Industrial Shifts Through the Mary Meeker AI Report: New Rules for Capital, Computing Power, and Competition

Preface: An Empirical Atlas of AI Prosperity

At a time when the artificial intelligence industry is rapidly heating up, data and capital are often the most reliable yardsticks for judging trends. Mary Meeker, a partner at Bond Capital, recently released a 340-page report, _Trends in Artificial Intelligence_. It is her first major industry trend report in six years and a document closely watched by the global technology and investment community. Meeker told Axios, "We have never seen user growth like ChatGPT's, especially considering market changes outside the United States; it shows the transformation of the global technology and distribution landscape." The charts and data in the report reveal the speed, cost, and competitive structure of AI growth, offering a rare high-resolution snapshot for understanding this technological transformation.

Industry Background: The Accelerating Engine of AI Infrastructure Investment

The report first reveals the infrastructure support behind AI's rapid expansion. From 2014 to 2024, capital expenditures by six major technology giants—Apple, Nvidia, Microsoft, Google, Amazon, and Meta—grew by 21% per year, while global data generation climbed at a compound annual growth rate of 28%. In this era, AI's value no longer resides solely in the algorithm layer, but increasingly centers on whether one can acquire data at scale, train models, and perform fast inference. Hyperscale data centers, high-speed networks, and the latest chips constitute the foundational base of the new AI industry.

At the same time, the scale of data required for model training has also expanded exponentially. According to Epoch AI's statistics, over the past 15 years, the scale of AI model training datasets has grown at an average annual rate of more than 250%. This metric reflects that possessing the most valuable data assets and data-processing capabilities has become a key competitive barrier for AI companies.

Market Impact: Plummeting Costs and Accelerating Commercialization

On the market-demand side, the report's most striking finding is the sharp decline in AI inference costs. From OpenAI's GPT-3.5-generation products to the GPT-4o series, and then to DeepSeek-V3 released in late 2024, the inference price per million tokens fell from the ten-plus-dollar range to below $1—a decline of 99.7% within two years. Meeker describes this trend as "zero-interest-rate-style inference." The drop in inference costs directly removes commercialization barriers in many potential scenarios, making high-frequency, real-time generation applications possible.

Another set of metrics also confirms the improvement in AI's ability to reach the market. The data shows that, in the ramp-up to $5 million in annualized revenue, AI companies take an average of 24 months, while SaaS companies take an average of 37 months. AI can deliver business value more quickly, on one hand because of the leap in performance of AI products themselves, and on the other hand because of the strong demand from enterprises for intelligent transformation.

Competitive Landscape: The Lead of Closed-Source Models and the Bipolar U.S.–China StructureAt the model ecosystem level, closed-source models are widening the compute gap with the open-source camp. The report shows a lag of about 17 months between closed-source and open-source models in training compute intensity, with closed-source models growing compute capacity by 5.2x per year, compared to 3.6x for open-source models. This means that more advanced model clusters are increasingly concentrated in the hands of a few giants capable of bearing extremely high compute costs. Although the open-source community has always advocated a fair and open technical path, the skewed allocation of resources may lead to the gradual concentration of influence over the ecosystem.

Geographically, large AI systems are highly concentrated in China and the United States. The U.S. has close to 150 large models and AI systems, while China has also surpassed 100, clearly leading the U.K., France, Canada, Germany, and others. The dual-power model of the U.S. and China in data scale, engineering capability, and investment intensity is defining the global AI industry chain. This landscape will affect other countries' technological autonomy, cross-border data rules, and security strategies, and will further intensify discussions about "AI sovereignty."

Enterprise Implications: Distinguishing Real Action from the Noise

For enterprises developing AI strategies, the report's insights are clear and practical.

First, the decline in inference costs means enterprises should reassess AI use cases that previously lacked ROI, such as real-time customer service, personalized recommendations, and code assistance. The marginal cost of AI capabilities is now entering a range that can be afforded at scale.

Second, the capital expenditure race among tech giants shows that AI depends on physical compute and energy networks. When choosing cloud service providers and model platforms, enterprises must assess the long-term compatibility of compute supply and potential lock-in risks.

Third, the gap between closed-source and open-source models requires enterprises to build a "dual-track" strategy: for core businesses and environments with high compliance requirements, high-end closed-source models can be selected; while in scenarios that demand high data security or deep customization, open-source alternatives—though lagging in capability—offer controllability and transparency.

Fourth, AI's energy consumption is becoming a key external variable. When deploying AI applications, enterprises should plan green energy solutions in parallel to avoid compliance issues caused by energy consumption and carbon emissions.

Future Outlook: Evolution over the Next 12 Months to Three Years

In the next 12 months, inference costs are highly likely to continue declining, bringing more small and medium-sized enterprises into AI application innovation, with agent-based services becoming the next hot application frontier. Meanwhile, large model companies will further bundle ecosystems through APIs and cloud services, while chip supply and electricity supply will become new bottlenecks constraining industry growth.

Over the next 24 months, the performance gap between open-source and closed-source models may remain or even widen, but if hardware efficiency and training techniques achieve new breakthroughs, the open-source camp could also shorten its catch-up cycle. At the same time, the "twin powers" status of China and the United States in terms of AI system numbers will solidify, prompting surrounding countries and regions to align their industrial layouts more closely with their technological endowments.Looking ahead to a three-year horizon, the AI industry may transition from the current "model arms race" to "competition in services and ecosystems." The threshold for capital investment will rise even higher, and the number of players able to simultaneously command complex systems of algorithms, data, computing power, and energy may be extremely small. More importantly, whether AI can, like the internet and mobile technology, deliver sustained and visible contributions to productivity across all industries is the ultimate criterion for testing the industry's value.

References

This article is based on an analysis of the report released by Mary Meeker and Bond VC. Original article: 10 Charts That Define the AI Boom, According to Mary Meeker. All facts and data presented in this article are cited from this report.