Executive Reference Architecture

Sovereign Hybrid AI: Strategy Blueprint

Prepared for: Chief Technology Officer Date: September 3, 2026 Version 3.0 Supersedes: v2.0

Executive Summary

On the morning of September 3, 2026, ChatGPT, Claude, and Grok failed simultaneously, with confirmed status-page outages at all three providers and downstream casualties including the Cursor coding agent. Early reporting points to a shared dependency on Microsoft Azure as a contributing factor. Concentration risk in the AI stack is no longer theoretical; it is operational, dated, and board-relevant.

This document presents a strategic blueprint for a sovereign hybrid AI model across three layers:

  1. Control plane: an AI Gateway that routes all AI traffic, including agentic traffic, to internal models or external providers under policy, failover, and governance.
  2. Execution plane: autonomous AI Workers (Mariete) that turn model capability into finished business outcomes, deployed inside the enterprise's own boundary.
  3. Infrastructure plane: Tier 1 model serving on infrastructure you control, with three procurement paths compared in this version: self-built cloud VPC, capex on-prem, and HPE GreenLake / HPE Private Cloud AI consumed as a service.

The goal remains substitutability over independence: any model, provider, agent, or infrastructure substrate swappable via configuration, not rewrites. Two honest notes frame everything: sovereignty transfers accountability for availability rather than conferring it, and below roughly 2 to 5 million tokens per day of sustained volume, self-hosting is an insurance premium rather than a savings.

1Strategic Rationale

The September 3 outage matters less for its duration (hours) than for its structure: three competitors, running on substantially shared cloud infrastructure, failed together. A resilience strategy that ignores correlation domains is cosmetic.

Sovereign AI does not mean building everything from scratch. It means:

  • Running critical models in environments you control.
  • Owning the data plane and model customization.
  • Failing over, or degrading gracefully on your own terms, during provider outages.
  • Avoiding lock-in at the application, agent, and infrastructure layers.
  • Knowing, and limiting, your blast radius when any single provider, cloud, or model family fails.
TierWorkload TypeModel SourceAvailability Posture
Tier 1, SovereignCustomer-facing, revenue-critical, data-sensitive, regulatedSelf-hosted open-weight models, in a different failure domain from your primary API dependencyOwned SLO (target 99.9%+), funded and drilled
Tier 2, CommodityInternal productivity, non-real-time, exploratoryExternal commercial APIs, at least two providers across at least two cloudsBest-effort, with tested fallback to Tier 1
Tier 3, ProhibitedRegulated data classes that must never touch external APIs; unapproved shadow AITier 1 only, or blocked at the gatewayEnforced, not aspirational
Degraded-mode contract

Every Tier 1 workload pre-registers: (a) its fallback model, (b) the minimum quality score that model must pass on the golden set, (c) the user-facing disclosure shown while degraded, and (d) whether the workload fails open (weaker model serves) or closed (graceful unavailability). Failover is a governed decision, not a routing reflex.

2High-Level Reference Architecture

Four planes, each independently substitutable:

Business outcomes / applications
        |
 [ Execution Plane: Mariete AI Workers ]  <-- recipes, MCP connectors, outcome reporting
        |  (provider router enforces approved endpoints)
        |
   [ Control Plane: AI Gateway ]  <-- policy, failover, guardrails, budgets, agent governance
      /        \
 Tier 1        Tier 2 (multiple providers, multiple clouds)
 (self-host,     |
  separate       +-- Provider A (cloud X)
  domain)        +-- Provider B (cloud Y)
      |
 [ Infrastructure Plane: Cloud VPC / HPE GreenLake PCAI / on-prem capex ]
      |
 [ Data & Knowledge Layer: vector DB, pipelines, registry, eval sets ]

 Cross-cutting: Evaluation & Quality Plane | Governance Plane
  • Control plane (AI Gateway): unified OpenAI-compatible API; routing, failover, guardrails, budgets, observability.
  • Execution plane (new in v3.0): AI Workers that plan, use tools, verify, and deliver finished work. Mariete's provider router is itself a policy-enforced gateway layer; in this architecture it sits behind (or alongside) the enterprise gateway, so model endpoints remain yours.
  • Infrastructure plane: where Tier 1 models physically run; three options in section 4.
  • Evaluation & Quality Plane: golden datasets, benchmark harnesses, regression gates. Substitutability is only real if a replacement model is proven before it serves.
  • Governance Plane: policy, license/provenance records, audit trails, budgets, all as versioned code.

