How AI Infrastructure Is Measured

One thing to understand before reading further

sourc.dev does not tell you what is best or worst. It does not rank by value judgment. It records what the data shows.

If you ask a question and the data supports an answer, you get one. If the data does not support it, that is stated clearly. What is most valuable for you is for you to determine. The interpretation is always yours.

The market moves faster than the tools built to measure it

The AI infrastructure market is moving faster than any traditional analysis framework was built to handle. These are not estimates. These are verified observations from the sourc.dev archive.

None of these values are manually estimated. Each is computed from verified, sourced data and updated automatically. The archive grows every day. What it recorded yesterday cannot be reconstructed tomorrow.

What the data can answer

The same data answers different questions depending on who is asking. Here are examples drawn directly from the archive.

We can generate more or less any fact the data supports. If we cannot answer, we do not have the data — and we say so. What you see on this page is not all you can get.

For the investor

For the CTO

For the developer or creator

These are examples. The data behind sourc.dev can answer hundreds of variations of these questions — combinations of price, capability, integration, compliance, geography, and market position.

What you see on this page is not all you can get. The depth is in the data, and there is more under the hood than what the public display shows.

Contact → for institutional access, API licensing, or custom intelligence.

The knowledge graph

Most AI data platforms store lists. A list tells you that Cursor exists, that Claude 3.5 Sonnet exists, that Anthropic exists. It tells you what each costs and how each benchmarks.

A list cannot tell you that 34 tools depend on Anthropic. It cannot tell you that if Anthropic changes its pricing tomorrow, those 34 products are directly affected. It cannot tell you that 3 of those tools have no alternative — Anthropic is their only certified LLM dependency.

sourc.dev is not a list. It is a map.

Every entity tracked — models, tools, SaaS platforms, APIs, companies — is a node in a directed, weighted, and temporal knowledge graph. The connections between them are not inferred or estimated. They are verified, sourced, and dated.

Why relations are weighted differently

Not all connections are equal. A certified integration — where both parties publicly document the relationship — is fundamentally different from a community mention where one party references the other informally. They are weighted accordingly.

This weighting matters because it determines how much each connection contributes to ecosystem scoring. When sourc.dev records Anthropic's Blast Radius Score as 34, it means 34 tools have certified — not just mentioned — dependencies on Anthropic's models. That is a precise claim. The weight system is why it is precise.

Why the history of the graph matters as much as its current state

A relation that existed in 2024 but was removed in 2025 is not deleted. It is marked deprecated, with a date and a source documenting the change.

When did a tool add support for Claude 3.5 Sonnet? When did it drop GPT-3.5? Which AI integrations formed in the month after a major pricing change? These questions can only be answered if someone was recording the graph at the time those changes happened.

We were. And we will be tomorrow.

What the graph looks like today

How the system works

sourc.dev is not a manually maintained catalogue. It is a system of interconnected processes that continuously finds, verifies, enriches, monitors, and preserves data about the AI infrastructure ecosystem. Each process has individual value. Together they are self-reinforcing — and together they produce something no single process could produce alone.

Discovery — finding what exists

When a new entity enters the graph, it is automatically checked against every existing entity for potential connections. If a verified relation points to an entity not yet tracked, that entity becomes a candidate and goes through the same process. The graph expands with the ecosystem — not by manual mapping, but by following verified connections wherever they lead.

Discovery finds it. But without the rest of the system, a discovered entity is just a name. What gives it meaning is what comes next.

Linking — verifying how things connect

Every relation in the graph requires a primary source — an official page, a documented partnership, a marketplace listing, a press release. Relations are not inferred from similarity or assumed from co-occurrence. They are verified against evidence and dated when that evidence was found.

Linking is what turns a list of entities into a map. It is also what makes Blast Radius Score meaningful — because every connection counted in that score was individually verified.

Enrichment — describing what exists

Entities are continuously enriched with verified attributes — pricing, benchmark scores, authentication signals, distribution presence, compliance flags, lifecycle stage. Each attribute is sourced independently and timestamped. Nothing is assumed from the category an entity belongs to.

Enrichment is what gives metrics something to compute on. The sourc Value Index, the Price Deflation Index, the Integration Density Score — none of these exist without enrichment filling the underlying data.

Pipeline — keeping it current

A daily automated pipeline monitors pricing changes across tracked entities. When a price changes, the new value is recorded with its source and date. The old value is preserved. Nothing is overwritten.

Every pipeline run adds to the archive. Every day that passes is one more day of verified, timestamped pricing history that cannot be reconstructed retroactively. The pipeline does not just keep data current — it builds the time-series that makes trend analysis possible.

Radar — detecting what changes

The ecosystem does not only change on a daily schedule. New integrations form. Products launch and deprecate. Partnerships are announced. Prices shift without warning. Radar monitors for these changes across entity pages, official documentation, and structured data sources.

When something material changes, it is recorded — not replaced. The radar ensures that no significant change in the ecosystem passes unrecorded. Every observation that radar captures becomes part of the permanent archive.

Notifications and feeds — communicating what happened

Changes detected by radar and pipeline do not stay buried in raw data. They are surfaced as intelligence signals — plain-language statements of what changed, when it changed, and by how much.

A signal is always derived from verified data. It is never editorial. It describes what the data shows — not what it means for any particular decision. That interpretation remains with the reader.

The archive — the only source for what was

For historical values, trend analysis, and time-series data, sourc.dev's own archive is the authoritative source. Not because it was decided to be — but because no other source has it.

A price recorded on November 1 2024 with a verified source URL cannot be found anywhere else after the fact. It exists only because it was captured at the time. The same is true for every deprecated relation, every enriched attribute, every pipeline observation since the platform launched.

