The Information Frontier Bench finds that AI agents trying to reach live data are blocked mainly by payment and identity gates rather than model quality — a gap the paper says dwarfs the difference between frontier models by roughly an order of magnitude.
The model can read the page, but it cannot get past the login screen, the API key request form, or the invoice. A paper posted July 13, 2026 to brickroad.network, turns that everyday friction into a measurable claim. The paper introduces the Information Frontier Bench (IFB), which it calls a “living benchmark,” to test how far AI agents can actually reach into canonical economic verticals when let loose with real access constraints — not just how well they parse what’s already in front of them.
The headline finding, in the paper’s own words: “reach is access-bound, not model-bound: the gap across access regimes dwarfs the gap between state-of-the-art models by roughly an order of magnitude.” And when agents do fail, the paper reports the failure is “gate-shaped” — they get stuck at payment and identity checkpoints, not at finding or reading the data. For Brickroad’s agents, the scarce resource isn’t better parsing or bigger context windows, it’s infrastructure that lets an autonomous buyer actually transact for access.
gate-shaped, not model-shaped
The distinction matters because it reassigns where value should accrue. If reach were model-bound, the frontier labs building GPT-, Gemini-, and Claude-class systems would be the natural chokepoint, and pricing power would sit with whoever has the best weights. Instead, the paper’s framing puts the bottleneck downstream, at the payment rail and the identity check that gates an API call or a paywalled dataset. That is a different market than the one annotation shops and synthetic-data vendors have been building for — it points toward infrastructure that lets an agent, not a human, authenticate and pay in real time.
a frontier that keeps expanding
The paper’s second argument is that there is no near-term ceiling forcing consolidation among data holders. It cites Lloyd’s estimate of roughly 10^90 storable bits in the observable universe as a permissive physical ceiling — a figure the paper itself flags as “striking, if hard to independently verify” in spirit, since it assumes converting all matter-energy in the universe to computronium with no error-correction overhead. Against that, it puts today’s global datasphere at roughly 10^24 bits and cites the IDC Global Datasphere series growing from about 33 zettabytes in 2018 to roughly 175 zettabytes in 2025, with extrapolations to about 400 zettabytes by 2030 — a growth rate the paper pegs at 10–30% a year with “no saturation visible.” Separately, it cites Hilbert and López’s 2007 census putting compressed humanity-stored bits at roughly 3×10^20, a figure that sits an order of magnitude below IDC’s byte-equality count for the same period, a discrepancy the paper attributes to different ways of counting duplicate bits rather than resolving it outright.
What actually binds supply, the paper argues, isn’t physics but engineering. It cites the Landauer limit — about 3×10^-21 joules per irreversible bit-write — and says real machines run roughly a billion times less efficient than that floor. It also estimates today’s global measurement throughput, across cameras, sensors, and telescopes, at roughly 10^20 bits per second, which by its own math would take about 10^24 years to fully scan the Earth. None of these numbers are independently verified in the piece; they are presented as the paper’s own calculations, and the trade press should treat them as such.
what this means for buyers of data
If the paper’s access-bound framing holds up under scrutiny, it reshuffles who competes for margin in the training-data economy. Frontier labs already pay for tokens through licensing deals and annotation pipelines; this paper argues the next competitive edge is whichever intermediary (presumably Brickroad can) can broker machine-to-machine payment and identity verification at scale — a lane currently occupied by no single dominant player in the dossier’s own account. The paper describes its harnessed-agent framing — retrieval, tool use, planners, judges, memory, all chained together — as the dominant deployed pattern “by 2026,” a nod to what it calls a possible “sibling to Sutton’s bitter lesson”: how the computational graph is built may matter less than whether one can be built at all.
what to watch
The authors say they are releasing the bench, its trace contract, and supporting infrastructure so outside groups can extend it “without compounding load on production endpoints” — an offer that, if taken up, would let the claims be tested independently. For a market that already prices scarcity in tokens and annotated examples, the more provocative question the paper raises is whether the next scarce input is not data at all, but the checkout flow.
We find that reach is access-bound, not model-bound: the gap across access regimes dwarfs the gap between state-of-the-art models by roughly an order of magnitude. And failure is gate-shaped: when agents fall short, they overwhelmingly fail at payment and identity gates, not at discovery or parsing.