3Core Components

3.1 AI Gateway (Control Plane)

Recommended tools: LiteLLM (open-source, self-hosted, virtual-key budgeting), Portkey (governance depth), or Kong AI Gateway (if Kong already runs your API estate). Also evaluate cloud-native control planes (AWS Bedrock, Azure AI Foundry, Google Vertex AI) and enterprise platforms such as TrueFoundry. Note that HPE Private Cloud AI now ships a unified model gateway as part of its stack (section 4.3); if you adopt it, decide explicitly whether it serves as the enterprise gateway or is federated under your existing one. Avoid two uncontrolled gateways.

Capabilities:

  • Dynamic routing by task, user, tenant, cost, latency, data sensitivity, and provider health.
  • Automatic failover across providers and failure domains, under degraded-mode contracts.
  • Guardrails: prompt-injection detection, PII redaction, content filtering before data leaves the enterprise.
  • Agentic-traffic governance: session and tool-call mediation, step-level budgets, loop/runaway detection, kill switches. (On HPE Private Cloud AI, NVIDIA Agent Toolkit and OpenShell provide analogous runtime controls natively.)
  • Full observability: tokens, latency, cost, quality signals, audit logs per request and per agent run.
  • Policy enforcement as versioned code.

3.2 Execution Plane: Mariete AI Workers New in v3.0

Gateways and models do not produce business value by themselves; finished work does. The execution plane is where autonomous workers turn capability into outcomes (research, analysis, documents, follow-ups, workflows). v3.0 integrates Mariete as the reference execution layer, based on its published enterprise offering ("Mariete in the Box"):

  • The Worker: each task runs behind a real cloud computer (file system, shell, browser, tool suite), so output is usable deliverables, not chat answers.
  • Provider router: a per-customer routing layer points the Worker at the model endpoint your policy approves and enforces the boundary. This maps directly onto the AI Gateway: Mariete workers consume Tier 1 and Tier 2 endpoints through the same policy plane as applications.
  • Recipes + MCP connectors: reusable named workflows connected to your actual tools and data, matching the blueprint's agentic-governance requirements.
  • Deployment modes (aligned to this blueprint's tiers):
    1. Your approved cloud + your LLM provider (Azure OpenAI, AWS Bedrock, Google Vertex, or a self-hosted open-weight endpoint), in your account and region, no new vendor in the data path. Recommended mode for Tier 1/Tier 2 coexistence.
    2. Mariete managed cloud, fastest activation, suitable for Tier 2-style workloads while Tier 1 is built.
    3. Fully local / air-gapped, open-weight models on customer hardware, mapping to Tier 3 and pairing naturally with the HPE GreenLake option in section 4.3.
  • Commercial model (verified from mariete.ai/enterprise, Sept 2026): outcome-based pricing rather than seats or per-token: Blueprint (free readiness assessment), Pilot from $1,200/mo (single team, 2-3 outcomes, six-month minimum), Outcomes from $3,500/mo (company-wide, monthly outcome reporting), Outcomes: Regulated from $7,500/mo (fully local/air-gapped, confidential-computing configurations, compliance artifacts, audit cooperation, priority SLAs). Operating motion: Plan (week 0), Build (weeks 1-2), Deploy (activation day plus continuous stewardship).
  • Positioning note for this audience: Mariete's published target is SMB and mid-market; for large-enterprise adoption, treat it as the execution-layer pattern (and vendor, where scoped) rather than the whole strategy. The pattern, workers behind a policy-enforced router delivering measured outcomes, is what matters architecturally.

3.3 Internal Model Serving (Tier 1, Sovereign)

Serving stack: vLLM as default; evaluate SGLang (12 to 33 percent higher throughput on shared-context workloads such as RAG and agents) and TensorRT-LLM for maximum NVIDIA performance. TGI is legacy; do not start new deployments on it. On HPE Private Cloud AI, NVIDIA AI Enterprise, NIM, and NeMo are pre-integrated as the serving and fine-tuning stack.

