The AI Epoch: Where Does Value Reside? Asset Definition Shifts from "Object" to "Executable Boundary

The core of the traditional asset definition, as articulated by International Accounting Standard IAS 38, defines an asset as "a resource controlled by the entity as a result of past events and from which future economic benefits are expected to flow to the entity." This "resource ontology" assumes that an asset is primarily an identifiable "object" with intrinsic value. Control is equated with exclusivity and is primarily established through legal ownership. For instance, IAS 38 explicitly states that sufficient control over the future economic benefits arising from a skilled staff team and training is usually lacking, and therefore they do not constitute intangible assets; customer relationships are treated similarly, highlighting the limitations of the old paradigm for non-physical assets.

Furthermore, measurability, a crucial threshold for asset recognition, often requires that costs can be reliably measured. This leads to a systematic bias: many internally generated "true assets" vital to an organization (e.g., brands, processes, customer relationships) are expensed due to the difficulty of reliably measuring their costs, whereas externally purchased ones are easily capitalized. Separability/tradability is considered central to "identifiability," and liquidity serves as a proxy for value. Simultaneously, the old paradigm strictly distinguishes between stock (assets) and flow (services), treating "access rights" as expenses rather than assets. For example, the IFRS Interpretations Committee ruled in March 2019 that "access rights" to cloud-based software do not constitute intangible software assets. Ultimately, value is simplified to the discounted future cash flows and scarcity rents derived from scarce objects. Whether based on labor theory of value, marginal utility theory, or DCF models, all revolve around this core logic, jointly forming the foundation of tax, accounting, and corporate law.

The impact of AI on the traditional asset paradigm is first manifested in its near-free generative capabilities and non-rivalrous replication, completely removing the premise of "scarce objects." When content, code, and model outputs can be infinitely replicated at zero marginal cost, the output itself is no longer scarce, undermining the basis of asset exclusivity in the traditional sense. DeepSeek's public release of its R1 model weights in January 2025, with performance comparable to top inference models, directly challenges the old notion that "owning weights equals owning an asset." This implies that the most valuable objects may not be owned, or may even be open-sourced, compelling value to migrate to other layers such as compute supply, data pipelines, distribution channels, evaluation systems, or compliance and liability.

Secondly, the migration of data value and the increasing prominence of accounting's "asset blind spots" are evident. Data itself is difficult to make exclusive, has extremely low replication costs, and its value is highly dependent on context and combination during use. Data without boundaries is closer to a liability (storage, compliance, privacy risks) than an asset. More concerningly, for even the smallest organizations in the AI era, true accumulation lies in reusable workflows, repeatable evaluators (evals), transferable memory, and auditable provenance. However, these "process capitals" are precisely excluded from the scope of assets by accounting standards like IAS 38, creating systemic "asset blind spots.

Furthermore, active systems (agents) introduce new depreciation logic and sources of value. Agents act both as capital and labor; their value requires continuous operation, monitoring, and re-validation to be maintained. Model upgrades, data drift, or interface changes can quickly render them obsolete, making linear depreciation inappropriate. When output is no longer scarce, true scarcity migrates to "verification, provenance, and liability," and the value of an asset increasingly equates to the value of its surrounding "executable trustworthy boundaries." Concurrently, Birch and Muniesa, in "Assetization," point out that "assetization" itself is a politico-technical process, not a natural fact. Tokenization (e.g., the BlackRock BUIDL fund surpassing $2 billion by the end of 2025) pushes this process to its extreme, meaning that for the first time, asset boundaries can be written as code, automatically executed, and combined, where the "creation" of assets replaces "discovery.

Under the new paradigm, the ontological definition of an asset is reconstructed as: "a set of executable boundaries established around a flow, comprising an exclusive boundary, an independent verifier, and an attributable liability-holder." Its value equals "the potential of the flow × the executability of the boundary × the credibility of verification × the degree of liability attribution." This means that while the old paradigm focused on "Is this my resource?", the new paradigm asks, "Can I establish exclusivity, verification, and assume responsibility for this flow?" The four constituent elements of assets are thus rewritten: control is no longer limited to legal ownership but becomes diverse "executable exclusivity"; measurement shifts from "reliable cost measurement" to "verifiability"; separability transforms into "orchestratable/licensable," with access rights becoming the primary form; and scarcity migrates from the object itself to complementary control layers that augment verification, distribution, trust, and responsibility.

