Aiolum Browse stories
All stories
Hosted inference API for open models · United States

Wally by RunAnywhere: Product, team and how it was built

Small teamFundedB2BDeveloper tools

Wally by RunAnywhere (Hosted inference API for open models, United States): Paid API with ten open models, October 2026. Evidence grade E: no revenue figure was checked. Team: 2 founders in 2025; 5 people on the Y Combinator page, November 2026. Launched 21 September 2026; still running.Added

Outcome
Paid API with ten open models, October 2026
Evidence
E no revenue figure was checked
Launch
21 September 2026
Status
Still running
Team
2 founders in 2025; 5 people on the Y Combinator page, November 2026
Outside funding
Y Combinator, Winter 2026 batch; amount not found
Group
Developer tools and coding agents
Main channel
Not stated
No revenue figure

What is not verified

No revenue figure was found. The pages read give prices and terms but no revenue, customer count or named paying customer. A reader could look for a payment-processor profile, a founder interview or a later press piece with figures. This story is grade E: the business is real and its product, price and customers are described from public pages, but no revenue figure was checked. Aiolum shows it as an idea worth studying, not as proof that it makes money.

Wally is a hosted inference service, and inference means running a trained AI model to produce an answer. Wally runs open-weight models, whose trained parameters anyone can download, on its own servers. A developer calls them through an OpenAI-compatible API, a web interface that programs use to send requests and read replies. RunAnywhere, Inc. makes Wally and also publishes open-source SDKs for running models on phones and laptops.

Sanchit Monga and Shubham Malhotra started RunAnywhere in 2025, according to the company's About page. The company went through Y Combinator in the Winter 2026 batch, and its first products ran models on devices. The company's blog says it launched Wally on 21 September 2026, and its Y Combinator launch post is dated 29 September 2026. On 22 October 2026 the company published a catalog of ten models, and customers buy prepaid credits through Stripe.

No source gives a revenue, customer or user count for Wally, so this story has grade E. Prices, dates and the launch claims come from the company's own pages. The Y Combinator page gives a team of 5 people. The page does not say how many customers pay.

Wally by RunAnywhere: Paid API with ten open models, October 2026. Hosted inference API for open models, United States, team of 2 to 10, evidence grade E.
The story in one card. Share it with the address of this page.
01

Why did Wally by RunAnywhere work?

Four causes, in the order they mattered. Each one is a reusable lesson.

01

The founders claim Wally serves the same models faster than other hosts

In the launch post, the two founders claim that Wally serves GLM-5.3 Flash at 380 tokens per second with a 274 ms first token. They cite Artificial Analysis, a benchmark site, for 28 September 2026. These are the founders' claims, not independent checks. Speed matters in loops such as coding agents, where each tool call waits for the first token.

02

It reuses engine work from the on-device products

The founders write that they spent nine months on engines for Apple and Qualcomm chips. They also write that none of that work was specific to phones, and that they pointed it at data-center GPUs. The company's blog says an agent pipeline writes the kernels, which are the low-level programs that run a model on a chip. The company says its engineers direct and check that work.

03

Customers pay per token from prepaid credits

The pricing page lists a rate per million tokens for each model. Customers buy credits through Stripe in a console, and each request debits the balance at the listed price. A request fails once the balance runs out. The terms add that the credits of a new account start with a US$5 grant.

How it was built

These details come from the company's own pages and blog. They are grade C. [2, 6, 12]

  • The API follows the OpenAI chat-completions format at inference.runanywhere.ai/v1, so existing OpenAI client code works after a change of base URL.
  • A command-line tool, wally, connects coding agents such as OpenCode, Claude Code and Codex to a model. Sign-in is in the browser with Google or GitHub.
  • Payments use Stripe. The privacy policy names PropelAuth for sign-in, Resend for transactional email and Vercel for the marketing site.
  • The company says AI agents write the kernels and the serving stack, and its engineers direct and verify their work.

Tools named in the sources

Each one opens its page in the Tools tab, with every story that uses it.

AstroBPropelAuthCVercelCStripeCResendC
02

Who tried this before Wally by RunAnywhere, and why did they fail?