Model policy: maintain a small curated portfolio by capability class, reviewed quarterly:

ClassRoleCurrent examples (Sept 2026; re-verify quarterly)
Frontier open-weightTier 1 primary for complex workloadsDeepSeek-V4 family, Qwen 3.5 397B, GLM-5.2, Kimi K2.6
Mid-tier workhorseDefault for most Tier 1 trafficMistral Medium 3.5, Gemma 4 31B, Qwen3.6 27B, GPT-oss 120B, NVIDIA Nemotron family
Small/efficientHigh-volume simple tasksQwen3.5 4B/9B, Gemma 4 12B, GPT-oss 20B
Fine-tuned variantsDomain adaptation via LoRA/QLoRA (or NVIDIA NeMo on PCAI)Per business unit

Hardware: H200 (141GB) is the 2026 value default; B200/GB200-class for larger models and better cost per token at scale; A100 is no longer a sound default. Reserved capacity for the Tier 1 floor; spot only for dev/batch.

3.4 External Commercial APIs (Tier 2, Commodity)

Minimum two providers on two different clouds, all behind the gateway. Weight enterprise governance posture, status-page transparency, and contractual availability, not just benchmark rank.

3.5 Data & Knowledge Layer

  • Vector DB: pgvector; Qdrant, Milvus, or Weaviate at scale. On HPE PCAI, the Alletra MP X10000 data layer with metadata/governance automation is the integrated equivalent (HPE claims up to 20x time-to-first-token improvement via KV-cache-aware inference on Nemotron 70B in its benchmark testing).
  • Pipelines: Airflow, Dagster, or Prefect; HPE Data Fabric adds MCP support for Airflow and an enterprise AI inventory (available October 2026).
  • Model registry: MLflow, W&B, or self-hosted hub, with provenance and license metadata mandatory.
  • Evaluation datasets: versioned golden sets per workload, owned by the Evaluation & Quality Plane.

3.6 Security, Compliance & Legal

  • Identity via OIDC/SSO; workload identities for agents. Encryption everywhere; keys in KMS/HSM.
  • Internal models in private subnets; external egress via controlled proxy with DLP.
  • Immutable audit records of prompts, completions, and agent tool calls.
  • Confidential computing (new in v3.0): for the strictest Tier 1/Tier 3 workloads, NVIDIA Confidential Computing (integrated into the HPE AI Factory through HPE Services) protects models and data during execution with cryptographic attestation. Mariete's Regulated tier likewise lists confidential-computing configurations. This closes the last "data in use" gap in the sovereignty story.
  • Licensing & provenance: legal pre-approval per model family; board-level provenance policy decided in advance.
  • Regulatory: EU AI Act deployer obligations apply now (August 2026 high-risk milestone in force); map each Tier 1 workload. HPE positions its Sovereign AI Factory against government-grade baselines (NIST 800, STIG, FIPS, CMMC), relevant for regulated and public-sector buyers.

4Deployment Topologies & Infrastructure Options

4.1 Topology patterns (correlation-aware)

TopologyDescriptionUse CaseCorrelation Check
Cloud-onlyAll components in your cloud VPC with dedicated GPU instancesFastest deploymentWeak if that cloud also hosts your Tier 2 providers' substrate; mitigate with multi-region and a second provider elsewhere
Hybrid cloud + on-prem/GreenLakeGateway and data in cloud; model serving on owned or GreenLake-consumed infrastructure in your facilityStrict residency, existing data centersStrong: Tier 1 sits outside the cloud failure domain entirely
Multi-cloudGateway cloud-agnostic; internal models on any cloud with GPU capacityAvoid infrastructure lock-inStrong if Tier 1 and Tier 2 are deliberately placed on different clouds
EdgeLightweight serving on edge devices with central managementLow latency, disconnected sitesStrong for site-local continuity; manage fleet drift

4.2 The three procurement paths for Tier 1 infrastructure

Option A

Self-built cloud VPC

GPU instances in your cloud account, self-assembled stack. Fastest (weeks, if quota available). Opex at cloud GPU rates. Weakest correlation posture.

Option B

On-prem capex (self-built)

Purchased GPU servers in your data center. Maximum sovereignty; 4-9 months to production; full ops burden including hardware lifecycle.

