Conway Research has launched an invite-only beta of Underdog, a personal AI assistant that runs entirely on a user's own device. The software currently supports Macs and Windows 11 PCs, with versions for Linux, iPhone, and Android in development. The announcement, reported by TechCrunch on October 6, positions Underdog as a privacy-first alternative to cloud-based assistants and a departure from ad-supported or subscription-based revenue models.
A Privacy-First, On-Device Architecture
Underdog encrypts the account keys for services such as email that users choose to connect, and all AI inference is processed locally rather than in a data center. This approach reduces exposure of sensitive personal information and removes reliance on rented cloud computing. The design reflects the company's core emphasis on privacy and user control over their own hardware.
Founder Sigil Wen developed a proprietary inference engine called Husky to support this local processing. Husky is designed to improve speed by reducing data movement between a computer's CPU and GPU. Underdog currently uses a 27 billion parameter inference model fine-tuned from Qwen 3.8 27B, and Wen said it performs comparably to Claude Opus 4.6 on certain benchmarks for everyday tasks such as shopping research and math homework.
An Unusual Revenue Model
Underdog is being offered free at launch and will not carry advertising, according to Wen. Because the AI runs on users' devices, the company avoids the large inference costs associated with cloud-based assistants. Instead, Conway Research is exploring a plan to collect small fees on transactions that the assistant processes over the Stripe payments network.
Stripe co-founder Patrick Collison is an angel investor in Conway Research, lending credibility to the payments-focused approach. Andreessen Horowitz led the company's first round through partner Chris Dixon, with participation from Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund. The round size was not disclosed, and no transaction fee percentage has been published.
Founder Background and Market Context
Wen taught himself to code and moved to Silicon Valley at 17, where he lived in an AI hacker house with Andrej Karpathy. He later worked alongside Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown, and he is currently a Peter Thiel Fellow. This background places him within a network of influential AI builders and investors.
Underdog enters a consumer AI market where major assistants are pursuing different answers to the question of who pays. Meta's Muse sits on an advertising business, ChatGPT charges a subscription and a purchase fee, and Instinct has raised capital without publicly naming a price. Underdog's model removes both advertising and subscription, relying instead on transaction fees that grow only when users spend.
Revenue Potential and Challenges
Analysts note that a transaction-based fee can generate significant revenue only if the assistant handles a large share of a user's spending. At a 1% cut of average annual online retail spending, the fee would total roughly US$36 per year, or about US$3 per month. That falls far below a typical US$20 monthly AI subscription, so Underdog may need to become a default payment method.
Wen's own examples include shopping research and math homework help, activities that do not always end in a transaction. The money would arrive only when the assistant moves from advising to actually paying on the user's behalf. This makes Underdog less a narrow shopping tool and more a fintech bet that users will trust it with everyday purchases, including larger spending categories beyond retail.
Underdog's invite-only beta represents a distinct experiment in consumer AI, combining on-device privacy with an unconventional transaction-based revenue model. Its success will depend on technical execution, user trust, and whether the assistant can handle enough real-world purchasing to replace subscription or advertising income. For now, the launch provides an early look at how AI assistants might be funded without compromising user privacy or adding subscription fatigue.