Four companies in our database of failed startups sold developer infrastructure or AI inference to other builders. The values below are from that database and are not independently verified.

Died 2011 to 2017Burned US$7.0M

Parse

The promise

A backend service for mobile developers, with sign-in, push messages and storage in hours instead of weeks.

Why it died

Facebook bought it in 2013 for about US$85M and shut it down in 2017. The page names weak unit economics under freemium pricing.

Acquired, then shut down
Died 2008 to 2012Burned US$18.0M

DotCloud

The promise

A platform that hides the work of running cloud servers, so developers focus on code.

Why it died

The company pivoted to Docker, its own container technology. The page says the original service was overshadowed by it.

Pivot
Died 2018 to 2023Burned US$1.0M

Surge Cloud

The promise

A Polish cloud provider with lower latency, GDPR-native compliance and lower prices than the large US clouds.

Why it died

The page says it underestimated the capital and the competitive moat needed to challenge the large cloud providers.

Capital intensity
Died 2017 to 2025Burned US$118.0M

Rain AI

The promise

Neuromorphic chips for AI inference with far lower energy use than GPUs.

Why it died

The page says the company burned US$118M on a hardware platform before it had a software ecosystem that developers could use.

Hardware without a software base

The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.

What changed

Each of the four sold the infrastructure under other people's products. Each met a larger provider or a cost that small revenue could not carry. Facebook bought Parse and later shut it down. DotCloud moved to Docker, its own container technology. Surge Cloud and Rain AI needed more capital than they had.

Wally differs in what it owns. The launch post says Wally serves open models on the company's own GPUs. It does not build chips or a whole cloud. Its risk is also different. The launch post says Artificial Analysis lists 19 providers serving the same GLM-5.3 Flash weights, and the terms say prices can change.

03

Can one person copy Wally by RunAnywhere now?

Possibly, but only with funding and engine skill.

Hosting open models fast needs GPU capacity and kernel engineering, so a solo founder with no capital would struggle to copy it. A reader can study the sequence instead. Build open-source tools first, find what users still lack, and add a paid service for it.

Where it is still open: New open-weight models that other hosts serve slowly or not at all, and customers that need the same speed on their own GPUs.

  1. 1Publish open-source tools for a narrow job and watch which requests the tools cannot meet.
  2. 2Add a paid hosted service with a public price per unit of use.
  3. 3Connect to the tools your customers already use, such as coding agents, with one command.
04

How much revenue does Wally by RunAnywhere make?

Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.

A paid hosted API launched in September 2026, with prices and no revenue figure. Most rows are company statements
FactValueGradeHow we knowSources
Launch of Wally21 September 2026CThe company's blog post of 22 October 2026 says Wally launched on 21 September. RuntimeWire published its launch article on 21 September 2026.[6, 11]
Start of the company2025CThe About page lists 2025 as the year the founders started RunAnywhere. The Y Combinator page also says founded in 2025. The terms name RunAnywhere, Inc.[1, 4, 8]
Team5 peopleCThe Y Combinator company page gives Team Size 5. The About page names two founders and adds that a small team of engineers works with them.[4, 1]
FundingY Combinator, Winter 2026 batch; amount not foundCThe About page says the company went through Y Combinator in Winter 2026. No source read gives an amount.[1, 4]
Models in the catalogten, on 22 October 2026CThe company's blog post of that date lists ten models, with fast versions that have their own model ids.[5]
Price range, inputUS$0.042 to US$2.925 per million tokensCPricing page. Eve has the lowest input rate and Kimi K3 the highest. The highest output rate is US$15.00 for Kimi K3.[3]
Free credit for a new accountUS$5CThe terms call it a promotional welcome grant. The launch post says US$5 free on signup, no card.[8, 4]
How customers payprepaid credits through StripeCThe refund policy of 19 September 2026 says purchased credits are non-refundable except for a billing error, a duplicate charge or an outage.[7]
Service statuspreview, with no uptime commitmentCThe terms effective 26 October 2026 say the service is a preview with no uptime or support-response commitment.[8]
Data retentionzero by defaultCThe product page says Wally does not store prompts or replies. The console playground keeps recent conversations.[2]
GitHub stars of the on-device SDK repository10.3k stars, 416 forksCRead on the repository page on 11 November 2026. The repository holds the open-source SDKs, not the hosted service.[10]
Time from launch to the October 2026 fact1 monthcalcCalculated: 21 September 2026 to 22 October 2026.[6, 5]

How to read this table. Most values are statements on the company's own pages or on the Y Combinator page, so they have grade C. The star count was read on GitHub. The "calc" value is computed on this page. No value has grade A or B.