Option A: Cloud VPCOption B: On-prem capexOption C: HPE GreenLake / PCAI
What it isGPU instances in your cloud account, self-assembled stack (K8s, vLLM, gateway)Purchased GPU servers in your data center, self-assembled stackTurnkey AI factory co-engineered with NVIDIA (compute, networking, storage, AI software), in your data center or colo, consumed via GreenLake, or purchased outright
Time to prodWeeks (if GPU quota available)4-9 months (procurement, facilities, build)Weeks to ~2 months; pre-configured, right-sized configurations (from 8x H200 NVL DL380a class up to GB200 NVL4 and RTX Pro 6000 Blackwell worker nodes; multi-node inference to 256 GPUs)
Cost shapeOpex, cloud-rate GPUs; reserved for the floorLarge capex, then low opex; utilization risk is yoursSubscription opex without full capex; capacity scales under the GreenLake contract; HPE markets token-cost control and GPU-utilization optimization
SovereigntyGood (your account/region), but shares the cloud failure domainMaximum; fully inside your perimeterMaximum physical sovereignty plus managed operations; Sovereign AI Factory variant targets government-grade compliance
Ops burdenYours (the 4-8 FTE problem in full)Yours, plus hardware lifecycleShared: HPE operates the stack to a service level; your team focuses on models, evals, and policy
Agentic readinessBuild it yourselfBuild it yourselfNative: NVIDIA Agent Toolkit, OpenShell secure runtime, local agent registration, Zerto rogue-agent rewind
Correlation postureWeakest (same substrate risk)StrongestStrongest (same physical independence as B, without the build time)
Best forSpeed, uncertain demand, Phase 1Steady high-volume demand, existing facilities/staffEnterprises wanting on-prem sovereignty with cloud-like consumption and without a 6-month build
Recommendation

Start Phase 1 on Option A for speed, with Tier 2 providers on a second cloud. In parallel, evaluate Option C (HPE GreenLake Private Cloud AI) as the Phase 2/3 Tier 1 home: it collapses the on-prem build from quarters to weeks, transfers part of the ops burden that gates this program, and its native agentic-governance stack aligns with section 3.1's requirements. Option B remains right for organizations with existing data centers, staff, and steady multi-year volume. In all cases the AI Gateway and the Evaluation & Quality Plane stay vendor-neutral so the substrate beneath them remains substitutable, including away from HPE later.

Mariete + GreenLake pairing

Mariete's fully local/air-gapped mode (open-weight models on customer hardware, Regulated tier with confidential-computing configurations and compliance artifacts) maps directly onto GreenLake-resident infrastructure: HPE supplies the sovereign substrate and model serving; Mariete supplies the execution layer; the gateway enforces policy across both. For Tier 3 workloads this is the cleanest end-to-end sovereign stack in this document.

5Strategy Blueprint: Phased Implementation

Phase 0, Assessment & GovernanceWeeks 1-4
  • Inventory AI use cases, including agentic workflows; classify into Tiers 1, 2, 3.
  • Correlation-domain analysis: map every AI dependency (models, providers, clouds, agents, auth, data) and identify shared failure domains.
  • Define data sensitivity levels and external-sharing rules.
  • Form the AI governance board (CTO, CISO, DPO, legal, business leads); decide the model-provenance position.
  • Select initial open-weight models; evaluate against the golden dataset.
  • New in v3.0: run the Mariete Blueprint (free readiness assessment) or equivalent to baseline outcome candidates; issue an RFI to HPE on GreenLake Private Cloud AI sizing and commercial terms alongside cloud-GPU quotes, so the Phase 1 infrastructure decision compares all three options on real numbers.
Exit criteria: tier classification approved; correlation map signed off; golden dataset v1; three-option infrastructure comparison on the table.
Phase 1, Foundation & GatewayMonths 1-3

Entry: Phase 0 exit criteria; at least one senior MLOps/inference engineer hired or contracted.

