NVIDIA’s AI Grid vision is clear: inference should run where latency, cost and data rules fit best, and telcos already own the sites and connectivity to put it there.
Just two things missing: a way to remove the GPU cap from your smaller sites, and the business layer needed to commercialize your AI Grid. With hosted·ai you get both.
hosted·ai’s GPU Hypervisor brings full GPU virtualization to your NVIDIA sites. It pools GPUs, gives every tenant secure, isolated vGPUs, and multiplies the workloads, revenue and ROI each GPU supports.
Illustrative four-GPU edge site
In most GPU clouds, each model is pinned to its own GPU. The number of paying tenants at a site is capped by its GPU count, which is capped by power, especially at cable hubs and RAN sites. Models that outgrow their GPU have nowhere to go, and capacity a model doesn’t use is stranded.
Same four GPUs, pooled
With hosted·ai, every GPU joins a pool. Tenants run on vGPUs drawn from the pool, not on a particular card.
Every model has a guaranteed floor, and compute follows demand: busy models take the cycles idle models aren’t using and hand them back as traffic falls. Memory is partitioned per tenant with hard limits, so sharing stays secure.
In most GPU clouds, each model is pinned to its own GPU. The number of paying tenants at a site is capped by its GPU count, which is capped by power, especially at cable hubs and RAN sites. Models that outgrow their GPU have nowhere to go, and capacity a model doesn’t use is stranded.
With hosted·ai, every GPU joins a pool. Tenants run on vGPUs drawn from the pool, not on a particular card. Every model has a guaranteed floor, and compute follows demand: busy models take the cycles idle models aren’t using and hand them back as traffic falls. Memory is partitioned per tenant with hard limits, so sharing stays secure.
hosted·ai runs on the RTX PRO 6000 and GB300 sites NVIDIA’s AI Grid specifies. No new hardware.
Dynamo, NIM, TensorRT-LLM and partner placement engines keep their jobs. We add utilization, products, billing and demand.
Unmodified CUDA. Tenants see a GPU or an API, never the hardware.
Multi-Instance GPU, time-slicing and Multi-Process Service each give up flexibility, isolation or fairness. Multi-tenant inference at the edge needs all three. hosted·ai delivers by pooling GPUs and dynamically scheduling multi-tenant workloads to maximize output per GPU.
| Approach | Models per GPU | Isolation | Changing shares | Busy model uses idle compute |
|---|---|---|---|---|
| Dedicated GPU | One | Full | Buy a GPU | No |
| MIG | Fixed slices | Hardware | Destroy and recreate | No |
| Time-slicing | Several | None for memory | Round robin | Yes, unprotected |
| MPS | Several | Address space only | First come | Yes, can starve others |
| hosted·ai GPU Hypervisor | Up to 10 vGPUs per card | Memory isolated per tenant | Continuously, by demand | Yes, with a guaranteed floor |
hosted·ai turns your NVIDIA sites into GPU-as-a-Service products that sell, bill and pay out, under your brand. It also connects your AI Grid to a global demand network, bringing you buyers for tokens and rentals from day one. NVIDIA’s reference design doesn’t change.
Buyers
Sell tokens, virtual GPUs, on-demand, short-term and long-term capacity from the same fleet, dedicated or shared. Change the mix as demand moves, site by site.
The token exchange and hosted·ai marketplaces bring buyers at launch, served on your capacity under your brand. Your own customers always come first.
Every token and GPU hour is metered per tenant, site and GPU. hosted·ai collects from exchange buyers and pays you out automatically.
Enterprise demand for edge inference is the growth vector for AI Grid. Inference platforms need capacity now, at peaks and in regions where they have none. hosted·ai enables the token exchange, an inference marketplace between operators, so you serve your clients, fill idle capacity, and have access to more when your sites are full.
Operators invented roaming. The home network keeps the customer; the serving network gets paid. The token exchange does the same for inference.
Book your AI Grid review: a 30-minute call on your sites, your workloads and the products your grid could sell, covering commercials, technology or both.
We use cookies for analytics and advertising. Privacy policy