Asset categories will also expand to include competitive physical capital (e.g., GPUs), non-rivalrous cognitive capital (e.g., weights, data), and increasingly important process capital (workflows, evaluators, SOPs), relational/trust capital (reputation, delivery records), and active capital (agent systems). Among these, non-rivalrous cognitive capital only becomes an asset when coupled with a control layer; otherwise, it is a public good or a liability. Valuation metrics will shift from traditional "cost/discounted cash flow" to "capability × reuse rate × trustworthiness × liability premium." Markets will increasingly adopt metrics such as revenue-per-employee, capability assessments, and cost-per-task. Depreciation logic will also be rewritten, transitioning from linear amortization to a superposition of "capability depreciation" (replacement), "drift depreciation" (model updates, data distribution changes), and "validation depreciation" (verifier expiry), where an asset's "useful life" is determined by its update cadence.

Ultimately, "assetization" is viewed as an explicit and programmable process, rather than a mere "discovery." AI enables evaluation standards to define boundaries, provenance to define property rights, agents to be the asset ontology, and contracts to be execution. This implies that the opportunity for research OPCs lies in "designing boundaries that others have not yet conceived," rather than merely searching for undervalued assets. This series of fundamental shifts collectively forms the core judgment of a new asset paradigm revolution, moving from traditional resource ontology to one based on "flows" and "boundaries.

On one hand, growing evidence points to a drastic change in asset definition. International Accounting Standard IAS 38 demonstrates its limitations in several areas, such as not recognizing employee skills and processes as assets. Yet, in the actual market, OpenAI achieved a valuation of approximately $500 billion in October 2025 through a secondary share sale of about $6.6 billion, far exceeding what its tangible book assets could explain, underscoring the dominant role of intangible assets in the modern economy (Ocean Tomo data shows intangible assets accounting for approximately 90% of the S&P 500's market capitalization, although the exact figures and methodologies are debated). Furthermore, BlackRock's tokenized money market fund BUIDL surpassed $2 billion in assets and accumulated over $100 million in dividends by the end of 2025, indicating that asset boundaries are evolving towards programmability and automatic execution, a stark contrast to the traditional "object"-based form of assets.

However, we must be wary of simplifying the "revolution" into mere technological optimism. Significant counter-examples and institutional inertia persist. Nvidia became the first company to surpass a $5 trillion market capitalization in October 2025, proving that competitive physical capital (chips, manufacturing, energy) remains a source of immense value, rendering any assertion that "physical assets are no longer important" untenable. Concurrently, the institutional revolution is not yet complete: the IASB's intangible assets project was still in the "gathering feedback, no decision made" stage as of February 2025, and accounting standards, tax laws, and legal definitions of "what constitutes an asset" have not changed. This implies that while a conceptual shift has occurred, institutional recognition is still underway, and the time lag between the two presents both opportunities and risks.

At the macroeconomic level, there are also unverified complexities. Acemoglu's (2024) estimates suggest that AI's boost to ten-year TFP might not exceed 0.71% and could widen the capital-labor gap. If macro-productivity does not significantly improve, the "asset redefinition" might be more a capital market narrative than a universal statistical fact. Furthermore, Michael Burry's questioning of depreciation assumptions for large tech companies reveals that AI assets might be systematically underestimated in terms of depreciation, with their lifespan potentially much shorter than accounted for. The emergence of open-source models (like DeepSeek R1) also demonstrates the invalidity of "the model itself is an asset" in certain contexts. These counter-examples collectively remind us that the interplay and tension between old and new paradigms constitute a more realistic picture of the current asset landscape.

At the business level, the logic of valuation and moat building is shifting from "what one owns" to "what one can establish boundaries around." When profits from information and capability replication approach zero, the value pool migrates to "control layers" such as verification, evaluation, provenance, compliance, liability assurance, and distribution. This means that enterprise revenue pricing models will shift from traditional "per seat/per token" to "per outcome/per unit value." Capital structure decisions face new considerations: "build (capex)" versus "rent (opex)" is no longer a simple cost trade-off but concerns the ability to accumulate truly moated "process capital" and "trust capital" while mitigating high depreciation risks.

At the organizational level, enterprises are increasingly evolving into "containers of process capital and trust capital," rather than merely collections of physical assets. This aligns with the perspective of Coase's rewritten boundaries in FR-12: whether to outsource depends on "whether external capabilities can be owned" and "whether internal verification and liability must be proprietary." A sharp paradox is that an organization's most core assets (e.g., team skills, internal processes, customer relationships) are precisely not recognized as assets under existing accounting standards. Therefore, organizations must, through institutional design such as extensive documentation, building evaluators, setting permissions, signing contracts, and establishing memory systems, "partially assetize" these intangible elements that cannot be fully legally demarcated, to solidify their value and build internal moats.

