IAS 38 Capitalization Rules Classify Non-Deterministic Pre-Training as Research, Forcing Foundation Model Compute Directly into Current-Year Operating Expenses

When a traditional technology company writes a million lines of deterministic code, the accounting department treats the engineering salaries as an asset, capitalizing the cost over years. Foundation model developers expected the same treatment for the massive compute clusters used to train artificial intelligence. The difference is that traditional code compiles predictably, while a neural network's pre-training run is an expensive statistical experiment with no guaranteed outcome.[3]

That single technical distinction has triggered a collision between modern machine learning and the International Accounting Standards Board. Under the rules governing intangible assets, specifically IAS 38, the billions of dollars spent on graphics processing units cannot be treated as a long-term investment. Instead, auditors are forcing those costs directly onto the current-year income statement.[7]

The financial mechanics alter the apparent profitability of the entire artificial intelligence sector. Capitalizing a $100 million training run spreads the expense over a five-year useful life, hitting the ledger at $20 million annually. Expensing it immediately wipes $100 million from the current year's bottom line, turning otherwise profitable enterprises deeply negative.[2]

This accounting reality stems from how international standards define the boundary between exploring an idea and building a product. The framework was designed for pharmaceuticals and traditional software, where the transition from research to development is clearly marked by a successful prototype or a chemical patent.[1]

For artificial intelligence, that boundary is fundamentally blurred by the mechanics of deep learning. A model does not exist in a testable state until the vast majority of the computational budget has already been consumed, leaving accountants with no prototype to evaluate.[8]

The IAS 38 Divide: Research vs. Development

The International Financial Reporting Standards dictate how companies account for internally generated intangible assets. IAS 38 strictly separates the creation process into a research phase and a development phase. Costs incurred during research must always be recognized as an expense when they are incurred, with no exceptions.[7]

To move costs into the development phase—and thus capitalize them as an asset on the balance sheet—a project must meet six strict criteria. The most formidable hurdle is demonstrating the technical feasibility of completing the intangible asset so that it will be available for use or sale.[4]

In traditional software engineering, technical feasibility is established early. Once the architecture is defined and the core algorithms are proven, writing the remaining modules is a matter of execution. The company can reliably demonstrate that the software will function as intended before the bulk of the coding budget is spent.[2]

As PKF Littlejohn outlines in their assessment of the standard, the burden of proof rests entirely on the developer. "An entity must be able to demonstrate the technical feasibility of completing the intangible asset," the firm notes, adding that this requires showing how the asset will generate probable future economic benefits.[4]

If a company cannot prove that the software will work, the accountants must assume it is still an experiment. Until all six criteria are met simultaneously, every dollar spent on salaries, cloud computing, and data acquisition remains an operating expense.[5]

The Non-Deterministic Nature of Pre-Training

Foundation models fail the technical feasibility test because their creation is inherently non-deterministic. When engineers initialize a neural network with billions of parameters and begin feeding it terabytes of text, they cannot guarantee what capabilities will emerge. The model might achieve human-level reasoning, or it might suffer from catastrophic forgetting and collapse entirely.[8]

This uncertainty persists throughout the pre-training phase, which consumes roughly 80 percent of the total compute budget. A training run can take three months and cost upwards of $60 million in server time, during which the model's final performance remains unknown.[8]

KPMG's analysis of AI software investments highlights this exact friction point. Because the outcome of training a novel algorithm on a massive dataset is unpredictable, the process aligns perfectly with the IAS 38 definition of research: original and planned investigation undertaken with the prospect of gaining new scientific or technical knowledge.[1]

You cannot demonstrate the technical feasibility of a statistical black box until you open it. By the time the pre-training run concludes and engineers can finally verify that the model works, the money has already been spent. The costs were incurred during the research phase, permanently locking them out of capitalization.[6]

This creates a paradox for artificial intelligence startups. The most expensive part of building their core product is legally classified as an exploratory science project, even when the engineering team is highly confident in the underlying transformer architecture.[3]