Wally prices per million tokens, read from the pricing page of runwally.com on 11 November 2026 (grade C)
PlanPer month, billed yearly (US$)Per month, billed monthly (US$)Limit
DeepSeek V4 Flash 0731input 0.293, output 0.88, cached 0.0093 per million tokens
DeepSeek V4 Pro (Fast)input 1.68, output 3.38, cached 0.14 per million tokens
DeepSeek V4.1 Flashinput 0.275, output 1.20, cached 0.006 per million tokens
DeepSeek V4.1 Flash 1017input 0.195, output 1.00, cached 0.006 per million tokens
DeepSeek V4.1 Flash (Fast)input 0.275, output 1.20, cached 0.006 per million tokens
Eve (decision model)input 0.042 per million tokens; no output or cached rate
GLM-5.3 (Fast)input 1.40, output 4.40, cached 0.26 per million tokens
GLM-5.3 Flashinput 0.15, output 0.50, cached 0.03 per million tokens
Kimi K3input 2.925, output 15.00, cached 0.30 per million tokens
MiniMax M3input 0.30, output 1.20, cached 0.06 per million tokens

The pricing page says that Wally uses prepaid credits, spent per million tokens. The table shows the input rate, the output rate and the cached input rate in US dollars. There is no monthly plan. The terms say that new accounts receive a US$5 welcome grant and that the service is a preview.

05

When did Wally by RunAnywhere launch, and how did it grow?

From a founders' company in 2025 to a paid hosted API in September 2026. Each entry carries its grade.

  1. 2025CThe two founders start RunAnywhere to run AI models on devices that people already own. [1]
  2. Winter 2026CThe company goes through Y Combinator and open-sources its on-device SDKs. [1]
  3. 21 September 2026CThe company's blog says Wally launched on this day, and RuntimeWire published an article about it. [6, 11]
  4. 29 September 2026CThe founders post a launch announcement on the Y Combinator site. It says sign up and get US$5 in free credits. [4]
  5. 22 October 2026CThe company publishes the Wally catalog of ten open models. Fast versions have their own model ids. [5]
  6. 27 October 2026CA company blog post says DeepSeek V4.1 Flash Uncensored is coming to Wally, with reviewed access for authorized security work. [9]

What could go wrong for Wally by RunAnywhere?

The risks are the mirror image of the causes above.

What is not verified

No revenue figure was found. The pages read give prices and terms but no revenue, customer count or named paying customer. A reader could look for a payment-processor profile, a founder interview or a later press piece with figures.

A preview service with a short history

The terms effective 26 October 2026 say the service is a preview, and that the catalog and pricing can change. Wally launched about seven weeks before this page was written.

Logo wall, counter and badges

The "Works with" lists on the home pages name nine coding agents and twelve platforms with links. They do not say that any of them pays. The 10.3K counter on the site counts GitHub stars of the on-device SDK repository. "Backed by Y Combinator" matches the Y Combinator page, and "SOC 2 Type II Audit in progress" is the company's own line. The About page lists press pieces from early 2026, mostly about on-device work, which were not read.

Security model with reviewed access

A company blog post of 27 October 2026 announces DeepSeek V4.1 Flash Uncensored on Wally for authorized security work. The Security programme page of runwally.com says a person at RunAnywhere approves each account. The acceptable use policy allows tests only on systems the user owns or has written permission to test. The terms say approved accounts pay 50% of the published price from 26 October to 30 November 2026. The model is not on the pricing page of 11 November 2026. The post calls the access review KYC, and the founder writes that he is not a security expert.

More in Developer tools and coding agents

Other businesses in the same group, best evidence first. All 116 stories in the group, 10 with grade-A revenue.

