Precision AI

Adobe’s Answer to the Data Wall

Andrew Ross Sorkin asked Dario Amodei a direct question at the New York Times DealBook Summit this past December: was he surprised by where AI sits today. Amodei said no. Sorkin, visibly thrown by the answer, pushed him to explain why anyone wouldn’t be surprised by a technology that barely existed in public consciousness a few years earlier. Amodei’s answer was almost mundane in its simplicity. He and his eventual co-founders had documented the scaling laws of AI more than a decade earlier, and the relationship had held ever since: more compute and more data, fed into models with only minor architectural tweaks, produced steady, broad-based improvement across coding, science, medicine, law, and finance. He called the incremental innovations driving each leap “tiny little tweaks.” He had been watching that curve compound for roughly twelve years. Nothing about 2026 surprised him because the trendline had been visible the entire time.

That framing is useful, but incomplete. Scaling laws describe a rigid relationship between compute, model size, and data. Betting against the continuation of that exponential curve sounds contrarian—and in the narrow sense of capital expenditure on chips and data centers, it probably is.

But there is a cleaner way to look at this that doesn’t require betting against Nvidia or hyperscaler capex. Exponential capability growth has always assumed exponential growth in training data. That assumption has broken. The internet’s stock of high-quality, human-generated text is finite, and the major labs have already scraped what is economically and legally available. A curve cannot compound indefinitely on an input that has stopped growing at the same rate.

While frontier labs are forced into a grueling, expensive pivot toward workarounds like synthetic data and reasoning models to cheat this “data wall,” the game has fundamentally changed for everyone else.

Companies without a frontier lab’s checkbook were never going to out-spend their way to the frontier on compute alone. They have to win differently: by extracting disproportionate value out of proprietary data they already control, rather than trying to out-scale rivals who can simply buy more chips. I call this precision AI.

Adobe is arguably better positioned to capture this prize than the market gives it credit for. Its real advantage may not come from winning foundation model benchmarks, but rather from two existing strengths: its deep integration within corporate workflows, and a compliant legal framework that competitors will struggle to match.

What Vectors Demand That Pixels Don’t

Most AI tools work by guessing really well. A chatbot predicts the next likely word in a sentence. An image generator resolves a blurry canvas of static into a finished photo, pixel by pixel. In both cases, a small mistake is forgivable. If one pixel in a sunset photo lands a shade too pink, nobody notices.

Vector graphics don’t offer that same forgiveness, because a vector file isn’t really a picture. It’s a set of instructions, the same way a recipe is a set of instructions rather than a finished cake. A vector tells a computer to draw a line from one exact point to another, curve it a specific way, and fill the inside with a precise shade of blue. Because those instructions are math rather than a grid of colored dots, the same logo can be a tiny app icon or a forty-foot billboard and look equally crisp at both sizes. Stretch a photograph that large and it turns to mush. Stretch a vector and nothing changes but the size.

That difference in how an AI actually has to learn the skill needs to be understood. Picture two students learning to draw a circle. One only ever watches videos of someone sketching a circle freehand and learns to imitate the wobble of the pen. The other studies the real geometry behind it: a center point, a radius, and the formula connecting them. The second student can produce a perfect circle at any size, on command, forever. The first is always just one more careful approximation. An AI trained only on flattened images of finished logos and icons is the first student. To produce real, editable vector files, it needs to train on the actual construction steps, the points and curves themselves, not just a picture of what those points eventually added up to.

Here’s the part that complicates the picture, though. It turns out an AI doesn’t need a mountain of that construction data to write some version of the underlying code. A vector file is just text, written the same basic way a webpage’s code is written, full of words and numbers a model can read and produce directly. Google’s Gemini models can already do this: ask for a simple rocket-ship icon in plain language, and Gemini will write out actual vector code rather than a flattened image, because it’s drawing on the same general skill it uses to write a Python script or an HTML page. That’s a real capability, and Google didn’t need anything close to Adobe’s vault of licensed illustrations to build it.

But ask for more than a basic shape and the cracks show.