[miktam — preface]

Companion to Every Company Can Be a Palantir Now. That essay made the institution-level case; this one extends it to the individual. Same physics, smaller perimeter. The argument is strategic; the hardware that tests it sits on my desk.

— miktam

This essay argues that local, sovereign intelligence can now outperform centralised, cloud-based models on specific tasks — not by being more powerful, but by asking better questions. The shift is architectural, and it’s measurable.

The Physics of the Local Engine

Let’s start with hardware. NVIDIA spent decades solving a hard physics problem: maximising compute and memory bandwidth between CPU, GPU, and system memory. A single Nvidia B200 GPU delivers around 8 terabytes per second of memory bandwidth — an order of magnitude beyond typical consumer hardware.

Apple’s approach is different. By unifying CPU, GPU, and Neural Engine into a single system-on-chip with shared memory, Apple Silicon enables highly efficient local inference. While it does not match datacenter GPUs in raw performance, it makes powerful, zero-cloud compute accessible on consumer hardware.

We Didn’t Upgrade the Model. We Upgraded the Question.

Once hardware is sufficient, the challenge shifts to orchestration — and this is where the counterintuitive result lives.

A flat query (“which of these seven listings has a sea view?”) handed to a capable model produced five false positives out of seven candidates. The same task, restructured as a pipeline — one listing, one bounded question, results joined in deterministic code — produced zero. Same hardware. Same model family. Different architecture of the question.

This is the spot→verify pattern: filter candidates with a SQL query, adjudicate each one with a single bounded model call, join the results in code. No cross-document reasoning in latent space. No hallucinated comparisons. The model sees one document and answers one question. Everything else is code.

The implication is uncomfortable for the “bigger model wins” assumption: on this task, orchestration outperformed a more powerful model asking an unstructured question. We didn’t need a better model. We needed a better question.

The Leviathan vs Data Sovereignty

The real shift is not hardware — it is data.

Cloud-based models require sending data to centralised providers. Local models invert this: all data remains under your control.

The key question becomes: can a model with full access to your private context outperform a more powerful model that lacks it?

In many cases, it already does.

The Real Constraint

The constraint is not the model. It is the question.

A frontier model without access to your private data operates with partial visibility — that part is well understood. But a frontier model with your private data and an unstructured question still fails. The false positives don’t come from ignorance; they come from being asked to do too many things at once. Cross-document reasoning, comparison, and synthesis in a single prompt is where both cloud and local models degrade.

The structural advantage of local isn’t just privacy. It’s that you control the architecture. You decide what each model call sees. You decide what gets joined in code. A sovereign individual with well-designed question architecture and private data beats a cloud model with the same data and an unstructured prompt. That’s the measurable claim.

The Shift

General intelligence is becoming commoditised.

As it does, value shifts toward:

  • proprietary data
  • high-quality context
  • control over both
  • and the architectural discipline to ask the right question

Without the fourth, the first three don’t compound. Private data fed into a flat, unstructured prompt is still a flat, unstructured prompt. The sovereign individual isn’t just someone who keeps their data local — it’s someone who knows how to structure a question so a smaller model answers it correctly.

This is not entirely new. But it is becoming operationally real, and the evidence is now measurable rather than theoretical.

The trade-off between privacy and capability is no longer absolute.

Every company can be a Palantir.

And increasingly, every individual can build one — if they learn to ask better questions.