Frontier question · Model systems

Can experimental data become a more durable moat than the model?

AI4S.fyi synthesis · Current answer

Experimental data may create a durable advantage, but exclusivity is not the same as irreplaceability.

Whether it improves a key task, how costly it is to reproduce, and whether the platform can keep producing useful new data matter more than raw volume. Models, data, platforms and team experience may also reinforce one another.

Updated 2026-09-10. The four assets are compared without assuming a fixed decay order. Interview views and open-source choices reveal strategy, but do not independently validate long-term commercial outcomes.

Why it matters

Whether the durable asset is the model, the data, the instrument or the operating experience determines what these companies should do with their capital — and determines what an outsider should read into an open-source release. When a company gives away its models, it is either being generous, or telling you where it thinks the value is not.

Four candidate assets, ordered by how fast each loses scarcity · 9 entries

Fastest

The model

2 key
1 further

Proprietary data

2 key
1 further

The platform

2 key
1 further

Operating experience

1 key
0 further

Slowest

Counts are not weights. The ordering is a hypothesis drawn from the interviews, not a measured result — and the ordering itself is what the disagreement is about.

Asset one · fastest to commoditise

The model

Weights, architectures, and the capability they encode.

AssessmentRadical's open-source actions demonstrate strategy, not that models must commoditize fastest. Lila's emphasis on model value remains a company self-report.

Confidence: moderate · 2 key entries that directly contradict each other

Key evidence

Radical AI — "Models aren't the moat. Experiments are."

→ Data is the moatExpert testimony + public artifacts
2026.08

Bearing on this question

Krause expects models to commoditise and has open-sourced TorchSim, MATRIX and LitXBench accordingly — TorchSim has since been spun out into an independent non-profit. This is the strongest evidence on the page precisely because it is a costly signal: the company gave away work it could have kept.

Does not establish

Acting on a belief is not evidence the belief is correct. The proprietary experimental data layered on top of the public models is not released, so the claim cannot be checked from outside.

Open interview →

Lila Sciences — "The model itself is the thing of value"

→ Model is the moatCompany self-report
2026.07

Bearing on this question

The direct contradiction, from the most highly valued of the three. Lila describes the laboratory platform as the token generator that feeds the model, making the data instrumental and the model terminal — the exact inversion of Radical's position.

Does not establish

Nothing public tests either position. Both companies are describing bets, not results.

Open interview →

Further evidence

1 entry · corroborating

2026.02
CuspAI interview — Welling expects most models to be open source within five years, with perhaps one or two proprietary exceptions
Testimony
Data

Asset two

Proprietary experimental data

Measurements the company owns and no one else has.

AssessmentOpenBind-0 supports the scientific value of coverage, but not a commercial moat for proprietary data. Replication cost, continued supply, rights and substitutes also matter.

Confidence: low to moderate · 2 key entries

Key evidence

OpenBind-0

→ Data is the moatPhysical dataset · Open model and data
2026.08

Bearing on this question

717 new experimental structures, with gains concentrated where they changed training coverage. The cleanest public demonstration that the value of a data asset is about which regions it covers, not how large it is — which means a data moat can be deep and narrow, and a competitor with different coverage is not behind, just elsewhere.

Does not establish

The 2021–2025 model comparison is not fully controlled, and the loop from identified gap to next dataset is not closed.

Open evidence →

Lila Sciences — ten trillion verified reasoning traces

→ RefinesCompany self-report
2026.07

Bearing on this question

The largest claimed data asset in the field, and the least inspectable: the split between real physical experiment, simulation and model-generated tool traces is not disclosed, nor is the deduplication method. If a data moat is real, this is the biggest one — and there is no way to size it from outside.

Does not establish

No public link between any part of this corpus and any measured model improvement.

Open interview →

Further evidence

1 entry · bounds the claim

2026.06
MinerU.Chem — extracting existing records cannot fill regions never documented; a hard bound on how large a data moat can be built from literature alone
Public system
Refines

Asset three

The physical platform

The ability to generate data at all: instruments, integration, orchestration.

