Sovereign AI infrastructure
More capacity
from the GPUs you already run
Sovereign AI infrastructure has limited room. Capacity is finite on your own GPUs, in a European cloud and on a neocloud. Akamas Inference Optimizer is AI-driven configuration for LLM inference. It tunes the full inference stack under your real traffic, on every GPU type you run. Each GPU serves more tokens within the latency targets you set.
- More tokens and more tenants per GPU
- Latency targets held in every configuration
- Runs on-prem or in your cloud account: models and data stay with you
2026 to 2030
- Demand: 24× Token consumption 2026 to 2030.
- The walls: 6× Memory price June 2026 vs June 2025.
- The room inside: Same GPUs More tokens from capacity you already paid for.
Why sovereign
Why are organizations moving to sovereign AI infrastructure?
More organizations now run their own models instead of calling an external AI provider. Some run on-premise AI on their own GPUs. Others use a European cloud, a neocloud, or a mix of all three. Three forces push them there.
All three lead to the same place: every token now runs on GPU capacity you are responsible for.
Control and jurisdiction
Prompts and answers carry customer data, source code and business knowledge. Organizations want them in their own data center or cloud account. That keeps them under their own rules and jurisdiction. In Europe, it means AI that runs under European control.
Independence from a provider’s terms
When your AI runs at an external provider, its price, terms and availability are theirs to change. A self-hosted alternative keeps that decision with you. A change in someone else’s pricing stops being your problem.
Regulation
Prompts and answers often contain personal data, which GDPR governs. NIS2 makes essential and important entities manage risk across their technology suppliers. In finance, DORA does the same for ICT providers. Every external AI provider is one more supplier to assess.
The problem
Why does sovereign AI infrastructure run out of room?
Control comes with a fixed amount of capacity, in a market where demand grows faster than anyone can build. New capacity takes hardware, power and permits, whether you build it or a European provider does, and GPU supply sets the pace. Overprovisioning was a cloud habit. On prem, it’s a capital expense with a waiting list.
- Driver
- Agentic AI
- Effect
- New capacity costs more
Where the capacity goes
Why do default settings waste the GPUs you paid for?
The fastest place to grow is the capacity you already have. Default settings leave part of every GPU’s capacity unused.
Inference is a full-stack problem. Every layer between the hardware and your users has its own settings, and they interact. The defaults were not chosen for your model, your GPUs or your traffic. Tuning one layer at a time misses the best combination.
Tuning just 8 of the 40 vLLM settings already gives 48k+ possible configurations. At about an hour per test under load, trying them all would take more than five years of GPU time.
Read more on AI workload optimization for GPUs and Kubernetes and on why GPU utilization stays low in LLM inference.
Hardware and GPU
GPU type, interconnect, GPU sizing and sharing between models.
Kubernetes
Pod requests and limits, replicas and autoscaling policy.
Inference server
Batching and concurrency, KV cache, parallelism, prefill and decode split.
Model
Quantization, attention backends and context length.
Routing
KV-cache-aware routing and model selection.
Workload
Prompt and answer lengths, concurrency and the latency targets each service must meet.
How it compares
How do you add capacity to sovereign AI infrastructure?
There are three ways to serve more tokens. They work together: tuning makes every GPU you buy or rent serve more.
| Buy more GPUs | Rent more sovereign cloud GPUs | Tune the GPUs you run | |
|---|---|---|---|
| What it takes | Hardware, power, space and permits | A contract with a European cloud or neocloud | Experiments on the GPUs you already run, under your traffic |
| What limits it | GPU supply, memory prices and data center build time | The GPU types the provider has available | How much capacity your current settings leave unused |
| What you pay for | New hardware, up front | GPU hours or reserved capacity | Software. No new hardware |
| Stays sovereign | Yes | Yes, with a European provider | Yes, runs on-prem or in your cloud account |
| Effect on the other two | None | None | Every owned or rented GPU serves more tokens |
Hybrid
How do you get the most from a hybrid GPU mix?
Many sovereign setups mix paths. Your own GPUs carry the steady load. A European cloud or neocloud covers the peaks with whatever GPUs it has available.
Every GPU type needs its own configuration, and the mix changes with demand. The same settings copied to different hardware give very different results. No team can re-tune each combination by hand.
The Akamas AI engine finds the best configuration for each GPU type under your load, automatically.
Demand moves through the day. The peaks are where capacity runs short.