Faster, cheaper serving and post-training of open-weight AI models, sold as a per-token API · United States · Funded

Eigen AI

Evidence A
Sold for $643M

Started small, grew after funding · sold 11 months after launch

Its founders wrote some of the methods that the industry already used to run models cheaply, such as AWQ for 4-bit weights. They turned that research into a paid serving API. They proved the speed on a public benchmark and partnered with a cloud company that needed exactly that speed. Eleven months after its first public post, the partner agreed to buy it.

4 grade-A facts · 4 that failed before it
No-code platform to build and govern AI agents for companies · United States · Funded

StackAI

Evidence A
Sold for $74.6M

Started small, grew after funding · sold 37 months after launch

It aimed at regulated enterprises that needed AI agents on their own servers, with audit logs. The company said in 2025 that hundreds of organizations used it. Asana bought it in May 2026. Asana's quarterly report states the price.

3 grade-A facts · 4 that failed before it
Paid library of Agent Skills for AI coding agents · China

Skillry

Evidence A
$7.7K in 30 days

Solo founder · 2 months after launch, revenue flat

AI coding agents can build a working page or deck, but in the founder's words their visual output often looks generic. Skillry sells finished visual styles, one task per Skill, shows real output before the sale, and adds new Skills almost every day under one subscription. When sales began on 28 September 2026, they held at about US$430 a day for 17 days.

5 grade-A facts · 4 that failed before it
Paid kits of skills, workflows and subagents for AI coding agents · Vietnam

AgentKit (formerly ClaudeKit)

Evidence A
$82.2K verified revenue

Solo founder · 12 months after launch, revenue falling

It launched in October 2025. By then, Anthropic's Claude Code had made skills, slash commands and subagents a normal way to steer a coding agent. Most developers were still writing their own. A founder who presents themself as a CTO with 18 years of experience sold a tested set as a one-time purchase with a money-back promise. Verified revenue reached US$16,040 in March 2026 and US$82,195 in total, but it fell to under US$2,000 a month by September 2026.

6 grade-A facts · 4 that failed before it
Voice-controlled Mac workspace for AI coding agents · USA

CNVS

Evidence A
$50.6K verified, all time

Solo founder · 4 months after launch, revenue falling

Said over US$36K in a monthRecords US$39.6K by end of JulyHeld

It turned a new working habit into a native Mac app. The habit was running several AI coding agents at the same time. The app had voice control and a strong visual style, and it charged once instead of every month. The founder sold it to his own audience through daily live streams, X and a Discord community. Stripe recorded more than three quarters of the revenue in the first seven weeks.

6 grade-A facts · 4 that failed before it
Desktop app that compresses the input of AI coding agents · Netherlands

Headroom

Evidence A
$2K MRR

Solo founder · 6 months after launch, revenue growing

The compression engine already existed as a free, open-source Python tool, but it was hard to install. Garm Lucassen packaged it into a desktop app. The app installs with a double click, connects the coding agents by itself, and removes its changes when it quits. The engine's creator suggested that he build it and keep the Headroom name. Developers who hit the usage limits of their AI plans pay a small monthly price to make those limits last longer.

13 grade-A facts · 4 that failed before it

Built with the same tools

Businesses whose sources name tools that Wally by RunAnywhere also uses: Resend, Stripe, Astro, Vercel. Every tool has its own page in Tools.

AI listing assistant that writes Vinted and eBay titles and descriptions from item photos · Netherlands

AutoLister AI

Evidence A
$376 MRR, Stripe-verified

Solo founder · 14 months after launch, revenue flat

Second-hand sellers list many items, and each listing needs a title and a description that buyers will find. The tool removes that typing for a small fee: the cheapest plan is €3.99 a month for 75 listings, about €0.05 a listing. Stripe data shows revenue that grew every month from December 2025 to September 2026 and then stayed at about the same level.

5 grade-A facts · 4 that failed before it
AI quote and invoice software for tradespeople · Germany

AngebotFIX

No revenue figure
Paid plans live

Solo founder · 23 months after launch

German tradespeople must soon send electronic invoices to business customers, and many still write quotes in Word or Excel in the evening. AngebotFIX lets them dictate the quote on site. It sells this at a lower price than the established trade software. The founder answers support questions himself by WhatsApp.

