New Appropriability Playbooks in GenAI Era: Rethinking how Moats, AI Stack and Models work together

Patents and classic IP still matter—but in foundation‑model markets, the durable moats look different: weights secrecy, tacit know‑how, compute access, safety tooling, eval/benchmark control, and distribution. Leaders should update their appropriability playbooks accordingly.

Let’s understand models first…

As of May 2025, ChatGPT had 400 million weekly active users, with 122 million daily users, including 1.5 million enterprise customers. OpenAI earned $3.7 billion from ChatGPT in 2024 and reached $10 billion annualized revenue by June 2025. In the U.S., 28% of employed adults now use ChatGPT at work, up from 8% in 2023, demonstrating rapid enterprise adoption. Reportedly, over 80% of Fortune 500 companies integrated ChatGPT within the first nine months of its release. Among professional users, adoption is strong: 64% of journalists, 63% of software developers, and 65% of marketers regularly use ChatGPT.

Similarly, as of June 2025, Gemini Pro is used by 27 million enterprise users globally and over 52 million individual users. Enterprise usage accounts for 63% of total usage. As of late March 2025, Gemini logged 350 million monthly active users and 35 million daily active users, while presented data shows ChatGPT at 600 million MAU and Meta AI at 800 million MAU. Recent data also shows the Gemini app surpassed 450 million monthly active users, with daily usage growing by over 50% quarter-over-quarter.

Having said that, these models offer different degrees of openness—where openness is measured by how transparent a model is, i.e. making its source code, parameter weights, documentation, training process, etc., available for other developers to replicate and run on different compute resources as stand-alone models. OpenAI’s models are closed—where source and weights are proprietary, while Meta’s LLaMa has partial openness. Yet, LLaMA still doesn’t meet all the 10 criteria to be an “open source” model.

What does this mean for businesses?

Two key dynamics emerge in the GenAI area:

1. Patents’ Declining Centrality for Foundation Models

Patents often fall short at protecting core advances in foundation models. Instead, firms rely on alternative mechanisms of appropriation:

• Trade secrets and tacit knowledge: Foundation models are complex to build, but once model weights and documentation are available, others with sufficient compute can replicate or deploy them. Thus, weights, training datasets, and documentation around preprocessing become critical differentiators. Access to these resources allows more actors to fine-tune and adapt models for novel applications. Developers themselves are also high-value assets: the tacit knowledge embodied in experienced engineers is indispensable. Research from Stanford’s Foundation Model Transparency Index shows modest improvement in disclosure, but overall transparency remains low (around 60%), suggesting that full openness in AI development is unlikely.

• Control over model weights and fine-tuning pipelines: Companies pursue different “openness paths” depending on their strategy. Closed models like OpenAI’s ChatGPT restrict access to weights and internals, ensuring functionality remains under developer control. This limits scrutiny and customization by downstream actors, potentially reinforcing embedded biases or privileging certain interests. By contrast, Meta’s LLaMA models release weights and documentation under controlled licenses, which expand opportunities for fine-tuning while still constraining redistribution and commercialization under stipulated terms.

• Proprietary data orchestration and domain-specific tuning: The way firms collect, clean, and structure proprietary data—and then fine-tune foundation models for specific verticals—has become a central appropriation strategy. Examples include BloombergGPT, which leverages curated financial text corpora, Med-PaLM (Google DeepMind), tuned on clinical datasets, and Cohere for AI, which emphasizes multilingual and enterprise-specific applications. These efforts illustrate how orchestration and vertical tuning transform general-purpose models into defensible, domain-specific assets.

2. Complementary Assets Shape Value Capture

In the GenAI economy, profits accrue less to model inventors than to those who control the assets that make models usable, scalable, and trustworthy. These are the complementary assets that turn raw innovation into commercial advantage:

• Compute stacks. Training and deploying large models requires advanced hardware and high-performance computing (HPC) infrastructure. Leaders like NVIDIA capture disproportionate value by controlling GPU supply, while hyperscalers (AWS, Microsoft Azure, Google Cloud) leverage massive data centers to bundle compute with storage and orchestration. Public compute initiatives (e.g., NAIRR in the U.S., PAICE in Canada, EuroHPC’s AI Factories) attempt to level the playing field but remain dwarfed by private fleets.

