DigitalCarbon thinks the flat web is a temporary condition

Look at any real estate listing, any Airbnb page, any product page on an e-commerce site. You get a strip of photographs, and you page through them trying to reconstruct the thing in your head. Does the kitchen connect to that living room? How big is this jacket's collar, really? What's behind the photographer?

The internet shows you flat pictures of a three-dimensional world, and you do the spatial reasoning yourself. We've lived with this so long that it barely registers as a limitation. Every attempt to fix it has run into the same wall: getting the real world into 3D was expensive. You either paid a specialist with dedicated camera hardware to scan the space, or you paid a 3D artist to model the product by hand. Both paths cost too much to apply to the long tail of listings, products, and places that make up most of the web.

DigitalCarbon is going after that wall. The company, part of Y Combinator's Summer 2024 batch, turns ordinary photos and videos into photorealistic, interactive 3D scenes. You capture a space or an object with the phone already in your pocket, and its AI pipeline reconstructs it as a scene you can move through, orbit, and edit. The company says the results render at more than 100 frames per second on everyday devices. There is no rig to buy and no 3D artist to hire.

The pitch compresses to a sentence: if a phone video is all it takes to produce a usable 3D scene, then everything that has photos today can exist in 3D tomorrow.

The technique that made this possible

For decades, the standard way to get reality into a computer was photogrammetry: take many overlapping photos, find matching points between them, and stitch together a textured mesh of polygons. It works, which is why surveyors and VFX studios have used it for years. It is also slow, brittle about reflective and thin surfaces, and labor intensive to clean up. The output usually looks like what it is, a model of the thing rather than the thing.

The research world spent the early 2020s replacing that pipeline. Neural radiance fields showed that a neural network could learn a scene's appearance from photos and synthesize new viewpoints with startling realism, but rendering was far too slow for real products. The breakthrough that matters for DigitalCarbon arrived in 2023 with a technique called 3D Gaussian splatting. Instead of polygons or a neural network you query pixel by pixel, a scene becomes millions of tiny translucent blobs, each with a position, shape, and color. Those render fast on ordinary graphics hardware while preserving the photographic quality that made radiance fields exciting in the first place.

Fast matters more than it sounds. A photorealistic scene that renders at 10 frames per second is a tech demo. The same scene at 100 frames per second is a product: it can live inside a listing page, respond instantly to a user dragging the view, and run on the laptop or phone of whoever happens to open the link. DigitalCarbon's bet is that Gaussian splatting crossed the threshold where real-world 3D stops being a specialist artifact and starts being ordinary web content, and that someone needs to build the boring, reliable machinery that takes people from "I have a video" to "here's your scene."

Capture with the camera you already have

The deliberate choice in DigitalCarbon's design is what it asks of the person doing the capture: almost nothing.

Earlier generations of virtual tours were built around dedicated hardware. A tripod-mounted 360 camera, a trained operator walking room to room, a processing pipeline tuned to that one device. The results were good, and the economics confined them to the top of the market. A brokerage selling a $2M house would pay for a scan. The owner of a modest rental, a small hotel, or a hiking trail nobody monetizes would not.

DigitalCarbon starts from the phone instead. Walk through the space the way you'd film it for a friend, or circle a product on a table, and upload the footage. The reconstruction happens on their side. The demos on the company's site make the range concrete: an Airbnb interior you can walk through in first person, a backyard, a garden, a trail. For objects, a handful of smartphone photos becomes a product model a shopper can spin and inspect.

That capture decision is also a distribution decision. When the input device is something billions of people carry, the number of people who can produce 3D content stops being a function of hardware sales. The constraint moves to software quality: how reliably the pipeline handles shaky handheld footage, awkward lighting, and all the other sins of amateur capture. That's where DigitalCarbon wants the hard problems to live, on its side of the API rather than the user's.

Tours, product pages, and an API underneath

DigitalCarbon ships the obvious applications itself. An immersive tours product covers real estate and hospitality, the industries that have wanted affordable walkthroughs for years. A product showcase tool covers e-commerce, where a spinnable 3D model answers the questions a photo strip can't. Both are in early access.

The more interesting layer sits underneath: a developer API for the reconstruction itself. Send images or video, get back an interactive 3D scene you can embed. That framing treats 3D capture the way modern infrastructure companies treat payments or messaging, as a capability other software calls rather than a destination users visit.

The API is what makes the long-tail use cases plausible. The company points at drone inspections and digital twins alongside tours and product pages, and those markets share a shape: a lot of real-world structure that gets captured repeatedly and viewed by people who weren't there. Think of an insurance adjuster reviewing roof damage from a drone pass, or a construction manager comparing this week's site capture against last week's. Nobody is going to send a specialist with a rig to all of those places on a schedule. A drone or a phone plus an API call can be everywhere.

Founders who already built this kind of company once

DigitalCarbon was founded in 2024 in San Francisco by Guru Rao and Michael Nguyen, and their background explains the shape of the product. Both were founding employees at AssemblyAI, the YC-backed company that turned speech recognition into an API developers could call without thinking about the models underneath.

That experience maps onto the current problem almost exactly. AssemblyAI sat on top of a research field moving at a sprint, speech models kept leapfrogging each other, and its job was to absorb that churn so customers just got better transcription over time. 3D reconstruction is at the same stage now. Gaussian splatting is barely two years old as a published technique, variants and improvements land constantly, and anyone who hard-codes today's method will be outdated in a year. The durable position is the one AssemblyAI occupied: own the customer relationship and the pipeline, and swap better research into the middle as it appears.

The team seems to understand that, because they describe themselves less like a SaaS startup and more like a research lab with a product attached. Coverage of their YC batch notes they're exploring how diffusion models can improve radiance field reconstruction, which targets the technique's main practical weakness. Splatting likes generous, careful capture footage. Real users will give it twelve seconds of shaky video and one side of the room. If generative models can fill the gaps and produce a complete scene from sparse, imperfect input, the capture bar drops again, and every drop in that bar widens the set of people who can produce 3D content at all.

The wedge is boring, and that's the point

It would be easy to dress this company up in metaverse language, and to its credit, DigitalCarbon doesn't. Tours and product pages are unglamorous wedges. They're also how a technology like this actually gets adopted: concrete use cases where 3D measurably beats photos at an existing job, selling into industries that already pay for visual content.

A buyer who can walk a listing filters out mismatches before the showing. A shopper who can inspect a product from every angle returns it less often. These are old, well-understood value propositions. What changed is the input cost. When the scan required a rig and an operator, the math only worked at the high end. When it requires a phone video, it works for nearly everything, and a format that was a premium feature starts to look like a default.

That's the quiet claim embedded in DigitalCarbon's product: the flat web was a cost constraint, not a preference. Photos won because they were the cheapest faithful way to show a real thing on a screen. The company is betting that the cheapest faithful way is about to be a 3D scene reconstructed by AI from the same photos you were taking anyway, served fast enough that viewing it feels like nothing. If that's right, the strip of listing photos you flipped through this morning is a format on borrowed time.