Luma AI.

01 // The SceneThe lab that ships models like other people ship tweets.

A small team in a Palo Alto building is, right now, shipping a video model that thinks before it renders. They have done this twice in the last six months. They will probably do it again before you finish this sentence.

It is May 2026, and Luma AI has the kind of quiet that only well-funded labs have. The team is in the high two hundreds. The cap table includes NVIDIA, AMD, Andreessen Horowitz, and a Saudi sovereign-backed AI firm called HUMAIN that just wrote a check for $900 million. The product, Dream Machine, has thirty million registered users. The newest model, Ray3, is the first reasoning video system in the world to render natively in 16-bit HDR.

None of that is the interesting part. The interesting part is that almost none of those users found Luma through an ad.

02 // The ProblemWhy generative video is the hardest thing AI has tried yet.

Text is easy - tokens in, tokens out. Images are harder, but the universe of a single frame is still flat. Video is the cliff. It demands physics, continuity, character consistency, lighting that obeys the laws of a room, and motion that does not break the spell. Most generative video, until very recently, broke the spell about three seconds in.

The skeptical reader has a fair question. Why does this matter to anyone outside a VFX shop? Because video is the substrate of how humans actually consume information now. TikTok, Reels, Shorts, YouTube, in-app ads - the modern internet is mostly motion. If generative AI cannot make video that holds up, it cannot make the dominant medium of the internet. That, more than poetry or pictures, is the trillion-dollar problem.

Luma's bet is that you do not solve video by training a slightly bigger image model. You solve it by teaching a model to understand the physical world - light, geometry, gravity, the way a coat falls when someone sits down. They call this unified general intelligence. It is the kind of phrase that sounds either profound or insufferable depending on the day.

03 // The Founders' BetThree founders. Two prior lives. One unfashionable idea.

Amit Jain spent his pre-Luma years at Apple, where he was a systems and machine learning engineer on two things you have probably touched: the first LiDAR sensor in the iPhone, and the Passthrough feature on the Apple Vision Pro. He understands cameras, depth, and the gap between a sensor and a perception. He co-founded Luma in September 2021 with Alex Yu, a graphics and neural rendering researcher, and Alberto Taiuti, who came from the AR and computer vision side.

For the first two years, Luma was a 3D capture company. They built tools around NeRF and Gaussian Splatting that let you photograph a thing with an iPhone and walk around it on the web. Nice tech. Modest business. The kind of company a16z funds because the team is good, not because the market is obvious.

Then in 2024, Luma did the unfashionable thing. They pivoted away from the product that worked and toward the problem that did not. Video. They reasoned, correctly, that the same model machinery that captures 3D worlds could be turned around and asked to generate them. The bet was that video was just 3D plus time, and that the company with the best world model would win the moving image.

04 // The ProductWhat Dream Machine actually does (and what it does not).

Dream Machine, launched in June 2024, is a web and iOS app that turns a sentence or an image into a video clip. You type "a Labrador in a wetsuit catching a wave at golden hour." A few seconds later, you have a video of a Labrador in a wetsuit catching a wave at golden hour. The first time it works on you, it is mildly disorienting. The second time, you start trying to break it.

Underneath, Luma ships a family of models. Ray, then Ray2, then Ray3 in September 2025 - the first reasoning video model to natively output 16-bit HDR. Then Ray3 Modify, which lets a director keep an actor's original take and re-direct the performance. Then Ray3.14, in January 2026, which delivered native 1080p, four-times-faster generation, and pricing roughly a third of what it cost a quarter earlier. Photon and Photon Flash handle stills. UNI-1.1 handles the reasoning underneath.

What it does well: short narrative clips, character consistency across cuts, lighting that obeys a scene, camera moves that feel directed rather than drifted. What it does not yet do, despite the marketing, is replace a film crew. Anyone who tells you otherwise is selling you something.

05 // The ProofThe receipts, in case the rest sounded like a pitch deck.

Numbers are the part of a profile you can argue with. Here are Luma's, drawn from public filings and announcements through early 2026.

The customer roster reads like a polite hostile takeover of the creative industry. Adobe integrated Ray3 into Firefly and Firefly Boards in September 2025 - the company's first third-party model deal. Amazon put Ray2 inside Bedrock. Monks, the S4 Capital agency, signed on as a global launch partner. So did Galeria and Strawberry Frog. HUMAIN Create, the Saudi initiative, made Luma its model of record across MENA.

06 // The Mission"Unified general intelligence." Yes, really.

Luma's stated mission is to build unified general intelligence that can generate, understand, and operate in the physical world. Most companies put aspirational sentences in a deck and quietly bury them. Luma puts theirs on the website and shapes the roadmap around them.

The bet is that the path to AGI does not run through chatbots. It runs through models that have to predict what happens next in a scene - which is to say, models that have to learn physics from pixels. Whether they are right is the central open question of the next five years. If they are, video is the wedge into something much larger than entertainment. If they are not, Dream Machine is still a very large business.

It is the kind of bet that only sounds reasonable in hindsight. Right now it sounds slightly mad. Which, historically, is the correct tone for an interesting AI company.

07 // Why It Matters TomorrowThe studio is becoming a model.

If Luma is right, the unit cost of a high-quality video clip is on a curve that ends near zero. Ray3.14 already cut per-second pricing by two-thirds in a quarter. The implication is not that filmmakers go away - good directors will always be scarce - but that the long tail of video work, the explainer, the ad, the storyboard, the pre-viz, the social cut, becomes something a small team can ship in an afternoon.

It also implies a strange new industrial geography. A Palo Alto lab is now co-building a two-gigawatt AI campus in Saudi Arabia. That is the power draw of a small US city, dedicated to making video out of language. Whatever you think about that sentence, it is the kind of sentence that did not exist eighteen months ago.

08 // Back to the SceneThe quiet, revisited.

Walk back into that Palo Alto building. The team is still small. The models keep shipping. The thirty million people who use Dream Machine are mostly not VFX professionals - they are kids, marketers, small business owners, somebody's grandmother making a birthday video for a grandkid she has not seen since Christmas.

That is the part nobody quite predicted. Not the headline numbers, not the funding round, not the partnership map. The part is this: a generative video model, the kind of thing that lived in research papers two years ago, is now a tool that ordinary people reach for when they want to show somebody something they cannot quite describe. Luma did not invent that desire. They just built the model that quietly meets it.

The lab keeps the lights on. The ticker keeps moving. The next model is, by now, probably already training.