Abstract

BOND's May 2025 "Trends – Artificial Intelligence" report signifies a pivotal moment in technological advancement, underscoring that Artificial Intelligence (AI) is evolving at a pace and scale that is unprecedented, even when compared to prior technological revolutions like the internet. The core technical concepts revolve around the rapid proliferation of Large Language Models (LLMs) and generative AI, which are demonstrating remarkable capabilities in user adoption, usage intensity, and the sheer capital expenditure (CapEx) they command for infrastructure. Key technical approaches highlighted include the escalating compute requirements for training frontier AI models contrasted with the dramatically falling costs of inference per token. This economic shift is democratizing access and fueling a surge in developer activity and innovation. However, this "AI Gold Rush" is characterized by intense competition, emerging and sometimes uncertain monetization pathways, the significant momentum of open-source models, and the ascent of China as a formidable AI power. The report meticulously documents AI's expansion beyond digital realms into the physical world—transforming industries from automotive to agriculture—and its profound impact on the nature of work, creating new roles while reshaping existing ones. Geopolitically, AI has become a new chessboard, with nations, particularly the USA and China, vying for technological supremacy, which has far-reaching implications for economic leadership and national security. The significance of these trends lies in their collective power to reshape global industries, economies, and societal structures, making an understanding of AI's trajectory essential for navigating the future.

Section 1: Introduction – The AI Tsunami: A New Era of Unprecedented Change

The contemporary technological landscape is being redefined by Artificial Intelligence (AI) at a velocity and magnitude that challenge conventional understanding. As BOND's May 2025 "Trends – Artificial Intelligence" report articulates, the AI revolution is not merely an incremental advancement but a paradigm shift occurring at an unparalleled speed and scale. To contextualize, Vint Cerf, a "Founder of the Internet," famously described internet time in 1999 as akin to "dog years"—one year feeling like seven. The BOND report suggests that AI's current trajectory makes even that accelerated pace seem modest, asserting that "machines can outpace us" and the "pace and scope of change related to the artificial intelligence technology evolution is indeed unprecedented" (p. 3). At the heart of this revolution are Large Language Models (LLMs) and generative AI—systems capable of creating novel content, from text and images to code and complex data analysis. The report highlights the almost instantaneous global adoption of tools like OpenAI's ChatGPT, which achieved 1 million users in just five days (p. 26, 59), a feat that took foundational technologies like the iPhone years. This isn't just about speed; it's about the breadth of impact, touching technical, financial, social, physical, and geopolitical domains simultaneously (p. 3). This article delves into the key trends identified in the BOND report, dissecting the data-driven insights that paint a picture of AI's transformative journey.

Section 2: The Pillars of AI's Explosive Growth: Users, Data, and Capital

The meteoric rise of AI is built upon three interconnected pillars: an explosive growth in users, the exponentially increasing datasets that fuel AI models, and the colossal capital expenditure (CapEx) being poured into AI infrastructure.

User Adoption and Usage:

The adoption of AI tools, particularly consumer-facing generative AI like ChatGPT, has been "unprecedented" (p. 53). The report indicates that leading US-based LLMs reached 800 million weekly active users by April 2025 (p. 5, 56), a staggering figure achieved in a remarkably short timeframe. The sheer scale of AI adoption is visualized in the following figure:

Figure 1: Leading USA-Based LLM Weekly Active Users (Millions), October 2022 - April 2025. This chart illustrates the growth of ChatGPT weekly active users. Data per OpenAI. (Source PDF, p. 56).

This rapid uptake is not confined to specific demographics but spans across age groups, with usage intensity also on the rise (p. 81-84). For instance, daily time spent on ChatGPT by US active users saw a +202% increase over 21 months (p. 83)

Data as the New Oil:

AI models, especially LLMs, are data-hungry. Their performance and capabilities are directly correlated with the volume and quality of data they are trained on. The report underscores that "ever-growing digital datasets that have been in the making for over three decades" (p. 3) form the bedrock of current AI advancements. The global data generation continues to explode, providing a richer and more diverse pool of information for training increasingly sophisticated AI.

Capital Expenditure (CapEx):

The development and deployment of advanced AI necessitate significant investment in physical infrastructure. This CapEx is primarily directed towards building and equipping massive data centers with specialized hardware, such as Graphics Processing Units (GPUs) and custom AI accelerators (e.g., Google's TPUs, Amazon's Trainium chips). The "Big Six" US technology companies (Apple, NVIDIA, Microsoft, Alphabet, Amazon (AWS), & Meta Platforms) have dramatically increased their CapEx, which reached $212 billion in 2024, a 63% year-over-year increase (p. 5, 102). This user growth and the data it generates are mirrored by massive capital expenditure, as seen in the following figure: !Placeholder for visual described as: Bar and line graph of Big Six US Tech Company CapEx vs. Global Data Generation, 2014-2024

Figure 2: Big Six USA Public Technology Company CapEx Spend ($B) versus Global Data Generation (Zettabytes), 2014-2024. The chart displays CapEx as blue bars and Global Data Generation as a red line. Data per Capital IQ & Hinrich Foundation. (Source PDF, p. 97)

The chart illustrates a +21% annual growth in CapEx for these companies over ten years, closely tracking the +28% annual growth in global data generation. This symbiotic relationship is clear: more users generate more data, which fuels the development of better models, attracting more users and justifying further massive CapEx investment. Hyperscalers (large data center operators) are at the forefront, with their cloud revenue growing +37% annually over ten years (p. 99), reflecting the surging demand for compute resources.

