TechCrunch Disrupt 2026

Nader Khalil & Sydney Sykes on What AI Can't Commoditize

Nader Khalil & Sydney Sykes· Director of Developer Tech and Global Head of VC Partnerships, NVIDIA at NVIDIA
·~31 min·English·TechCrunch
AgentsAI SafetyAI InfrastructureOpen SourceBusiness Strategy
TL;DR

Two NVIDIA leaders argue that when AI lets anyone build anything, the only durable advantages are the ones it cannot copy: velocity, genuinely hard problems, proprietary and physical data, and the founder's own judgment.

01Fundraising

Raise for Compute, Not Headcount

<strong>The money is no longer for hiring.</strong> When AI collapses the cost of building, venture capital exists to buy compute and tackle genuinely hard problems, not to grow a payroll.

today, if you don't need money for hiring the way you do five years ago, what do you need it for? You need it for compute for sure.

— Sydney Sykes, TechCrunch Disrupt 2026
Key Insight
The shift from headcount to compute quietly changes who wins a round: a lean team with a clear compute thesis now reads as stronger than a large team with a vague plan, because payroll is no longer the proxy for ambition.
02Operating Principle

The Highest-Velocity Team Wins

<strong>The highest-velocity team wins.</strong> When the landscape shifts this fast, your own priors are the trap, so the edge goes to whoever can discover and discard ideas fastest.

the team that has the highest velocity is the team that's going to win.

— Nader Khalil, TechCrunch Disrupt 2026
Key Insight
He names the tension himself: stability and velocity pull against each other, and he comes down on velocity, because it is the churn, not the polish, that keeps uncovering the opportunities a shifting landscape throws up. The quieter implication is that a team optimizing for a stable, reliable product may be optimizing for the wrong thing this cycle.
03The VC Reality

A VC Won't Turn Your Ship Around

<strong>The founder and team are the engine, not the investor.</strong> A VC can add go-to-market, capital, and compute, but it cannot rescue a weak team, which is why the sharpest VCs now chase talent and step out of the way.

a VC is rarely going to be the game changer for a company.

— Sydney Sykes, TechCrunch Disrupt 2026
Key Insight
Her point quietly reframes what a VC is for: not the engine of a company but a supplier of go-to-market, capital, and compute to a team that can already build. It is why, as she describes, the sharpest investors now scout research conferences for talent rather than backing the proven ex-CEO.
04Agent Security

An Agent Should Only Do Two of Three Things

<strong>Give an agent any two powers, never all three.</strong> An agent can touch your data, reach the internet, or write and run new code; allow all three and it can exfiltrate anything, so the limit must be enforced in a runtime the agent cannot reach.

It can access the internet and it can build new capabilities for itself and run them. And you should only let an agent do two of those three things.

— Nader Khalil, TechCrunch Disrupt 2026
Key Insight
Treating an agent as a system rather than a model relocates safety from the lab to the runtime. His framing is pointed: the public conversation fixates on model alignment, yet the controls that actually contain a deployed agent sit a layer lower, in its tools, permissions, and a runtime it cannot rewrite.
05Build Decisions

Start Closed, Then Specialize

<strong>Closed models are a discovery tool you graduate out of.</strong> Use a general frontier model to find what is possible, and once a task is clearly in-domain, move it to an open-source model you can make faster, cheaper, and your own.

Once you know that a task is in domain, then it makes sense to move to an open source model. You're going to get performance improvements. It could be faster. It could be cheaper.

— Nader Khalil, TechCrunch Disrupt 2026
Key Insight
Framing the migration as a lifecycle dissolves the open-versus-closed loyalty debate. Closed models become the exploration layer you pay a premium for only while you are still learning what you are building; the moment a task is understood, keeping it on a general API is just renting generality you no longer need.
06Data Moats

The Unclaimed Data Is Physical

<strong>The next moat is the data no one recorded.</strong> The internet's text and images are already harvested, so the open frontier is the industrial, real-world data that was never collected because no one thought to.

there's a lot of valuable forms of data. The physical world is probably one of the areas Nvidia is most interested in.

— Nader Khalil, TechCrunch Disrupt 2026
Key Insight
Pointing at the physical world as uncollected implies the next moats are less likely to come from scraping more of the internet than from instrumenting places it never reached. If that holds, the edge may tilt toward teams with real-world operations to record, over pure-software plays drawing on the same public corpus as everyone else.
07Durability

Only Hard Problems Are Moats

<strong>Easy problems get copied no matter how fast they reach revenue.</strong> Durability comes from solving something hard that genuinely matters, because an important problem that is also difficult leaves few others able to catch you.

it really comes down to just one problem, one question. Is this a hard problem to solve? If this is an easy problem to solve, then a lot of people are going to solve it.

— Sydney Sykes, TechCrunch Disrupt 2026
Key Insight
Fast revenue is a trap the old funding playbook still rewards. The implication is that the metrics that close a round, quick traction, and the metrics that build a company, a hard and important problem, have come apart, and a founder now has to choose which of the two games they are actually playing.
08Forward Look

The 18-Month Roadmap Is Dead

<strong>A long, detailed plan now reads as naivety.</strong> No one can see a year of AI ahead, so the signal investors want is not a roadmap but proof that you build and iterate now.

I went to a pitch recently where a founder showed an 18-month road map and the VC was like yeah I'm out.

— Sydney Sykes, TechCrunch Disrupt 2026
Key Insight
Killing the long roadmap is not only about speed. It signals that the artifacts that used to prove competence, a detailed multi-year plan, now actively hurt a pitch, because they advertise that the founder has misjudged how fast the ground is moving underneath them.