Data-Infrastructure Closed Loops: Who Closes the Telemetry-to-Training Loop, and How
Headline finding
In data infrastructure today, the legitimate "deployment telemetry trains the vendor's next product version" loop is the exception, not the rule. The two pure cases are Salesforce Einstein (default-on, opt-out, recently softened to self-service in Spring 2026 after backlash) and Palantir Foundry (opt-in by design, mediated through the Ontology). Snowflake and Databricks both publicly disclaim training on customer data; their loop is the opposite shape — customer-data-stays-put, vendor brings the model. dbt, Tableau, and Power BI collect product telemetry (events, usage), not customer data, and offer opt-outs. The "loop is the moat" pattern Avedissian describes for robotics has weaker analogs in data infra than the framing implies, because contract terms and enterprise procurement actively block the loop closure that he names as the moat.
Implication for the personal-data-sovereignty thesis: the differentiation thins. RDCO's "user owns their data" pitch is not meaningfully differentiated against Snowflake, Databricks, dbt, or even Tableau, all of which already disclaim training on customer data in their enterprise tier. It IS differentiated against Salesforce (default-on telemetry training until Spring 2026) and against any consumer-grade SaaS where the loop is closed by acceptance of free service. The sharper RDCO wedge is not "we don't train on your data" (that's table stakes in enterprise data infra) but "we don't extract derived insights, query patterns, or usage analytics that constitute a parallel data asset for us" — the Datadog-style human-interface harvest, which is structurally different from model training.
Method
- 2 vault QMD passes (loop-is-moat thesis cluster, data-sovereignty terms)
- 5 targeted web searches (Snowflake/Databricks, Salesforce, Palantir, dbt/Hex, Tableau/Looker/Power BI)
- 2 web fetches (Salesforce VantagePoint analysis, Snowflake Cortex Analyst docs)
- Cross-referenced against Avedissian (May 9), Turing/Siddharth (Apr 30), Datadog/Natkins (Apr 28), and Mitohealth 5-layer loop (Apr 30) vault docs
Taxonomy: three distinct loop-closure shapes in data infra
The robotics-style "deployment telemetry trains the next model" loop is one shape. In data infra, it splits into three:
Shape A: Locked-in telemetry training (vendor harvests by default)
Mechanism: Master Subscription Agreement grants vendor rights to use customer data for "service improvement" / "model training" / "research and development." Default-on. Opt-out exists but is buried.
Canonical example: Salesforce Einstein (pre-Spring 2026).
- Auth: customer's existing Salesforce login; data is already in the platform
- Contract term: Master Subscription Agreement clause permitting use of "Customer Data" for "global predictive AI models"
- Data pipe: internal — data already lives in Salesforce's multi-tenant warehouse
- Opt-in/out: opt-out. Until Spring 2026, required filing a support case. Now self-service toggle at Setup → Einstein → Opt Out of Customer Data Access
- PII protection: Einstein Trust Layer masks credit cards, government IDs, ethnicity, financials before LLM
- Loop velocity: fast (data is already in-platform)
- Why this works for the vendor: they own the substrate, the auth, and the contract — three of three loop-closure surfaces
This is the closest SaaS analog to Avedissian's robotics loop. Notable that Salesforce got pressured into a self-service opt-out in Spring 2026 — a signal that the "default-on training" model is becoming socially expensive even when contractually permitted.
Shape B: Opt-in flywheel (vendor closes loop only with explicit customer cooperation)
Mechanism: customer explicitly chooses to share derived assets (RL traces, deployment outcomes, fine-tuned weights) back to the vendor in exchange for product improvement, often as part of a paid co-development or beta program.
Canonical example: Palantir Foundry + AIP.