At the product level, AI-era products are evolving from mere "feature packages" to "verifiable capability packages." Product design requires built-in evaluation mechanisms, observable costs, transparent provenance and assumptions, and the ability to retract failures. This means that evals (evaluations), provenance (source records), and audit logs are no longer auxiliary functions but rather "first-class citizens" for third-party recognition of product value, forming crucial components of the product's own "boundaries." This shift highly aligns with the underlying logic of "generator/verifier separation" in FR-11, emphasizing the central role of verification mechanisms throughout the product lifecycle. For research OPCs themselves, the true path to asset accumulation lies in process capital, trust capital, active capital, and boundary design capabilities, rather than simple possession of models and data, which aligns with the themes of FR-09 (knowledge distribution) and FR-10 (talent definition).

For research OPCs, clarifying and quantifying their "control layer assets" is crucial. First, an "Assetization Register" should be established. For each significant output or capability, it should be registered using the quartet "Flow → Exclusive Boundary → Verifier → Liability Holder," and its category among the five asset types (competitive physical/non-rivalrous cognitive/process/trust/active) should be marked. If any element in the quartet is missing, the item should not be considered an asset but rather an expense or public good. Secondly, when allocating resources or deciding to outsource, an "assetization test" should be conducted: for any candidate asset, question its exclusivity, verifiability, orchestratability/licensability, the presence of complementary control layers, and whether there is a clear liability holder. For items weak in all five, long-term accumulation investment should be avoided; for those strong in the control layer, self-building should be prioritized to counter the old paradigm's misjudgment of "data/model as asset.

Treating verification and provenance as core assets to manage is an urgent priority. Every deliverable should be accompanied by reproducible validation scripts, a complete list of sources and assumptions, failure/retraction records, and version and liability signatures. This not only embodies the value boundaries of the new conceptual paradigm but also serves as a hard threshold for meeting increasingly stringent compliance requirements and market trust pressures. For example, the SEC has enforced "AI value claims," where unverified claims can lead to legal liability. Concurrently, research OPCs must strictly differentiate between "rented capabilities" and "proprietary processes." Models, computing power, external APIs, etc., should be explicitly listed as rental costs, not proprietary assets; whereas workflows, evaluators, memory, and provenance ledgers should be considered proprietary process capital, managed with version control, made transferable and reproducible, and their reuse rate and verifiability updated regularly.

To address the risks introduced by new asset forms, research OPCs should establish a "capability depreciation ledger" to track the replacement speed, interface changes, and data drift of models, tools, and evaluators used. Each critical capability should be tagged with an "update-driven lifespan" to hedge against concentrated risks. Betting on a single frontier model is equivalent to holding a highly depreciating asset; therefore, maintaining model substitutability and locking value into the process and trust layers rather than the weight layer is crucial. Furthermore, external narratives should only use "verifiable facts," actively avoiding "AI-washing" risks, and replacing vague "AI empowerment" labels with evaluation results, reproducible case studies, and unit economic data. Ultimately, research OPCs should view "assetization" itself as a core design capability, actively designing authorization and licensing mechanisms, composable interfaces, call metering, provenance records, and reputation accumulation mechanisms to transform non-rivalrous information flows into capturable cash flows, and strive to develop a unique "boundary design handbook.

Please note that some data and views in this report are yet to be further verified. For instance, while Ocean Tomo's data stating "intangible assets account for 90% of the S&P 500's market capitalization" is widely cited, its precise percentage, year sequence, and methodology still require cross-referencing with original research, and the methodology itself is subject to public debate. The $2 billion size and $100 million dividend data for the BlackRock BUIDL fund originate from a single CoinDesk report; it is recommended to cross-verify this with official data from BlackRock or Securitize. While DeepSeek R1's "public weights" have been confirmed by multiple sources, the specific open-source licensing terms (e.g., whether it's an MIT license) still need to be checked against the original model card. Furthermore, Michael Burry's allegations regarding the depreciation deficit of large tech companies are currently an investor's perspective rather than an audited conclusion, but the issue he raises—that AI asset lifespans may be overestimated—warrants continued attention and verification.

This report emphasizes that the IASB's intangible assets project is still in its early stages, gathering feedback as of February 2025, with no final decisions made. Therefore, any statement claiming "accounting standards have changed" is inaccurate. There is a significant time lag between conceptual shifts and institutional recognition, which may present opportunities for early movers but also carries risks associated with institutional inertia. Finally, a core risk warning: treating "data/model as an asset" as an established fact, rather than critically examining it and designing boundaries around it, could lead to significant misallocation of resources and missed opportunities for genuine value accumulation.