The archive is append-only. No row is ever deleted or modified. It is the reason every other process in this system has permanent value — because every discovery, every verified link, every enriched attribute, every pipeline run, every radar signal becomes part of a record that compounds every day.

The combination

Discovery finds it. Linking verifies it. Enrichment describes it. Pipeline keeps it current. Radar catches what changes. Notifications communicate it. The archive preserves it all — permanently, append-only, irreproducible.

Take away any one process and the system works less well. Keep all seven and the system is self-reinforcing: each process makes the others more valuable, and the archive that results is more valuable than the sum of its parts.

That is what makes sourc.dev an observatory rather than a directory.

How data is handled

Pricing — why sourc.dev tracks what it tracks

sourc.dev tracks one price per entity: the monthly price at which a professional user gains full access to the product's AI capabilities — including all models, integrations, and extensions that the knowledge graph describes. This is called price_full_access_usd.

Free tiers are not tracked. Free tiers are capability-limited, quota-limited, and feature-limited. They do not describe the product the relations graph documents. Pricing that does not unlock what the graph describes is not comparable to pricing that does.

If a product's full capabilities are genuinely free — open source, no paywall, no quota — the price is $0.00. That is the correct answer.

Freshness — why every data point has an age

Data ages. A price verified three months ago may no longer be accurate. A relation documented last year may have been deprecated. Every data point carries a verification date — and is flagged when too long has passed since last verification.

Stale data is visible as such. It is not hidden. A stale value with a visible date is more useful than no value — because it tells you something real, with an honest caveat attached.

Quality gate — why some metrics show Calibrating

A metric is not published until its underlying data meets minimum thresholds. Below the threshold, the metric shows Calibrating.

This is not a placeholder. It is not an estimate. It is an honest statement that the data is not yet sufficient to publish a reliable number. A number that cannot be defended should not be shown.

Unknown is not false

If a data point has not been verified, it appears as unknown. Not as zero. Not as false. Not as absent.

Unknown means unverified. False means checked and confirmed absent.

A product with unknown SAML support is not the same as a product with verified no SAML support. That difference is never collapsed.

Built to trust with capital

This platform was not built to look rigorous. It was built because verified, sourced, timestamped intelligence on the AI infrastructure market was needed — for real investment decisions, with real capital.

The standard is simple: trustworthy enough for capital allocation — published. Not trustworthy enough — not published.

Every night, that standard is measured across seven dimensions and expressed as a single number — visible on this page.

How entities are ranked — and why relations come first

Every entity on sourc.dev has a tier. That tier is not editorially assigned. It is computed from data.

Ranking is by relations — not by size, revenue, funding, or reputation — because relations are the only objective measure of ecosystem weight in the AI infrastructure market. An entity that 34 other products depend on carries more ecosystem weight than one that 2 products depend on. That is a data observation, not a value judgment.

This is why Anthropic ranks as ecosystem infrastructure. Not because it was decided editorially. Because 34 published tools have certified dependencies on Anthropic's models. If Anthropic changes its API tomorrow, those 34 products are directly affected.

Tier is not permanent. It is computed from current data and changes as the ecosystem changes. An entity that gains certified integrations moves up. One that loses them moves down. The data decides — always.

Thresholds calibrated against current entity distribution. Reviewed quarterly as the platform grows.

What is not published — and why

Transparency includes being explicit about what is left out.

The goal is not to tell you what to do. The goal is to give you the most accurate, sourced, and structured picture of what is — so you can decide for yourself.

The metrics

Price Deflation Index

Measures how much cheaper a model has become relative to the GPT-3 Davinci baseline of June 2020 — the first widely commercially available large language model at scale.

The 2020 baseline gives every model a common, immovable reference point. Consistency over time is more valuable than perfection at a single point.

Current example: GPT-4o PDI 4.2 — 95.8% cheaper than the 2020 baseline. Gemini 1.5 Flash PDI 0.17 — 99.8% cheaper.

Value Density Score

Measures one dimension of intelligence per dollar — specifically, how much MMLU-measured reasoning capability each dollar of input cost purchases.

This is a single-dimension metric. A model with a lower score may still be the right choice when speed, context length, or instruction-following matters more than cost per benchmark point. The limitation is documented because it should be known.

Current example: Gemini 1.5 Flash VDS 1,052 — GPT-4o VDS 35.5.

Integration Density Score

Measures how deeply embedded an entity is in the AI infrastructure ecosystem — weighted by the strength of each integration type. A certified integration contributes more than a community mention. Both are counted. Neither is hidden.

Current example: LangChain IDS 38 — LlamaIndex IDS 12.

Blast Radius Score

Measures ecosystem dependency. How many other published entities have a certified dependency on this one. If this entity changes pricing, degrades, or becomes unavailable — how many AI products are directly affected.

Current example: Anthropic Blast Radius Score 34. A pricing change reaches 34 certified tools directly.

sourc Value Index

Measures value per dollar, adjusted by entity type. How much AI capability does each dollar of monthly spend deliver — within each domain.

Version 1 is intra-domain only — valid for comparing tools against tools, and models against models. A normalized cross-domain version is in development. Partial metrics are labeled as partial.

Current example: Cursor sourc Value Index 4,650 — coding tools domain.

Velocity Index

Measures the rate of new certified integrations over a rolling 90-day window. Shows the direction of ecosystem position change over time.

In development. Active Q3 2026 when sufficient relation history is available.

Versioning — why history is never rewritten

When a formula changes, it receives a new version number. Historical values computed under the old formula are never overwritten. Both coexist in the archive with their respective version tags.

The archive is append-only by design. No row is ever deleted or modified. This is not a technical preference. It is a business rule — because the archive is the asset.

Changelog

Frequently asked questions

Citations and contact

For citations in published work:

For questions about methodology, corrections, institutional access, or licensing:

Contact →