The crossover year
Inference cost, model quality, and regulatory weather all converged in 2026. The frontier-vs-sovereign decision is now a planning decision, not a research one.
For most of 2023 and 2024, the frontier labs were the only place serious enterprise reasoning could happen. By the second half of 2025, the gap had narrowed to a handful of workload classes. In 2026 it closed for everything except creative ideation and large-context coding.
What changed is not only the open-weight models, it is the supply chain underneath them. Sovereign GPU capacity in Europe is now bookable on the same procurement cycles as ordinary cloud.
How the 84 workloads are scored
Four axes, one published rubric, and a model you are meant to argue with.
Each workload type is scored on sensitivity (regulatory and commercial), reasoning intensity, latency tolerance, and strategic value of the underlying data, then normalised to a 0 to 100 sovereign-suitability index.
The index is modelled rather than measured. Public benchmarks (MMLU, GPQA, SWE-Bench Verified, LegalBench) inform the reasoning-intensity axis and are cited where they do; we do not re-run them and we do not treat vendor-supplied numbers as evidence. The point of publishing the rubric is that you can recalculate any score with your own inputs.
Higher = stronger case for self-hosted, open-weight deployment. n=84 enterprise workloads, weighted by sensitivity, latency, and reasoning intensity.
The placement map
Legal, HR and finance go home. Creative and engineering stay with the frontier. The middle band is where the routing layer lives.
Of the 84 workload types in the map, 51 score above 60, which is where we think the sovereign case is strong enough to act on. Eighteen score above 85, where we struggle to construct an argument for keeping them on a frontier API at all.
At the other end, 19 workloads scored below 35. These are dominated by software engineering and creative work, where the marginal capability of Claude Opus and Codex still pays for itself.
The middle band of 14 workloads is where governance and routing earn their keep. A workload-aware gateway is no longer optional architecture.
“By Q4 2026, the question is no longer whether to self-host, it is which two-thirds of the workload to bring home first.”
What to do in the next 90 days
Six concrete steps, ordered by leverage. The first three cost almost nothing and unlock the rest.
Run a frontier-API audit. Most enterprises cannot list every prompt their employees send to Claude or ChatGPT this week. Until you can, every other decision is guesswork.
Stand up a sovereign endpoint behind your existing gateway, even at zero traffic. The presence of the route changes the conversation with every product owner.
Pick one high-suitability workload (legal drafting is usually the cleanest pilot) and move it. The internal proof point is worth more than any external benchmark.
- 1. Headline figures are Rindogatan models, directional benchmarks to be calibrated to a specific institution, not survey statistics.
- 2. Partner data points are drawn from publicly published research (e.g. Snowflake's Modern Marketing Data Stack, Databricks' State of Data + AI) and cited for direction only.
- 3. Regulatory references: EU AI Act, Reg. (EU) 2024/1689; GDPR, Reg. (EU) 2016/679; DORA, Reg. (EU) 2022/2554; NIS2, Dir. (EU) 2022/2555.
- 4. Sovereign deployment modelled on European sovereign infrastructure.