As artificial intelligence changes how companies compete, accountants may increasingly confront an unusual valuation problem: some of the information creating a company’s competitive advantage may be economically significant without appearing as a separate asset on its balance sheet.

For much of the AI boom, attention has focused on models, computing power, and algorithms. But as sophisticated AI models become more widely accessible, an increasingly important differentiator may be the proprietary data companies have accumulated over decades.

Recent research by Dwarkesh Patel and Jerry Han illustrates why. Comparing AI advances from 2019 to 2025, they estimated a 12x compute-efficiency gain from data improvements versus a 3.7x gain from model improvements at the compute scale they tested.¹ While the researchers caution against extrapolating too broadly, the finding raises an important question: What if one of a company’s most valuable AI assets is not the model but the data?

A Moonshots discussion with Peter Diamandis put the idea more provocatively: “The next 10x may be sitting in the data” already inside an organization.²

When Everyone Has AI, What Remains Proprietary?

Companies can increasingly access powerful foundation models rather than develop them from scratch. If competitors can license similar AI capabilities, simply “having AI” is unlikely to create a durable advantage. The differentiator may instead be what each company can teach those models.

Consider a manufacturer with decades of equipment failures, warranty claims, and maintenance records; an insurer with claims and behavioral data; or a financial institution with millions of historical transactions and customer interactions. Such datasets can contain knowledge that competitors cannot easily purchase or recreate.

McKinsey describes this concept as “privileged data,” which is information that becomes a competitive moat when it enables AI outcomes competitors cannot readily reproduce.³ The advantage can compound as product usage generates additional proprietary information that further improves the AI.

For accountants, however, the relevant question is more than whether a company has a large volume of data. Instead, the most important question is whether that data represents an economically meaningful resource and how that value should be understood in financial reporting, valuation, and transaction analysis.

Not Every Dataset Is a Data Moat

A company can own petabytes of information with little competitive significance. Accountants and valuation professionals can assess potentially valuable proprietary data across six dimensions:

• Exclusivity: Can competitors obtain substantially similar information elsewhere?
• Relevance: Does the data materially improve a commercially important prediction, decision, product, or workflow?
• History and depth: Has the dataset accumulated over enough time or transactions that recreating it would be costly or impractical?
• Rights and control: Does the company have the legal and contractual ability to use the information for AI, analytics, or commercial purposes?
• AI readiness: Is the information sufficiently accurate, structured, labeled, and accessible to produce useful outcomes?
• Learning effects: Does use of the company’s products continuously generate new proprietary information that makes those products better?

These factors help distinguish ordinary operational data from information capable of producing future economic benefits.

Why This Matters to Accountants

The accounting treatment creates an important paradox. Under U.S. Generally Accepted Accounting Principles, costs associated with many internally developed intangible assets are generally expensed as incurred unless specific guidance permits capitalization. As a result, a company may spend years developing proprietary data through normal operations without separately recognizing its economic value on the balance sheet.

Acquisition accounting can produce a different result. Under Accounting Standards Codification 805, an acquirer recognizes identifiable intangible assets separately from goodwill when the applicable recognition criteria are met and measures them at acquisition-date fair value. Databases are specifically among the types of intangible assets that can qualify for separate recognition.

That means accountants may encounter proprietary data in several contexts, including purchase price allocations, fair value measurements, M&A diligence, impairment analyses, management forecasts, and disclosures around intangible-asset disclosures.

The distinction also matters when evaluating company performance. Better proprietary data may allow AI to forecast demand more accurately, price risk more effectively, identify equipment failures earlier, or automate specialized work. Those improvements can ultimately affect revenue growth, margins, customer retention, pricing power, and the durability of future cash flows. Those are the same assumptions that often underpin valuations and accounting estimates.

There is also a controls dimension. As companies use proprietary datasets to drive AI-enabled decisions, accountants may increasingly need to understand whether the underlying information is complete, accurate, appropriately governed, and legally usable. If AI-generated outputs feed financial forecasts, estimates, or reporting processes, data quality and lineage are no longer purely technology issues.

The Asset the Balance Sheet May Not Show

The Financial Accounting Standards Board has an active research project considering potential improvements to the accounting and disclosure of intangible assets, including internally developed intangibles.⁴

AI may make the gap between accounting recognition and economic value increasingly visible. Accountants have long dealt with businesses whose brands, customer relationships, intellectual property, and human capital contribute significant economic value beyond what is directly reflected on the balance sheet. Proprietary data may increasingly belong in that discussion.

The question for accountants is therefore broader than “How much data does this company have?” Instead, it is “What proprietary information does the company possess, what economic benefits does it generate, and how should those benefits affect the judgments accountants make about valuation, transactions, controls and financial reporting?”

As AI models become more widely accessible, the model itself may become less differentiating while the proprietary knowledge supplied to it becomes more important. Some of the most valuable assets of the AI economy may already exist inside companies even when they remain difficult to see on the balance sheet.

This article was originally published in Accounting Today. 

1. Dwarkesh Patel and Jerry Han, “Pretraining progress is mostly coming from data,” dwarkesh.com, September 8, 2026.
2. “Should we slow down AI progress? | MOONSHOTS #288,” Moonshots with Peter Diamandis, podcast, September 11, 2026.
3. Dago Diedrich, Evan Williams, Tanguy Catlin, and Tim Fountaine, “From AI table stakes to AI advantage: Building competitive moats,” McKinsey Quarterly, May 15, 2026.
4. “Accounting for and Disclosure of Intangibles,” Objectives of Research Projects, webpage, Financial Accounting Standards Board.