Better Models Don’t Kill Moats: Talking AI with Intel Capital’s Avi Bharadwaj
Every company building AI right now is wrestling with the same question: if the frontier models keep getting better, more general-purpose, and more accessible, what actually makes a company defensible? It’s the question that keeps founders up at night and determines where billions in venture capital get deployed.
Avi Bharadwaj is one of the people deploying that capital. As an Investment Director at Intel Capital, he focuses on the software infrastructure layer of AI, backing companies like Scale AI, Bria, TrueFoundry, and Twelve Labs. In this episode of Talking AI, he sits down with Matt Paige to share exactly where he sees moats forming, where he thinks they’re illusions, and how the landscape is shifting faster than most people realize.
It’s a Mistake to Think Better Models Kill Moats
Avi’s opening thesis is direct. Better models don’t destroy defensibility. They commoditize one layer of a much bigger puzzle. The model is one component of a sophisticated software platform, and as that layer gets commoditized, defensibility shifts to the other layers. He identifies five specific places where it’s showing up for application companies: unique and proprietary data (not just more data, but data frontier models don’t have), workflow and system of action (touching multiple enterprise systems across 50 to 100 steps), product reimagination (redesigning how work gets done rather than adding screens), integration depth (building connectors frontier labs won’t build), and trust and compliance (deploying at enterprise scale with the guardrails large organizations require).
Build Things That Improve as the Model Improves
One of Avi’s sharpest insights is his warning against building for the gap. Many startups identify something frontier models can’t do today and build a company around that capability gap. The problem is that gap is ever-shrinking. Instead, Avi advises founders to build products and platforms that get better as the underlying models improve. That’s what makes a company durable rather than a temporary bridge.
He also draws a meaningful distinction between bottom-up and top-down adoption. Frontier model companies have won standalone, individual-adoption use cases like coding and design. But they haven’t cracked the broader enterprise use cases that require senior sponsorship, workflow complexity, compliance, and integration depth. That’s where startups still have the advantage.
The Chatbot Era Was Brief. Agents Are First-Class Citizens.
The conversation shifted from “how do we build chatbots that answer questions from our data” to “how do we build actual claims processing and revenue cycle management with agents.” That shift has profound infrastructure implications. Agents need memory architectures, enterprise-specific reinforcement learning, and systems designed for hundreds of tokens per second rather than the five tokens per second humans read at. The entire search ecosystem, the UX patterns, the speed assumptions were all built for human users. Agents are a fundamentally different user class.
How a VC Actually Uses AI Every Day
Avi breaks the VC workflow into four buckets: seeing, picking, winning, and supporting. AI has transformed the first two. He has an agent on Claude Cowork that reads research papers, technical blogs, Hacker News, Reddit threads, and Discord channels overnight, identifies what companies are being discussed, looks them up on PitchBook, and enters them directly into his CRM. By morning, he has a report of what’s emerging and what the technical community is saying. Financial modeling that used to take hours now takes minutes using Claude Cowork within Excel.
World Models and Emergent Abilities
Avi is personally most excited about world models. His framing is that language is an abstraction of reality, and a lot is lost in that translation. Everything you can see, feel, touch, and smell gets simplified when it’s encoded into language. World models could open up something currently not possible at this level of abstraction. He’s particularly curious about emergent abilities. Just as no one expected LLMs to be good at reasoning or coding, he wonders what unexpected capabilities might emerge as world models scale. Could they understand causality? Could the laws of nature be encoded and emerge from training?
The Coolest and Most Overrated Things About Being a VC
The episode closes with Avi’s candid take. The coolest part is the front row seat to the smartest founders building amazing things, and the constant intellectual curiosity that the job demands. The most overrated part is the power people ascribe to VCs. At best, he says, VCs are sidekicks helping founders achieve their goals. At worst, they’re detractors. Most VCs aren’t as powerful as the outside world thinks.
Key Moments
00:00 — Matt’s intro: the moat question that defines this era
01:41 — “It’s a mistake to think better models kill moats”
03:25 — Workflow and integration moats
06:27 — Jack Dorsey’s “Hierarchy to Intelligence” thesis
09:55 — From data scientist to AI investor
13:12 — How a VC uses Claude Cowork agents daily
18:58 — Specialized models vs. the ever-shrinking gap
22:45 — “Build things that improve as the model improves”
25:43 — Agents are first-class citizens now
31:41 — “Don’t use models like if-else loops”
35:08 — World models and emergent abilities
38:34 — When will robots hit mainstream?
43:10 — Coolest and most overrated things about VC
Watch or Listen to the Full Episode
Watch or listen to the full episode to hear where an AI investor writing checks into the infrastructure layer actually sees defensibility showing up, and why building for the gap between what models can and can’t do is a losing bet.
Connect with Avi Bharadwaj on LinkedIn | Learn more about Intel Capital
