AI Software — Capex vs Opex Under Ind AS 38 + Section 32 (India)

Who this page is for

Three audiences:

1. Indian listed entities preparing financial statements under Ind AS, where AI software classification affects intangible asset balances + amortisation + deferred tax
2. Indian mid-market CFOs + finance heads scoping AI spend treatment for monthly accounting + year-end audit
3. Statutory auditors + ICAI practitioners evaluating client AI spend classification under the SEBI/MCA January 2026 AI disclosure mandate

The short answer: most AI spend is opex. SaaS subscriptions, foundation-model API consumption, fine-tuning third-party APIs, training costs — all opex by default. Capex applies narrowly: purchased perpetual software licences, internally-developed AI capability meeting Ind AS 38 paragraph 57 criteria, and certain customisation scenarios with customer control.

The side-by-side

Spend category	Typical treatment	Why	Ind AS 38 / Section 32 reference
SaaS subscription (monthly / annual)	OpEx — period expense	No control / intangible asset created; recurring access right ceases on non-renewal	Ind AS 38 §17 (control test) + Section 37 IT Act
Multi-year SaaS prepayment	Prepaid expense — amortised over commitment period	Pre-paid OpEx, not an intangible asset	Ind AS 38 (no asset recognition) + Section 37
Foundation model API consumption (OpenAI, Anthropic, Google)	OpEx — period expense	Pay-per-use consumption; no asset created; no control of underlying model	Ind AS 38 (no asset) + Section 37
Fine-tuning on third-party foundation model (via OpenAI / Anthropic / Google APIs)	OpEx	Resulting capability inseparable from provider’s platform; criterion (c) of Ind AS 38 §57 fails	Ind AS 38 §57 (fails ability-to-use test) + Section 37
Fine-tuning on open-weight model self-hosted (Llama / Mistral / etc.)	OpEx by default; CapEx if all 6 criteria met	Customer-controlled outputs possible; capitalisation requires all six Ind AS 38 §57 criteria met	Ind AS 38 §57 + Section 32 (40% WDV if capitalised)
Purchased perpetual software licence (on-premise)	CapEx — capitalise + amortise	Control of asset; future economic benefit extending beyond one period	Ind AS 38 §10-17 + Section 32 (40% WDV)
Internally developed AI capability (custom model, in-house build)	OpEx during research phase; CapEx in development phase if criteria met	Ind AS 38 §54 (research) vs §57 (development); 6-criteria test applies	Ind AS 38 §54-§57 + Section 32 (40% WDV if capitalised)
Post-deployment maintenance + updates	OpEx	Recurring maintenance of existing asset; does not extend useful life	Ind AS 38 §69 + Section 37
Staff training to operate AI	OpEx	Training costs cannot be capitalised under any framework	Ind AS 38 §69(b) (explicit exclusion) + Section 37
Cloud reserved-instance or committed-use prepayment	Prepaid expense — current + non-current portions	Multi-period operating expense, not an asset	Ind AS 38 (not an asset) + Section 37 (over period)
Cloud labour at delivery (engineers building on cloud infra)	OpEx default; CapEx allocable to internally-developed asset if Ind AS 38 §57 met	Direct attributable cost can be capitalised if asset criteria met	Ind AS 38 §66

The Ind AS 38 paragraph 57 six-criteria test (for any CapEx claim)

For any internally-generated intangible asset (including internally-developed AI capability), all six criteria must be demonstrably met:

1. Technical feasibility of completing the intangible asset
2. Intention to complete and use or sell it
3. Ability to use or sell the intangible asset
4. Demonstration of how the asset will generate probable future economic benefits
5. Availability of adequate technical, financial, and other resources to complete development
6. Ability to reliably measure expenditure attributable to the intangible asset

In practice, the criterion that most often fails for AI fine-tuning is criterion 3 — ability to use or sell. If the fine-tuning lives on a third-party API platform and the entity cannot extract the resulting capability for use independently of the provider, criterion 3 fails — and capitalisation is precluded regardless of how well the other five criteria are met.

For internally-developed AI capability on open-weight models (Llama / Mistral / etc.) hosted on the entity’s own infrastructure, all six criteria can be satisfied — capitalisation becomes possible.

