Audit your siteBook a call
Brand identity + product

Thaw: a brand for serverless GPUs that cost nothing while they wait

Book a call, get a price range

Thaw runs AI models on GPUs only while they're working: every model sits frozen at $0 when idle, and the first request thaws it onto a GPU in 0.41 seconds. We built it from scratch: the name, a world of frozen machines with one running hot, a wide wordmark, the console and the CLI, and the brand on posters, a hoarding and a conference booth.

Project
An Alterra Studio project: we built the company's brand, product and launch from scratch
Stage
AI infrastructure, Series A stage
Services
Naming, world and key art, wordmark, product UI and CLI, posters, out-of-home and event pieces, animated board
Tools
Imagine (our in-house design app), Nano Banana Pro, Kling, HyperFrames
Brief

The problem and the brief

Inference startups sell on speed and scale, then open on an abstract render. Not one shows a cold model answering, or what it costs to sit idle.

Thaw is serverless GPU inference. Every model a team deploys sits frozen at $0 while nobody calls it; the first request thaws it onto a GPU in 0.41 seconds, and it's billed per GPU-second while it works. Its buyer is the ML platform lead or CTO at a Series A or B AI product company, paying for GPUs that wait all night.

The brief, in a founder's words: "Our homepage talks about AI infrastructure and scale when an ML platform lead wants to see one cold model answer a request, with the cold-start time and the idle bill, and we lose them before they book a call."

What the buyer is up against

The research

17.7%

of a cloud's model-serving GPUs sat with rarely called models that served only 1.35% of requests.

Source: Alibaba, SOSP 2025

84 s

to load Llama 2 70B onto 8 GPUs from local SSD: one cold start, done the slow way.

Source: ServerlessLLM, OSDI 2024

83%

of enterprises running their own GPUs report 50% utilisation or less.

Source: VentureBeat Research, 2026

Three doubts an AI team brings to the first call

  • CTOWe still pay for GPUs that sit waiting all night.
  • ML platform leadCold starts will make our users wait and leave.
  • Their userThe app hangs on the first request of the day.
Logic

The strategy: show the thaw and the bill

Every screen shows one model going from frozen to serving and back, with the time and the cost on it. The world says it without words: a row of frozen machines, one running hot.

  1. Doubt 1 · Idle cost?We still pay for GPUs that sit waiting all night.

    Frozen means $0.00

    The console lists each deployment with its GPUs and its state. Frozen rows read $0.00 an hour; when a model goes back to sleep, a toast closes the account: 3 requests, 18 GPU-seconds, $0.01, back to $0.00 an hour.

    83% of enterprises running their own GPUs report half capacity or less. The idle bill is the line that sells.

  2. Doubt 2 · Cold starts?Cold starts will make our users wait and leave.

    The thaw, measured on screen

    When a request reaches a frozen model, the row counts the thaw in real time, 0.03 s of 0.41 s, while the machine's fins heat through to molten orange. Orange appears only in that state, so hot means one thing.

    A cold start the slow way took 84 seconds for one large model. Showing the 0.41 s answers the doubt directly.

    The console serving: thawed in 0.41 s, with the GPUs and the cost per hour
  3. Doubt 3 · First request?The app hangs on the first request of the day.

    One command, one answer

    In the CLI the whole story fits on one screen: deploy, a cold request, thawed in 0.41 s, a 200 in 1.19 s, then frozen again. Developers read the product in its own medium.

    17.7% of one cloud's serving GPUs sat with models that saw 1.35% of requests. Freeze those and pay only for the answer.

System

The design system

Ice grey, graphite and aluminium for everything at rest. Molten orange only for a machine that's running, so the colour itself means "working, and billing".

  • Ice grey #D3D8DAFrozen: grounds and idle surfaces
  • Graphite #2A2C2EType, ready machines, dark grounds
  • Aluminium #8D9194Cooling, secondary type
  • Molten orange #FF6B1AServing. Never a button, never the logo

The world. A row of identical machines asleep under frost. When a request comes in, one heats through to orange, does the work, and cools back to grey.

Wordmark. Wide, even-weight capitals traced as drawn, calm and machined, set big on posters and small in the console.

Type. Hanken Grotesk for what people read, Fragment Mono for what machines print: commands, times, costs.

What we left out. All 10 competitors put a picture where the product should be, 9 open on an abstract render, and 7 on a near-black screen. None uses orange. Thaw's first screen is the console.

Thaw key art: a row of frozen machines, one glowing hot orange
The 10-second board, made for LinkedIn and X.
The world in motion: one model thaws while the rest stay frozen.
Poster: stop paying for GPUs that wait
Poster: 0.41 s from frozen to serving
Poster: idle $0.00
A Thaw poster on a street hoarding
A graphite vest with the Thaw wordmark
A Thaw conference lanyard badge
The Thaw booth at a developer conference
The Thaw wordmark on ice grey and graphite
One machine running hot, glowing orange
Social

The brand in the feed

A mockup of how the brand would run on X. Not a live account.

Thaw avatarFollow

Thaw

@thawgpu

Serverless GPU inference. Frozen at $0 while idle, serving in 0.41 s on the first request. Billed per GPU-second.

thaw.runJoined October 2026

93 Following7,120 Followers

Pinned

Thaw @thawgpu · Oct 13

Meet Thaw. Stop paying for GPUs that wait.

Thaw @thawgpu · Oct 14

thaw invoke doc-extract-ft cold → thawed in 0.41 s → 200 in 1.19 s …and back to $0 when it's done.

Thaw @thawgpu · Oct 16

Idle: $0.00. Find us at the booth.

Poster: Idle $0.00, working billed per GPU-secondThe Thaw conference booth
Method

The category audit

We scored 10 funded serverless GPU startups on our startup homepage benchmark. Not one shows a cold start or the idle cost on the first screen.

Each homepage was scored on 12 criteria, 0 to 2 each, at 1440 and 390 pixels wide. The group averages 15.1 out of 24, above the 14.1 of our wider benchmark, and every one shows some customer proof.

The gaps are on the first screen. 0 of 10 show the product working there at both widths, 1 names who it's for, and 2 state a specific difference. Those are the gaps Thaw's homepage is designed to close.

Homepage score out of 100 (12 criteria, 0 to 2 each): 10 funded serverless GPU startups, October 2026
  1. Company A75
  2. Company B71
  3. Company C71
  4. Company D71
  5. Company E67
  6. Company F67
  7. Company G67
  8. Company H54
  9. Company I50
  10. Company J38
  11. Benchmark average, 100 startups59

Alterra Studio startup website benchmark 2026. Companies are anonymised.

The offer

What we'd ship

For a real AI infrastructure company, this is a Brand + Website Build.

Brand
Name check, world and key art, wordmark, palette and type, and the rule for the one orange
Homepage
A first screen that shows one model thaw and answer, with the time and the idle bill, at desktop and phone width
Product
The console and the CLI in the brand
Launch
The animated board, posters, out-of-home and the conference booth
After this

What's next

  1. Build the homepage with the thaw as the first screen at 1440 and 390, and score it on the same benchmark.
  2. Draw the mark a symbol that works at 16 pixels, for the app icon and the avatar.

The thinking behind it.

Guides on the decisions behind work like this, written by the team that did it.

All insights

Want this for your company?

Book 30 minutes. We'll tell you what we'd change first in your brand and on your first screen, and what it would cost.

Book a call, get a price range

Curious how your site compares? Run the same audit on your site

Take the 3-min scorecardBook a 30-min call