100 Signups, 95 Drop-offs: How Gamma Turned AI Slides Into a $50 Million Business

Gamma started in 2020, well before generative AI went mainstream, with a vision Noronha says has barely changed: reinvent presentations. Co-founder Grant Lee came to him in the depths of COVID, stuck on a park bench joining conference calls from a tiny phone because his wife had the only desk in the house. All he could see were other people's PowerPoint decks on Zoom. Noronha recognized the job immediately. He was spending his own days staring at Google Slides, or making them, and enjoying neither.

Slides, he argues, are the language of business — the way high-stakes ideas get communicated, including every pitch a Sequoia partner sits through. Ask people whether they enjoy making them, though, and you get what he calls an "ick reaction." The most common complaint: 90% of the time goes to formatting, 10% to content, and audiences still judge the look rather than the argument. Gamma's bet was not a nicer PowerPoint editor. It was a new format for visual communication.

Host Sonya Huang notes the outcome that makes the origin story worth telling: more than $50 million of ARR, cash-flow positive, a team of 30. Noronha is quick to say it never felt like an overnight success. Huang, who actually likes making slides, gets classified as "the 1%." The rest of the conversation is how a company that started as a format experiment became, in the show's framing, the Cursor for slides.

Did they know AI would arrive when they started? "Absolutely not." Noronha was among the first people with GPT-3 access in 2020 and ran the obvious tests: could the model summarize a document into bullets, or expand bullets into paragraphs? The answer was no. He filed the demo under neat and spent the next few years on a future-of-work thesis instead — remote, async, more interactive and webpage-like than slides. They landed in a frustrating middle: some product-market fit, not enough exponential growth to be a venture-scale business.

About two years in, runway was dwindling. Inflation and interest rates were climbing, and even their own bank, Silicon Valley Bank, crashed. Noronha describes existential angst about whether the company would make it. The thing that saved them, he says frankly, was generative AI getting good just in time. They had been checking back in on the models over those two years, but what caught his eye was not language models. It was images.

Stable Diffusion dropped that summer, and Twitter filled with people posting silly little things they had made. DALL-E was taking off too. Noronha realized presentations are mostly visual design, often decorating slides just to fill space. PowerPoint's key innovation, he argues, was clip art — maybe 30 years ago — and nobody had answered what the next clip art should be. Image models were the most compelling candidate he had ever seen.

Ironically, the image-model boom sent them back to GPT-3. The same prompts that had failed two years earlier suddenly worked. ChatGPT looks like an overnight success, Noronha reminds Huang, but OpenAI had been improving the models the whole time, especially through instruction tuning — making a model follow a request phrased the way a human would. With that, Gamma had the ingredients to stop being a tool for making presentations and start making the presentation for you.

Pre-AI Gamma had a cold-start problem hiding in activation. About 100 people would sign up, and 95 or so would fall off in the first few minutes. They landed in a new, innovative editor that was still just a blank page. The team pointed at features; users asked what they could actually make. Only the diehards powered through. The reason they seized on AI early, Noronha says, was to solve that blank-page problem and show something magical fast.

What they failed to appreciate until launch was that it was not Gamma's blank page being solved. It was the user's. Presentations intimidate people because a new file opens on "Enter presentation title" and a spiral starts: outline, story, hook, key moments, fonts, colors, imagery. "You get stuck in this spiral of thinking through all these things." AI that filled Gamma's empty canvas turned out to fill everyone else's too.

Users could go from a vague idea to a fully worked-out rough draft, and their job became editing instead of starting from scratch. Huang, who gets asked to give a lot of AI talks, says she now drops a document into Gamma and asks it to become a deck. Noronha credits daily dogfooding: the team uses the product internally, so breakage shows up the same day. Moving the starting line from a blinking cursor to a draft you can argue with is, in this telling, the whole product bet.

Huang asks the application-layer question: what sits between foundation models and users? Noronha starts with the medium itself. Gamma never wanted to be a PowerPoint clone, and users now describe it as "if Notion and Canva had a baby." The Notion half is block-based and writing-based, so you type instead of dragging boxes around a canvas. The Canva half is visual range — diagrams, charts, theming, and templates — except Gamma does that part for you. You type, and you do not need to know what will look right.

A huge share of the company's work is not prompts or fine-tuning. It is tuning the building blocks: the best diagrams, layouts, and visual themes, which become the arsenal AI is let loose on. Taste is a staffing choice. For a while the company was one-third designers — four of twelve — unusual for an early Silicon Valley startup. The philosophy, Noronha says, is "we do the great design so that you don't have to." They also run the data, watching where users get lost and using AI to grind through thousands of external slide decks for common layouts.

Making a presentation has two hurdles: storytelling and structure, then formatting and design. On the second, AI has tricks that are easy for a model and tedious for a person. One is not trying one design: try ten and pick the best, which an AI can do in about three seconds. Another is forcing every image onto the same color palette, trivial if you generate from scratch and nearly impossible if you are pulling from Google Image Search. Stack enough of those tricks, and you get output no human would actually produce by hand.

How do they know which of those ten options is best? Noronha's last job was in product at Optimizely, which pioneered A/B testing as a universal practice for marketers and product launches. That through line now runs Gamma. The day of the recording, Sam Altman was on stage announcing GPT-5, and someone on the Gamma team was already writing the code to plug it in. They run evals when a model lands, but they are in a creative domain with no single right answer. You cannot declare that GPT-5 beats Claude Sonnet by 2.7%.

