Big Technology

Renen Hallak on why your IP is becoming model weights

Renen Hallak· Co-founder & CEO at VAST Data
·~43 min·English·Big Technology
AI InfrastructureInferenceAgentsBusiness Strategy
TL;DR

VAST Data's Renen Hallak argues that as company value shifts from people and data to model weights, trust has to move from promises to math: encrypt the weights, run inference on-prem, and confidential computing can unlock the enterprise data AI has not touched yet.

01Core Mental Model

The Game Is Done

Hallak's inflection point is a machine solving a problem no human could, which he argues proves AI now creates new IP instead of remixing what we already knew.

For the first time, it's solved something that no human was able to solve and people were trying for the last 90 years.

— Renen Hallak, Big Technology
Key Insight
Notice how carefully he hedges: 'if they really did solve it.' The entire optimistic case rests on that one unverified result being real. Strip the conditional and the inflection point is just a claim.
02The Thesis

Life, Liberty, and Weights

As all IP turns into model weights, the classic right to property has to extend to them, or a handful of companies end up owning everything.

And property, you want this IP to be owned by everybody.

— Renen Hallak, Big Technology
Key Insight
If IP becomes weights, and weights copy perfectly and invisibly, then ownership cannot be enforced by law alone. It has to be enforced by the infrastructure, which is exactly the business he is in. A values question quietly becomes a product spec.
03The Problem

Force Against Force

Enterprises will not expose their data and labs will not leak their weights, so AI adoption is stuck in a force-against-force standoff.

It's a very clear force against force of enterprises with their data.

— Renen Hallak, Big Technology
Key Insight
This is a chicken-and-egg on trust: neither side moves first, so adoption stalls at a fraction of the available data. Whoever supplies the neutral guarantee in the middle captures the position both sides depend on.
04The Solution

Encrypt One Side, Not Both

You do not have to encrypt both sides: lock the model weights in hardware, run inference on the enterprise's own premises, and neither party sees the other's secret.

And you don't even need to encrypt both sides. Just one of the sides needs to be encrypted.

— Renen Hallak, Big Technology
Key Insight
'Encrypt one side' is a go-to-market insight as much as a technical one. You can convince a handful of model labs; you cannot convince millions of enterprises. Who is easy to sell to ends up dictating the architecture.
05Architecture

The OS Above the Models

Models become just another resource, and the operating system layer, not the model, is what manages them.

I think we're realizing now that the models are not above the operating system but rather the operating system needs to manage these models.

— Renen Hallak, Big Technology
Key Insight
Putting the OS above the models is a claim about who captures the value. Whoever manages the resource has historically owned the platform, so he is arguing the model is not the platform, the layer that governs the models is.
06The Flip

A Million Agents

In a few years an enterprise is a million fine-tuned agents plus its data, so the fear of leaking weights flips from the labs onto every company.

5 years from now, 10 years from now, that enterprise will have a million agents and the only thing that those agents are worth are the models that they've fine-tuned uh throughout their experience. And so that becomes the IP of the enterprise.

— Renen Hallak, Big Technology
Key Insight
When an enterprise's value becomes its weights, the security anxiety that lives at the labs today migrates to every company. That is a far larger market than protecting a few frontier labs, and it is the one that does not exist yet.
07What's Next

An App Store for Skills

Once continual learning works, each agent's experience becomes a modular skill you can lease from its owner, an app store for real-world abilities.

Can I lease that module from you? But you own that module because you spent the effort of training your agent on how to do carpentry.

— Renen Hallak, Big Technology
Key Insight
Leasable skill modules only work if you can prove provenance: which module came from where, trained on what. The app store needs a control plane that tracks lineage, which is the same OS layer he is selling. The vision and the product are one thing.
08Adoption

A Human in the Loop

Keep a human in the verification loop: checking an answer is far easier than finding it, so AI can do the hard part while people stay accountable.

If you keep a human in the verification loop and you let the AI find the fraud, then I don't think anyone would object because we have a human.

— Renen Hallak, Big Technology
Key Insight
Human-in-the-loop is framed as safety, but it is also what makes adoption legal today: the human is the accountable party a regulator needs. Removing that human later is not just a capability milestone, it is a liability question he flags and sets aside.