a16z Podcast

Jesse Zhang & Ashwin Sreenivas on Why the Moat Isn't the Model

Jesse Zhang and Ashwin Sreenivas· Co-founders of Decagon at Decagon
·~80 min·English·a16z
AgentsInferenceOpen SourceAI CompanyBusiness Strategy
TL;DR

Decagon's founders argue that in the agent era the durable moat isn't the model but the software and process around it — fine-tuned 'model factories,' encoded business logic, and the machinery that makes frontier capability deployable inside the enterprise.

01Core Mental Model

The False Trade-Off

A fine-tuned small model doesn't just trade intelligence for cost — on one narrow task it beats the frontier model on quality, speed, and cost at once.

So when we fine-tune smaller, dumber models, it's that they're just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models

— Jesse Zhang, a16z Podcast
Key Insight
The industry frames model choice as smart-but-expensive versus dumb-but-cheap. Decagon's data says that's a false axis: narrow the task enough and the 'dumber' model wins on every axis at once — which is why roughly 90% of their production workflow now runs on open source.
02How They Build

Decagon Labs Is a Model Factory

They built an internal pipeline whose only job is to shrink the gap between a new model's release and a fine-tuned, task-specific version running in production.

we find ourselves constantly training net new models and deprecating old ones that are no longer relevant because you know maybe the frontier has advanced a lot

— Ashwin Sreenivas, a16z Podcast
Key Insight
Most companies treat model choice as a one-time decision. Decagon treats it as a conveyor belt, because a model tuned last quarter is deadweight this quarter once the frontier moves. The durable asset isn't any single model — it's the speed of the factory that keeps replacing them.
03Industry Structure

Open-Source Share Is Falling — For a Good Reason

Even amid open-source hype, the share of open-source inference is dropping right now, because every new use case starts on frontier models and only migrates once it's proven.

at a certain point it's strictly better to use open source models because when your use case is solidified and you're in production at scale

— Jesse Zhang, a16z Podcast
Key Insight
Rising open-source usage and falling open-source share can be true at the same time. New frontier-model use cases are being spun up faster than old ones migrate down to open source — so the hype and the share numbers move in opposite directions, and neither reading is wrong.
04The App-vs-Infra Debate

Even AGI Needs Somewhere to Put Things

The 'labs will be the last startups' narrative misses that business logic lives in the application layer — and even AGI agents need software to store, retrieve, and reason over.

Even once you have AGI, agents are going to need somewhere to store work and pull information from and reason about things. I don't think software as a whole in any meaningful way is going away.

— Jesse Zhang, a16z Podcast
Key Insight
Their bet reframes the app-versus-infra question. A model smart enough to rebook three passengers off a cancelled flight still needs the encoded rule for how your company rebooks — and that rule lives above the model, not inside it. Labs generalize; the application layer is where vertical-specific logic accretes.
05The Org Method

Forward-Deployed: Eat Pain, Excrete Product

Forward-deployed engineers exist to learn brand-new workflows nobody has run before — but the discipline is to productize what they learn so the next ten customers get it for free.

forward deployed engineers eat pain and excrete product

— Ashwin Sreenivas, a16z Podcast
Key Insight
Palantir popularized forward deployment; Decagon's twist is treating it as a temporary state, not a business model. The phrase — from Palantir CTO Shyam Sankar — captures the rule: a forward-deployed engineer who never productizes isn't scaling a product, they're a consultant with better branding.
06The Product Magic

Duet: The Agent That Builds the Agent

Duet is a second, bigger, slower agent whose whole job is to write the procedures, tools, and tests for the customer-facing agent — and then monitor its conversations.

It's like a second agent that's much bigger, much slower, but its job is to do all the tasks I just described.

— Jesse Zhang, a16z Podcast
Key Insight
The magical part isn't that Duet automates support — it's that a general reasoning model, never trained on Decagon's authoring tasks, turned out to be good at all of them at once. That only became possible when frontier reasoning models, built mostly for coding agents, got good enough to also run the meta-work.
07The Durable Moat

The Moat Is Deployability, Not Intelligence

Even if every model were perfect, raw capability far outruns what enterprises can safely use — the moat is the software that makes a model governable, testable, and deployable inside a regulated org.

the capability of models today is far greater than they are being used for within the enterprise

— Ashwin Sreenivas, a16z Podcast
Key Insight
Assume perfect models tomorrow and Decagon's moat still doesn't evaporate. You can't hand a model the keys to the whole enterprise: you need to say what it may and may not do, let hundreds of experts shape its behavior, and test it against regulatory lines. That connective software is a glass box, not a black box — the reason a recent customer left a rival.
08The Human Impact

It May Kill Jobs, But Not Careers

Automating support 30% cheaper rarely means firing 60% of the team — there's more latent demand for good customer service than supply, so companies do more of it instead.

kill jobs but not careers in a way because like those jobs that are being done currently should not be done by humans

— Jesse Zhang, a16z Podcast
Key Insight
This is Jevons' paradox in the labor market: make a valued service cheaper and demand expands to fill the new supply. One customer handling 50,000 tickets a month didn't cut staff — they put support on every page and opened it to free users, because latent demand had always outstripped what they could afford to serve.