The capability matrix

Four platforms, from the buyer's chair

A company shopping for an enterprise intelligence operating system (OS) asks one question per row: can this platform do it, and what does it cost me to get it? Here is the honest grid: Metis beside Palantir, Databricks, and Snowflake, capability by capability, with the pros, the cons, and the risks of each.

This page is shared by link during evaluations and is not indexed. Every competitor cell is sourced from public filings, documentation, and independent research; where a value is our estimate, it says so.

Talk to sales Why it works

Capability by capability

The grid

Twenty-four rows. The six marked rows are the category-defining ones, and they are one chain in order: the decisions say what data matters, each variable is priced by what knowing it is worth, what is missing gets a name and an owner, the gaps feed the collection assets, the ceiling is measured, and the returns are stated before the work.

Product capability
Partial, or the customer supplies the hard part
Not offered
Category-defining: a link in the method's chain

The grid
Capability	Metis	Palantir	Databricks	Snowflake
The platform, capability by capability
Data engineering and pipelines	Product capability. Zetta compiles verified pipelines	Partial, or the customer supplies the hard part. Transforms; often sits on a data platform underneath	Product capability. The market's deepest	Partial, or the customer supplies the hard part. ELT and Snowpark; heavy transformation strains the economics
Warehouse / analytical store	Product capability. The store, over yours, or becoming it	Not offered. Not a warehouse; consumes one	Product capability. The lakehouse	Product capability. The managed benchmark
Streaming and real-time	Partial, or the customer supplies the hard part. On the platform, where the decision needs it	Partial, or the customer supplies the hard part. Consumes feeds; not a streaming engine	Product capability. Strongest of the four	Partial, or the customer supplies the hard part. Improving; warehouse-first
Dashboards and SQL analytics	Product capability. Dashboards over one definition per metric	Partial, or the customer supplies the hard part. Operational apps more than BI	Partial, or the customer supplies the hard part. Independently rated less mature	Product capability. Best-in-class concurrency
Semantic / operational layer	Product capability. Meaning scored and owned per decision	Product capability. The Ontology, their center	Partial, or the customer supplies the hard part. Unity Catalog, plus semantics homework	Partial, or the customer supplies the hard part. Semantic views are the customer's homework, per their docs
Model lifecycle: train, serve, monitor	Product capability. Models trained next to the data	Partial, or the customer supplies the hard part. Orchestrates models more than trains them	Product capability. The reference stack	Partial, or the customer supplies the hard part. Snowpark ML, narrower
Generative AI and agents	Product capability. Agents under per-action permission	Product capability. AIP, governed by the Ontology	Product capability. Mosaic AI and Agent Bricks	Partial, or the customer supplies the hard part. Cortex is live; tools and semantics are customer-built
Governed action: per-action permission and audit for agents	Product capability. Agents act under per-action permission with full provenance	Product capability. The Ontology's actions: approved operations, controlled and audited	Partial, or the customer supplies the hard part. Unity Catalog and AI Gateway govern tools; the policies are the customer's build	Partial, or the customer supplies the hard part. Budgets and roles; agent tools and guardrails are customer-configured
Applications: build and host	Product capability. Applications live on the platform	Product capability. Foundry apps and workflows	Partial, or the customer supplies the hard part. Databricks Apps, young	Partial, or the customer supplies the hard part. Streamlit and Native Apps, young
The method's chain: the category-defining six
1Decision-driven, not data-driven: requirements derived from the decision inventory	Product capability. The method's first step	Partial, or the customer supplies the hard part. Workflow-first delivery finds decisions by hand; no derivation method	Not offered. Consumption is decision-blind	Not offered. Consolidation is decision-blind
