Most AI companies are wasting 20% to 40% of their inference spend.

What I usually find

Inference bills grow faster than the products behind them, and almost nobody can say which line of that bill is waste. None of it shows up as a failure. It shows up as a cost line growing 15% a quarter while nobody can name the cause, compounding every month it goes unmeasured.

The waste itself is rarely exotic. It is almost always one of these four, and it persists because nobody has the capacity to go looking, not because nobody suspects.

• Frontier models doing trivial work

  Classification, routing and extraction running on the most expensive model in the catalogue because that's what the prototype used.

• The same context paid for repeatedly

  System prompts, few-shot examples and retrieved documents re-processed on every request, with a prefix cache either off or missing most of the time.

• Agent loops burning tokens on retries

  Reasoning steps and tool calls nobody instruments, where the cost per completed task is several times the cost per task that succeeds first time.

• Compute sized for training, running inference

  Idle floors, absent scale-to-zero, on-demand pricing where committed or batch-tier rates apply, and GPUs at high utilisation doing very little useful work.

The 72-Hour Margin Diagnostic

Three working days. One document, the Inference Waste Map, delivered on day three. It is yours without restriction: act on it internally, take it to your board, or hand it to another supplier.

1. Your real cost baselineAnnualised inference spend decomposed by model, workload and cost phase: prefill separated from decode, cost per token alongside cost per completed task.
2. The waste mapEach inefficiency found, what it costs you per year, and the billing evidence behind it.
3. A prioritised recovery roadmapP0 to P2, each with an estimated annual saving and the implementation method.
4. What it would cost you to do it yourselfAn honest engineering-effort estimate per item, so you can decide whether you need me at all.
5. Risk and compliance notesWhere an optimisation carries a quality, latency or compliance consideration, it's stated plainly rather than buried.

Low-risk to onboard

Your security team will ask before your engineering team does. Here is the whole answer.

Confidentiality. Everything I see stays confidential. Access begins only after I have signed your NDA.

Who does the work

I'm Evandro Zeferino, a software engineer and team lead in cloud infrastructure, with five years running production infrastructure and cost optimisation. EZ Consulting is that work pointed at a single problem: inference has become the largest controllable line in an AI company's cost of revenue, and almost nobody owns it.

Every finding in your report is analysed and vetted by me personally. There is no automated scanner, no offshore analyst team, and no junior consultant learning on your stack. If the numbers are wrong, they're wrong on my name.