2023 – 2024

Thunder Compute (YC S24)

How do you get strangers to trust each other to share expensive compute hardware?

A terminal window renting a GPU with one Thunder command: “Your GPU is connected, Bhargs. It’s that easy.”

I turned an early GPU marketplace concept into a product that strangers could trust. As Thunder’s first designer, I shaped the marketplace, its trust architecture, and the MVP that helped secure investment from Y Combinator and angel investors.

Type
Startup, backed by YC (S24)
Duration
Nov 2023 – Apr 2024
Team
2 founders and me
Role
Early product designer
Stage
Pre-funding

Overview

So, why does compute scarcity deserve to be solved?

Students, indie developers, and small startups are paying the price for AI compute scarcity. GPUs currently require large upfront commitments to cloud providers such as AWS and Azure, even though the hardware largely sits idle while they’re reserved.

Meanwhile, capable GPUs sat idle in homes and labs. Thunder’s idea was to connect the two. That opens a market, but crowd-sourced resources come with multitudes of trust-related issues.

  • Seed raised$700k

    From YC and angels

  • Timeline6 mo

    Concept to MVP

  • Core flows3

    Setup, renter, host

  • After handoffIntact

    Baseline flows kept

What did I work on?

  1. 01

    A trust model for a peer-to-peer market

    I replaced direct renter and host contact with a black box (a network-mediated system) that held both sides accountable and reduced the potential for fraud.

  2. 02

    A payments barrier caught early

    My research exposed crypto wallet setup as a primary adoption barrier, so we moved to traditional payments before building it.

  3. 03

    Foundational servicing models

    The baseline service flows supported the seed raise and largely survived the handoff to an external design team.

Research

Striking a balance between anonymity and trust.

Design during the founder stage was more systems-design forward than I’d expected. It didn’t involve much visual creation. My role was figuring out exactly what it takes for a stranger to hand over money or hardware, then designing the system that makes that complex interaction real.

When I joined, Thunder was a concept and a crypto payment model borrowed from Helium[1]Helium rewards people with tokens for contributing hardware to its network.. Hosts would earn tokens for lending compute, and renters would pay through a blockchain-backed mechanism. My first job was figuring out what problem we were actually solving, and for whom.

Instead of comparing features, I compared how each platform makes strangers willing to transact. AWS sets the bar for reliability but buries it in enterprise setup. Vast proves crowdsourced supply works but exposes strangers to each other directly. Helium and Render make contributions verifiable, but crypto onboarding scares off renters. That pointed Thunder to a clear position: peer-to-peer prices, with the network as the trusted middle.

Competitive analysis comparing Render Network, Vast.ai and AWS on structure, payments, trust and safety, and ease of use
Competitive research pointed Thunder to a clear position: peer-to-peer prices, with the network as the trusted middle.
Comparison of trust approaches across platforms

Trust model

We put GPU sharing in a black box.

Anonymity felt like a risk: anonymous users can commit fraud and vanish. But direct contact between strangers was the bigger one.

So I made renters and hosts anonymous to each other and fully accountable to the network. Credits are verified before a job runs, hosts are paid once completion is recorded, and neither side ever sees the other.

Before and after. Before, direct peer-to-peer: no guarantee the job runs, no guarantee the host is paid, no one is accountable. After, network mediated: credits verified before a job runs, jobs routed through the network with identities never shared, host paid once completion is recorded.

Service flows

I took extreme measures to simplify users’ roles in the Thunder network.

We took Thunder from concepts of an involved marketplace app to a simple terminal and web dashboard.

01

Thunder provides GPU compute to renters.

Green arrows show the renter’s actions. We distilled their responsibilities down to two:

  • purchase Thunder credits
  • run programs using Thunder credits
Renter service flow diagram
02

Hosts share GPU compute with Thunder’s network.

Teal shows the host’s actions. We distilled the host’s role to two actions as well:

  • connect their GPU to the Thunder network
  • get paid
Host service flow diagram: a host registers their GPU, the Thunder host app announces capacity to the network, jobs run and are recorded in a black box, and the host redeems a payout from the Thunder console

Payments

Funneling users through crypto would create a high-friction barrier.

The original model asked renters to set up a crypto wallet before running anything. Mapping the renter flow end to end made the cost visible: the setup steps outweighed Thunder’s price advantage, and users said they’d pay more elsewhere for a payment they already understood.

I raised this with the founders early, and the team moved away from crypto as the sole payment method before building it.

Renter flow showing crypto wallet setup steps

Learnings

What did I learn?

  1. 01

    Design the system before the screens.

    At the founder stage, design isn’t wireframes first. It’s deciding what has to be true for a stranger to hand over money or hardware, then building the system that makes it true. That question runs through all my work, from fraud prevention at Truist to here.

  2. 02

    Anonymity can build trust.

    Anonymous users felt like the risk, but direct contact between strangers was the bigger one. Keeping both sides anonymous to each other, and accountable to the network, made the market safer than introducing them.

  3. 03

    Raise friction before it’s built.

    Mapping the renter flow end to end showed crypto setup outweighing Thunder’s price advantage. Raising it with the founders early meant the team changed course before building it.

Where Thunder is now.

The MVP I helped establish helped the founders raise $700k and go full time on Thunder. As the company matured, it shifted away from crowd-sourced compute because of the network’s unreliability, but the baseline flows I designed have mostly stayed intact.

thundercompute.com ↗

Get in touch

Want to learn more about the project? Any thoughts?

bhargs@bhargs.com
Next project Disability accommodations for Canvas Student Explore