  • Deploy the AI Gateway in staging with observability and budget policies.
  • Stand up one mid-tier internal model (Option A cloud GPUs for speed).
  • Connect two external providers on two clouds, with failover configured.
  • New in v3.0: launch a Mariete Pilot (from $1,200/mo, six-month minimum) scoped to one team and 2-3 measurable outcomes, deployed in "your approved cloud + your LLM provider" mode so its provider router points at your gateway endpoints from day one.
  • Build an internal demo exercising failover and the degraded-mode contract.
Exit criteria: provider-to-internal failover by config change in under 15 minutes, observed in a drill; first Mariete outcomes reporting against baseline.
Phase 2, Production PilotMonths 3-5
  • Move one Tier 1 workload to the internal model as primary under its degraded-mode contract.
  • New in v3.0: select the Tier 1 infrastructure path. If Option C: stand up HPE Private Cloud AI, migrate the internal model to it, adopt its unified model gateway (federated under the enterprise gateway), and enable agentic controls (agent registration, OpenShell runtime, Zerto rewind).
  • Implement PII redaction and prompt-injection filters in path.
  • Chaos testing: external outage, internal overload, GPU node failure, and a correlated drill (primary cloud region plus one provider down together).
  • Fine-tune on domain data (LoRA, or NeMo on PCAI); measure against the quality floor.
Exit criteria: pilot workload meets SLO and quality floor for 30 consecutive days on the chosen infrastructure; correlated-failure drill passed; Mariete pilot outcomes reviewed and a scale/no-scale decision made.
Phase 3, Scale & ExpandMonths 5-9
  • Add internal models per the capability-class table; multi-AZ/region or multi-node (PCAI scales inference to 256 GPUs).
  • Cost-based routing: small internal models for simple traffic, frontier APIs for complex non-sensitive traffic.
  • RAG at scale (Alletra X10000 data layer on PCAI, or equivalent in cloud).
  • New in v3.0: expand Mariete from Pilot to Outcomes tier for company-wide workflows; move Tier 3 workflows to Mariete Regulated on GreenLake infrastructure (fully local/air-gapped, confidential computing).
  • Self-service onboarding for business units via the gateway, with inherited budgets and guardrails.
Exit criteria: majority of AI spend through the gateway; two or more business units self-serving; measured outcome reporting for at least five workflows.
Phase 4, Continuous ImprovementOngoing
  • Quarterly model review (mandatory): re-run golden sets against new releases; swap where quality/cost improves.
  • Quarterly infrastructure review (new in v3.0): re-compare cloud GPU rates, GreenLake contract terms, and utilization; the consumption model only wins if it stays economical.
  • Refine failover logic, load shedding, and degraded-mode contracts.
  • Semiannual disaster drills, including simultaneous loss of all external providers and loss of the primary cloud region.
  • Annual substitutability audit (new in v3.0): prove, by drill, that the gateway, execution layer, and infrastructure substrate can each be swapped without application rewrites. Vendor-neutrality that is never tested is a claim, not a capability.

6Operational Model & Team Structure

TeamResponsibilityKey Skills
Platform Team (AI Infra)Gateway, model serving, GPU clusters (cloud and/or GreenLake), observabilityKubernetes, MLOps, SRE
Applied AI/ML TeamFine-tuning, RAG pipelines, evaluation harnessML engineering, data science
AI Outcomes / Enablement NewMariete recipes, outcome definitions, business-unit onboarding, monthly outcome reportingWorkflow design, analytics, change management
Security, Legal & ComplianceGuardrails, audits, licensing, AI Act mappingSecurity engineering, privacy, tech law
Business UnitsConsume AI via gateway within budgetsDomain expertise

Sizing: 4 to 8 FTE at steady state for Options A/B; the GreenLake path (Option C) can shave 1 to 2 FTE off infrastructure operations because HPE operates the stack to a service level, redirecting that capacity to evals and outcomes. The binding constraint remains one to two senior MLOps/inference engineers; start that hire in Phase 0. First three hires: senior inference/MLOps engineer, ML engineer (evals/fine-tuning), platform SRE. The outcomes/enablement function can start as a shared PMO role during the Mariete Pilot.

7Cost Management & Economics

The honest frame: below roughly 2 to 5 million tokens per day of sustained volume, self-hosting on reserved cloud GPUs costs more than equivalent frontier-API spend; Tier 1 is resilience and sovereignty insurance first. The savings case activates at scale and with the right model mix.