2Relevance-driven: each variable priced by the value of knowing it	Product capability. Priced variable lists; unused data retired	Not offered. Not offered	Not offered. Not offered	Not offered. Not offered
3Existence gaps made visible: the missing data named, with an owner	Product capability. A catalog shows what exists; the profile shows what does not	Not offered. The Ontology models what exists	Not offered. Unity Catalog inventories what exists	Not offered. Horizon catalogs what exists
4Creating the data the gaps name: decision logs, collection assets, experiments	Product capability. The discipline the visibility feeds	Not offered. Models what systems emit	Not offered. Processes what systems emit	Partial, or the customer supplies the hard part. Sharing buys external data; internal gaps untouched
5The ceiling measured: availability scoring and the CIC score, before the spend	Product capability. The product's spine	Not offered. Not offered	Not offered. Not offered	Not offered. Not offered
6Returns stated before the work: the ROI chain, signed by finance	Product capability. A 400% ROI target wired to a measuring instrument	Not offered. Its filings disclaim outcomes	Not offered. Commissioned studies only	Not offered. Commissioned studies only
Delivery, deployment, the source, and scale
Forward-deployed delivery	Product capability. The execution side of the OS	Product capability. They invented it	Not offered. Field engineers sell, not run	Not offered. Partner-led
Speed to a first working result	Product capability. Days: one binary, one unit, scoring on day one	Product capability. The bootcamp: a working workflow on your data in days	Partial, or the customer supplies the hard part. The platform lands fast; value waits on an engineering program	Partial, or the customer supplies the hard part. Fast for managed analytics; AI value waits on the semantics homework
Deployment: on-prem and disconnected	Product capability. Your cloud, on-prem, or dedicated	Product capability. Their specialty, to classified and edge	Not offered. No on-prem posture	Not offered. Public cloud only, by their own documentation
Data sovereignty	Product capability. Complete sovereignty possible: your infrastructure, and you hold the source	Partial, or the customer supplies the hard part. Runs in your environment, to classified; the platform and its operation stay theirs	Partial, or the customer supplies the hard part. Data in your cloud account; the control plane is vendor-run	Not offered. Data lives in their managed public-cloud service
Source posture: who holds the code	Product capability. Source available: the enterprise receives the source	Not offered. Closed; no documented source delivery	Partial, or the customer supplies the hard part. Open components; closed platform	Not offered. Closed; open table formats only
Lock-in and exit	Product capability. Source available; deliverables in your terms; hand-over is a method step	Not offered. Documented: operational logic accumulates in the Ontology; nothing retained after the subscription	Partial, or the customer supplies the hard part. Tables leave in open formats; the compute, governance, and AI services do not	Partial, or the customer supplies the hard part. Iceberg opens the tables; the platform, SQL behavior, and Cortex stay proprietary
Cost predictability	Product capability. Priced per unit from the returns chain, agreed before the work	Not offered. Negotiated and opaque; no public rate card	Partial, or the customer supplies the hard part. Published rates; forecasting the bill is its own discipline	Partial, or the customer supplies the hard part. Published credits; a consumption volatility its filings acknowledge
Hand-over: your teams run it	Product capability. Teaching the unit's teams to run the loop is the engagement's last step	Not offered. Dependence on embedded engineers is the documented pattern	Partial, or the customer supplies the hard part. Yours to run, given the engineering bench it presumes	Product capability. Near-zero administration by design, for the analytics core
Ecosystem, integrators, track record at scale	Not offered. Being built	Product capability. A decade of the hardest environments	Product capability. The largest ecosystem	Product capability. Very large, analyst-led
The tallyEach column counted down: product capability, then partial, then not offered, out of twenty-four.