AssessmentHard to copy, but the difficulty is accumulated engineering rather than a technical barrier — and that kind of moat erodes. There is direct evidence in the interviews that it is already eroding.

Confidence: moderate · 2 key entries

Key evidence

CuspAI — "It's not rocket science" to design, surprisingly hard to build

→ Platform is the moatExpert testimony
2026.02

Bearing on this question

Welling says the platform architecture he wrote down at the start is still essentially the architecture; the moat is the data you can obtain and the engineering of actually building it. Coming from someone with every incentive to describe his platform as technically deep, the choice to describe it as merely hard is informative.

Does not establish

No throughput, hit-rate or uptime figures are disclosed, so the size of the advantage cannot be assessed.

Open interview →

Radical AI — the war stories

→ Platform is the moatExpert testimony
2026.08

Bearing on this question

Reverse-engineering instrument software with no API, designing custom actuators for hot alloy samples stuck to trays, discovering that mechanical and mechatronics engineering are two different hires. Krause now describes this accumulated unglamorous work as the company's real moat — "it's not about a robot in front of a tool."

Does not establish

Difficulty is not durability. Nothing here shows the next entrant will face the same difficulty.

Open interview →

Further evidence

1 entry · erodes this asset

2026.08
Radical AI interview — instrument vendors refused API access two years ago and have since begun providing support; direct evidence this particular barrier is thinning
Testimony
Erodes

Asset four · slowest to transfer

Tacit operating experience

What the people who run the system know and cannot write down.

AssessmentSlowest to transfer, and hardest to prove exists. The property that makes it durable — resistance to documentation — is the same property that makes it unverifiable from outside.

Confidence: high on its importance, low on whether it accumulates · 1 key entry

Key evidence

Radical AI — the 3M adviser

→ Experience is the moatExpert testimony
2026.08

Bearing on this question

An adviser with 35 years at 3M told the team that the hardest thing to capture in manufacturing is knowing which knob to turn at which moment — and that writing it as a formula does not make it reliable, because the right setting is not the same every time. This is the clearest available statement of why this asset resists both copying and verification.

Does not establish

Whether such experience accumulates into an institutional asset, or simply lives in individuals and leaves with them.

Open interview →

Counter-evidence and cross-pressure

Reasons the ordering above may be wrong

The Lila position already appears as key evidence under asset one, because on this question the counter-argument is not a footnote — it is one of the two main claims.

Radical AI — open sourcing built position rather than eroding it

→ ComplicatesCompany self-report
2026.08

The claim

Releasing the models returned more community feedback and more ideas than the company could have produced internally, and TorchSim grew into an independent organisation.

Why it matters here

It suggests the model/data dichotomy is too clean: giving the model away may be a way of acquiring something else. A moat framing that assumes assets are held rather than exchanged may be the wrong frame entirely.

Open interview →

Max Welling — experience may not accumulate either

→ ComplicatesExpert testimony
2026.02

The claim

Automation is highly vertical-specific: solving one problem completely does not carry to the next, because the instruments and the models both change.

Why it matters here

If operating experience is as vertical-bound as the automation itself, then the slowest-decaying asset is also the least portable — durable within one domain and worthless outside it. That would make every position on this page correct locally and wrong as a general claim.

Open interview →

AI4S.fyi synthesis · Current gap

No public case links a specific proprietary dataset to a specific measurable model improvement. Until one exists, the data-moat thesis rests entirely on the confidence of people who have already bet on it.

What would change our view

DataAny company publishing a causal chain from one batch of proprietary data to one measurable model gain, with the counterfactual.
ExperienceA natural experiment: what happens to a platform's output after key technical staff leave. This is the only cheap test of whether tacit experience is institutional or personal.
ModelThe date at which open models match proprietary ones in a scientific domain — a direct test of the five-year prediction, and the cleanest falsifier on this page.
PlatformInstrument vendors shipping usable APIs as standard, which would remove the integration barrier that currently constitutes much of the platform moat.