Underdog’s 'Free' Promise: A Sustainable Pivot or a VC Vanity Project?

Opinion: In the current AI gold rush, the pitch deck usually follows a predictable script: massive compute costs, venture-backed subsidies, and an eventual, inevitable pivot to a monthly subscription. Then comes Sigil Wen with Underdog, as TechCrunch first reported.

Wen, a Thiel Fellow and former collaborator with the likes of Andrej Karpathy and Aravind Srinivas, is positioning Underdog as the 'most private' AI assistant in the valley. The technical hook is impressive: the model runs entirely on-device via Wen's proprietary Husky inference engine, meaning the user's hardware—not a centralized data center—bears the compute load. As reported by TechCrunch, this architecture allows Wen to avoid the crushing overhead of inference payments that plague his competitors.

But as a skeptic of 'free' in Silicon Valley, I find the business model more precarious than the technical one. Wen tells TechCrunch that because his costs are 'so super low,' he doesn't need to charge a subscription. Instead, he is borrowing a fintech play. Backed by Stripe co-founder Patrick Collison, Underdog intends to take a tiny percentage of payment transactions the AI makes using Stripe’s secure rails—essentially an interchange fee.

On paper, this is a seductive vision. By eschewing ads and subscriptions, Wen claims the AI is aligned with the user, much like a bank or credit card provider. He argues in his 'AI manifesto' that users shouldn't have to surrender private information to access AI. By running a 27-billion parameter reasoning model (fine-tuned from Qwen3.8-27B) locally, Wen claims the tool is as capable for everyday tasks as Claude Opus 4.6 was six months ago.

However, the P&L reality is this: can a 'tiny percentage' of transaction fees actually sustain a company backed by the heaviest hitters in venture capital? Conway Research—the startup behind Underdog—isn't just funded by angels like Guillermo Rauch, Noam Brown, and Deedy Das. It has the full weight of Andreessen Horowitz (via Chris Dixon), Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund (a partnership between Anthropic and Menlo Ventures).

When you see that level of institutional capital, 'free' usually means 'subsidized until we find a way to charge you.' While on-device inference removes the marginal cost of a single query, it doesn't eliminate the massive costs of R&D, talent acquisition, and scaling a business. Relying on transaction fees assumes a high volume of commercial activity mediated by the AI. If users primarily use Underdog for 'math homework questions' or 'shopping research' without completing the purchase through the app, the revenue stream vanishes.

Wen is betting that privacy is a feature people will value enough to drive this specific type of commerce. It is a noble goal—Wen told TechCrunch he wants to build a product he'd be proud for his future children to use—but in the valley, nobility is rarely a sustainable revenue driver. Until we see the actual conversion rates on those Stripe transactions, Underdog looks less like a disruptive business model and more like a high-concept vanity project funded by the most patient capital in the world.