• Distribution platforms. Control over user access points—APIs, app stores, and integrations into productivity ecosystems—shapes adoption. OpenAI’s ChatGPT API and plugin ecosystem, Microsoft’s Copilot integrations across Office, and Adobe’s Firefly within Creative Cloud demonstrate how embedding AI into established workflows secures both revenue and lock-in.

• Governance and safety infrastructure. Standards and tooling for evaluation, alignment, and compliance define “acceptable” AI. Firms like Anthropic market safety as a differentiator, while OpenAI’s eval suites and Stanford’s Foundation Model Transparency Index influence policy debates. Whoever controls benchmarks and reporting frameworks gains soft power in shaping markets and regulation.

Because these assets require scale, capital, and institutional credibility, large incumbents are best positioned to dominate. But that does not leave smaller firms without options:

• Startups can specialize in verticals where proprietary data and domain knowledge matter more than scale—for example, BloombergGPT in finance or Med-PaLM in healthcare.

• Niche players can build “picks and shovels” around the ecosystem—tools for fine-tuning, compliance auditing, or lightweight inference—that piggyback on foundation models without competing head-on.

• Academic labs and SMEs can leverage fractionalized public infrastructure (e.g., PAICE in Canada, NAIRR in the U.S.) for experimentation and early proofs-of-concept before migrating to commercial clouds for scale.

Takeaway: In GenAI, control of the stack around the model—compute, distribution, governance—often matters more than the model itself. For smaller firms, the path to value capture lies in targeted differentiation, partnerships, and building complementary capabilities that incumbents either overlook or cannot efficiently deliver.

Practical insights on existing moats:

The current GenAI landscape shows how firms construct moats less through patents and more through opacity, data control, and governance infrastructure.

• Opacity as a feature: Developers reveal far less than the public or regulators would prefer. Even with modest progress tracked by Stanford’s Foundation Model Transparency Index (FMTI), transparency remains partial, with most leading models scoring under 60% on disclosure. Limited visibility into model architecture, training data, and safety practices means incumbents preserve control while downstream actors face information asymmetry.

• Data rights in flux: Litigation over training data—such as New York Times v. OpenAI/Microsoft—and the U.S. Copyright Office’s 2025 report on generative AI complicate the use of copyrighted material. This raises compliance costs for firms that rely on “grey area” web scrapes, while reinforcing the value of clean, documented corpora with clear licensing and provenance. As regulatory uncertainty grows, curated proprietary datasets become both a moat and a bargaining chip.

• Benchmarks and safety tooling: Control over evaluation metrics and safety infrastructure is another lever of influence. Proprietary benchmark suites (e.g., OpenAI’s evals) and alignment pipelines (RLHF/RLAIF) tilt purchasing decisions and shape what regulators consider “safe” or “high-quality” AI. Initiatives like Stanford CRFM’s FMTI push for standardized disclosures, but until such standards are widely enforced, the companies defining benchmarks wield significant market-shaping power.

  
  

Leadership playbook:

In this environment, leaders must update their appropriability strategies to reflect how moats are actually built in GenAI:

• Map your moat to the stack, not just the model: Sustainable defensibility comes from controlling the surrounding assets—exclusive or verified data rights, repeatable fine-tuning and alignment pipelines, evaluation and monitoring systems, and at least one distribution channel where you own user access.

• Adopt selective openness: Firms can use openness as a weapon: release model weights, APIs, or documentation tactically to attract developer ecosystems, while keeping high-leverage assets (compute, governance, data pipelines) proprietary. Meta’s LLaMA is an instructive example—positioned as “open” but bounded by restrictive licensing that shapes downstream usage.

• Prepare compliance as a strategy: In regulated industries, compliance is not just a box to tick—it is a differentiator. Treat provenance tracking, consent management, and red-teaming not as costs but as profit centers. The firms that make compliance seamless for customers will win conservative buyers (e.g., in finance, healthcare, or government), where trust often outweighs raw performance.