NewSovereign AI means your rules travel with your workload.Read the blog
← All blogs

Blog

Can I Borrow Your GPU Without Lending You My Data?

Abilash Menon

CEO & Co-founder · November 17, 2026 · 6 min read

The Printer Was Inside the Building

Remember the intranet?

If you were on the office LAN, you were inside the company. Inside meant trusted. The file server, the shared drive, the guy from accounting who replied-all to everything. And the printer in the hallway. Nobody knew who managed it.

Everybody knew its IP address.

Then we opened a second office.

The internet could connect the two buildings. The CISO had one follow-up question: “Who else is on the way?”

I’ve spent most of my career working on answers to that question. At Cisco I worked on MPLS forwarding. Customers thought they were buying labels. Nobody has ever fallen in love with a label. They were buying a boundary: their traffic, kept apart from everybody else’s, on a service someone was accountable for.

Carrier Ethernet and Layer 2 VPNs went further and made the branch feel like it was down the hall.

Then I went to 128 Technology to work on SD-WAN. For a lot of enterprises, routed and encrypted overlays across whatever links were handy became the default again. Broadband, LTE, MPLS, the cable modem behind the loading dock.

Networking never throws anything away. We add a layer and call it a transformation.

We used to ask whether a branch office could feel like it was down the hall. Now the branch office has GPUs.

“I Found the GPUs”

“I found the GPUs,” says the head of IT.

“Where?” asks the security chief.

“That’s the exciting part.”

“For which one of us?”

Same purchase, two different movies. IT sees the workload running next month at a better price. Security sees a new operator, a new jurisdiction, new admins, new logs, and a review that starts on page one.

They’re both right. That’s the gap I want to close.

The next advantage in cloud isn’t who has the most GPUs. It’s who lets you move your AI workload without rebuilding trust every time.

I want enterprises to be able to move an AI workload to a better GPU, a better price, or a better location without starting the security review from scratch. And when the primary link goes down on Saturday, I want to know exactly where the prompt went and where it could be read.

That’s what sovereign AI means. Not flags on a data center. Your rules, traveling with your workload.

In Fairness to the Hyperscalers

The big clouds earned their trust. They gave security teams a familiar place to live and a decade of audits to point at. Plenty of workloads should stay exactly where they are.

But AI made the shopping list longer.

The GPU you need, in the region you’re allowed to use, on the date you need it, is a surprisingly specific product. The cheapest GPU-hour can still produce an expensive answer once you add utilization, connectivity and data movement.

Sometimes the hyperscaler wins that math. You should be allowed to check.

The model endpoint is one line of code. That line has a lot of friends: identity, pipelines, observability, egress. None of them enjoy moving.

Meanwhile there are dozens of neoclouds that can win on capacity, price, location or hardware. What they can’t ship in a rack is a decade of trust.

Trust takes years. The GPUs are available Tuesday.

Friday: The Move

So let’s move a workload.

Say you run a customer-support assistant on a hyperscaler. A neocloud has the GPUs you need, at a better price, next month instead of next year. In the old world, that’s a quarter of work: new identity, new firewall rules, a public endpoint nobody loves, and a security review that opens with “who are these people?”

Here’s how I’d rather it go.

The application calls a private inference endpoint using the company’s own DNS. At the enterprise edge, a device uses SD-Ethernet to extend a dedicated Layer 2 segment of the company’s network to an isolated handoff at the neocloud.

The network picks a primary path and a backup, and only from routes the customer has approved.

That’s the piece we’re building at MaiaEdge. Our Path Border Controller (PBC) connects the two ends. Our Path Computation Engine (PCE) picks the paths.

The application keeps the same private service name and the same network policy it had yesterday.

To your application, the GPU is another room on your network. Somebody else owns the building. You still own the door.

The CISO has five questions. So do I.

  1. Data: Where can my prompts, outputs, logs and backups go, and how long can they stay?

  2. Jurisdiction: Whose rules apply at each step, and which operators and destinations did we approve?

  3. Operations: Who can administer, change, inspect or recover the service? Even in your own data center, do you know what root on the inference host can see?

  4. Keys: Who can authorize decryption, and where can my data become readable?

  5. Paths: Which routes can my traffic take, including when the primary route fails?

Show that all five meet your policy, with evidence, and you can defend the claim. Answer them with adjectives and you have a slide.

The CISO looks at the diagram. “Fine. What happens if that line goes down?”

I like this question. It usually means I get to unplug something.

Saturday morning, I pull the primary link.

We send another prompt. It takes the backup we approved. Not whichever path the network happened to find. The one with the CISO’s signature on it.

I pull that link too. This time, the request times out.

A timeout is an unusual thing to demo. I’m pretty happy with this one.

If your policy says the traffic stays in-country, the proof has to cover the carrier route underneath too. A path you can’t see is a path you can’t approve.

The link test tells us where the prompt went. The next test is where it can be read.

A prompt can take the approved road and still be readable by someone we never approved. That comes down to who holds the keys.

Here’s what I’d test with a neocloud. Can we protect the prompt all the way from the enterprise to the GPU running the model, so it becomes readable only inside the approved confidential environment? And if that environment can’t prove it’s the one we approved, does the customer’s key service say no?

Two things people like to blend together. Post-quantum key establishment helps protect recorded traffic against future quantum attacks. Key policy decides who can decrypt it right now.

Quantum-safe encryption with someone else’s keys is a very good lock on a door someone else can open.

Here’s what I want us to be able to show a customer: where their data becomes readable, and who is allowed to read it.

Same Test, Different Building

The inference factory.

Some teams skip running their own stack and rent inference as a service.

Then the question gets personal: I rented the GPU. Did I accidentally donate my prompts?

A prompt can leave exhaust: traces, debug logs, crash dumps. Attestation tells me what’s running. I still want to know what it logs.

The border.

A European company wants American GPUs for some of its workloads. The GPUs can be American. The data needs to check its passport.

Restricted workloads stay home. Approved ones cross on approved operators, with keys the company holds, after the lawyers confirm the legal basis.

In my last blog I said every nanosecond has a lawyer. Here, every byte does.

Where This Goes

I want the next GPU purchase to be a decision about capacity, cost and location.

If a neocloud can prove it meets the customer’s rules, it should get a shot at the workload. And if a better option shows up six months later, the customer should be able to take it.

MaiaEdge builds the path and the proof of where it went. The customer keeps the keys. That split is the design.

That gives the CISO a way to say yes to a new provider, and gives the neocloud a way to earn that yes.

If the workload moves on Friday, the rules should go with it.

The Printer, Revisited

The GPU can live in somebody else’s building. To your application, it can feel like another room on your network.

The difference this time is that we can show who has the keys, what can get out, and which door the traffic takes when the first one is locked.

We never asked this much of the printer.

We probably should have.

Subscribe for the latest news

Written by

Abilash Menon

CEO & Co-founder, MaiaEdge

Follow on LinkedIn →