They can, however, run experiments on millions of real users. That afternoon they planned to A/B test GPT-5 against Claude and Gemini, asking for ratings and measuring how much people edited, whether they exported, whether they shared, and whether a free user converted to paid. They have run probably hundreds of these tests, learning which models fit which tasks and which pieces of their complex prompts actually resonate. Huang calls it a data flywheel; Noronha agrees nobody is better positioned to build one than the Optimizely crowd.

Fine-tuning is not the main lever. Noronha says much of the hype comes from companies that sell fine-tuning, and in Gamma's experience it seems to hobble the model's actual intelligence. Open-source models, where fine-tuning is popular, have not come close to closed foundation models, so the daily drivers are Claude, Gemini, and GPT. Asked for a favorite, he is a Claude fan: it has a taste for what looks good, a creativity that "doesn't show up in any benchmark." Coding is the killer app everyone else is optimizing for, and Gamma is a visual expression platform, so those public scores get thrown out.

The most slept-on model is Gemini, and it is by far Gamma's most heavily used, because nothing beats Gemini Flash on intelligence per dollar. Earlier Claude versions sometimes beat newer ones as the family has optimized for coding. Reasoning models were not an unlock: the longer a model thinks, the less creative it gets, and DeepSeek was "a total dud" for this use case. Image generation has no single winner either — they support around 20 models, lean on Ideogram and Flux, and find GPT Image unusually strong on text and infographics. They tried Stanford's DSPy to auto-optimize prompts and dropped it, because prompts are treated as user experience a designer can edit, and because one prompt has to travel across models when one of them is broken.

The competitors that scare Noronha most are still PowerPoint and, to a lesser extent, Google Slides. Gamma is proud of millions of monthly active users, but he has heard each incumbent has 500 million — a billion slide users in all. They still see themselves as an insurgent. The luck, such as it is: those companies have been slow with AI, and agentic coding has soaked up so much attention that productivity tools have been neglected. Foundation models are the other threat. ChatGPT's agent can already make PowerPoints, "thankfully for us, not very good ones," though he knows that will change.

So Gamma walks a tightrope. They do not want to be a great PowerPoint builder, because PowerPoint will eventually become that. They want to be a great Gamma builder — their own visual storytelling medium. Differentiation has a cost: the format is unfamiliar, and some users get angry about export. At the same time they have to feel convenient, because the first 50 million users were early adopters and the next 100 million will be an early majority who need things to fit existing workflows, especially as the company moves from pure B2C into B2B.

Canva is inspiration more than bogeyman. Noronha says they proved product-led growth can work at enormous scale — the stat he has heard is that Canva hired no salesperson until $500 million of ARR — with an SMB-focused strategy and a huge template library. Gamma wants to follow that trail but not copy the product. Canva is a drag-and-drop design tool; Gamma's pandemic DNA is a writing-based tool, closer to Notion, where you need no design skill to make something beautiful. They will compete more head-on over time. He would rather forge a path than be a copycat.

A few years ago three or four AI slide tools were in the race; Gamma pulled ahead while others pivoted or stalled. Noronha points to a lean, paranoid mindset: other companies raised a lot and built big teams, while Gamma assumed the AI world would keep changing and that the trait that would matter was the ability to pivot. That meant a small team relative to traction, and a hard cap on inference spend. Competitors ran GPT-4 and DALL-E when those were tremendously expensive, the classic move of "selling dollars for 75 cents." Cash-flow positivity, he says, is what lets them keep moving faster than everyone else.

His advice for other application-layer founders starts with a unique perspective on the market. Gamma's is that they are not a PowerPoint builder; they are creating a new format to replace the slide deck, a lens that has guided product choices and sometimes angered users. He would be wary of yet another vibe-coding startup, a lane already crowded with companies doing the same thing. Better to find a neglected area, even one that goes against what foundation models are training for. Coding tools take off because the models are being optimized for coding; work on something adjacent that they are not optimizing, and you are less likely to get "smooshed by the giant dinosaurs."

The other habit is experimentation across models. Founders tend to lock onto one provider and assume it is the best. Innovation is so rapid, uneven, and unpredictable that you have to plan for a world where a different model is best on any given day of the week. Huang hears the Optimizely DNA coming out again. Then she turns to what 250 million Gammas on the platform actually are: not just the original slide use case, but the adjacencies users opened on their own.

They reached product-market fit on presentations, a wider world than it sounds — from visual TED-style talks to wordy consulting and investment-banking slides — and still see room to get better there. Users also pulled them into documents: not plain-text Google Docs, but PDF proposals, brochures, shiny customer reports, and white papers. If PowerPoint's market is 500 million people, Noronha thinks PDF is more like a billion. The next surprise was websites. People made a Gamma and realized it worked as a personal site or a small-company homepage, the same way Notion docs become mini websites.

Websites are sticky and they are a public face, so Gamma launched a sites product last year. The working definition is now visual storytelling platform: anytime you have something high-stakes to convey, and you will be judged by the cover, Gamma wants to be the place. Near-term, September was supposed to bring a rebuilt core editor with more visual range, more agentic editing you can talk to — this five-slide deck should be 20; expand section 3 — and an API. A sales team could pipe CRM notes into a templated pitch and skip an hour of deck-building. Investors keep telling him they have ten portfolio companies that need exactly that.