  • Internal models (Option A): reserved GPU capacity for the floor; autoscaling above it; quantization where quality floors allow; small models for high-volume simple traffic.
  • Internal models (Option C, GreenLake): subscription pricing converts capex to predictable opex and shifts utilization risk partly to HPE; HPE also claims up to 20 percent token-throughput improvement from its integrated stack (vendor benchmark; validate in your own pilot). Compare the 3-year TCO of A vs C on your measured Phase 1 volumes before committing in Phase 2.
  • Execution layer: Mariete's outcome pricing (Pilot $1,200/mo; Outcomes $3,500/mo; Regulated $7,500/mo) is deliberately decoupled from token volume, which makes workflow ROI legible: each outcome has a baseline, target, and monthly measurement. Contrast with per-seat copilot licensing, which bills whether or not work improves.
  • External APIs: per-team budgets enforced at the gateway; anomaly alerts (agents are the usual culprit).
  • Measure: cost per million tokens by tier and workload, monthly; cost per verified outcome, monthly (new in v3.0); break-even re-evaluated quarterly.

8Risks & Mitigations

RiskMitigation
Correlated provider/cloud failure (the Sept 3 pattern)Correlation-domain analysis; Tier 1 in a separate failure domain (GreenLake/on-prem strongest); Tier 2 across two clouds; correlated-failure drills
GPU scarcity / high costReserved capacity or GreenLake subscription for the floor; quantization; small-model routing
Open-weight quality lags frontierQuarterly model review; external APIs for high-quality non-sensitive tasks; quality floors with fail-closed defaults
Failover degrades answer qualityDegraded-mode contracts; golden-set gates; UX disclosure
Agentic workflows strand or overrunGateway-mediated tool calls, step budgets, kill switches; on PCAI: agent registration, OpenShell runtime, Zerto rogue-agent rewind; Mariete workers verify deliverables before returning them
Execution-layer lock-in NewMariete recipes sit above standard endpoints (OpenAI-compatible gateway, MCP connectors); annual substitutability audit proves the execution layer can be re-pointed or replaced
Infrastructure-vendor lock-in NewGateway and eval plane stay vendor-neutral; GreenLake is a consumption contract, not an application dependency; models remain portable open weights
License / provenance exposureLegal pre-approval per model family; registry metadata; board-level provenance policy
Regulatory non-compliance (EU AI Act, sectoral)Phase 0 workload classification; DPO on the board; compliance artifacts from Mariete Regulated and HPE's government-grade baselines where applicable
Mid-market vs enterprise fit of execution vendor NewMariete's published positioning targets SMB/mid-market; large enterprises should scope an enterprise pilot and validate SLAs, or apply the same execution-layer pattern with an in-house equivalent
Self-hosted availability worse than providers in year oneHonest SLO setting; staged migration; drills before dependence

9Measuring the Strategy

Report quarterly:

  • Time-to-switch provider: under 15 minutes, config-only, drill-verified.
  • Failover drill pass rate: including the correlated-failure scenario.
  • Percent of AI spend behind the gateway: target 100 percent; anything else is shadow AI.
  • Quality-floor pass rate: per Tier 1 workload, on golden sets.
  • Cost per million tokens: by tier and workload, trended.
  • Cost per verified outcome (new in v3.0): from Mariete monthly outcome reporting, per workflow.
  • Blast radius: maximum percent of AI capability lost to any single provider or single cloud failure. Target: Tier 1 workloads never below SLO.
  • Substitutability audit result (new in v3.0): annual proof that gateway, execution layer, and infrastructure substrate are each swappable.

10Conclusion

The September 3, 2026 outage was a warning delivered gently: hours of downtime, not weeks. The disciplined response is a sovereign hybrid architecture built on substitutability: a vendor-neutral gateway, an execution layer that turns models into measured outcomes (Mariete, deployed inside your boundary), and Tier 1 infrastructure chosen deliberately among cloud speed, on-prem control, and HPE GreenLake's middle path of sovereign infrastructure consumed as a service.

None of these layers is a permanent commitment; each is audited annually for swap-ability. That is what makes the strategy anti-fragile rather than merely well-diagrammed. The phased plan allows incremental adoption with clear entry and exit criteria. The time to act is now, before the next correlated failure is measured in days.


