Data for the Real World: The Internet Has Been Read — the race to measure what was never written down

This is the eleventh article in this fourteen-part series reading Y Combinator's Fall 2026 Requests for Startups one request at a time. The requests so far have mostly asked founders to apply intelligence we already have to problems we already understand. This one is different. "Data for the Real World," written by Austin Tindle and Diana Hu, asks founders to go get the raw material itself — to build the instruments that will feed the next generation of models, because the current generation has largely exhausted its diet.

What the request actually says

The argument opens with an asymmetry. AI has become remarkably good at learning from data, and we now have superhuman models for code, language, and images — domains where the training data was already digital, abundant, and free to copy. "But for the physical world?" the authors write. "We're still working with sparse data from remote sensors designed for humans, not AI."

That last clause is the sharpest observation in the request. The world is not unmeasured; it is measured for the wrong consumer. A weather station reports what a human forecaster needs to glance at. A pressure gauge on a boiler reports what an inspector needs to log. None of this was designed to be dense, continuous, or machine-consumable at training scale. The request's thesis is that two curves have now crossed — improving foundation models and plummeting sensor costs — making dense physical-world data collection feasible for the first time.

The authors write from inside the problem. Tindle is the founder of Sorcerer, a YC company whose autonomous weather balloons collect atmospheric data that, per the request, the US government uses to improve forecasts. The text also cites Gecko Robotics, which sends robots into hard-to-reach places to collect data and build predictive models. Then it widens the aperture: the world's biggest industries — energy, agriculture, logistics, construction — "rely on limited data and intuition-based models."

And then it says the quiet part plainly: "More real-world data enables precise modeling. And once you can model a system, you can control it. Steering hurricanes. Reversing desertification. Cooling the planet." The request is not asking for dashboards. It is asking for the measurement layer of a control loop, and it is candid that control — not observation — is the destination.

Why now

The timing case rests on a supply problem and a demand problem arriving together.

The supply problem is the data wall. Frontier language models were trained on something close to the entire useful internet, and the internet does not grow at the rate model appetites do. The domains where AI went superhuman are exactly the domains where data was a byproduct of human activity — every commit, post, and photo was training data someone else paid to create. The physical world produces no such exhaust. Its data must be deliberately collected, which is why it remained sparse while text became infinite.

The demand problem is embodied AI. Robot foundation models need action-paired sensorimotor data — synchronized joint angles, forces, camera frames — and industry accounts describe each teleoperated demonstration consuming minutes of skilled operator time to yield a single training trajectory. There is no web-scale corpus of the physical world behaving. Whoever builds one owns something that cannot be scraped, only replicated at comparable cost — which is the classical definition of a moat, and increasingly the only kind of data moat left.

The cost curves matter too. Reported figures around the cited companies illustrate the shift: Sorcerer describes balloons that stay aloft for months and collect on the order of a thousand times more data per dollar than conventional radiosondes, which fly for hours and are discarded. When the cost of a measurement falls by three orders of magnitude, the sensible architecture flips from sampling to saturation.

What is actually hard

The physics of cheap sensing is the easy part. The hard parts are economic and institutional.

• Sensor economics are unforgiving. A dense network means thousands or millions of units, each of which must be manufactured, powered, connected, and eventually retrieved or written off. The per-unit cost that looks charming in a pilot becomes the entire P&L at coverage scale, and revenue usually arrives only after coverage exists — a capital-intensity trap software investors are not built for.
• Ground truth is scarcer than data. A million readings are worth little until something links them to outcomes: this vibration signature preceded that failure, this soil profile produced that yield. Labels in the physical world arrive slowly, on the calendar of harvests and corrosion, and no annotation vendor can speed them up.
• Coverage is a chicken-and-egg problem. A weather model wants the whole atmosphere; a partial constellation improves forecasts only marginally until it crosses a density threshold. The value function is convex in coverage while the cost function is linear, which makes the middle of the buildout the most dangerous place to be.
• Privacy and sovereignty bind harder than in software. Dense sensing of farms, cities, and infrastructure is surveillance by another name if governed badly, and much of the most valuable data — airspace, weather, defense-adjacent infrastructure — sits under regulators who move at their own pace. The request's own examples sell to governments, which is both validation and a warning about sales cycles.

Who is attempting it, and what it takes

Beyond the two companies named in the request, the pattern is visible elsewhere. WindBorne Systems, another balloon constellation company with Stanford roots, pairs its fleet with an in-house AI forecast model reported to steer the balloons themselves toward data gaps — a closed loop where the model directs its own data collection. Gecko Robotics is reported to operate a fleet of roughly 250 inspection robots and recently won a US Navy contract to assess warship hulls. In embodied AI, companies such as Physical Intelligence bootstrap manipulation models from video and then buy the expensive real-world trajectories demonstration by demonstration. The common shape: hardware that collects, a model that consumes, and a feedback loop where the model's needs dictate where the hardware goes next.

That shape is also the recipe. Building in this space means being three companies at once — a hardware manufacturer with real unit economics, a data operations company that turns raw sensing into labeled, calibrated, sellable truth, and a modeling company that converts the dataset into decisions someone will pay for. The failure mode is stopping at the first: selling sensors is a commodity business; owning the only dense record of how a system actually behaves is not. The request's examples all charge for outcomes — forecasts, inspections, predictions — with the dataset compounding quietly underneath.

Where Gwen stands

Sensor networks, balloons, and robots are not Gwen's lane, and this paper will not pretend otherwise. Gwen builds and hosts websites and applications from plain-language descriptions, and does marketing, research, and operations work — digital labor, delivered through long-lived missions in a customer workspace. Nothing in that touches a hurricane.

But the request's underlying lesson — that proprietary end-to-end data is the durable asset — applies in Gwen's domain directly. Underneath Gwen's work runs a routing rail that sends each task across many AI models by measured quality and cost, with caching and continuous evaluations. Every mission produces telemetry about which models do which work well, at what price, under what conditions. That measurement is itself a proprietary dataset, accumulated task by task, and it is the digital analogue of what Tindle and Hu are asking founders to build for the physical world: not observations about a system, but the operating record of one — dense enough, eventually, to steer by.