1 grade-A fact · 2 that failed before it
Open-source social media scheduler for people and AI agents · Israel, company registered in Hong Kong

Postiz

Evidence A
$3.2M/yr run rate

Solo founder · 27 months after launch, revenue growing

Said US$245K a monthRecords US$263K MRRSaid less than the records

The founder took a twenty-year-old product category with no open-source entry and open-sourced all of it. This gave him millions of Docker downloads and 37,000 GitHub stars without a marketing budget. Revenue stayed small for a year. It multiplied when he turned the product toward people who automate their posting with n8n workflows and, from late 2025, with AI agents. Agents need an API-first scheduler and keep paying because they post without human effort.

6 grade-A facts · 5 that failed before it
Voice notes and dictation app · Spain

Letterly

Evidence A
$2.6M/yr run rate

Team of 2 to 10 · 29 months after launch, revenue flat

The first apps that combined OpenAI's Whisper and ChatGPT for voice notes appeared in 2023, and the founder found them "not implemented well enough". His team built a simpler one, and collected early users and feature requests through lifetime deals on AppSumo in 2024. In 2025 it grew with advertising and paywall tests measured in Adapty's subscription data.

5 grade-A facts · 5 that failed before it

Grew the same way: product

Businesses in other groups whose main growth factor was also product, best evidence first.

iPhone app that turns voice memos into carousels and posts · Georgia

Draft AI

Evidence A
$259K/yr run rate

Solo founder · 14 months after launch, revenue falling

A carousel needs both writing and design. Draft AI asks for a minute of speech instead and returns finished slides, plus scripts, threads and hooks from the same recording. It charges before the first result, so everyone who uses it has subscribed or started the trial of a paid plan.

6 grade-A facts · 4 that failed before it
AI hairstyle and hair-colour try-on app for iPhone · Croatia

Hairly AI: Hairstyle Try On

Evidence A
$19.3K in 30 days

Team of 2 to 10 · 11 months after launch, revenue growing

People who plan a haircut or a new colour cannot see the result before the salon has done it. The app shows it on the user's own photo in seconds. It uses Google's Gemini model to analyse the hair and fal.ai to generate the image, as its privacy policy states. Revenue was close to zero until March 2026. In April 2026 the seller shipped a redesigned onboarding, a new paywall, a coin system and translations. From then revenue rose every month, to US$17,882 in October 2026 up to 26 October, as TrustMRR reads it from the billing account.

7 grade-A facts · 4 that failed before it
AI food-label scanner app · United Kingdom (company registered in England and Wales)

Goodie AI

Evidence A
$178.9K verified revenue

Team of 2 to 10 · 19 months after launch, revenue falling

It turns a label that most shoppers cannot read into one personal number. The diets are ones that a general calorie app does not cover: low FODMAP, seed-oil-free, gluten-free, vegan. The store page says that analysis needs a subscription, so every active user pays. Verified revenue reached US$23,122 in its best month, February 2026, and US$178,939 in total, read from the app's billing account.

7 grade-A facts · 4 that failed before it
QR code generator with AI-generated art and scan tracking · France

QR Code AI

Evidence A
$14.7K MRR, Stripe-verified

Team of 2 to 10 · 40 months after launch, revenue growing

It launched in June 2023 as a generator of artistic QR codes made with Stable Diffusion and ControlNet. No large QR code company offered them at the time. The art brought the first users. The company then added the plain QR code types, link tools and analytics that businesses search for. It also built hundreds of pages that rank in Google for each type.

7 grade-A facts · 4 that failed before it

Same pattern: Open source first

The founder published the code for free. Developers who did not want to run the code themselves paid for a hosted version. All 49 stories with this pattern, 8 with grade-A revenue.

API gateway for AI models · USA

LLM Gateway

Evidence A
$1M gross in 12 months

Team of 2 to 10 · 16 months after launch, revenue growing

Every team that builds with AI models ends up with several provider accounts, keys, invoices and SDKs. One gateway with open code replaces them, and a developer can read the code before sending traffic through the hosted version. In 2026 the founders added DevPass, a flat monthly price for coding agents. Its allowance is up to three times the price in model usage. The verified monthly receipts grew from US$57,994 in January 2026 to US$249,922 in August 2026; TrustMRR does not split them by product.

