CNBC

Aravind Srinivas on why the harness — not the model — is the product

Aravind Srinivas· CEO of Perplexity at Perplexity
·~61 min·English·CNBC
AI InfrastructureInferenceOpen SourceBusiness StrategyAgents
TL;DR

Perplexity CEO Aravind Srinivas argues the AI product is no longer the model but the harness around it: open-weight orchestrators that route to a frontier model only when a task needs it, run cheaply per watt on hardware you own, over proprietary data you keep.

01Core Mental Model

The Harness Is the Product

The single best model no longer wins on its own — value has moved to the orchestration system that wraps a model in tools, context, and an eval loop.

the model alone is no longer the product. It is the harness, the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools.

— Aravind Srinivas, CNBC
Key Insight
Once a harness can call any model, the frontier model demotes to a sub-agent — a tool the orchestrator invokes, not the whole product. That is why Srinivas thinks the durable advantage belongs to whoever owns the harness, not whoever ships the top-ranked model this month.
02The Technical Bet

Bake the Routing Into the Weights

Instead of a separate router that picks models blindly, Perplexity post-trains an open model so the skill of escalating to a frontier model lives inside the weights.

embed that routing skill deeply inside the model weights itself instead of building like standalone routers that have no knowledge of what the model is capable of inside the agent harness.

— Aravind Srinivas, CNBC
Key Insight
A bolt-on router chooses a model without knowing what that model can do, so it guesses. By training the escalation skill into the weights, the orchestrator itself decides when it is out of its depth — turning the open-versus-frontier question from a config setting into a capability the model learned.
03The Open-Model Bet

Opus-Grade at a Third of the Cost

A post-trained open model (GLM, ~700B params with ~40B active) paired with a frontier advisor reaches Opus-grade quality for roughly one-third of what Perplexity paid Opus.

we post trained it on task that our customers care about in our product and our agent harness and then deployed it in production today and have lowered the cost like one-third to what we're paying Opus

— Aravind Srinivas, CNBC
Key Insight
The claim is not the tired one that open models are cheaper. It is that a cheap open base, post-trained on the tasks that actually pay and allowed to escalate, matches the best closed model where it counts. Srinivas is careful to note GLM generalizes to production tasks — not just the benchmarks that get hill-climbed.
04The Metric That Matters

Count the Value, Not the Tokens

The scoreboard is not raw token volume — a search query and an agent finishing a paid job both count as tokens but differ wildly in worth, so Srinivas optimizes for value per watt per user.

I actually think what matters more is the valuable tokens more than the number of tokens.

— Aravind Srinivas, CNBC
Key Insight
This is his quiet rebuttal to the panel's headline — Benchmark's Peter Fenton predicted 90%-plus of tokens will come from open-weight models within 18 to 24 months. Srinivas's point is that raw token share is the wrong yardstick; a Gemini search and an agent closing a deal are not the same token, and value per watt folds cost, power, and usefulness into one number.
05The Hardware Future

AI Is the New Computer

Put a unified-memory box like Nvidia's DGX Spark on every desk and the local model becomes the computer — running the harness, calling your apps, and reaching to a server-side frontier model only as a remote tool.

DGX Spark is such an amazing piece of hardware because it unifies the memory across the CPU and the GPU inside that local device unlike a data center where the memory between the CPU and the GPU are not unified.

— Aravind Srinivas, CNBC
Key Insight
The pitch to a nervous non-tech CFO is not run a data center — it is buy a PC for each employee, admin-controlled like any corporate laptop. By reframing local AI as a fleet of managed endpoints instead of infrastructure to operate, Srinivas turns sovereignty from a risk into a familiar procurement decision.
06The Open-Source Case

Affordable or It Doesn't Diffuse

Open weights can be served at the price of compute instead of the labs' markup — which Srinivas frames as the only way AI's benefits reach small businesses instead of staying locked inside a few frontier labs.

If you want the benefits of AI to be widely distributed to small businesses in America and American allied countries, then you really need AI to be a lot more affordable.

— Aravind Srinivas, CNBC
Key Insight
Srinivas ties affordability straight to policy: he is not worried about a backlash against Chinese open models because he expects Nvidia's Nemotron line to close the gap, and he argues America should compete on capabilities rather than export controls. His template is Linux and Android — open software everyone patched and trusted, not locked away.
07Data Sovereignty

Your Tokens Are Your Moat

An enterprise's value lies in its proprietary tokens and tacit knowledge, so Srinivas argues you should own your weights, deployment, harness, and improvement loop instead of relying only on frontier labs.

the value of any enterprise is in the unique proprietary tokens you have.

— Aravind Srinivas, CNBC
Key Insight
This lands Srinivas in the louder faction — Palantir's Alex Karp and Benchmark's Peter Fenton have made the same argument publicly. The technical version: your switching cost between models is only the price of re-hydrating your context, so owning that context in your own weights and harness loop keeps it compounding for you instead of leaking to a vendor who also sells your rivals.
08The Economic Reframe

The Bottoms-Up Boom

A single-person, roughly $3M-a-year company spending about $800K a year on agents shows where AI value lands: bottoms-up with small businesses, not top-down enterprise seats.

The amount of money we make from a company with sub 10 employees is sometimes even higher than companies where we've sold hundreds of seats.

— Aravind Srinivas, CNBC
Key Insight
Srinivas's close inverts the enterprise-sales playbook: value should accrue to the user, so the winner is whoever delivers the most output per token — which lets a tiny firm out-spend a big seat deal. The SaaS per-seat model, in his telling, is legacy; the real market is the $3 trillion in small-business payroll AI can make more productive.