
Neoclouds have hit the headlines for many wrong reasons of late. First, frontrunners like Coreweave have suffered from large selloffs in recent months while their biggest backer – Nvidia’s Jensen Huang – has resigned from the world’s largest AI company.
Nvidia’s better-than-expected results just today lifted many technology stocks though it has become clear neocloud companies, so called because they focus heavily on AI, have to think beyond providing graphics processing units (GPUs) for rent.
Not that the demand is falling, but news of Huang backing neoclouds to keep his GPUs in demand will worry some investors over the long term. For neoclouds, it’s always good to have an offering that others have not developed yet.
Over the long run, they have to build services on top of GPUs and capture more value from the hefty investment, says Anindo Sengupta, vice-president of AI, cloud-native and neocloud/service providers at Nutanix, which provides software to cloud providers including neoclouds.
In the same way older hyperscalers have developed in the past, neoclouds should provide managed services and offer improved visibility into GPU usage, he tells Techgoondu in the latest Q&A.
“It’s not just about the hardware anymore,” he explains, in an interview earlier this month, before the Nvidia leadership crisis. “The neoclouds that deliver the full stack will turn today’s demand into a lasting business.”
Q: What are neoclouds doing differently that make them important to the AI gold rush in recent years?
A: Neoclouds have emerged as a new category in the recent few years to focus on purpose built AI cloud infrastructure. Neoclouds raised the capital, secured the GPUs and built their data centres around them.
Specialising in GPUs has allowed neoclouds to play a critical role in the growth of AI, because demand for GPU keeps rising alongside AI usage. AI has progressed from chat interfaces to agents that reason through a problem, and now to AI co-workers that work alongside people.
Each step consumes far more compute than the last. Neoclouds were ready to absorb that demand, because GPUs are their core business rather than one line in a hyperscaler’s broad catalogue.
But to sustain that growth over the long run, neoclouds can no longer rely on just adding more GPUs. They need to build services on top and capture more value from the investment.
As organisations around the world take AI into production, customers are all asking for the same things: managed services, visibility into what each workload consumes and costs, and control over their data – especially as data sovereignty, residency and privacy become boardroom priorities across APJ (Asia-Pacific and Japan).
It’s not just about the hardware anymore. The neoclouds that deliver the full stack will turn today’s demand into a lasting business.
Q: Neoclouds have been riding the wave of growing AI demand in the past couple of years. What happens now with the recent selloff in neocloud stocks, as well as bonds for AI data centres?
A: The recent selloff reflects a fair question: how will billions of dollars of GPU spending turn into durable revenue? Demand for AI compute has not gone away. What investors want to know is whether neoclouds can earn more from that capacity than it costs to finance and run.
That depends on the economics of the hardware itself. A GPU is an expensive asset with a short shelf life, because every new generation of chips outperforms the last.
A provider that only rents out GPU hours needs that hardware busy around the clock, and it has to hold its prices as more capacity comes online. Capacity utilisation is a real problem, and an idle GPU is pure cost.
Hyperscalers faced a version of this years ago. They turned infrastructure into a business by building a rich software layer on top of it, so customers bought services rather than servers. Neoclouds now need to make the same move.
The ones that get more value out of every GPU they own by shifting from pure infrastructure to high-margin platforms will come out of this period stronger.
Q: Will open-weight AI models like DeepSeek make AI workloads less costly to run and therefore reduce the seemingly unquenchable thirst for AI compute?
A: Open-weight models will make AI more cost efficient to run, but I don’t expect them to cool demand for compute. The cost per token keeps falling, yet usage is growing much faster than costs are coming down.
A simple question to a chatbot uses very few tokens. Ask an agent to research and write a 10-page report, and it can use a hundred times more.
It’s the Jevons paradox: in the 1860s, the economist William Stanley Jevons observed that more efficient steam engines led Britain to burn more coal, not less.
We saw this at Nutanix when our engineers started using AI for coding. Usage exploded, and so did the bill. Instead of restricting use, we invested about US$20 million in our own GPU cluster to run open-weight models.
We expect it to handle up to 80 per cent of our internal AI needs, with frontier models kept for the most demanding work, and we expect the investment to pay for itself within a year. Our engineers can keep using AI as much as they need, without every request adding to a per-token bill.
That is the real effect of open-weight models. They change where AI runs and what each task costs, not how much AI organisations want to use.
Getting it right means knowing which model each workload uses and what it costs, so every job runs on the right model in the right place – whether on-premises, at the edge, or in a partner cloud.
Q:Given these challenges, what does Nutanix offer to neoclouds and other similar AI data centre operators to be ready for the future?
A: We help neoclouds and other AI data centre operators go from metal to margin. Most have already invested heavily in GPUs. What many of them lack is the software layer that turns that hardware into services customers will keep paying for.
Our platform provides that layer. With Nutanix Service Provider Central, an operator can onboard many customers onto shared infrastructure, give each one its own resources and quotas, and keep them securely isolated, while customers run and manage their own workloads.
With this Nutanix full stack software offering, Neocloud operators can offer GPU-as-a-Service, Kubernetes-as-a-Service, VM-as-a-Service and Tokens-as-a-Service from the same AI Cloud management platform.
We are also tackling utilisation so our customers have better operational efficiency and visibility. Our acquisition of Ryax brings AI-aware scheduling that places each workload on the most cost-effective hardware, and we are integrating this directly into our platform so operators waste less of the capacity they have paid for.
For the enterprises those operators serve, Nutanix Agent Gateway gives them one place to see and control which agents use which models, and what those cost.
The GPUs are the metal; and the services, utilisation and governance around them will be where the margin comes from.