- Auth: Foundry workspace identity
- Contract term: Foundry contracts vary by deployment; the loop runs through the Ontology layer rather than raw data — when a customer makes a decision in AIP, that action and its outcome are written back to the customer's own Ontology, not Palantir's central model. Palantir's improvement comes from selling the pattern (Ontology-Aware Generation) to the next customer, not from harvesting the previous customer's data
- Data pipe: customer's tenant only; cross-tenant model improvement is product/architecture iteration, not data harvesting
- Opt-in/out: opt-in by structure. Each customer's loop closes within their own boundary
- Loop velocity: slow on the cross-customer dimension; fast within a single customer
- Why this works for the vendor: the moat is the Ontology pattern and the deployment muscle, not aggregate customer data. Palantir's flywheel is "we get better at building Ontologies" not "we get better LLMs from your data"
Secondary example: Turing. Per [[06-reference/2026-04-30-jonathan-siddharth-turing-superintelligence-loop]], Turing positions its data-deployment loop as the moat — but the loop closes through Turing's engineers deployed into Fortune 500 enterprises, generating data Turing then sells to frontier labs. Customer signs a services contract, not a data-share agreement. The loop is mediated by labor, not by terms-of-service-grant. This is closer to a consultancy-flywheel than a true SaaS telemetry loop.
Shape C: Vendor improves itself, not its model, from product telemetry
Mechanism: vendor collects anonymized product-usage events (which buttons clicked, which queries ran, which errors hit) to improve the product UX. Customer's actual data is never touched. This is the dominant pattern.
Examples and contract specifics:
Snowflake (Cortex Analyst, Cortex AISQL). Explicit policy: "Cortex Analyst does not train on Customer Data. We do not use your Customer Data to train or fine-tune any Model to be made available for use across our customer base." SQL queries logged in Query History (customer-visible); conversation history not logged. Loop closure is through Snowflake Trail telemetry on pipelines and agents — operational metrics, not training data.
Databricks. Same posture: customers use their own enterprise data to fine-tune via MosaicML-stack within their workspace; data does not leave the customer's tenant for Databricks's central training. DBRX was trained on Databricks' own/licensed corpora, not on customer data.
dbt Labs. Telemetry on by default for dbt Core ("anonymous usage stats"); opt-out via
dbt_project.ymlorDO_NOT_TRACKenv var. dbt Cloud caches client data ephemerally during query execution; employees access only with manager approval and business need; "never with the purpose of enriching dbt Labs." Explicit contractual disclaimer of the harvest pattern.Tableau. Basic Product Data sent by default to
prod.telemetry.tableausoftware.com; enterprises block at the firewall. Product telemetry only; no customer-data flow.Power BI. Telemetry on by default; opt-out via
ENABLECXP=0installer flag. Per-user, not per-org, which is itself a tell about how seriously Microsoft treats the opt-out.
The structural pattern
In SaaS data infra, the loop that Avedissian names as the moat is contractually disclaimed by the leaders. Snowflake and Databricks compete on not harvesting customer data. The reason is procurement: Fortune 500 buyers won't sign deals where the vendor can train on their proprietary data. The "we don't train on your data" clause is now table-stakes in enterprise data-infra MSAs.
This means the loop closes elsewhere:
- Through the product UX (telemetry on which features get used → roadmap)
- Through Snowflake/Databricks's own model training corpora (Arctic, DBRX) trained on public/licensed data, then sold back to customers as in-platform inference
- Through pattern accumulation in the vendor's professional services / ecosystem (Palantir's Ontology playbook, Turing's deployment muscle)
The contract terms that would enable the robotics-style loop are precisely the terms enterprise data-infra vendors had to give up to win enterprise deals.
Cross-check: does this contradict the vault thesis cluster?
It partially does, and that's the most useful finding.
The vault thesis (Turing piece, Mitohealth 5-layer piece, Avedissian piece) treats "the loop is the moat" as a near-universal AI-era pattern. In enterprise data infrastructure, contract terms and procurement reality have shut the loop in its strongest form. What remains is opt-in flywheels (slower, smaller telemetry pool) and product-UX telemetry (doesn't actually train the next model, just informs the next roadmap).
