AI benchmarks are driving billion-dollar GenAI valuations

Generative AI companies are using benchmark scores to build powerful market narratives around large language models (LLMs), turning technical evaluations into tools that help justify massive investment and claims of frontier status, according to a new study.

Published in Big Data & Society, the study titled “Making GenAI valuable: Benchmarks, singularities, and the enrichment economy” argues that benchmarks now do more than compare model performance. They help make GenAI models appear exceptional, scientifically credible and closer to artificial general intelligence, giving them a market value that cannot be explained only by data extraction, platform power or future revenue expectations.

Benchmarks become valuation tools in the GenAI market

The study examines how benchmark scores have moved from a technical role in computer science into the center of the GenAI economy. Benchmarks are standardized tests and datasets used to evaluate AI systems. In the current market, they are also public signals used to rank models, shape investor confidence and present some systems as leaders in the race toward advanced AI.

The authors argue that this shift became clearer during recent controversies around frontier AI models. DeepSeek’s launch shook markets after the company presented its model as matching leading systems at lower cost. Meta faced criticism over Llama 4 after concerns emerged that its public performance was shaped by benchmark optimization. OpenAI, Google, Anthropic and other firms have also used benchmark comparisons to frame model releases as major advances.

These cases show that benchmarks do not simply measure AI performance. They help create market value by giving investors, developers and the public a way to compare systems that are otherwise hard to assess. Most outsiders cannot inspect proprietary model architectures, training data or internal testing. Benchmark charts and rankings fill that gap by making performance appear measurable, comparable and scientifically grounded.

Leading LLMs often share similar foundations. Many are built around transformer architecture, trained on massive datasets and refined through related methods. Their technical differences can be difficult to interpret, even for experts. Benchmarks make these differences visible. A small gain on a widely watched evaluation can be presented as evidence that one model family has moved ahead of another.

The study uses the concept of the enrichment economy to explain this process. In traditional enrichment markets, value is created by attaching narratives of rarity, authenticity or distinction to objects such as artworks, luxury goods or heritage items. The study applies this idea to GenAI, arguing that benchmarks enrich large language models by presenting them as singular, elite and non-standard objects.

Unlike heritage goods, GenAI models are not mainly enriched through stories about the past. They are enriched through stories about the future. Benchmark scores position models on a path toward artificial general intelligence, or AGI. That future-oriented story helps make present models appear more valuable, even when their long-term revenue potential remains uncertain and individual versions quickly lose their frontier status.

The authors argue that existing frameworks explain only part of this economy. Surveillance capitalism captures how AI firms depend on data. Platform capitalism explains infrastructure dominance and user lock-in. Assetisation explains attempts to turn models into future revenue streams. But none fully explains how particular models gain high value through public claims of exceptionality. Benchmarks provide that missing link by turning technical scores into stories of leadership, scarcity and future promise.

Scientific authority strengthens claims of AI progress

The study traces benchmark culture back to earlier computer science practices, where standardized evaluations were used to make AI research more credible, testable and accountable. Early benchmarks helped researchers compare systems on shared tasks and datasets. Over time, benchmark performance became a dominant way of defining progress in machine learning.

In the GenAI era, this culture has expanded and intensified. AI companies often present new models through research-style documents, preprints, technical reports, charts and leaderboards. These practices give commercial releases the appearance of scientific breakthroughs. The study argues that this scientific language is central to how model value is constructed.

The authors show that benchmarks now perform two roles at once. They create comparability by placing models in the same ranking system. They also create distinction by making one model appear exceptional within that group. A model can be framed as part of an elite collection of frontier systems while also being presented as singularly advanced.

The study also highlights how benchmarks have moved beyond narrow technical tests. Earlier evaluations often focused on specific tasks such as language understanding, image recognition or information retrieval. Newer GenAI benchmarks increasingly draw from areas associated with human intelligence, expert reasoning and academic performance, including mathematics, graduate-level questions and broad tests of reasoning.

This shift helps companies connect model performance to claims about human-like intelligence. When a system performs well on a difficult exam-style benchmark, the result can be framed as evidence that AI is approaching expert-level reasoning. The study warns that this can overstate what benchmark results actually prove. A model may perform well on a specific test without possessing broad understanding or general intelligence.

The authors also point to problems carried over from standardized intelligence testing. Such tests have long been shaped by contested assumptions about ability, knowledge and hierarchy. When they are adapted for AI evaluation, they may reproduce narrow views of intelligence and favor forms of knowledge that are easy to score at scale.

Benchmark saturation adds another challenge. As models improve or become optimized for specific tests, benchmarks can lose their ability to separate systems meaningfully. Once top models perform similarly, new evaluations are introduced to create fresh competition. This produces a fast cycle of model release, testing, ranking and replacement.

Leaderboards intensify that cycle. Public evaluation platforms can strongly influence perceptions of which model is leading. Some leaderboards rely on user preferences, allowing people to compare outputs from different models. The study notes that these systems have become important market devices, but they also raise concerns about selective testing, private evaluation, model tuning and the potential for benchmark gaming.

The authors argue that the result is a market environment marked by constant production and constant evaluation. AI firms release new models and variants at high speed, while benchmark results are used to claim progress. In this setting, evaluation can become difficult to separate from marketing. A benchmark score may reflect a real capability, but its public meaning depends on how the test is selected, framed and circulated.

AGI narratives turn small gains into major market signals

The study claims that benchmarks gain power because they are tied to narratives about artificial general intelligence. A score increase is not presented only as a technical improvement. It becomes a sign that a model may be moving closer to a future where AI can perform a wide range of economically valuable human work.

The authors identify three narratives that support this process: saturation, surpassing and emergence. Saturation occurs when models perform so well on a benchmark that the test no longer clearly differentiates them. Rather than weakening the story of progress, this can strengthen it by allowing companies and supporters to suggest that certain tasks have been solved.

Surpassing refers to claims that models outperform humans on specific tests. These claims can be powerful market signals because they suggest that AI has crossed a human performance threshold. The study cautions that this does not necessarily mean a model has general intelligence. It may exceed humans on a narrow benchmark because of training, pattern recognition or test-specific optimization.

Emergence refers to claims that large models display unexpected abilities not directly programmed into them. This narrative is especially valuable because it suggests that scale may produce surprising new capabilities. The authors argue that apparent emergence can also result from the way benchmarks are designed, from hidden training data or from selective reporting. What looks like a sudden leap may sometimes reflect incremental optimization.

These narratives help turn benchmark results into evidence of a larger technological destiny. They make GenAI models appear close to a future that is still uncertain but economically powerful. This future orientation helps explain why some models attract huge attention and investment even when their current business value remains unclear.

The study does not dismiss GenAI valuation as simple hype. The authors argue that benchmarks have real effects as they influence markets, guide research priorities, shape public trust and help determine which companies are seen as leaders. They also affect the way AI progress is defined, because firms have incentives to optimize models for visible tests.

If benchmarks shape valuations and trust, then benchmark design, transparency and interpretation become matters of public importance. Evaluation systems should not be treated as neutral technical tools. They are part of the economic infrastructure of AI.

The findings raise questions for investors, policymakers and researchers. Investors may need to treat benchmark claims with more caution, especially when scores are presented as proof of durable advantage. Policymakers may need to scrutinize how private companies define AI progress. Researchers may need to examine how evaluation systems shape both knowledge production and market power.