6 grade-A facts · 4 that failed before it
AI assistant and agent plugin for the Obsidian notes app · United States

Copilot for Obsidian

Evidence A
$17.9K MRR, $914K all time

Team of 2 to 10 · 42 months after launch, revenue falling

The plugin lives inside Obsidian's community plugin store. The founder says that the store gave it organic growth with no ads or paid creators. The downloads reached 300,000 by the end of 2024, 1 million by February 2026 and 2 million by October 2026. The code is free, so there was nothing to sell until the company added a managed layer on top. The layer is hosted models, web search and tools that work without any setup. One-time lifetime plans, first Believer and then Supporter, brought large payments in single months. TrustMRR, reading the company's Stripe account, shows US$913,644 of revenue in total. Monthly recurring revenue was US$17,899 from 1,356 subscriptions on 27 October 2026, after a fall of about 51% from the best month, February 2026.

5 grade-A facts · 5 that failed before it
Open-source chat app for a team of AI agents, with paid cloud machines · India

MausBot

Evidence A
$5.5K in 30 days

Team of 2 to 10 · 2 months after launch, revenue growing

xAI released Grok Bot, a messaging app of AI agents, in beta on 11 August 2026. The two founders published their own open version about two hours later, under the name OpenGrokBot. They pitched it to people who wanted the same idea on Claude or Codex instead of Grok. Payment records show the first payment on 27 September 2026 and US$5,503 in the 30 days to 25 October 2026.

8 grade-A facts · 5 that failed before it

Sources

The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.

  1. C RunAnywhere, About page (runanywhere.ai/about)
    Read in full on 11 November 2026. Company page. Source for the start in 2025, the founders, the Y Combinator batch and the press list.
  2. C Wally by RunAnywhere, home page (runwally.com)
    Read in full on 11 November 2026. Source for the product description, the data retention claim and the Works with lists.
  3. C Wally, pricing page (runwally.com/pricing)
    Read in full on 11 November 2026. It loads as plain HTML. Source for the rate table.
  4. C Y Combinator company page: RunAnywhere, with the Wally launch post
    Read in full on 11 November 2026. Source for founded 2025, team size 5, the launch post of 29 September 2026 and its claims.
  5. C RunAnywhere blog, 22 October 2026: Wally now serves ten open models
    Read in full. Written by Sanchit Monga. Source for the ten-model catalog.
  6. C RunAnywhere blog, 22 October 2026: Why we build inference for everyone
    Read in full. Written by Sanchit Monga. Source for the launch date of 21 September and the move from on-device to hosted models.
  7. C RunAnywhere refund policy, effective 19 September 2026
    Read in full. Source for prepaid credits through Stripe and the US$5 welcome grant.
  8. C RunAnywhere terms of service, effective 26 October 2026
    Read in full. Source for the legal entity, the preview status and the welcome grant.
  9. C RunAnywhere blog, 27 October 2026: Cheap Tokens Don't Matter Without Access
    Read in full. Written by Sanchit Monga. Source for the security model with reviewed access.
  10. C GitHub: RunanywhereAI/runanywhere-sdks
    Read the repository page on 11 November 2026. Source for the star and fork counts of the on-device SDK.
  11. C RuntimeWire, 21 September 2026: RunAnywhere launches Wally, says its open-model inference outruns larger providers
    Read the headline, the subtitle and the date. The article body loads by script and was not read, so only the date is used.
  12. C RunAnywhere privacy policy, effective 26 October 2026
    Read in full. Source for the subprocessors Stripe, PropelAuth, Resend and Vercel.
  13. C Database of failed startups, local export
    Record pages read on the website on 11 November 2026. AI-written summaries, not independently verified. Source for the failed predecessors.
  14. C Wally, Security programme page (runwally.com/security)
    Read in full on 11 November 2026. Source for the gated security model and the review of each access request by a person.
  15. C RunAnywhere acceptable use policy, effective 26 October 2026
    Read in full on 11 November 2026. Source for the rule that security tests need ownership or written permission.

Run this company? Send corrections or a verified figure to aiolum@livyden.com

Back to all stories