The Compute Cost Collision

The scale of this accounting classification is unprecedented because of the hardware required for modern machine learning. Traditional software research involves paying a small team of engineers to build prototypes over a few months. AI research involves renting tens of thousands of specialized processors running at maximum capacity.[8]

A frontier model in 2026 requires an estimated 10^25 floating-point operations to pre-train. Securing that compute capacity costs hundreds of millions of dollars, representing the single largest cash outflow for an artificial intelligence developer.[8]

Forcing these massive figures through the current-year income statement destroys short-term profitability metrics. A company generating $50 million in genuine software-as-a-service revenue looks wildly successful until a $120 million pre-training expense is deducted in a single quarter, resulting in a massive net loss.[2]

The sheer volume of capital required means that a single failed training run can devastate a company's financial standing. Because the expense is recognized immediately, there is no mechanism to smooth the impact of a model that suffers from mode collapse and has to be discarded.[1]

Global FinX notes that this dynamic is particularly challenging for the IT sector, where software capitalization has historically shielded profit margins. Investors accustomed to the smooth, amortized expenses of the cloud computing era are now confronting the volatile, cash-heavy reality of foundation model development.[2]

The inability to capitalize pre-training compute also affects corporate valuations. Because the trained model cannot be recognized as an asset on the balance sheet, the company's book value remains artificially low, complicating debt financing and forcing reliance on venture capital.[3]

Where Capitalization Finally Begins

The accounting treatment shifts dramatically once the base model is complete. After pre-training yields a functional, testable neural network, the project finally crosses the technical feasibility threshold required by IAS 38. The research phase ends, and the development phase begins.[5]

Subsequent processes, such as reinforcement learning from human feedback and supervised fine-tuning, are highly deterministic. Engineers know exactly what the base model is capable of, and they are simply refining its behavior to follow instructions or serve a specific commercial application.[8]

The costs incurred during this refinement stage—including human annotator salaries, specialized fine-tuning compute, and application interface development—can be safely capitalized. Opagio's guidance on AI investments confirms that once a clear path to commercialization is established, these subsequent expenditures qualify as intangible assets.[6]

However, this provides little relief for the developers of frontier models. Fine-tuning typically consumes less than 20 percent of the total computational budget. The vast majority of the financial burden remains stranded in the expensed pre-training phase.[8]

This distinction creates a stark divide between companies building foundation models and those building applications on top of them. An enterprise software firm fine-tuning an open-source model for legal document review can capitalize almost its entire engineering budget, preserving its profit margins.[1]

Strategic Implications for AI Builders

The strict application of IAS 38 is quietly reshaping the structure of the artificial intelligence industry. Because pre-training costs must be expensed immediately, only the largest technology conglomerates can absorb the multi-billion-dollar hits to their income statements without triggering a shareholder revolt.[8]

A startup attempting to train a frontier model from scratch faces an impossible financial narrative. Even if they secure the necessary capital, their audited financial statements will show catastrophic losses year after year, making an initial public offering nearly impossible under traditional valuation metrics.[3]

Venture capitalists are increasingly aware of this dynamic, structuring their investments to account for the immediate P&L destruction. They understand that a massive operating loss in the first year is not a sign of operational failure, but rather a strict artifact of international accounting standards.[2]

This accounting reality accelerates the trend toward industry consolidation. Smaller developers are incentivized to abandon pre-training entirely, pivoting instead to fine-tuning existing open-weight models where the costs are lower and the accounting treatment is far more favorable.[2]

Token Capex argues that IAS 38 remains the definitive framework for internally generated AI intangibles, despite calls for modernization. The International Accounting Standards Board has shown little appetite for rewriting the rules simply because machine learning requires expensive hardware.[7]

Until the underlying mechanics of neural network training become predictable enough to guarantee success before the compute is spent, the financial reporting will reflect the engineering reality. Building a foundation model is not software development; it is an act of high-stakes scientific research, and the balance sheet will treat it exactly as such.[8]