Own the answer: A decision framework for closed, open, cloud, and on-prem AI.

The enterprise AI conversation has moved. For two years the question was which model is smartest. In 2026 it has become a harder, more valuable one: who owns the weights, where do they run, and who controls the data that shapes them? That shift turns model selection from a technical preference into a board-level decision about cost, control, and risk.

Most AI firms answer it with a religion: open is better, or closed is safer. That is the wrong instinct. The honest answer for most organizations is a blend: closed frontier models for the highest-complexity, lowest-volume reasoning tasks, and fine-tuned open-weight models running on infrastructure the client controls for the high-volume, regulated, or IP-sensitive production workloads (order-to-cash, procure-to-pay, claims, revenue cycle, and the like).

This paper sets out a workload-by-workload framework for making that call with evidence rather than ideology. It covers the four deployment choices (closed API, open-weight, cloud, and on-premises), the two questions that actually decide between them, the sector triggers that make on-prem the default rather than a preference, and the cost logic that turns a metered subscription into an owned asset. It closes with how Weaver sizes the right mix in Measure, a fixed-fee, two-to-three-week engagement, before a client commits capital to any path.

On 29 June 2026, Palantir and NVIDIA announced a joint engine for running NVIDIA's open Nemotron models inside sovereign, classified, and air-gapped environments. Government agencies can train the models on their own data, run them on infrastructure they control, and retain ownership of the resulting model weights. The framing from Palantir's leadership was blunt: the real risk is proprietary insight migrating into the weights of closed models, enterprise alpha absorbed into someone else's next release and effectively sold back to competitors.

That is not a government-only pattern. It is the same conversation Upper MidMarket and Enterprise buyers are now having internally: Healthcare and Life Sciences on documentation and governance, Insurance on claims and underwriting, Manufacturing on procurement and supply-chain data. Sovereignty, cost, and control have become board-level questions, and the organizations that answer them with evidence will pull ahead of those still asking which model tops a public leaderboard.

The urgency is grounded in two numbers. First, 95% of enterprise generative-AI programs fail to deliver measurable business value. MIT's Project NANDA found the divide is driven not by model quality but by approach, and that organizations buying from specialist partners succeed roughly twice as often as those building alone. Second, the market is compounding fast: the enterprise AI market is approaching $65 billion in 2026, and the AI-agents segment is forecast to grow from roughly $12 billion in 2026 to $251 billion by 2034. The cost of getting the architecture wrong compounds just as quickly.

The deployment decision is usually described as four options. They are better understood as two independent questions: what kind of model (closed or open weights) and where it runs (public cloud or on-premises / private). Any workload sits somewhere on both axes.

Closed API: frontier models accessed as a metered service; the provider owns the weights (Claude, GPT, Gemini). Best fit: highest-complexity, lowest-volume reasoning where frontier capability matters more than unit cost or data residency.

Open-weight: downloadable model families you can fine-tune and own (Llama, Nemotron, Mistral, Qwen, DeepSeek); permissive licences are common. Best fit: high-volume, well-defined production workflows where cost, latency, control, and IP ownership dominate.

Cloud: model runs in a managed public-cloud environment: fastest to stand up, elastic capacity, usage-based cost. Best fit: variable or spiky demand, rapid experimentation, workloads without hard data-residency constraints.

On-prem / air-gapped: model runs on infrastructure the client controls, up to a fully isolated network; the client owns the weights and the data path. Best fit: regulated, IP-sensitive, or high-volume steady-state workloads where control and predictable cost outweigh elasticity.

Model families and vendors are named as market context. Weaver is platform-agnostic and endorses no hardware, infrastructure, or model vendor.

Every vendor will push you to pick from this grid. That is the wrong first move. The choice becomes almost mechanical once you answer two questions about the workload, not the model.

This is a question of trust. A workflow that triages a claim, screens a contract, or approves a payment cannot run on a probabilistic guess. The higher the cost of a wrong answer, the more the architecture must favor deterministic, governed, auditable outputs, regardless of which model sits underneath.

This is the question of cost and control, and it has two edges. There is the metered cost of every API call, and there is the strategic cost of your proprietary process knowledge (how you underwrite, how you triage, how you run procurement) accumulating inside a model you do not own. Once both questions are answered for a given workflow, the closed-versus-open and cloud-versus-on-prem calls follow naturally.

For most organizations, the right architecture is not one model but a portfolio matched to workloads. The reasoning that runs a business day to day is mostly narrow and well-defined, and for those workflows, a fine-tuned open model matched to the task can rival a general-purpose closed model on accuracy once tuned on the client's own data, while beating it on cost, latency, and control. The frontier closed models earn their place on the small set of high-complexity, low-volume reasoning tasks where raw capability is worth the premium and the meter.

High-volume, regulated, or IP-sensitive workflows → fine-tuned open-weight model on on-prem or private infrastructure. Control, unit cost, and IP ownership are the drivers.

High-complexity, low-volume, novel reasoning → closed frontier API in the cloud. Frontier capability justifies the premium; low volume keeps token spend contained.

Spiky or experimental demand with no hard residency constraint → cloud (open or closed), elastic. Elasticity and speed-to-stand-up matter more than steady-state unit cost.

Any workflow where a wrong answer carries material cost → deterministic ML plus GenAI, governed and audited. Outputs must be governed and auditable regardless of the model underneath.

In several regulated sectors, the deployment choice is partly made for you. These are not preferences, they are compliance triggers that move on-prem or private deployment from nice to have to baseline.

Healthcare & Life Sciences: HIPAA. Prompts, logs, and embeddings can all capture protected health information, and any vendor touching PHI needs a Business Associate Agreement. Self-hosted and air-gapped deployment keeps PHI on-premises and materially simplifies the compliance surface across compute, storage, and network.

Insurance: NAIC Model Law. The NAIC Insurance Data Security Model Law, adopted in 28-plus jurisdictions as of early 2026, requires demonstrable control over how customer data is processed and where it lives. On-prem or private deployment makes that control auditable rather than assumed.

Manufacturing & Industrials: IP sensitivity. Procurement terms, supplier data, and process know-how are competitive assets. The exposure is less regulatory than strategic: this is exactly the proprietary knowledge that should not accumulate inside a model you do not own.

Every call to a closed model is metered. At production volume (claims, transactions, or documents per month) that token spend compounds, and at the end of it the organization owns nothing. Enterprise leaders have begun to say so publicly: in July 2026 Palantir's CEO criticized token-based pricing as fundamentally misaligned with the value enterprises actually receive, as rising costs push buyers toward lower-cost open-weight alternatives.

An open-weight model, fine-tuned on the client's own data and running on infrastructure the client controls, inverts the equation. The upfront cost is capex (hardware, tuning, deployment), but the marginal cost of each subsequent inference approaches zero, and every hour of tuning is accumulating inside an asset the client owns. At production volume, an owned asset beats a metered subscription on total cost of ownership within quarters, not years. The point at which it does depends on volume, workload complexity, and infrastructure choice, and that crossover is precisely what our Measure engagement is designed to calculate.

Don't guess, and don't pick a side on principle. The right architecture is a blend, and the mix is specific to your workloads, your regulatory triggers, and your volumes. Weaver will tell you, workload by workload, where an open model on your own infrastructure pays for itself, and where a closed model is still the right call. Either way, you own the answer.