01
DEPLOYMENT
An on-ramp edge node installed in your data center, deployable in HA, with standard NNI and cross-connect handoffs.
ON-RAMP EDGE NODEHA PAIRSTANDARD NNICROSS-CONNECT
Hyperscaler-grade access to your GPUs, under your brand
White-label connectivity infrastructure for GPUaaS and NeoCloud providers. One private gVPC per customer reaches any cloud, any data center, any site and the internet, provisioned in minutes, end to end.
ANY SCALE · ANY LOCATION · ANY TIME
THE UTILIZATION PROBLEM
The GPUs are bought and paid for. The road to them decides how many of their hours you ever sell.
5%
AVERAGE GPU UTILIZATION TODAY
70%
UTILIZATION A CLUSTER NEEDS TO BREAK EVEN
$330K
LOST A MONTH WHEN IT RUNS AT 55%
Minutes
TO PUT A CUSTOMER ON YOUR GPUS
GPU utilization: Cast AI, 2026 State of Kubernetes Optimization Report, April 2026, tens of thousands of Kubernetes clusters on AWS, Azure and Google Cloud. Cluster economics: ModulEdge, June 2026, a 1,024 GPU H100 cluster, citing American Compute. Reach figures and the partner network are covered below.
THE CHALLENGE
Most providers hand customers SSH access, and stop there.
Everything the enterprise asks for next is a networking project.
$ssh user@gpu-cluster
✓connected · 8×H100 ready
>private path to our cloud VPC
✗one-off project · weeks of lead time
>IPsec tunnel to HQ
✗one-off project · weeks of lead time
>secure link to a partner cluster
✗one-off project · weeks of lead time
>another cloud
✗one-off project · weeks of lead time
#ssh is where your product stops
- 01Does not scale
- 02Engineering time that does not differentiate you
- 03Cannot compete for enterprise workloads
>one gVPC per customer
✓live · under your brand
>private path to our cloud VPC
✓live · provisioned in minutes
>IPsec tunnel to HQ
✓live · provisioned in minutes
>secure link to a partner cluster
✓live · provisioned in minutes
>another cloud
✓live · provisioned in minutes
#one gVPC, every path
- 01One on-ramp, one gVPC per customer
- 02Your engineers stay on GPUs
- 03Enterprise-grade access under your own name
Everything your customers want from the GPU, they can have.
Give them the way in, and nothing is out of reach.
The public clouds spent years building their own on-ramps. You should not have to.
Endless options. Whatever comes next, the AI Access Platform will carry it.
Humanoid robot training, Autonomous driving, GPU fleet monitoring, Medical imaging, Factory digital twins, LLM training and fine tuning, Drug discovery, Real time generative video, Financial market analytics, Low latency inference
WHAT IT DELIVERS
Five things your customers get on day one.
TODAY▶WITH THE PLATFORM
01
AI Access Platform
One white-label connectivity layer, run as your own. A private gVPC per customer reaches any cloud, data center, site or the internet, under your brand end to end.
TODAY▶WITH THE PLATFORM
02
Security: nothing touches what is yours
Policy-driven Security Groups replace per-site VPNs and firewalls. Every workload reaches only what it needs; full encryption, zero added hops.
TODAY▶WITH THE PLATFORM
03
Observability: know where everything stands
Live rate, loss, and completion on every link. Anything moving where it should not is flagged in real time.
TODAY▶WITH THE PLATFORM
04
Routing: one language, not ten
A single routing layer spans every cloud and carrier: Direct Connect, ExpressRoute, cluster BGP, and the wavelength beneath. The path is yours to set, and to prove.
TODAY▶WITH THE PLATFORM
05
Elastic: any scale, any time, pay for the job
Provisioned in minutes, gone at teardown. Hundreds of TB moved in a few hours, at 100G and 40G bursts, with cheap persistent links in between.
WHO IT IS FOR
Built for the provider, felt by the enterprise.
GPUAAS AND NEOCLOUD PROVIDERS
Offer private, enterprise-grade access under your own name without first building a network division. Connectivity stops deciding whether an enterprise workload lands with you or a competitor.
YOUR ENTERPRISE CUSTOMERS
Private, compliant paths from their sites and clouds into your GPUs, provisioned in minutes, under your brand, with the isolation and audit their security teams require.
GETTING IT ON YOUR NETWORK
An installation and onboarding, not a network build.
02
OPERATIONS
Yours to run, and this is the point: the platform sits on your network, not over the top of it. Optionally, InsidePacket and the underlying carrier operate it for you end to end: routing, security, observability, billing and day-2 support. Either way, your engineers stay on GPUs.
ROUTINGSECURITYOBSERVABILITYBILLINGDAY-2 SUPPORT
03
COMMERCIALS
White-label and usage-aligned: costs scale with the connectivity revenue you bill, with no capex program competing with GPU spend. Scoped to your platform in a working session.
WHITE-LABELUSAGE-ALIGNEDNO CAPEXWORKING SESSION
WHERE IT IS USED
Five enterprise use cases.
Every one of these needs private paths across several clouds, sites and partners at once. Today each is a one-off networking project, quoted and built per account. On the platform the customer turns it on in minutes, and tears it down when the job ends. Figures show the service model at work in representative cases.
