Baseten
Two Gumroad engineers trained a fraud model on a laptop. Running it built Baseten.
Verdict: Baseten's homepage promises the "open frontier," while Grid, launched November 4, also routes coding to Anthropic and OpenAI models.
A Narrative Authenticity Framework analysis of Baseten. NAF score 7.4 out of 10. Weakest components: Friction and Real Aspiration, 7.0 each.
A little over a decade ago, Gumroad had a credit card fraud problem.
Two of its early machine learning engineers, Tuhin Srivastava and Phil Howes, built a classifier to catch it. They had grown up together in Australia, and they trained the model on a laptop CPU using structured data.
That part, by their co-founder's account, was easy. The hard part was making the model run in production, fast and at scale. Two data scientists had to teach themselves full-stack engineering. The first thing they had to learn, Srivastava later said, was how to serve the model behind an API. The systems they built went on to screen hundreds of millions of dollars in transactions.
In late 2019, Srivastava and Howes teamed up with Amir Haghighat, Gumroad's head of engineering, and Pankaj Gupta, an Uber engineer, to start a company around that lesson. They named it after the base-ten blocks children use to learn arithmetic.
People told them it was stupid. You have to train a model before you can serve it, so why start with serving? For three years, Haghighat says, "the company wasn't working." Srivastava told Forbes revenue was "basically zero."
Then GPT-4 and Stable Diffusion arrived, and the rest of the industry learned what the founders had learned at Gumroad. Cursor, Notion, Abridge and OpenEvidence now run on Baseten.
You won't find the laptop on Baseten's homepage or its About page. The story lives in a 2025 blog Q&A, a 2022 investor post and a few podcasts.
Where Baseten stands
The early years were slow. Greylock co-led the 2019 seed. The first product shipped in May 2021 after more than 18 months of building. Contrary Research puts 2023 revenue at about $2.7 million.
Then the curve bent. The March 2024 Series B valued Baseten at a little over $200 million. In January 2026, a $300 million Series E led by IVP and CapitalG, with NVIDIA participating, put it at $5 billion. In June came a $1.5 billion Series F at $11 billion and $13 billion across two tranches. On October 30, Bloomberg reported Baseten is set to raise about $2 billion at up to $30 billion. Baseten hasn't announced that round.
The June release reported revenue up about 20x year over year. Sacra estimates the run rate at roughly $600 million by March 2026; Baseten hasn't confirmed a revenue figure.
The customer list reads like the AI application layer. Cursor, Notion, Harvey, HubSpot, Lovable, Abridge and OpenEvidence all appear in Baseten's materials. A Baseten case study says Patreon saves nearly $600,000 a year running Whisper on Baseten without hiring an in-house ML infrastructure team.
2026 has been a year of expansion. Baseten bought Parsed, a post-training startup, and Blaxel, which builds sandboxes for agents. It launched the Loops training SDK, a Frontier Gateway that lets closed-weight labs serve their models under a white label, and a research arm called Base Labs. In September it began serving open models to OpenAI enterprise customers. Sheila Vashee, previously CMO at Figma, joined as CMO on September 25. On November 4 it opened Grid in preview, a router that sends each coding task to an approved model and tracks the spend.
Competition is crowded. Together AI and Fireworks AI sell overlapping inference clouds, Modal and Replicate chase developers, and AWS, Google Cloud and Azure all sell inference at hyperscale.
The brand line today is the homepage hero: "The open frontier is built on Baseten." The About page has no origin story and names no founders.
Why the brand line falls short
"The open frontier is built on Baseten" describes where the market is heading. It says nothing about what goes wrong for a buyer who doesn't use Baseten.
The bigger problem is the word open. Baseten's own 2026 moves have stretched it thin. In May it launched a gateway for closed-weight labs. It is preparing to run Microsoft AI's MAI-Thinking-1, which it describes as sitting in "the space in between" open and closed. Grid supports models from OpenAI and Anthropic. Srivastava himself says customers mix open and proprietary models depending on the task.
None of that is a mistake. It does mean "open" no longer describes what Baseten does, and Together AI, its closest rival, already describes itself as a believer in open and transparent AI.
Meanwhile, Baseten's leaders keep reaching for a sharper word: owned.
- Greylock's May 2026 video with Srivastava is titled From Rented to Owned Intelligence.
- The Parsed acquisition was announced as a way "to Enable Companies to Own Their Intelligence."
- The MAI-Thinking-1 post promises "control back to our customers who want to own their intelligence."
- On Alex Heath's Sources podcast in September, Srivastava argued every company will want to own its AI.
The homepage touches the idea only in passing, in body copy about "data you own." Blackbird's post on the Series F put it in one line: "In AI, you can rent your intelligence or you can own it." That line came from an investor. Baseten should be the one saying it.
The NAF scorecard
| Component | Score | One-line read |
|---|---|---|
| Friction | 7.0 | The cost-and-control pain is named in case studies and Grid, not the hero |
| Lived Stake | 7.5 | The founders lived the problem at Gumroad; the brand keeps it in a blog Q&A |
| Distinctive Qualification | 8.0 | Elite customers and steep growth, with headline numbers that move undated |
| Real Aspiration | 7.0 | Owned intelligence lives in interviews while the hero says "open" |
| Composite | 7.4 |
That puts Baseten level with FluidStack on the SeekArgus leaderboard and just behind compute peers Nebius and Nscale, both at 7.8. Patagonia's 9.5 is the reference point.
Friction: 7.0
Baseten does name a costly problem. It just doesn't lead with it. Its Sully.ai case study says inference costs "scaled linearly" with growth until they became a bottleneck, and reports a 90% cost cut after moving to open models on Baseten. The new Grid launch promises to "cut AI coding spend." Both sit below the fold. A buyer who lands on the homepage reads about an open frontier, uptime and scale, and nothing about what their current setup is costing them.
