Sovereign LLM-as-a-service on AMD Instinct™ MI350P.A three-tier model catalog with per-tenant endpoints on one LIQID UltraStack server, in your facility and jurisdiction.
Bet on a GPU cluster before tenants sign, or miss the demand for private endpoints.
In-country LLM endpoints their data never leaves.
Fast, 70B and reasoning traffic rise and fall at different times.
Large GPU clusters come before paying tenants do.
Buying for peak leaves capacity unused most of the day.
Tenant data has to stay in your facility and jurisdiction.
Each one gets its own endpoint and API keys.
One 10-GPU LIQID UltraStack server serves a three-tier catalog: a fast small model, a balanced 70B and a premium reasoning model. Each tenant gets its own endpoint, and each tier runs on its own GPUs, one model per GPU.
Llama 3.1 8B replicas serve the low-cost, high-volume tier.
Llama 3.3 70B, one full model per GPU, for general-purpose endpoints.
gpt-oss-120b, one model per GPU, serves the reasoning tier.
No cluster bet: one server, three tiers, and capacity that grows with paying tenants.
Add GPUs and a new job, not a new cluster.
Fast, balanced and premium endpoints with per-tenant API keys.
4 × Llama 3.1 8B fast tier · 3 × Llama 3.3 70B balanced tier · 3 × gpt-oss-120b premium tier
Neocloud / CSP product lead. First tenants outgrow a 4-GPU starter.
Shared tiers plus reserved GPUs for enterprise tenants that want dedicated capacity.
10 × shared catalog tiers (8B, 70B, gpt-oss-120b) · 10 × dedicated tenant models, one per GPU
VP Product / GM, AI cloud. Enterprise tenants asking for dedicated, in-country capacity.
The full catalog plus dedicated tenant models, in your jurisdiction.
10 × catalog tiers (8B, 70B, gpt-oss-120b) · 20 × dedicated tenant models, one per GPU
CTO / sovereign cloud program. Customers want sovereign AI without a hyperscaler.
Two questions settle the size of the opportunity: how many tenants want private endpoints today, and which models do they want?
*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.
Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models.