Every AI request has a cost. We audit the path.
Seidia audits UK businesses' AI inference spend across Azure OpenAI, model APIs and cloud GPUs. We fix the waste by hand, then turn the fixes that repeat into software that acts on its own.
Pre-seed · forming at UCL, London · nothing built yet
- Smaller models£11,000
- Batch API£4,800
- Frontier models£7,500
- Prompt cache£1,900
Four of these five fixes can run as software. Those become the product.
AI bills grow in places nobody audits.
Teams adopt models faster than finance can follow. The waste is rarely one big mistake. It is the same five leaks, paid for every month, under someone else's budget line.
| Where it leaks | What the audit does | Can it be automated? |
|---|---|---|
Every request goes to the biggest model. | Model right-sizingMatch each workload to the smallest model that does the job, and route the rest. | Yes: routing |
Overnight work is billed at real‑time rates. | Batch APIsMove non-urgent jobs such as summaries, tagging and evals to batch processing. | Yes: scheduling |
Context windows are padded with whole documents. | Context-window bloatSend the passages that matter, not every file a user touched. | Case by case |
The same instructions are paid for on every call. | Prompt cachingCache shared system prompts and reference material instead of re-sending them. | Yes: caching |
No one can say which team spent what. | Visibility and budgetsAttribute spend per team and workload, then set budgets and controls that hold. | Yes: enforcement |
Start with services. Earn the right to build the product.
The first offer is a hands-on AI spend audit, modelled on how Duckbill HQ approached AWS bills and applied to inference. Each audit pays its way and shows us where the money actually leaks.
- Scope: every AI line item.
- AI APIs, Azure OpenAI and Microsoft Foundry, and cloud inference, mapped by team, model and workload.
- Deliverable: findings, then fixes.
- A ranked list of savings levers with the fix for each, implemented alongside the client's engineers rather than left in a report.
- Price: fixed fee or share of savings.
- The client chooses. Either way, our incentive is a smaller bill they can see.
Proposed first offer. Scope, deliverables and pricing are still being tested.
Cost of compute is an acquisition market.
Cloud cost tools and compute-efficiency companies have repeatedly been bought by the companies that own the compute or the customer: cloud providers, chipmakers and GPU clouds.
Cloud FinOps
| CloudHealth | VMware |
| Cloudability | Apptiolater IBM |
| Cloudyn | Microsoft |
| Spot.io | NetApp |
| Kubecost | IBM |
Compute efficiency
| Granulate | Intel |
| Run:ai | Nvidia |
| Deci | Nvidia |
| OctoAI | Nvidia |
| Neural Magic | Red Hat |
Inference optimisation, 2026
| Eigen AI | Nebius~$643M reported |
| Inferize | Nebius$100–130M reported, ~10 months after founding |
| SqueezeBits | Rebellions |
| Moondream | Prime Intellect |
Acquirer on the right. Compiled from public reporting; deal values are as reported and not independently verified.
Dashboards got squeezed. Tools that act got bought.
Pure spend dashboards lost ground to the cost tools the clouds now ship themselves. The companies that lasted either took automated action or did deep, hardware-aware optimisation. Seidia is built to end up there: audits first, then software that acts.
Audits, then software that acts, then hardware-aware advice.
Each stage funds and informs the next. We only automate what audits prove repeats.
Hands-on AI spend audits for UK businesses.
Now · services
Starting with mid-market firms whose AI spend is growing. Every audit is paid, and every finding is logged.
Productise the fixes that repeat.
Next · after 10–20 audits
Routing, prompt caching and budget enforcement as software that takes action automatically, never a passive dashboard.
Hardware-aware AI advice.
Later · hardware track
On-prem versus cloud inference, and what hardware to run open models on. Grows from our hardware-software integration consulting.
What we are deciding now
The strategy is proposed, not committed. These are the open questions the first audits will answer.
- Which segment first: UK mid-market on Azure OpenAI, direct API users, or startups.
- The exact shape of the first audit: scope, deliverables and pricing.
- The minimal tooling needed to run audits efficiently.
- How the hardware track connects to the cost-of-compute track.
The risks, stated plainly.
Every investor will ask these, so we answer them first.
Model prices keep falling.
Our answer: we pitch control and waste, not a fixed percentage off the bill. Our bet is that cheaper tokens bring more usage, and more need to govern it.
Clouds keep shipping native cost tools.
Our answer: that is why dashboards were squeezed. We aim to take action, and to work across providers rather than inside one console.
Consulting does not scale.
Our answer: audits are the research phase, not the business. After 10–20, recurring fixes become software.
Deep inference-optimisation expertise is thin today.
Our answer: audits are our learning loop, the founder's systems background shortens it, and our mentors bring enterprise Azure practice.
A systems team, starting close to the metal.
A small founding team forming Seidia at UCL in London. The idea grew out of the founders' own market research into AI workflow spending controls.
The founder
Electrical engineering background, from hardware description up to backend services.
- C / C++
- Verilog
- Systems programming
- Kotlin / Spring Boot
- Python
Mentors
Mentored through a UCL programme by practitioners from Avanade, the Microsoft and Accenture joint venture focused on enterprise Azure. They help us learn how enterprises manage AI spend today.
Where we start
UK-first and London-based, expanding to Europe and the US later. Hardware-software integration consulting runs alongside as a second track.
Talk to the founders.
We are speaking with investors, mentors and UK businesses whose AI bills are growing. If you are one of them, write to us and we will send the full thesis.
Pre-seed · London · 2026