- Demand: Demand moves through the day. The peaks are where capacity runs short.
- The mix: Your own GPUs carry the steady load. The peaks go to whatever GPU types the European cloud or neocloud has that day, and every type runs on its defaults.
- Tuned: The Akamas AI engine finds the configuration for each GPU type under your load. Your own GPUs carry more of the day, and less is left to rent.
How it works
How does Akamas find capacity in GPUs you already run?
Sovereignty decides where your models run. Akamas decides how well they run there. Akamas Inference Optimizer runs experiments on your own GPUs, under your own traffic. Its AI picks each next configuration from what earlier runs measured. It keeps only configurations that meet your latency targets. Akamas Inference Optimizer is built on the Akamas Studio engine.
-
Set the goal and the limits
For example: maximize tokens per second per GPU, keeping time to first token p95 under 1.5 s.
-
Run under your real load
Each experiment runs your model on your GPUs, with traffic shaped like your production traffic.
-
Search the whole stack
Settings across the inference server, GPU sharing and Kubernetes change together, experiment after experiment.
-
Review and apply
You get the best configuration found, with every limit held, ready to apply to production.
What changes
What does tuning change in a sovereign AI setup?
Sovereignty is the reason to build. Efficiency is the reason it pays back.
On your own GPUs, idle capacity is money already spent. If you sell sovereign AI capacity, it’s capacity you can’t sell.
See the measured results on the Akamas Inference Optimizer page
More from every GPU
GPU capacity can’t be overcommitted the way CPU can. Tuning is how you fit more users, models and versions on the GPUs you have, or more tenants if you sell capacity.
Latency you can commit to
Time to first token and time per output token are limits held in every experiment. You know what a configuration delivers before it serves users or backs an SLA.
Hardware purchases pushed back
Every point of efficiency is capacity you don’t have to buy, wait for or rent. On owned GPUs, it delays the next purchase. On rented GPUs, the same traffic needs fewer of them.
Every new model or GPU, tuned in days
A new model, a new GPU type or a new workload changes the best configuration. Re-optimize in days instead of months of hand tuning.
BLOG
FAQ
Questions about sovereign AI infrastructure
Sovereign AI is AI that an organization or a country runs under its own control. Its models, data and infrastructure stay under that jurisdiction’s rules. It can run on owned GPUs, in a European cloud or on a neocloud. The goal is independence from outside providers for the AI that matters most. Read the full definition.
Sovereign AI infrastructure is GPU capacity an organization controls under its own jurisdiction. That capacity runs AI models without sending data, prompts or models to an outside provider. Sovereign AI providers, telcos and enterprises that keep models in-house build this capacity.
Three reasons. Organizations want prompts, answers and models in their own data center or cloud account. They want independence from an external provider’s pricing, terms and availability. And GDPR, NIS2 and DORA make every outside AI provider one more supplier to assess.
New capacity takes hardware, power and permits. GPU supply and memory prices set how fast it arrives. Meanwhile token demand keeps growing. The capacity you already have has to serve more.
Not by itself. But new GPUs and a new serving stack start on default settings, and defaults rarely fit your models or your traffic.
Every point of efficiency is hardware you don’t have to buy or wait for. On owned GPUs, idle capacity is money already spent.
Akamas Inference Optimizer tunes how your models are served, across the whole inference stack. That covers the inference server, GPU sharing and Kubernetes. It runs experiments under your real load. It keeps the configuration that serves the most tokens per GPU within your latency targets.
On your infrastructure. Akamas Inference Optimizer runs on-prem or in your own cloud account. It tunes the inference layer of your sovereign AI platform. Models, prompts and measurements stay under your control.
Yes. It works for on-prem LLM deployments, European clouds and neoclouds. The Akamas AI engine finds a configuration for each GPU type under your load. On owned GPUs, freed capacity serves more users and delays the next hardware purchase. On rented GPUs, the same traffic needs fewer of them.
Yes. Akamas is the autonomous optimization platform for Kubernetes and AI infrastructure. It is headquartered in Milan, Italy, with an office in Boston, US. Founded in 2019, Akamas builds on more than 20 years of performance engineering.
The Akamas Inference Optimizer page shows the measured studies. Each one lists its model, its GPUs and the latency limits it held. See the results.


How much capacity is hiding in your racks?
Tell us your models, your GPUs and your latency targets. We’ll show you how much more the GPUs you already run can serve.