AquilaClouds Kestrel
Know what your AI costs, what it returns, and who authorized it.
Kestrel gives finance, engineering and AI leaders a single record of AI spend across frontier providers, privately hosted models and your own GPUs. It attributes every call to a team, an application and an owner.
It routes each request to the right model and deployment, enforces who may use which model and with what data, and gets more throughput from your GPUs. Then it shows what all of it returns to the business.
TRUSTED BY LEADING ORGANISATIONS WORLDWIDE
The problem
Enterprise AI now runs in three places, and each one is managed separately.
-
Frontier APIs
Spend on OpenAI, Anthropic, Google Gemini and others grows faster than finance can forecast, attribute or cap it. Regulated data often can't go there at all.
-
Self-hosted GPUs
Organizations buy GPUs for sovereignty or cost, then leave them underused because nothing schedules, tunes or meters them.
-
Sovereign and private environments
Regulators require them, but they usually run as a separate project with their own tools and their own audit gaps.
The result is familiar:
- There is no single number for AI spend, so budgets are set blind and overruns surface late.
- There is no reliable record of what a model was asked, which model answered and who approved it.
- Moving a workload between a frontier API, an owned GPU and a sovereign environment means rebuilding it.
- Leaders can't say whether AI investment is paying off.
42%
~30%
47%
30–40%
Capabilities
Seven capabilities in one AI control layer
See all of your AI spend in one place
Token observability and cost attribution
One token ledger across every provider and deployment.
- Real-time token consumption and cost across OpenAI, Anthropic, Google Gemini, Vertex AI, Azure AI Foundry and on-premises models.
- Spend broken down across twelve dimensions, including provider, model, user, project, API key, account, subscription and deployment.
- Attribution by department, application, model family, project and user group, with showback and chargeback.
- Baseline thresholds and proactive alerts that stop runaway spend.
- API key tracking that exposes shadow AI.
Pay less for every task
Token and model optimization
Lower the cost of each request without lowering the quality of the answer.
- Analyze and compress prompts and chat history to remove unnecessary input tokens.
- Use prompt caching where it pays off.
- Route each query to the most cost-effective model that meets the task's accuracy goal.
- Compare models and providers on cost, quality and latency.
- Combined with GPU optimization, this targets savings of 30% or more of the AI budget.
Govern AI with an audit-ready record
Policy-based routing and AI governance
Governance enforced by the platform, not by convention.
- Route every request by data classification across three lanes: frontier models for long-context and agentic work, cost-efficient models for most traffic, and a sovereign lane that keeps restricted data in-country with no egress.
- Escalate a failing sub-task to a stronger model automatically.
- Set who may call which model and provider, and what data may leave.
- Keep a hash-chained, tamper-evident audit ledger of every decision and tool call: the question, what was retrieved, which model version answered, what it returned and who reviewed it. Auditor access is recorded too.
- Change policy through a signed policy edit, not a code release.
Get more from your GPUs
GPU runtime optimization
The same model on the same GPUs, tuned for the traffic you actually run.
- An optimizer runs beside each server and tunes live serving settings against real traffic.
- Each change is measured and kept only if it improves results. Otherwise it is reverted automatically.
- It never sits in the request path. The gain shows up as fewer GPUs needed to meet the same service level.
Get more from your GPUs
GPU right-sizing
Pick the cheapest configuration that still meets your performance target, instead of provisioning for the worst case.
- Choose model, quantization, GPU type and GPU count against the workload's own service-level objective.
- Benchmark once on your own hardware, then reuse the results for every sizing decision.
- Right-sizing typically saves 30 to 40% on GPUs before a single request is served.
- Each recommendation is the lowest-cost configuration that still meets the workload's throughput and latency targets.
Model-as-a-Service for private AI
- One OpenAI-compatible API for every internal team or customer.
- Quotas, service tiers and provisioned capacity.
- A signed model catalog with certified launch profiles.
- Per-request metering of tokens in, out and cached, ready for internal chargeback or customer billing.
Prove the return
AI ROI and value measurement
After "what is AI costing us?" comes "what is it returning?" Connect AI cost to measured value by role and persona, using a method finance can inspect, so you can decide where to invest more and where to cut.
- The full cost side: license cost per seat, token cost measured from the ledger, and the hidden costs of change management, enablement and support.
- The value side, measured from real signals such as time on task, acceptance, deflection and escalation, and converted to money using value bands you can adjust.
- Three scenarios, conservative, base and optimistic, instead of one vendor-supplied ROI number.
- Common early findings: licensed seats that were never activated, often a quarter to a third of a large program, and returns that differ sharply by role.
Customer stories
What customers say
Robot Toolworx builds AI for industrial maintenance and repair in aviation, maritime and automotive.
“Our partnership with AquilaClouds helps us deliver AI solutions with the reliability, security, predictability, and cost control our customers require.”
Innominds runs AI programs for some of the world’s largest enterprises and ran Kestrel internally before offering it to clients.
“The enterprises that win with AI will not be the ones that spend the most. They will be the ones that know exactly what they are spending, on which models, and on whose authority.”
FAQ
Questions we hear most
Does Kestrel replace my model providers?
No. Kestrel works across the providers you already use, such as OpenAI, Anthropic, Google Gemini, Amazon Bedrock and Azure AI Foundry, and across the models you host yourself.
Can sensitive data stay in my environment?
Yes. Policy-based routing sends restricted data only to private, on-premises or sovereign environments, with egress controls per lane.
Do I need my own GPUs?
No. Token observability, model optimization, governance and ROI work with provider APIs alone. GPU runtime optimization and right-sizing add value when you run models on your own hardware.
How does Kestrel relate to Andromeda?
Andromeda is Aquila Clouds’ platform for multi-cloud and SaaS cost management. Kestrel is a separate AI suite that complements it. Organizations can use either one, or both together.
