·Fintech Accounting
AI Startup Accounting in Estonia: What to Do With the GPU Bill
An AI startup's largest expense line after payroll is usually compute, and how that compute is classified decides what the company's gross margin looks like to investors. Inference that serves paying customers is cost of revenue; training runs and experiments are R&D. Get the split wrong and the margin story in the pitch deck collapses in diligence. This post walks through AI startup accounting on an Estonian OÜ: GPU cost classification, capitalising versus expensing model development, why Estonia's 0% tax on retained profit suits compute-hungry companies, grant and Horizon Europe accounting, usage-based revenue recognition, and the investor reporting a seed+ AI company should produce.
Is GPU spend cost of revenue or R&D?
The honest answer is: both, and the split matters more than almost any other accounting decision an AI company makes. Compute that directly serves paying customers, inference for production API calls, hosting for deployed models, belongs in cost of revenue and therefore reduces gross margin. Compute burned on training runs, fine-tuning experiments and internal research belongs in R&D operating expense and leaves gross margin untouched. Two companies with identical bank statements can show a 40-point gross margin difference purely on this classification.
Investors know this, which is why diligence teams at seed and Series A test the split. The defensible approach is mechanical, not narrative: tag compute at the infrastructure level (separate projects or accounts for production inference versus training), let the ledger inherit those tags, and document the policy once. When your gross margin is challenged in a partner meeting, "here is the cloud billing breakdown, mapped line by line" ends the conversation in a way that "we estimate roughly 60/40" never does.
Should model-development costs be capitalised or expensed?
Accounting frameworks allow development costs to be capitalised as an intangible asset when strict criteria are met: technical feasibility, intention and ability to complete, probable future economic benefit and reliably measurable cost. In practice, most early-stage AI work is research and experimentation that fails these tests and must be expensed as incurred. Capitalising aggressively makes this year's P&L look better but creates an asset that must be amortised and can be impaired the moment a model is superseded, which in AI can be a quarter later.
Our default advice for seed-stage AI companies is to expense model development unless there is a specific, durable asset with a clear revenue link, and to be able to explain the choice in one sentence to an investor. VCs generally prefer a clean, conservative P&L over a balance sheet inflated with self-built intangibles. What matters for the raise is not accounting cosmetics but that the treatment is consistent, documented, and survives a diligence question.
How does Estonia's 0% tax on retained profit help an AI company?
Estonia charges corporate income tax of 22% only when profit is distributed. Profit that stays in the company, and gets reinvested in GPUs, training runs, and headcount, is taxed at 0% until distribution. For an AI startup this is a structural advantage: every euro of margin the product generates can be recycled into compute without a tax leak in between. A profitable AI company in a classical tax system pays corporate tax first and reinvests what is left; an Estonian OÜ reinvests the gross amount.
The same logic applies to the venture case. VC-funded companies rarely distribute profits at all before exit, which means the Estonian regime effectively defers corporate tax for the entire growth phase. Combined with a €1 minimum share capital, e-Residency for remote founders, and a single mandatory annual report to RIK, the administrative footprint of the structure is small relative to what it delivers. The discipline it does demand is proper monthly bookkeeping, because the annual report and any future diligence stand on it.
What about grants and Horizon Europe funding?
AI is one of the best-funded areas in European grant programmes, and Estonian companies regularly draw on Horizon Europe calls, national innovation support and ecosystem programmes around Startup Estonia. Grant money is welcome, but it is not revenue: it is typically recognised as deferred income and released to the P&L as the subsidised costs are incurred, and every grant comes with cost-eligibility rules, co-financing shares and reporting deadlines that the bookkeeping must support from day one.
The practical requirement is separate cost tracking: grant-funded compute, salaries and subcontracting must be identifiable in the ledger, usually per project, with timesheets where personnel costs are claimed. Retrofitting this at report time is painful and audit findings can lead to clawbacks. If a grant is in the pipeline, the project accounting structure should be set up before the first euro is spent, and this is exactly the kind of setup a specialist accountant does in an afternoon and a generalist discovers a year too late.
How do you recognise usage-based revenue?
Most AI products bill on usage: per token, per API call, per seat-plus-consumption. The accounting principle is that revenue is recognised as the usage happens, which aligns nicely with billing in arrears. Prepaid credit packs are the case to watch: cash received for credits is deferred revenue until the customer actually consumes them, and unused credits stay on the balance sheet as a liability. If credits expire, the expiry policy determines when the remaining balance can finally be recognised.
Usage-based models also blur the MRR concept, so investor metrics need care: committed spend, consumption run rate and net revenue retention on a consumption basis tell the real story. The ledger should be able to produce revenue by customer and by month directly from metering data, reconciled to the payment processor. That single pipeline, metering to invoice to ledger, is what makes both the auditor and the Series A analyst trust the revenue line.
What do investors expect when burn is compute-driven?
A seed+ AI company's burn has a different shape from a classic SaaS startup's: compute can rival or exceed payroll, and it scales with experiments, not headcount. Investors therefore expect burn reporting that splits compute from people and shows unit economics per unit of usage: cost per thousand API calls, margin per customer, and how training spend maps to roadmap milestones. A single blended burn number hides exactly the thing an AI investor is trying to underwrite: whether marginal usage is profitable.
TechAccounting builds this into the monthly close for AI clients: P&L with a defensible COGS/R&D compute split, burn and runway from actual cash, budget versus actual, grant project tracking, and ESOP payroll handled through TSD. Fixed-fee accounting starts at €499 per month, with Compliance at €1,500 and Fintech & AML at €3,000 for regulated setups, and consultations at €150 per hour. The goal is simple: when your lead investor asks how the GPU bill flows through the books, the answer is already in the data room.