Module evidence surface
NVIDIA Stack
GPU scale for deterministic risk infrastructure.
SAA runs ARIN22 evidence lanes on NVIDIA-class infrastructure where batch scale matters. The deterministic answer remains CPU-bound and hardware-agnostic; NVIDIA provides throughput, orchestration and physical-simulation surfaces around it. SAA is an NVIDIA Inception member and Innovation Lab member; this page records internal deployment status, not an external benchmark or endorsement.
Integration discipline
NVIDIA where scale matters. ARIN22 where correctness is proven.
The stack is deliberately split: GPUs carry batch throughput, simulators and agent infrastructure. The risk number remains deterministic, replayable and bounded by the published ARIN22 evidence envelope. The August 2026 acceptance battery is the sharpest expression of this split: GPU compute produced the brute-force reference; the CPU-bound deterministic kernel reproduced it to a 0.0156% median error under review-only discipline.
What it enables
A scale layer around the deterministic kernel.
This page does not claim a deeper NVIDIA partnership than exists. It shows what is internally live, what is roadmap, and which numbers are self-run evidence rather than external certification.
Scale the kernel.
H100-class batch lanes carry Enterprise-Wave throughput while the deterministic risk-core keeps correctness portable across CPU and GPU execution contexts.
Govern the agents.
NIM, NeMo Guardrails, Retriever and Evaluator support narrative synthesis, refusal checks, citations and release-managed reliability signals.
Simulate physical risk.
Earth-2 and PhysicsNeMo support climate, weather, infrastructure and cascade surfaces for physical-to-financial risk workflows.
Keep the boundary.
FLUX is limited to illustrative non-data-bearing imagery. Data charts and dependency graphs render from computed series, never image models.
Service mesh
16 services across four stack categories.
Internal live means self-hosted via NIM containers or active NVIDIA-hosted cloud API endpoint inside the SAA stack. Roadmap means targeted integration, not a current production claim.
Core AI
5 / 5 liveExecutive-summary generation, structured report synthesis and multi-model consensus. Specific model selections are internal.
Stress-test council workflow: entity classification, fast / deep scenario analysis, consistency checks and summary fan-out.
Filters agent output for safety, factuality and regulated-language boundaries before Governor release.
Enterprise RAG and citation layer so narrative claims link to underlying source material.
Agent observability, trace collection and replay support for council audit logs.
Physical simulators
2 / 4 liveClimate and weather feeds for catastrophe, infrastructure and physical-risk stress workflows.
Physics-informed simulation layer for critical-infrastructure cascades and physical-financial coupling.
Roadmap target for self-hosted high-throughput weather forecasting.
Roadmap target for km-scale climate downscaling and asset-level physical risk.
Infrastructure
1 / 4 liveIllustrative, non-data-bearing report imagery only. No data chart is generated by an image model.
Roadmap target for self-hosted LLM / embedding serving when council load justifies a dedicated GPU pool.
Roadmap target for disaggregated low-latency inference scheduling beyond 100 concurrent council instances, marked as an engineering estimate.
Roadmap target for SENTINEL voice alerts and optional control-room interfaces.
Data and evaluation
3 / 3 liveData-curation pipeline for regulator filings, news streams and risk-domain corpora.
Synthetic and adversarial fixtures for agent red-team testing. Not represented as realised client outcomes.
Continuous agent evaluation signals consumed by the Governor. Recalibration enters through versioned releases, not silent runtime mutation.
Hardware posture
H100 measured today. B200, GB200 and GH200 remain target platforms.
The clean line matters: H100 carries the measured Innovation Lab validation lane. Next-generation platforms are collaboration targets until a measured SAA evidence lane exists.
H100 / H200 SXM
H100 carries the measured validation lane. H200 is a same-class serving target for Triton and TensorRT-LLM once council load justifies a dedicated GPU pool. An H100 lane additionally generated the 100M-path reference corpus behind the August 2026 acceptance battery; the deterministic kernel then matched that reference within the published envelope running on commodity CPU.
B200 / GB200 NVL72
Target platform for Earth-2 downscaling and multi-agent fan-out. Collaboration-ready, but not deployed as a measured public evidence lane today.
GH200 Grace Hopper
Evaluation target for memory-bound Entity Memory, graph workloads and batched deterministic recompute. Not a current public evidence claim.
Evidence boundary
STAC-inspired workload. Self-run evidence. No certification claim.
The NVIDIA stack supports a STAC-archetype engineering workload, kept separate from canonical ARIN22 latency lanes and from any external STAC certification.
| Lane | Measured surface | Status | Boundary |
|---|---|---|---|
| STAC-A2 inspired Greeks | 310 M paths, 2.48 B valuation paths, 595.2 B path-asset-step ops in 14.875 s on 8xH100. Max lane: 166.72 M valuation paths/s. | REVIEW_ONLY | Self-run. Not STAC-audited. Production listed-option Greek parity pending. |
| STAC-M3 inspired tick analytics | 60 M tick updates, p99 0.501 ms / p999 0.626 ms on Evidence Wave 2026-05-20. | MEASURED | Separate from canon tick lane 2026-05-30: p99 0.44-0.46 ms · tail observation 0.653 ms (not a p999 claim). |
| STAC-T1 adjacent pre-trade | 60 M evaluations, p99 1.110 ms on Evidence Wave 2026-05-20. | REVIEW_ONLY | Separate from canon pre-trade lane 2026-05-30: p99 0.85-0.93 ms / p999 1.8-2.5 ms. |
| Enterprise Wave validation | 8,800 cases per backend, 8.8 B paths per backend, 0 execution failures. | MEASURED | Internal evidence. Full SHA, hardware fingerprint and replay bundle released under NDA. |
| H100 oracle reference corpus (Aug 2026) | 40 cases × 100M Monte Carlo paths per case · per-case evidence SHA-256 · seeds recorded — reproducible by seed | MEASURED (reference lane) | Self-provisioned H100 · served as brute-force reference truth for the estimated acceptance battery: deterministic kernel matched the reference 160/160 within the 5% envelope on CPU (REVIEW_ONLY, pending production re-run — see Data Room I-19) · not an NVIDIA benchmark or endorsement |
Honest framing: workloads are built to STAC archetypes and self-run on 8xH100. Independent STAC benchmarking is explicitly pending. Captured reproducibility material lives in the read-only evidence surface at /arin22-demo.
Design partner entry
Bring your NVIDIA stack into the evidence lane.
SAA is pre-revenue and accepting a limited design-partner cohort across insurance, banking, sovereign risk, critical infrastructure and asset management. Engagements are bounded, evidence-led and claim-gated.
NVIDIA, NIM, NeMo, Earth-2, PhysicsNeMo, Triton, Dynamo, Riva and FLUX are trademarks of NVIDIA Corporation. Status reflects SAA internal deployment or target-platform status as of publication.