Skip to content

CONCEPT Cited by 1 source

Send the model to the data (not data to the model)

Definition

Send the model to the data is the enterprise-AI deployment thesis that for the bulk of an enterprise's data — the private data that constitutes its competitive moat — the model executes inside the enterprise's firewall against local data, rather than the data being shipped to a hosted frontier-model API call.

Canonical statement on the wiki

Alex Gallego (Redpanda founder) canonicalises this as Redpanda's enterprise-AI infrastructure thesis (Source: Gallego 2025-04-03):

"Most large enterprises do not want to share data because that's the moat. Our intuition has been that the future will be about sending model to the data, not data to the model for the bulk of the data that an enterprise processes."

"Smaller, open models like Llama3, Gemma3, DeepSeekV3, Phi-4 provide state of the art performance for most tasks, and do it often at good enough quality in a single GPU. Beyond the price/performance improvement, it gives enterprises control of exactly what data is shared outside its firewall."

"Beyond the (often 10x less) cost advantage of purpose-specific, fine-tuned models on your data, there's no better custodian for your data than you."

Why this works in 2025

Three substrate trends Gallego cites make this thesis viable (vs 2023's assumption that enterprises would have to accept frontier-API dependency):

  1. Small open models reached "good enough" for most tasks. Llama3, Gemma3, DeepSeekV3, Phi-4 achieve state-of-the-art on the task classes most enterprises care about (classification, extraction, summarisation, structured output) — frontier models are no longer required for the typical internal use case.
  2. Single-GPU inference is tractable. Quantised versions of the above run on a single enterprise-class GPU — no specialised multi-GPU serving cluster required.
  3. ~10× price-performance for purpose-specific fine-tuned models vs frontier-API calls at comparable per-task quality (Gallego's asserted number, no benchmark cited).

Deployment consequences

  • BYOC (Bring Your Own Cloud) becomes load-bearing. See systems/redpanda-byoc. The data plane runs inside the customer's VPC; the model runs there too.
  • Data-plane-atomicity matters more. Once the model-inference layer is inside the customer firewall, any control-plane dependency in the data path becomes a compliance/sovereignty hazard. See concepts/data-plane-atomicity.
  • Frontier models remain valuable for orchestration. Gallego doesn't argue frontier models are obsolete — only that they're best used to orchestrate small local-GPU minions, not to process bulk private data. See concepts/frontier-model-minion-delegation.
  • concepts/digital-sovereignty — regulatory / geopolitical framing of the same constraint (residency, operational autonomy, infrastructure control). Model-to-data is a technical realisation of the digital-sovereignty demand.
  • concepts/shift-left-privacy — the orthogonal pattern of moving privacy controls earlier in the pipeline; model-to-data shifts the compute location rather than the policy point.

Seen in

Last updated · 470 distilled / 1,213 read