When acquiring becomes politically costly, incumbents increasingly hollow startups instead: license-and-hire ‘absorption’ emerges as a distinct quasi-acquisition that internalizes tacit AI capability without formal takeover. The model shows no single sourcing route dominates universally—outcomes hinge on contractibility, timing of a dominant design, and how much incumbents’ legacy products are exposed to substitution.

Main Finding

Which sourcing route (build, partner, acquire, absorb, or wait) wins in the generative-AI transition is shaped by contractibility, regulatory cost of integration, dominant-design timing, and substitution exposure. In particular, a license‑and‑hire “absorb” form (quasi‑acquisition) emerges endogenously as the favored way to internalize tacit AI capability when acquisition becomes costly: it transfers capability and talent without buying the whole firm, and it is structured to stay below merger review. When acquisition scrutiny rises, firms shift from acquisitions toward absorption (not toward partnering), and the total volume of owned capability transferred (acquire + absorb) remains roughly conserved. A one‑time fall in building costs revives building only if it arrives before a dominant design hardens. Waiting is relatively safe for firms whose legacy products are not substitutable by the frontier technology, but firms with high exposure lose substantial value (though usually not solvency).

Key Points

• Population-level dynamics matter: each incumbent’s choice reshapes partner and target pools and the regulatory price others face, so sourcing is ecological, not a single‑firm static decision.
• Contractibility is decisive (P1): where capability can be contracted (API, codified), firms mostly partner; where capability is tacit (team-bound), they tend to absorb or acquire.
• Absorb as a distinct mechanism (P2): license‑and‑hire deals internalize tacit capability economically and increase in attractiveness when conventional acquisition costs rise due to market concentration and regulatory scrutiny. Absorb can emerge endogenously from the model.
• Conserved internalization (P2): as acquisition becomes harder/costlier, firms shift to absorption such that the total owned capability transferred remains approximately conserved; if absorption is blocked, the same friction pushes firms toward partnering.
• Timing matters for building (P3): a sudden, population‑wide fall in building cost (e.g., open weights) increases building uptake only if it arrives before a dominant design locks in; after lock‑in it has much smaller effect.
• Exposure determines the cost of waiting (P4): waiting erodes value for firms whose offerings are directly substitutable by the new AI frontier; firms with low exposure (or those that have already partnered in) survive waiting with value losses rather than insolvency.
• Policy experiments: raising acquisition scrutiny tilts behavior toward absorption but does not create it; taxing absorption similarly slows it only slightly, because the form is effective at avoiding formal merger review.
• Empirical grounding: the model reproduces the documented 2022–2026 sequence (partnerships first, absorption/quasi-acquisition wave in 2024–25, acquisitions alongside) and places real deals plausibly on the simulated trajectories.

Data & Methods

• Modeling approach: history‑friendly agent‑based model (ABM) built from a general "transition" core (sense → screen → seize → transform → select loop for each firm) plus four independently switchable mechanisms specific to the AI episode. The architecture ensures that with all four switches off, the model reduces exactly to the general core.
• Core mechanics:
  • Firm position pos = q · sqrt(C), where q is a weighted index of firm attributes (knowledge, technology, adaptability, legitimacy) and C is capacity.
  • Market allocation follows a replicator/substitution dynamic with a sharpening exponent that increases as a dominant design forms (increasing returns).
  • Firms choose among routes using a bounded‑rational softmax/choice rule; heterogeneity is introduced through latent factor initialization.
• Four switchable mechanisms (each toggled on/off in experiments):
• Acquisition cost that rises with market concentration (emulates regulatory scrutiny/price of integration).
• Absorb route: license the target’s technology + hire its team, leaving a hollowed target (quasi‑acquisition).
• Open‑weight shock: one‑time, population‑wide drop in building cost (to model public release of model weights).
• Demand‑side substitution: maturing AI erodes legacy revenue of non‑adapters in proportion to product exposure.
• Experiments and identification:
  • Mechanism knockouts: switch each mechanism off to test whether headline patterns collapse or persist (causal attribution).
  • Timing experiments: introduce the open‑weight shock at different stages relative to dominant‑design formation.
  • Policy simulations: increase acquisition scrutiny or tax absorption-like transfers to test behavioral shifts.
• Validation and anchoring:
  • History‑friendly validation: model reproduces qualitative sequence and patterns of 2022–2026 (e.g., partnerships first, absorb/quasi‑acquisitions next, concentration dynamics), without fitting to firm‑level parameters.
  • Real deals and timeline used to parameterize and ground scenarios (examples: Microsoft–OpenAI partnerships and restructurings, Databricks–MosaicML, Microsoft–Inflexion license/hire, Meta–Scale‑AI stake; DeepSeek open weights as cost shock; Chegg as an example of substitution impact).
• Output measures: route shares over time (build/partner/acquire/absorb/wait), market concentration (HHI), total owned capability internalized (acquire + absorb), firm values and exits, sensitivity across random seeds.

Implications for AI Economics

• For antitrust and regulation:
  • Merger review pressure can produce regulatory arbitrage: firms respond to higher acquisition costs by using quasi‑acquisitions (absorb) that are designed to avoid merger thresholds. Regulators relying only on conventional HSR/merger metrics may miss substantial capability consolidation.
  • Simple taxation or retroactive scrutiny of absorption‑style deals slows but does not eliminate them; enforcement needs to adapt (e.g., broaden review definitions, require disclosure of license‑and‑hire arrangements, adjust reportable thresholds or criteria to capture capability/talent transfers).
  • Policymakers should monitor not only share transfers via formal acquisitions but also capability/talent flows via hiring and licensing arrangements.
• For firm strategy:
  • Route choice should be driven by contractibility: prefer partnering where capability is codifiable and accessible via APIs; prefer absorption (or acquisition) when capability is tacit and team‑bound.
  • Building is most attractive early, before a dominant design locks in; firms that can build cheaply early may capture outsized returns, but late building faces high barriers.
  • Waiting can preserve solvency for low‑exposure firms but destroys value for high‑exposure firms; strategy should consider substitution exposure explicitly.
• For market structure and industry evolution:
  • Absorption can preserve the effective internalized capability volume even as acquisition activity is squeezed, implying that headline merger statistics may understate effective consolidation of frontier capabilities.
  • Rapid, population‑wide shocks (like open weights) can reconfigure strategic choices but their effectiveness depends on timing relative to path‑dependent lock‑in.
• For researchers:
  • The history‑friendly ABM approach is effective for counterfactual exploration in fast transitions with strong endogenous feedbacks (scarce partners/targets, fast consolidation, regulatory shocks). Mechanism knockouts are a practical way to attribute aggregate patterns to micro mechanisms when traditional identification is infeasible.

Limitations (as the authors note) - The model is not a point forecast or a fine‑tuned fit to individual firms; it is a theory‑building, generative exercise. - It does not model displacement of labor, nor does it embed language models inside agents; the simulated firms are abstract decision agents. - Results are qualitative and mechanism‑based; quantitative magnitudes depend on parameter choices and calibration choices anchored to documented events rather than micro‑estimation.

Reference - Liu, Y., Cao, S., & Liu, Z. (2026). Crossing into AI: When Incumbents Build, Partner, Acquire, Absorb, or Wait — A History‑Friendly Agent‑Based Model of the Generative‑AI Market Transition. DOI: https://doi.org/10.21203/rs.3.rs-10682104/v1