01
Dynamic LLM training and inference
Elastic burst bandwidth for moving datasets and model weights between training and inference clusters across clouds. 600 TB lands in a 14-hour window at 100G, with a 1G persistent link held for the run; you pay for the window, not the year.
600 TB14 H WINDOW100G
02
Multi-cloud secure segmentation
Zero-Trust Security Groups isolate workloads across every cloud, data center, and site the customer connects. One endpoint can sit in several overlapping groups while blast radius stays per-group.
PER GROUP BLAST RADIUS
03
Disaster recovery and high availability
Continuous replication and automatic failover between geo-diverse clusters, no manual reconfiguration. A dedicated 10G link syncs state and model weights while the gVPC reroutes traffic the moment the primary fails.
10G SYNCAUTOMATIC FAILOVER
04
Collaborative multi-tenant access
Isolated ingress per data contributor, with separate read-only tiers for auditors or partners, no shared blast radius. Data moves to compute only after automated security scans, in fully isolated gVPC segments.
READ ONLY TIERSCANNED BEFORE MOVE
05
Edge-to-core low-latency routing
Aggregation and intelligent load-balancing of massive, distributed edge traffic into the right GPU cluster. Thousands of concurrent streams, with end-to-end latency targets under 50 ms.
UNDER 50 MS
WHY NOT BUILD IT, OR RESELL AN OTT PLAYER
The outcome of building it, without the build.
Building a comparable on-ramp internally means standing up multi-cloud interconnects, policy-based segmentation, encryption, telemetry, and intelligent routing as a standing engineering commitment, while the core business is GPU compute, not networking.
Reselling a labeled, over-the-top connectivity player instead trades that build cost for something worse: the provider's own brand disappears from the customer relationship, margin flows to the OTT player, and the provider loses control over the one layer, connectivity, that increasingly decides whether an enterprise workload lands with them or a competitor.
THE OUTCOME WITHOUT THE BUILD
The AI Access Platform gives the provider the outcome of building it themselves without the build: their brand, their customer relationship, their margin. InsidePacket's infrastructure underneath.
TIME TO A COMPARABLE SERVICE IN MARKET
BUILD IN HOUSE
ASSEMBLE FROM VENDORS
AI ACCESS PLATFORM
WITHIN 90 DAYS
TODAY1 YR2 YR3 YR4 YR5 YR6 YR
Bar length = time to a comparable service in market for those building it. Your deployment is an edge node in your facility, not a build: the platform beneath it is already in production (September 2025, US and EU, quarterly releases). Build and assemble durations are planning assumptions.
WHAT ENTERPRISES NOW EXPECT
The demand side is already there.
In a 2025 global survey of IT leaders, 58% of platform connections were specifically hybrid and multicloud, 62% ranked lock-in avoidance a top priority, and 71% called private connectivity essential for compliance. Six expectations now define the buy.
58% hybrid and multicloud
62% lock-in avoidance a top priority
71% private connectivity essential
Meet these six and the hard workloads follow. Multi-cloud segmentation, cross-cloud replication, partner extranets and edge aggregation stop being projects the enterprise has to fund and become paths it can turn on.
One control plane
One policy model across clouds, data centers and sites, instead of one per cloud.
Minutes, not months
Self-service and API provisioning, connections for as little as a day.
Low latency for inference
Private routes between data, GPU capacity and users.
Private paths
Cut egress fees and internet exposure at once.
No lock-in
Place workloads by price, region and strength, and move them when needed.
Compliance-grade segmentation
Zero-trust access, in-country routing, isolated partner extranets.
The gap this product fills
GPU clouds and neoclouds face exactly these expectations with no self-service on-ramp layer of their own. The largest still provisions private links by account manager, LOA and a cross-connect the customer orders, and reaches enterprises through partner fabrics. Across the market, agent access to networks today is beta, partial or absent; complete, production agent operability exists, as of August 2026, only on the platform beneath the AI Access Platform. Their build focus is GPUs, power and contracts, and rightly so. The on-ramp has to come from the network side.
THE READ
The buildout alone does not decide this market. Whoever packages distributed infrastructure into an on-ramp owns the customer relationship and the margin; the provider that does not risks selling capacity beneath someone else's platform.
THE STRATEGIC FRAME
Who owns the on-ramp to intelligence?
Public clouds did not win enterprise workloads on processors alone. They won on reach and operability: private connectivity in major metros, a consistent interface and a common way to consume networking, storage and identity. When AWS launched Direct Connect, it looked like a useful feature. It became part of the moat.
The same shift is beginning around AI infrastructure. Capacity now lives across hyperscale regions, GPU clouds, colocation halls and private clusters, and demand arrives in bursts.
The question behind this product is the strategic one: who owns the on-ramp? Providers that cannot expose their own and their partners' assets through a common, programmable model risk becoming interchangeable suppliers beneath someone else's customer relationship.
From "Who Owns the On-Ramp to Intelligence?", Eli Karpilovski, Founder and CEO, August 2026.
Read the letter
The letter publishes with the October announcement.
Ask us for an early copy at contact@insidepacket.com.