An enterprise intelligence OS that cannot see what data is missing, let alone create it, is a data platform wearing the name. The other three columns score near zero on the whole chain, and the last row is the honest one in the other direction: they have the ecosystems and the track records, and Metis is building both.

Returns

Three-year ROI, from customer evidence, not vendor studies

A vendor's own commissioned studies cannot be used to calculate return on investment: they are composite-model marketing built from interviews with selected adopters, and none of it appears in a contract. The candles below use customer-side evidence only; where a candle is our estimate, it is drawn dashed and says so.

Metis: The spread is narrow by construction: the price is set from the returns chain, so the engagement is scoped until the expectation clears the target. Realized history: being built.
Palantir: Estimated from customer behavior: documented exits at one end, documented multi-year expansions at the other. The vendor publishes no ROI figure and its filings disclaim outcomes.
Databricks: Estimated from independent, platform-agnostic surveys: bimodal. The modal pilot returns nothing; a disciplined minority reaches the top wick.
Snowflake: Same evidence base as Databricks: the surveys measure programs on the platform and do not separate the two.

For the record, and out of the business case: the vendors' commissioned composites claim 417 to 482% (Databricks; Forrester, Nucleus) and 354 to 612% (Snowflake; Forrester). Palantir publishes no number at all.

The base rates behind the wide candles

Customer-side measurements of AI programs at large, platform-agnostic. This is the market the target is priced against.

Generative-AI pilots with no profit-and-loss impact95%
MIT Project NANDA 2025, review
Organizations reporting any earnings impact from AI39%
McKinsey 2025, n = 1,993, self-reported
AI programs reporting negative ROI14%
IBM 2024, n = 2,413, self-reported
Agentic-AI users reporting significant ROI10%
Deloitte 2025, n = 1,854, self-reported
Organizations meeting the high-performer bar6%
McKinsey 2025; differentiator: workflow redesign

The theory's reading: every bar is the same failure measured differently — instruments funded above data that could not carry them, with no returns chain written first.

Through the returns chain, Metis is anchored on ROI: the engagement is scoped against what the unit actually needs, signed before the work, and measured after it. That anchoring is the buyer's protection — it is what makes the project that flops, or the spend that becomes an embarrassing waste, structurally hard to have.

Cost

What each platform costs, in words

Order of magnitude only; every vendor figure is negotiated and workload-dependent. Stated in words because this page carries no price.

MetisScoped per unit

The price is derived from the returns chain the engagement signs; entry is one unit and its priority decisions.

PalantirSeven to nine figures a year

Reported seven-figure entries; the average across its top twenty customers is close to nine figures (FY2025 filing).

DatabricksSix to eight figures a year

Consumption-priced; the company reports hundreds of customers above seven figures of annual consumption.

SnowflakeSix to eight figures a year

Capacity commitments of one to four years; its filing reports hundreds of customers above seven figures.

Pros, cons, risks

The honest column summaries

Metis

Pros: The whole chain: decision-driven and relevance-driven, existence gaps made visible and closed by creating the data, the ceiling measured before the spend, returns stated before the work. Source available with full sovereignty; rapid, unit-scoped deployment; the method handed to your teams.
Cons: Youngest of the four; ecosystem, integrator bench, and track record still being built; sales-led, no self-serve.
Risks: Young-vendor execution risk, mitigated by the unit-scoped entry, the signed returns chain, and holding the source. The CIC score must prove its predictive value engagement by engagement.

Palantir

Pros: The strongest operational-semantics layer shipped; unmatched deployment reach; a mature forward-deployed bench; the fastest proof-of-possibility.
Cons: Negotiated, opaque pricing; presumes a multi-year institutional commitment; complexity and training burden acknowledged in its own filings.
Risks: Deep dependence: closed source, no source delivery, your operational logic accumulating in their abstractions. Documented reputational drag in public-sector deals. No ROI commitment of any kind.

Databricks

Pros: The deepest engineering and model stack; open data formats so the tables are portable; a genuine agent stack under one governance plane; the largest talent pool.
Cons: Cost operationally sensitive and hard to forecast; a heavy skills burden by independent research; no answer to which decisions the spend serves.
Risks: Consumption grows whether or not decisions improve. No on-prem posture at all. Closed platform behind the open formats. No ROI commitment.

Snowflake

Pros: Managed simplicity at analyst scale; the market's best cross-organization sharing; high-concurrency SQL with near-zero administration.
Cons: Public cloud only, documented as such; heavy transformation strains the economics; the AI layer's semantics and tools are the customer's homework by its own docs.
Risks: Multi-year capacity commitments ahead of demonstrated value. Switching costs beyond the open tables. No ROI commitment.

Where every cell comes from

Competitor cells are compiled from public filings (Palantir FY2025 10-K, Snowflake FY2026 10-K), vendor documentation and releases, and independent research (BARC, G2, MIT Project NANDA, McKinsey, IBM, Deloitte, and the reported record of exits and expansions), researched 2026-08-30. Estimates are labelled as ours. The Metis column states the designed product and its founder-stated targets; the 400% figure is a target, not a guarantee.

Score it on your own decisions

The grid is the argument in general. Bring the decisions that matter and the data you have, and the first thing we do is score them; the second is show you what moves the score.

Talk to sales See the platform