The Datadog/Natkins piece [[06-reference/2026-04-28-semi-structured-datadog-moat-human-keyboard]] is consistent with this finding but at a different layer: Datadog's moat is human-interface lock-in, NOT customer-data harvesting. Datadog doesn't train models on customer telemetry; it charges customers to store their own telemetry and locks them in via dashboard muscle memory. That's a fourth shape: Shape D — interface lock-in masquerading as a data loop.
Implications for RDCO
Personal-data-sovereignty thesis: re-assess
The May 10 thesis (RDCO differentiates by "user owns their data, vendor doesn't harvest") needs sharpening:
- Against enterprise data infra (Snowflake, Databricks, dbt, Tableau): weak differentiation. They already disclaim training on customer data. Saying "we don't either" is parity, not advantage.
- Against Salesforce-class SaaS: strong differentiation. Default-on training was the norm until Spring 2026 and the social cost is rising.
- Against consumer SaaS / free-tier: strong differentiation. Free service IS the consent for data harvest in that contract shape.
- The actual sharp wedge: not "we don't train on your data" but "we don't extract derived assets — query patterns, schema metadata, usage analytics, dashboard topology — that constitute a parallel data asset we benefit from at your expense." This goes beyond model training to cover the Snowflake/Databricks Shape C and the Datadog Shape D harvest patterns.
Sanity Check angle (NOT derivative)
The vault Sanity Check ban on derivative pieces is satisfied here. Original re-frame: "In data infrastructure, the loop the AI thesis crowd celebrates is the loop the contracts forbid." The robotics framing assumes vertical-integration freedom that doesn't exist in B2B data; the contract-terms reality is the missing factor in the loop-is-moat thesis cluster. Worth queueing.
What to track
- Does Salesforce's Spring 2026 self-service opt-out cause measurable opt-out rates, and do other vendors follow? (Tells us whether the harvest model is dying in real-time.)
- Does Palantir's opt-in-by-Ontology pattern get copied? (Tells us whether the Shape B compromise is the sustainable middle ground.)
- Does any data-infra vendor try to build a cross-tenant training loop with explicit customer revenue-share? (Would be a genuinely new shape — closest analog: Foursquare's location-data-as-product play.)
Follow-up questions surfaced
- What does the Salesforce Master Subscription Agreement actually say about "Customer Data" use rights for AI training, verbatim? (VantagePoint piece references it but doesn't quote.)
- Has any enterprise customer publicly negotiated a data-sovereignty rider into a Snowflake/Databricks contract? Procurement intelligence on what's actually negotiable.
- What's Hex's actual telemetry posture? They sit between dbt and the warehouse and their notebooks contain analyst intent — extremely high-signal training data if they harvest it.
- Does Glean / Notion AI / Dropbox AI train on customer corpora? These are the closer enterprise-collaboration analogs to the robotics loop and weren't covered in this pass.
- What's the right RDCO contract-clause language to make "no derived-asset extraction" enforceable, given that Snowflake-style "no model training" disclaimers leave Shape C harvesting wide open?
Sources
Vault:
- [[06-reference/2026-05-09-avedissian-loop-is-moat-robotics]] — parent piece
- [[06-reference/2026-04-30-jonathan-siddharth-turing-superintelligence-loop]] — closest SaaS analog
- [[06-reference/2026-04-28-semi-structured-datadog-moat-human-keyboard]] — Shape D (interface lock-in)
- [[06-reference/2026-04-30-mitohealth-founder-5-layer-agent-native-company-loop]] — loop-thesis cluster
- [[06-reference/2026-04-10-jaya-gupta-anthropic-moat]] — moat-theory cluster
Web:
- Snowflake Cortex Analyst docs — explicit "does not train on Customer Data" policy
- Salesforce VantagePoint analysis — default-on Einstein training, Spring 2026 self-service opt-out
- Salesforce official docs — Global Model Opt-Out Process
- Palantir Foundry / AIP overview — Ontology-mediated feedback loop architecture
- dbt Labs Privacy Policy + telemetry docs — anonymous usage stats, opt-out via env var
- Tableau Basic Product Data docs — telemetry endpoint, firewall-block opt-out
- Microsoft Power BI / Fabric community —
ENABLECXP=0per-user opt-out