Section 32 IT Act — the tax depreciation overlay

Where Ind AS 38 capitalises a software asset, Section 32 IT Act applies for tax depreciation:

• Computer software falls in the Block of Assets category attracting 40% depreciation on Written Down Value (WDV) basis
• Tax depreciation calculation is independent of book amortisation under Ind AS 38
• Book amortisation under Ind AS 38 is typically straight-line over useful life (2-5 years for purchased software licences); tax depreciation is 40% WDV
• The differential creates a timing difference under Ind AS 12 (Income Taxes) → deferred tax line item
• For internally-developed AI capability capitalised under Ind AS 38, tax treatment depends on classification — if the asset is “computer software”, 40% WDV applies; otherwise the general intangibles regime applies

The common misclassifications BatchWise sees

Misclassification	Why it happens	Consequence
Multi-year SaaS prepayment expensed in year 1	Bookkeeper treats full payment as period cost	Year 1 P&L overstated; subsequent years understated; current-asset balance understated
Custom AI model build expensed instead of capitalised	Finance team treats all R&D as opex by default; doesn’t run the §57 test	Section 32 depreciation claim lost; permanent tax disadvantage
Fine-tuning on third-party API capitalised in error	Finance team capitalises any “AI development” cost regardless of platform	Audit risk — capitalised asset that doesn’t meet §57 criterion 3
Reserved-instance cloud commitments expensed upfront	Bookkeeper treats full commitment as period cost	Year 1 P&L overstated; mismatch with usage; prepaid asset balance missing
Staff training on AI tooling capitalised	Finance team includes training in “AI implementation cost”	Audit risk — §69(b) explicit exclusion violated
Configuration/customisation paid to SaaS vendor capitalised	Finance team treats one-time setup as asset creation	Audit risk — no control over the underlying SaaS

The AI Spend & Tax Optimisation methodology sub-domain 4 covers the reconciliation discipline + the deferred-tax implications.

The SEBI/MCA January 2026 AI disclosure mandate angle

The mandate effective for listed entities from January 2026 requires disclosure of material AI use including model documentation, audit-evidence sufficiency, and governance framework. The statutory auditor for FY 2026-27 cycles will explicitly test:

• Whether AI software in financial statements is classified consistent with Ind AS 38 control + §57 criteria
• Whether SaaS / API consumption is correctly opex-classified (not erroneously capitalised)
• Whether internally-developed AI capability meeting §57 is correctly capitalised (not erroneously expensed → loss of Section 32 claim)
• Whether deferred tax under Ind AS 12 correctly reflects book vs tax timing differences
• Whether multi-year prepayments are correctly classified as prepaid asset (not opex in year 1)

For listed entities preparing for FY 2026-27 statutory audit, AI spend classification review under Ind AS 38 + Section 32 is part of the audit-readiness work — best done before the FY closes, not at audit.

Practical Indian context

Most Indian mid-market companies (₹50-2,000 crore revenue) consume AI primarily via:

1. Foundation model APIs (OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock) — opex
2. SaaS tools with embedded AI (Notion AI, Slack AI, GitHub Copilot, etc.) — opex
3. Custom integrations and prompts built on the above — opex (no controllable asset)
4. Fine-tuning on third-party APIs — opex (criterion 3 fails)

In this profile, virtually all AI spend is opex. The capex question arises only for:

• Entities purchasing perpetual on-premise software (rare for AI)
• Entities building genuinely internal AI capability on open-weight models hosted on their own infrastructure (rare; mostly TCS / Infosys / HCLTech-scale entities investing in proprietary platforms)
• Entities making multi-year prepayments to cloud or SaaS vendors (treat as prepaid asset, not capex)

The simplifying default for Indian mid-market is: opex unless explicitly demonstrating Ind AS 38 §57 satisfaction in writing.

Cross-references

• AI Spend & Tax Optimisation methodology — sub-domain 4 covers Ind AS 38 + Section 32 in full
• Section 195 TDS on foreign AI vendors — relevant where capitalised/expensed AI software is sourced from foreign vendors
• RCM on foreign SaaS — 18% IGST RCM applies regardless of capex/opex classification
• AI cost allocation — cost allocation logic affects which entity records the expense or asset