headline-grabbing part, but converting shared hardware into a cloud with self-service access, multi-tenancy, governance, metering and billing is a long, messy engineering job. Operators do not want a shiny cluster sitting around as a very costly science project.
Recognition Is Useful, But It Is Not a Balance Sheet
Deloitte's Fast 500 ranks North American companies on fiscal-year revenue growth from 2022 through 2025, and this is the program's 32nd year. A third consecutive appearance says Rafay has maintained rapid growth over a meaningful stretch, which is better than a one-quarter AI boom story. It also follows the company's inclusion on the 2026 Inc. 5000 and an award from The Fast Mode.
Still, investors and industry watchers should keep their coffee cold before declaring victory. The release does not disclose Rafay's precise Fast 500 rank, revenue growth percentage, revenue base, profitability, customer concentration or retention. Fast-growth lists can identify commercial momentum, sure, but they do not tell us whether growth came from a tiny starting point, whether contracts are recurring, or whether the cost of winning customers is getting ugly. Those details matter a lot once the AI infrastructure spending frenzy starts demanding returns.
The Monetization Layer Is the Real Pitch
Rafay's platform is built to let an operator offer bare metal, virtual machines, Kubernetes, SLURM and token-metered model APIs from a common GPU pool. In plain English, it is trying to help infrastructure owners sell useful AI capacity rather than merely rent hardware by the hour. That distinction is not marketing fluff. A provider that can package different compute and inference services, set controls and bill cleanly has a better shot at raising utilization and protecting margins.
The newer Token Factory, Token Router and Token Budgets features push further into inference economics. Rafay says operators can publish AI models as token-metered services and offer frontier or open-source models through one endpoint under a single spending policy. If that works as advertised, it addresses a genuine headache: customers increasingly want flexible model access, while operators need some discipline around cost, security and runaway consumption.
Owning GPUs is not the same thing as running an AI cloud people can actually buy from.
That is the practical crux of this story. The hyperscalers have long had deep internal software benches. Regional providers, sovereign cloud projects and newer GPU operators often do not. Rafay is chasing the gap between hardware procurement and a commercially usable service catalog. Frankly, it is a much more believable niche than pretending every new AI infrastructure company will become the next giant platform.
Sharon AI Supplies a Serious Data Point
The clearest number in the announcement is a September agreement under which Sharon AI signed a five-year deal to orchestrate up to 150,000 GPUs on Rafay. That is a substantial headline figure and shows the company can be considered for large-scale deployments. It also lines up with management's claim that sovereign and neocloud operators are a major source of revenue growth.
But “up to” deserves a red circle around it. The notes do not say how many GPUs are currently active, the timing of any rollout, the contract value, minimum commitments or how much revenue Rafay recognizes as capacity comes online. A five-year agreement can be strategically important without translating into near-term financial fireworks. Anybody treating 150,000 as deployed, paid-for capacity is getting ahead of the evidence.
What Could Complicate the Run
- AI infrastructure budgets may tighten if customers struggle to monetize their own AI services.
- Large cloud platforms and hardware vendors have every incentive to expand competing orchestration, marketplace and billing tools.
- Sovereign-cloud projects can involve long procurement cycles, policy shifts and deployment delays.
- GPU supply, power availability and data-center buildouts can bottleneck demand even when software interest remains strong.
Rafay's third Fast 500 appearance does not settle any of those risks. What it does show is that the company is riding a real operational need in the AI buildout: capacity is only valuable when somebody can consume it without a swarm of engineers manually stitching together access, controls and invoices. That is a solid place to be, provided Rafay can keep turning recognition and large framework agreements into durable, disclosed economics.