Turn your fiber into the on-ramp every GPU cloud sells through.
The AI Access Platform, run as your wholesale product. Each GPU and neocloud provider becomes a tenant on it, which means their own brand, their own portal and their own customers. The scope is the access layer alone, on your network, live within 90 days.
ONE PLATFORMMANY PROVIDERSEVERY BRAND THEIR OWN
gVPC PER CUSTOMER
gVPC PER CUSTOMER
gVPC PER CUSTOMER
The compute is landing. The road to it is not built.
Capital is going into AI compute faster than any operator has seen it move, and every build lands as its own island. The providers behind those islands cannot give an enterprise what the public clouds gave themselves: a private path in, from any cloud, data center or site, in minutes. So they hand out SSH access and stop there, and everything the enterprise asks for next becomes a one-off networking project quoted per account.
GPU UTILIZATION, AGAINST WHAT A CLUSTER NEEDS
100%1 Cast AI, 2026 State of Kubernetes Optimization Report, April 2026, across tens of thousands of Kubernetes clusters on AWS, Azure and Google Cloud. 2 ModulEdge, June 2026, a 1,024 GPU H100 cluster, citing American Compute. Two different populations, shown on one scale to give the distance a size.
WHY THE GAP IS THERE
Weeks OF LEAD TIME PER PRIVATE PATH TODAY, QUOTED PER ACCOUNTEvery enterprise that cannot reach the cluster privately is utilization that never arrives. Every week of lead time on a cross-connect is a week the capacity waits. The compute is already bought and paid for, so the road to it decides how many of its hours ever get sold.
The distance between 5 and 70 is the demand for an access layer, and it is not a networking problem the GPU providers are going to solve. Their build focus is compute, power and contracts, and rightly so. The on-ramp has to come from the network side.
PROOF OF THE MARKET, AND A WARNING
A listed NaaS Platform pure play has already read it this way and put an on-demand GPU pool on top of the access platform it spent a decade building. Treat that as proof of the demand rather than as competition. It is over the top: it runs on fiber it does not own, and the brand and the margin travel up to it.
WHAT MAKES THIS DIFFERENT FOR AN OPERATOR
Not a NaaS Platform build. One access layer, many brands, live in a quarter.
You are not buying this to serve your own GPUs. You are buying it to become the layer every GPU and neocloud provider in your market sells through. Three things separate that from any other network product on your roadmap.
NEITHER LEVEL SEES OURS
One platform. Every provider. Each under its own brand.
Tenancy runs two levels deep. You stand the layer up once and operate it as your wholesale product. Each GPU and neocloud provider joins as a tenant, which means its own portal, its own domain, its own catalog and its own pricing, then brands the whole thing again to its own enterprise customers.
Your name is on the contract with the provider. Their name is on the contract with the enterprise. Adding the second provider is an onboarding, not a build.
The scope is the access layer, and nothing else.
No dedicated internet access. No data center interconnect. No wavelengths. No catalog of services to build, price, quote and support. One job: be the private path from an enterprise into a GPU or neocloud provider, turned up in minutes.
Everything a full catalog would have cost you in product, billing and operations is out of scope by design. That is what makes the number a quarter instead of five years.
Start with a few sites. Grow when a provider asks.
A high availability on-ramp edge node pair, in facilities you already run, handed off over a standard NNI. It sits on your network rather than over the top of it, and it is live within 90 days.
The footprint follows provider demand instead of a capital plan. No capex program sits on your roadmap waiting for volume that has not arrived.
WHAT YOU HAND EACH PROVIDER
Six things each GPU provider can sell on day one.
This is the product a GPU or neocloud provider puts in front of its own enterprise customers, and it carries their brand end to end. Nothing here is a project they have to scope, and nothing here is a service you have to build.
A gVPC per enterprise customer
Where a cloud VPC is private to one region of one cloud, a gVPC is private to one enterprise customer and reaches everywhere that customer works: any cloud, any data center, any site and the internet.
Security Groups instead of per-site VPNs
Policy-driven groups replace per-site VPNs and firewalls. Every workload reaches only what it needs, with full encryption and zero added hops. One endpoint can sit in several overlapping groups while blast radius stays per group.
Live rate, loss and completion
Telemetry on every link, at flow level. Anything moving where it should not is flagged in real time, which is the evidence an enterprise security team asks for before it moves a regulated workload.
One routing layer, not ten
Direct Connect, ExpressRoute, cluster BGP and the wavelength beneath, spoken as one language across every cloud and carrier. The path is theirs to set, and to prove.
Elastic capacity, billed by the window
Provisioned in minutes, gone at teardown. Hundreds of TB at 100G and 40G bursts, with cheap persistent links in between. Their customer pays for the window, not the year.
NEO, so an agent can order it
Their customer describes two endpoints in plain language and NEO discovers, prices and turns up the path. Everything the portal does is also an API, aligned to MEF LSO Sonata and Cantata and TM Forum TMF622 and TMF641.
Every screen, every API response and every invoice in that list carries the provider's brand. The enterprise never learns your name, and it never learns ours.
SCOPE
The whole product, in one table.
Most platform conversations with an operator start with a catalog and end with a programme. This one does not have a catalog. Here is everything that is in, and everything that is deliberately out.
