The GPU-as-service cloud market is evolving rapidly. By now, it’s common knowledge that many so-called neoclouds aren’t efficiently utilizing their GPUs. Often, they buy too many GPUs per customer in an effort to cover all necessary processing, only to end up with the costly components sitting idle, particularly in inference workloads with bursty traffic.
That’s the problem startup hosted.ai is out to solve. “A customer who needs half a GPU now, three in a few hours and ten in the overnight batch has to buy ten GPUs, because they must provision for the peak,” said hosted.ai Co-founder and CEO Ditlev Bredahl, in a recent interview with Futuriom. “The other 80% to 90% of what is deployed sits idle.”
The heart of hosted.ai’s solution is software that virtualizes GPUs so that they can be shared by multiple tenants in an oversubscribed way, with tenants using portions of a single GPU as workloads increase or decrease with processing needs.
It’s a model whose time has come, Bredahl said. “A customer may have many GPUs, but a GPU never has more than one customer,” he noted. “Nowhere else in infrastructure does this happen. CPUs are virtualized, RAM is ballooned, storage is thin provisioned, and bandwidth is oversubscribed because providers know not everyone uses everything at once.”
“VMware for Generative AI”
Hosted.ai’s founders term what they’ve developed as virtualizing GPUs in the way VMware virtualizes CPUs. They’ve had plenty of tech experience to draw on: CEO Bredahl and James Withall, now hosted.ai’s COO, built UK2, a European web hosting platform akin to what GoDaddy has been in the U.S.
After selling that company to Lloyds Development Capital (LDC), the private equity arm of Lloyds Banking Group, in 2011 for about US$77 million, Bredahl and Withall, along with the execs who would become the co-founders of hosted.ai, went on to build a cloud management platform called OnApp. Their mission was to provide software to CDNs, telcos, carriers, and MSPs that enabled them to create their own public clouds to compete with AWS. According to Bredahl, OnApp garnered more than 3,000 service-provider customers and was eventually sold in 2021.
Getting More Processing Out of GPUs
Hosted.ai hasn’t released any independent testing results, but it claims some dramatic efficiency improvements due to its software, as measured on hosted.ai’s own neocloud, packet.ai: A GPU billing 730 hours per month for single tenancy GPUs can bill 1,825 hours a month on hosted.ai’s platform. “That is 2.5X the sellable hours per card, and in pure on-demand scenarios 2X to 4X the revenue per GPU,” Bredahl told us. “Utilization moves from the 10% to 15% typical of inference to the 80% to 90% range the rest of infrastructure runs at. We call the goal zeroidle.”
Getting better capacity out of GPUs also means being more efficient with power, which hosted.ai sees as the real constraint for AI. Filling GPUs to 80% or 90% capacity adds a bit to the power bill but multiplies the revenue-generating capability significantly.
Virtualization and orchestration of GPU resources also helps service providers get up and running faster. “A provider is online in days rather than weeks,” Bredahl said. “We’ve done the heavy lifting and bring a neocloud in a box.”
The Mesh: Selling GPU Capacity
According to hosted.ai, its platform can free up about 70% of existing GPU capacity. That capacity becomes a salable item. And hosted.ai takes charge of selling ittar Alliance for GPUs. “Our line to providers is: You bring the capacity, we bring the cloud and the customers,” said Bredahl
So far, hosted.ai expects to generate between $250 million and $350 million in Mesh revenue this year, with over $500 million in the pipeline. Not bad for a service that started in March 2026.
Hosted.ai also has a plan called Coldstart, in which it helps neoclouds to start up in locations where land, power, and shell are available but vacant. “We find the anchor tenants, put the GPU financing in place with our financing partners, source the hardware with NVIDIA and its OEMs, deploy the stack, run the cloud and the 24/7 NOC, sell the capacity through the Mesh…. and manage the end customers,” said Bredahl. “The site owner keeps the lights on and sees a check. We take a management fee.”
Hosted.ai sees a bright future for service providers looking to virtualize GPUs and make money on the capacity this frees up. “Demand has barely started,” Bredahl said.
Startup Profile: hosted.ai
Headquarters location: Fremont, California
Founders: Ditlev Bredahl (CEO), Narendar Shankar (Chief Commercial/Revenue Officer), Julian Chesterfield (CTO), James Withall (COO)
CEO: Ditlev Bredahl (ex-packet.ai, GPUaaS.com, KickAss Capital)
Target markets:Service providers, neoclouds, and telcos looking to provide rapid, flexible, and billable GPU services with maximum utilization.
Prominent investors: Creandum, People Ventures, z21 Ventures, Golden Sparrow, Tekton, Repeat, Hersir Ventures, Nivesha Ventures
