A San Francisco startup teaching the internet to make movies the way it makes memes - in a sentence, in a second, in public.
On any given afternoon at Pika's office in San Francisco, someone is teaching a neural network to imagine a cake getting punched. By dinnertime, that cake will be on X. By Friday, an agency in Seoul will be billing a client for it.
That is the loop Pika has built. Researchers ship a model. The internet stress-tests it with the strangest possible prompts. The team watches, listens, and ships the next one. The cadence is closer to a games studio than a typical AI lab, which is partly the point. Pika is not trying to win a benchmark. It is trying to make the next clip you can't stop watching.
By the time you read this, Pika has shipped five named model versions in eighteen months: 1.0, 1.5, 2.0, 2.1, and 2.2. Each release brought a new trick - longer clips, sharper detail, user-supplied characters, keyframe transitions. None of them tried to be everything. All of them tried to be fun.
Writing is free. Photographs are free. Music has been one click away from your phone since the iPod. Video, somehow, kept its velvet rope. It needed cameras, lights, editors, time. Mostly time.
In late 2022 the rope started fraying. Text-to-image models like Stable Diffusion and Midjourney made still pictures a public utility. The question that followed was obvious to anyone watching: when does this happen to motion? The answer, less obvious, was: somebody has to choose to do it, on purpose, with focus, while the giants are looking elsewhere.
Demi Guo and Chenlin Meng met inside the Stanford AI Lab, which is the kind of detail that sounds like a press release until you read their CVs. Guo went through Harvard for a computer science master's, then Stanford for a PhD, working with Ron Fedkiw and Chris Manning. Meng was working on diffusion models with Stefano Ermon before diffusion was the room everybody wanted to be in.
Their bet had two parts. First, that diffusion-style techniques would generalize from images to short video faster than people expected. Second - and this is the unfashionable one - that the right wrapper around the model was not an API for film studios but a web app for everybody else. They left their PhDs. They started Pika in April 2023. They named it after a small alpine mammal known for moving very, very quickly.
Demi Guo
Co-founder · Chief Executive
Harvard CS, then Stanford AI Lab. Closed Pika's first $55M before her 27th birthday.
Chenlin Meng
Co-founder · Chief Technologist
Stanford AI Lab, advised by Stefano Ermon. Worked on diffusion models before diffusion was a word at parties.
The product is a text box and a result. Type a sentence, get a clip. Upload an image, get the image in motion. Drop a still from your last vacation, type "now make it Wes Anderson," and find out whether your last vacation was, in fact, Wes Anderson.
Under the surface, Pika ships features the way a games studio ships modes:
The flagship model. 1080p output, clips up to ten seconds, sharper motion and faces.
Give it two stills. It writes the ten seconds of motion between them. Keyframes for the rest of us.
Hand the model a character, a prop, and a location. It builds the scene around them instead of around your prompt's mood.
One-click physics: melt it, crush it, inflate it, cake-ify it. The reason your timeline broke last quarter.
The same models, behind a key. Used by agencies, indie devs, and platforms that don't want to train their own.
Standard, Pro, Fancy, Unlimited. Free to try, monthly to keep, and yes, "Fancy" is a real tier name.
Eighteen months, five models, two rounds, one viral effect. The timeline, for the file.
Most AI companies pitch a curve. Pika's curve is mostly fundraising, because that is what is public. Estimated revenue sits around $7.6M annually, which is not the headline. The headline is that a team of roughly 110 people, eighteen months in, is on the same shortlist as labs ten times their size.
Spark Capital (lead, Series B) · Lightspeed (lead, Seed/A) · Greycroft · Homebrew · Conviction · SV Angel · Jared Leto.
Web app, mobile-friendly site, an active Discord, and an API for the developers who would rather build than upload.
Ask Demi Guo what Pika is for and the answer is unfussy: anyone with an idea should be able to put that idea in front of other people, as video, today, without learning Premiere. Not as a replacement for cinematography. As an addition to writing.
That is a quieter mission than the ones AI companies usually advertise. There is no AGI claim, no civilization-scale framing. It is the kind of pitch a working filmmaker can argue with and a working marketer can buy with a credit card. Both have, repeatedly.
Named model releases in 18 months. The competitors, mostly, have shipped two.
People on payroll. Smaller than the marketing department at a typical streaming service.
Maximum clip length on Pika 2.2. Long enough for a joke, short enough to share.
Approximate valuation at the Series B. Not a unicorn. Not trying to be, yet.
The cake on the monitor is rendering for the seventeenth time. Someone in the Discord is asking for a Pikaframes button that takes three images instead of two. Someone in Seoul is sending the eleven-second cut to a client and pretending it took longer. Somewhere in the building, the next model is training.
The thing Pika has actually changed isn't the technology. The technology was coming either way. What Pika changed is the social fact of video - the assumption that making a clip required a crew. The crew is now a sentence. The sentence is now a tab in your browser. The tab now has a "Fancy" subscription plan, which is the most honest tier name in software.
If the next year is kind, Pika ships longer clips, sharper faces, better adherence to weird prompts, and an API that ends up under a few products you already use. If it is unkind, the giants ship something that does most of what Pika does, and the company has to keep being faster, weirder, and more willing to let users break things in public. The bet, either way, is on speed.