ASources & Verification

Incident and market context (verified September 3, 2026)

  1. Axios, "Widespread AI outage underway," Sept 3, 2026.
  2. 9to5Google, "ChatGPT, Claude, and Grok were all down in confirmed outages," Sept 3, 2026.
  3. PCMag, "Major AI Outage: Grok, Claude Are Down," Sept 2026.
  4. Anthropic, "A postmortem of three recent issues," Sept 2026.
  5. Onyx, Self-Hosted LLM Leaderboard, updated July 20, 2026.
  6. TrueFoundry, "A Definitive Guide to AI Gateways in 2026," June 2026.
  7. Spheron, vLLM vs TensorRT-LLM vs SGLang benchmarks, 2026.
  8. SitePoint, "Self-Hosted LLM Costs 2026"; René Zander, "Self-Hosted LLM vs API Cost: Break-Even Analysis (2026)."

Mariete (verified from mariete.ai, September 3, 2026)

  1. mariete.ai/enterprise ("Mariete in the Box"): deployment modes, provider router, recipes/MCP connectors, Plan-Build-Deploy motion, and published pricing (Blueprint free; Pilot from $1,200/mo; Outcomes from $3,500/mo; Regulated from $7,500/mo).
  2. mariete.ai homepage: Worker architecture (file system, shell, browser, tools), outcome-based positioning.

HPE GreenLake / Private Cloud AI (verified September 3, 2026)

  1. HPE press release, "HPE brings agentic AI into production with NVIDIA," June 16, 2026: unified model gateway, NVIDIA Agent Toolkit/OpenShell, Zerto rewind, multi-node inference to 256 GPUs, Alletra X10000 (20x TTFT benchmark claim), Data Fabric MCP support, NVIDIA Confidential Computing, DL394 Gen12 with Vera CPU.
  2. HPE Private Cloud AI product page and QuickSpecs: right-sized configurations, DL380a Gen12 with 8x H200 NVL, GB200 NVL4 (DL384b), RTX Pro 6000 Blackwell worker nodes, subscription includes hardware, software, and services.
  3. HPE, "Sovereign by Design" blog (Feb 2026) and Sovereign AI Factory page: data residency, DISA reference, NIST 800/STIG/FIPS/CMMC alignment.

All named models, hardware, and commercial terms carry a standing instruction to re-verify quarterly (section 5, Phase 4).

BChangelog

v2.0 to v3.0 (this version)

  • Added the Execution Plane (section 3.2): Mariete AI Workers, provider router, recipes/MCP connectors, three deployment modes mapped to the tier model, verified outcome-based pricing, and an enterprise-fit positioning note.
  • Expanded section 4 with a three-way infrastructure comparison: self-built cloud VPC vs on-prem capex vs HPE GreenLake Private Cloud AI, with recommended sequencing (A for Phase 1 speed; C evaluated for Phase 2/3 Tier 1; B for steady-state high-volume).
  • Mapped the Mariete + GreenLake pairing for Tier 3 fully-sovereign stacks.
  • Phases updated: Mariete Blueprint/Pilot added to Phase 0/1; infrastructure-path decision and PCAI adoption added to Phase 2; Outcomes/Regulated expansion added to Phase 3; quarterly infrastructure review and annual substitutability audit added to Phase 4.
  • Team model gained an AI Outcomes/Enablement function and GreenLake ops-burden adjustment.
  • Economics extended with GreenLake TCO guidance and cost-per-verified-outcome metric.
  • Risks extended: execution-layer lock-in, infrastructure-vendor lock-in, vendor enterprise-fit.
  • KPIs extended: cost per verified outcome; substitutability audit.

v1.0 to v2.0 (carried forward, abridged)

Named/dated Sept 3 incident with Azure correlation; Tier 3 defined; degraded-mode contracts; Evaluation & Quality and Governance planes; agentic gateway capabilities; TGI retired, SGLang/TensorRT-LLM added; quarterly-reviewed capability-class model table; H200/B200 hardware guidance; licensing/provenance and EU AI Act section; correlation-aware topologies; phase entry/exit criteria; team sizing; break-even economics (2-5M tokens/day); strategy KPIs.