- A high availability on-ramp edge node on your network
- A gVPC per enterprise customer, per provider
- Zero-trust Security Groups
- Live rate, loss and completion on every link
- One routing layer across clouds and carriers
- Elastic bandwidth, minutes to provision, gone at teardown
- A white-label portal, API and agent per provider
- Billing records per provider, per customer, per service
- Dedicated internet access
- Data center interconnect
- Wavelengths and dark fiber
- A full platform service catalog
- A network build or a new footprint
- A capex program
- GPU capacity, power or contracts
- Anything that sits between a provider and their customer
Out is not missing. Out is the reason this is a quarter and not five years. Every line on the right is either a product you already sell under your own name, or one nobody needed you to build for this. The access layer is the only piece the market is short of, so it is the only piece here.
HOW IT RUNS
An installation and an onboarding, not a network build.
Three questions decide whether this is a quarter or a programme: what goes in the ground, who runs it, and how it is paid for.
An edge node in a facility you already run
A high availability on-ramp edge node pair, handed off over a standard NNI or a cross-connect. White-box service nodes, commodity hardware, lightweight demarcation at the customer, and no forklift anywhere in your network.
It sits on your network, not over the top of it. Start with the sites your first provider needs, and add sites as providers ask for them.
Yours to run, or ours on your behalf
Run it with your own NOC and your own processes, the way you run the rest of your wholesale book. Or InsidePacket operates it for you end to end: routing, security, observability, provider onboarding, billing records and day-2 support.
Either way, your engineers stay on the network you already run and no new operating model lands on them.
Usage-aligned, with the margin on your side
A per site or unlimited platform license on commodity hardware, with operator margin in the range of 50 to 70 percent. Costs scale with the access revenue you bill your providers, so the layer pays for itself as they grow on it.
No capex program, and nothing competing with the rest of your capital plan. Scoped to your network in a working session.
You already sell access to competitors, price it, provision it and support it. A multi-tenant access layer is that muscle pointed at a new class of wholesale customer.
TIME TO MARKET
Five years, three years, or this quarter.
Building a comparable access layer internally means standing up multi-cloud interconnects, policy-based segmentation, encryption, telemetry and intelligent routing, and then keeping a standing engineering commitment behind all of it. Assembling one from vendors trades the build for an integration programme. Neither finishes inside the window.
TIME TO A COMPARABLE SERVICE IN MARKET
THE WINDOW: 2027 · AI ON-RAMPS BEING CLAIMED
What the 90 days actually contains
The edge node, the handoffs, the branding, and the onboarding of each provider's portal and APIs, which speak MEF LSO Sonata and Cantata and TM Forum TMF622 and TMF641.
The number is days rather than years because the deployment is an installation, not a build. The platform beneath it has been in production since September 2025 in the US and EU, with quarterly releases.
And what the second provider costs
An onboarding. Their brand, their domain, their catalog, their pricing and their billing records, configured on a layer that is already carrying traffic for the first provider.
The expensive one is the first, and it is the one priced in the 90 days.
Bar length is time to a comparable service in market for those building it. Your deployment is an edge node in a facility you already run, not a build. Build and assemble durations are planning assumptions, not measured outcomes.
WHY THIS LANDS WITH YOU
You already own the two things an over-the-top player has to rent.
The fiber and the metros are the first, and they are the obvious one. The second matters more: a wholesale motion that already sells access to competitors, prices it, provisions it, bills it and supports it. That is the exact muscle a multi-tenant access layer needs, and almost nobody outside a carrier has it.
| THE ACCESS LAYER | AN OVER-THE-TOP PLAYER | YOU, ON THE PLATFORM |
|---|---|---|
| Whose fiber the path runs on | Yours, rented from you | Yours |
| Whose brand the enterprise sees | Theirs | Your provider's |
| Where the access margin lands | With them | With you and your provider |
| In-country and regulated routing | Best effort, off-net | On-net, and provable |
| Who owns the customer relationship | They do, and they keep it | Your provider, with you underneath |
Scroll the table sideways on a narrow screen.
The two ways this goes wrong
Building a comparable access layer internally is a standing engineering commitment in a category that is not your core business, and it does not finish inside the window.
Reselling a labeled over-the-top service trades that build cost for something worse. Your brand disappears from the relationship, the margin follows it, and you lose control of the one layer that increasingly decides whether an enterprise workload lands on your network or on someone else's.
The AI Access Platform gives you the outcome of building it yourself without the build: your network, your providers, your margin, and a brand those providers put their own name on. InsidePacket's technology underneath, and nobody in the chain sees it.
Whoever packages distributed infrastructure into an on-ramp owns the customer relationship and the margin. For an operator the risk was never losing one GPU deal. It is carrying the traffic beneath someone else's platform, on fiber you paid for.
The compute is being built now, and the providers behind it are choosing their access layer in the next few quarters. Standing one up is a quarter of work on a network you already own. Watching one get chosen is permanent.
Scope a deployment for your network in one working session.
Bring your first candidate provider, the sites you would start with, and your wholesale pricing. You leave with a footprint, a timeline and a commercial shape you can take to a board.