Lived Stake: 7.5
Few infrastructure founders have an origin this clean. It scores above a 7 because Baseten tells part of it on its own blog, in a September 2025 Q&A with Haghighat that includes the fraud story, the "people told us this was stupid" reaction and the three lean years. It stops short of a 9 because the About page leaves all of it out.
Distinctive Qualification: 8.0
The proof is strong: Cursor, Notion, Harvey and OpenEvidence as customers, revenue up about 20x, and case studies with specific results. The gap is how the headline numbers drift. The June release said more than one billion inference calls a day across 87 clusters and 18 clouds. The September CMO announcement said 40 to 50 trillion tokens a day across 90 clusters and 20 clouds. Today's homepage says more than 16 billion daily inference calls and 20-plus cloud providers. None of them is dated on the site, so a reader can't tell growth from inconsistency.
Real Aspiration: 7.0
Baseten has a good vision of a changed world: every company owns the intelligence it runs instead of renting it by the token. It scores a 7 because that vision lives in CEO interviews, investor posts and press releases, while the front page talks about openness. A 9 would put one sentence at the top of the homepage describing a world where companies own the models they run, plus a plain definition of what "owning" means.
The story Baseten should tell
Training a model, or renting one, is the visible part of AI. Running it in production, at a cost you control, is where the business is won or lost. Baseten's founders learned that on a fraud problem years before it had a market, then paid for the belief with three years of near-zero revenue.
The origin also separates Baseten from its closest rivals. Together AI's founders came out of research labs at Stanford and ETH Zurich. Fireworks AI was founded by the team that ran PyTorch at Meta. Baseten's founders were data scientists at a small creator marketplace who had to teach themselves to ship their own models, which is close to where most of Baseten's customers sit today.
We learned it at Gumroad, years before it had a market. Training the fraud model was easy. Making it run fast, at scale, was the work. People told us building for that moment was stupid, and for three years the company didn't work. Then every AI company hit the same wall. The best of them have stopped renting intelligence. They own it, and they run it on Baseten.
Four moves to raise the score
1. Make owned intelligence the homepage promise (Real Aspiration 7.0 to 8.0)
Replace "The open frontier is built on Baseten" with the idea Srivastava already argues in public, and add one sentence that defines ownership in Baseten's terms: you control the weights and the data, and you run them in Baseten's cloud or yours. Link it to a signed Srivastava essay built from his Sources argument.
Prior-work check: Partial. The rented-versus-owned framing appears in Greylock's May 2026 video, the Parsed announcement, the MAI-Thinking-1 post and the September Sources interview. The homepage hero still reads "open frontier" and mentions ownership only in body copy, as of November 19, 2026.
2. Publish "The Inference Bill," a twice-yearly report from platform data (Friction 7.0 to 8.0)
Baseten sees where thousands of AI teams spend, and where that spend is moving. Report, in aggregate and anonymized, how much spend is shifting from rented APIs to owned and post-trained models, and what happens to cost per task when it does. Frame it as a bill and a loss of control, which keeps it credible with partners like OpenAI and Microsoft AI.
Prior-work check: Partial. Baseten publishes single-customer case studies, a savings calculator, a September 2026 essay on inference economics, and now a Grid dashboard that shows each customer its own coding spend. No public evidence of a recurring, dated, cross-customer report, as of November 19, 2026.
3. Rebuild the About page around Gumroad (Lived Stake 7.5 to 8.5)
Tell it in the founders' voice: the fraud classifier trained on a laptop, learning to serve it behind an API, the base-ten blocks behind the name, the people who called it stupid and the three years it didn't work. Sign it with all four founders' names.
Prior-work check: Partial. The story appears in Haghighat's September 2025 Q&A on the Baseten blog and in Greylock's 2022 investment post. The About page has no origin story and names no founders, as of November 19, 2026.
4. Publish a dated "Baseten by the numbers" page (Distinctive Qualification 8.0 to 8.5)
Put every headline metric in one place, each with an as-of date and a one-line method: daily inference calls, tokens per day, clusters, clouds and uptime. Update it monthly and point every press release to it.
Prior-work check: Not found. Figures appear only inside individual releases and on the undated homepage. No public evidence of a dated metrics page, as of November 19, 2026.
Projected composite: 8.3 (Friction 8.0, Lived Stake 8.5, Distinctive Qualification 8.5, Real Aspiration 8.0). These checks cover public sources only and can't see Baseten's internal plans.
Sources
- Baseten homepage and About page, viewed November 19, 2026
- Baseten: Introducing Grid, November 4, 2026
- Baseten: Q&A with Amir Haghighat, September 2025
- Baseten: Sheila Vashee joins as CMO, September 25, 2026
- Baseten: Announcing Baseten for Model Labs
- Baseten: MAI-Thinking-1 is coming to Baseten
- Baseten: Our partnership with OpenAI
- Baseten: Sully.ai case study and customer stories
- Business Wire: Baseten Series E, January 2026
- The Next Web on Bloomberg's report of a round at up to $30 billion, October 2026
- AI Magazine: Baseten at a $13 billion valuation
- Contrary Research: Baseten
- Latent Space wiki: Baseten
- Greylock: Baseten investment post, 2022
- Greylock video: From Rented to Owned Intelligence, May 2026
- AIM Research interview with Tuhin Srivastava
- Sources with Alex Heath, September 2026
- Blackbird: Baseten, the whole game
- Sacra: Baseten (estimated run rate, not confirmed by Baseten)
- WorkOS: Fireworks AI and the PyTorch team
Tim Donovan · SeekArgus