AWS INFRASTRUCTURE COST AUDIT
Most AWS overspend isn't in one place. It's spread across compute, storage, networking, AI inference, and idle infrastructure. We audit your entire AWS setup and tell you honestly where the money is going, what's worth cutting, and what to do about it.
Free of charge. Written report delivered within 3 business days of the discovery call. Read-only access only; we never make changes to your account.
WHERE THE MONEY GOES
These are the five places we look first. Most teams have at least two of them.
01
Every failed API call that retries is billed at full token cost. If your error handling is optimistic, with three retries and a logged failure, you are paying for the same tokens two or three times on every flaky request. In a high-throughput pipeline, this adds up fast and shows up nowhere obvious in your dashboard.
02
Agentic systems that pass full conversation history on every step compound quickly. A ten-step agent with a 4,000-token context is not spending 40,000 tokens. It is spending closer to 220,000 if the context grows linearly. Most teams don't model this until the bill arrives.
03
Retrieval-augmented pipelines that stuff the top-k chunks into the prompt regardless of relevance are paying for tokens the model ignores. The fix is usually a reranker or a tighter similarity threshold, not a model change. But it requires someone to look at the retrieval logs, not just the inference costs.
04
EC2 instances reserved for peak load sitting at 20% utilisation. S3 buckets accumulating data nobody queries. RDS instances sized for a future that hasn't arrived. NAT Gateway traffic that could be routed more cheaply. These are not engineering failures. They are the natural result of moving fast. They're also straightforward to fix once someone looks.
05
Opus 5 and GPT-5.6 are excellent. They're also 10–50× more expensive than the models below them for tasks that don't require their capability. Classification, summarisation, routing, and extraction are often running on frontier models because that's what the team started with. A routing layer that sends simple tasks to a smaller model can halve the inference bill without touching output quality.
HOW IT WORKS
01
A 20-minute call to understand what you're running, where you think the spend is concentrated, and what matters most. After the call, we request read-only access to your AWS account. We only read and never make changes. No production credentials beyond a read-only IAM role are needed.
02
We audit your compute, storage, networking, and data transfer costs, concentrating on AI workloads and the highest-spend services. We pull billing data, resource inventory, and usage patterns across the last 90 days and map spend against actual utilisation.
03
We review your provisioned AI services and LLM usage costs against how the product actually operates. This covers model selection, token usage, caching efficiency, batching strategy, and whether frontier models are doing work that cheaper models could handle.
04
The output is a clear, actionable document. It covers every finding with a specific recommendation against it, an implementation roadmap in priority order, and projected annual savings quantified per item. Where implementation work is needed, you can take it to your own team or we can introduce our development partner.
WHAT YOU GET
The report covers your full AWS infrastructure alongside a dedicated AI and LLM spend breakdown. The AI section examines model selection, token usage patterns, caching and batching efficiency, and whether you're paying frontier prices for tasks that don't need frontier capability.
Every finding comes with a specific recommendation, an implementation note, and the projected annual saving quantified in dollars. Findings are prioritised by impact so you know what to fix first.
The report is yours to keep regardless of what you do next. Take it to your own engineering team, implement it yourself, or ask us to introduce our development partner. There is no obligation either way.
COMMON QUESTIONS