Where the verdict flips

Three conditions. All three must hold.

The recommendation against on-prem is conditional. Three precise conditions reverse it. The default reading of each is FAILS for an SME; the flip reading is the narrow case under which an 8× H200 stack earns its capital. If any one of the three fails, do not build.

Three gates · default state, flip state, prevalence in an SME estate

Default · fails

Flip · holds

Typical SME

Workload · default

Agentic multi-step, novel reasoning at the edge.

Five-step chains compound error: 80% per-step → 33% end-to-end. Terminal-Bench gap of 14.8 pt is felt.

Open-weight trails 5–15 pt

Workload · flip

Bounded retrieval, summarisation, classification, modest code assist.

No high-fidelity agentic tool use, no edge reasoning. The open-weight gap to frontier collapses.

< 5 pt on relevant axes

Holds for ~ 1 in 4 most SMEs want agency

Compliance · default

Sovereignty is the operative need.

AU-resident weights and inference. Threat model does not demand physical air-gap.

Bedrock-Sydney suffices

Compliance · flip

Air-gap is the operative need.

Classified or regulated workload that mandates a physical break from public cloud — not a contractual one.

IRAP-PROTECTED+ binding

Holds for rare classified workloads only

Connectors · default

Notion, Gmail, Calendar, Word, Excel are core.

OAuth-gated SaaS routes data through US-domiciled cloud regardless of model residency.

Punctured · §V

Connectors · flip

Cloud SaaS is out of scope.

LLM restricted to on-prem data — file shares, local DBs, internal wikis, code repositories. Or full self-hosted stack.

Architecturally clean

Holds for ~ 1 in 5 most SMEs depend on these

∧

Build only if 01 and 02 and 03. Otherwise cloud-sovereign dominates on capability, cost, and coherence.

All three≈ 1 SME in 50

·Condition 01 evidence: the capability gap as a ladder of five benchmarks; three material, two immaterial; agentic compounding makes per-step matter
·Condition 02 evidence: cloud-sovereign at 30–50% of on-prem TCO closes the capability gap on workloads that do not require air-gap
·Condition 03 evidence: every named connector is OAuth-gated SaaS; the model's residency does not affect where the data sits
·The bandwidth floor still applies: even if all three flip, only the 8× H200 SXM HGX node meets the throughput bar at 500k context × 3 streams
·The verdict, restated: on-prem buys data locality and weight sovereignty — neither is the same as frontier capability

SourceOn-prem frontier LLM briefing, §Verdict and §Conditions under which the recommendation flips. Prevalence ratios are first-order estimates against the brief's framing — "most SMEs whose threat model does not demand air-gap" (compliance), "most SMEs want high-fidelity agentic work" (workload), and the named-connector default (estate). The flip is rare by design: the brief's centre of gravity is that the default verdict against on-prem is robust to all but a narrow case.