← Back to TIL
Second Gear / July 27, 2026

Open Models Sell Chips

Aggregation

World

0 critical / 1 happenings

Aggregation

Tech

1 critical / 0 happenings

Ideation

Ideation

Idea 1 Safety Cases for Robot Policy Updates

This company sells evidence packs that let a warehouse, factory, hospital, or insurer approve a new robot policy without pretending a one-time robot certification covers every model update. The primitive is a model-version safety case: simulation traces, real fleet telemetry, near-miss clips, operator overrides, and standards mappings compiled into a tamper-evident dossier for a specific robot body, site layout, task, and policy version.

Source Signals

Why Now

Open robot policies, world simulators, and synthetic data make deployment faster than the safety process around it. A facilities VP can buy robots this year, but every model update creates a new question from legal, insurance, and workers' comp: what changed, where was it tested, and what happens near people?

First Wedge

Start with autonomous mobile robots and mobile manipulators in warehouses. Pull logs from one fleet, replay high-risk events in simulation, map traces to ISO/ANSI safety requirements, and produce a monthly safety-case diff for the operator's insurer and internal safety lead.

Commercial Model

Warehouse operators and robot OEMs pay $3K-$15K per site per month, with higher-priced incident reconstruction when there is a claim or serious near miss. The budget exists because insurance premiums, customer audits, and worker-safety approvals can delay fleet expansion.

Defensibility

The company compounds a cross-OEM library of near-miss patterns, site morphology, task traces, and insurer-accepted evidence templates. Incumbents can certify hardware, and OEMs can log their own robots, but neither is naturally trusted as a neutral witness across mixed fleets and changing third-party policies.

Technical Risk

The hard part is making heterogeneous robot telemetry, simulation traces, and video evidence comparable enough that a safety reviewer trusts the output. False comfort is worse than no tool; the product needs calibrated uncertainty and clear boundaries.

Market Expansion

After warehouses, expand to hospitals, airports, retail floors, construction sites, farms, and defense logistics - any site where robots share space with people and where the buyer needs proof that a policy update did not silently change the risk profile.

Self-Critique

This could be too early if robot deployments remain pilots and insurers do not price safety evidence explicitly. It also risks becoming compliance consulting unless the telemetry connectors and evidence templates become repeatable software.

Next Experiment

In four weeks, partner with one AMR integrator and one broker. Reconstruct ten near-miss events from logs and video, generate a policy-version safety diff, and ask the broker which fields would change underwriting confidence or premium language.

Idea 2 Credit Maps for AI Infrastructure Commitments

This company sells a credit-grade map of AI infrastructure exposure: which data centers are real, which power contracts and interconnect queues back them, which chip leases or take-or-pay commitments sit underneath them, and which public companies or private-credit vehicles are economically tied to the same project. The primitive is not a data-center database; it is a physical exposure graph for the AI buildout.

Source Signals

Why Now

The AI trade is becoming a financing trade. Capex, leases, power, local incentives, and credit structures now matter as much as model benchmarks, while many investors still rely on earnings-call language and generic data-center market maps.

First Wedge

Sell to credit funds and bond desks a weekly verified watchlist for the top 100 AI infrastructure obligations: site status, power availability, permitting, counterparties, lease economics, debt stack, covenant changes, and contradictory public claims.

Commercial Model

Private-credit funds, infrastructure lenders, insurers, and rating-adjacent research teams pay $50K-$250K per year per seat group. The buyer already pays for Bloomberg, rating research, satellite data, legal review, and specialist channel checks because one bad exposure can erase years of subscription cost.

Defensibility

The system compounds proprietary mappings between public filings, utility queues, satellite/construction signals, permit records, equipment supply chains, and financing vehicles. Incumbent financial terminals have distribution, but they usually lag on messy physical verification; construction data vendors lack the credit graph.

Technical Risk

The hard problem is entity resolution across shell project companies, landlord-tenant leases, utility filings, equipment vendors, and private-credit vehicles. The product must show confidence levels and source trails rather than producing a single fake-precise exposure number.

Market Expansion

Start with AI data centers, then expand into semiconductor fabs, battery plants, defense factories, and climate infrastructure where capital-market risk depends on physical buildout, power, permitting, and hidden counterparty concentration.

Self-Critique

This can fail if the best data remains locked inside banks, rating agencies, or utilities, or if customers treat it as nice-to-have research rather than trading and underwriting infrastructure. It also faces strong incumbents if the product is only a prettier market map.

Next Experiment

In two weeks, build one exposure graph around Oracle/CoreWeave and one around a hyperscaler-backed power project using only filings, permits, utility dockets, satellite imagery, and news. Put it in front of three credit investors and ask what they would have paid to know earlier.