Section 3: The Shifting Economics of AI: Compute, Costs, and Developer Dynamism

The economic landscape of AI is undergoing a significant transformation, characterized by a dichotomy in compute costs: the cost of training frontier models remains high and is rising, while the cost of inference (using trained models) is plummeting. This shift has profound implications for model development, accessibility, and the competitive environment.

Training vs. Inference Costs:

• Training Costs :Developing state-of-the-art LLMs is a capital-intensive endeavor. The BOND report highlights that training costs for frontier AI models have grown approximately 2,400x over eight years, with some models now costing over $100 million to train, and projections suggesting$10 billion models could emerge by 2025 (p. 133). Conversely, the resources required for training cutting-edge AI models continue to soar, as shown by the data in the following figure: !Placeholder for visual described as: Scatter plot of Training Compute (FLOP) for Key AI Models, 1950-2025

Figure 3: Training Compute (FLOP) for Key AI Models, 1950-2025. This scatter plot shows the increasing computational power (FLOPs) required for training AI models over time. Data per Epoch AI. (Source PDF, p. 15)

This chart illustrates the exponential increase in floating-point operations (FLOPs), a measure of computational power, needed to train notable AI models over time, with a 360% annual growth rate over the last fifteen years.

• Inference Costs: Once a model is trained, the cost to run it for generating predictions or content (inference) is decreasing dramatically. This is driven by hardware improvements (e.g., NVIDIA's Blackwell GPU consuming 105,000x less energy per token than its 2014 Kepler predecessor, p. 137) and algorithmic efficiencies. The plummeting cost of AI inference, a key enabler of widespread AI adoption, is evident when compared to other key technology cost declines, as detailed in the following figure: !Placeholder for visual described as: Line graph comparing relative cost decline of ChatGPT 75-word response vs. Electric Power and Computer Memory

Figure 4: Relative Cost of Key Technologies by Year Since Launch - ChatGPT: 75-Word Response, Electric Power, and Computer Memory. This line graph compares the indexed cost decline of these technologies. Data per OpenAI, John McCallum, & Richard Hirsh. (Source PDF, p. 138)

Stanford HAI data shows inference prices for customers (per 1 million tokens) dropping by as much as 99.7% over two years for some models (p. 138).

Implications of Shifting Economics:

• Performance Convergence: As inference becomes cheaper, the performance gap between top-tier frontier models and smaller, more efficient alternatives is narrowing for many use cases (p. 145). This means developers and users can achieve similar results with lower-cost models, especially when fine-tuned for specific tasks.

• Developer Dynamism: The "cost collapse" in inference has made AI experimentation and productization feasible for a much broader range of developers, from solo entrepreneurs to small businesses (p. 145). This is evidenced by the explosive growth in AI developer ecosystems. For example, the NVIDIA ecosystem grew 6x to 6 million developers in seven years (p. 39), and Google's Gemini ecosystem saw a 5x year-over-year increase to 7 million developers (p. 40). The number of AI developer repositories on GitHub increased by approximately 175% in just sixteen months (p. 149). This dynamic aligns with Jevons Paradox: as AI inference becomes cheaper and more efficient, its overall usage and the demand for underlying compute resources increase, creating a perpetual cycle of innovation and adoption (p. 135).

Section 4: The AI Gold Rush: Monetization, Competition, and the Open-Source Wave

Despite the unprecedented growth in AI usage and capabilities, the path to sustained profitability for all players in the AI ecosystem remains a complex equation. The "AI Gold Rush" is characterized by rapid revenue growth for some, significant cash burn for many, intense competition, the disruptive influence of open-source models, and the strategic rise of China.

Monetization Strategies and Financial Realities:

AI companies are exploring various monetization avenues, including:

• Consumer Subscriptions: Freemium models with paid tiers for advanced features (e.g., OpenAI ChatGPT Plus, Google Gemini Advanced).

• Developer API Fees: Charging developers for access to powerful foundation models to build their own AI-powered applications.

• AI-Enhanced Products & Services: Integrating AI capabilities into existing enterprise software (e.g., Microsoft 365 Copilot, Salesforce Einstein) or offering specialized AI solutions for specific industries. While some new entrants like OpenAI are reporting substantial annualized revenue (e.g., $3.7 billion, a +1,050$5 billion against +$3.7 billion in revenue (p. 6, 174). Major tech incumbents are also seeing AI contribute significantly to revenue (e.g., Microsoft's AI business surpassing a$13 billion annual run rate, p. 211), but this is often coupled with increased CapEx and R&D spending, sometimes leading to compressed free cash flow margins (p. 175). Underpinning the drive for monetization is the rapid advancement in AI capabilities, with systems now outperforming humans on complex benchmarks like MMLU, illustrated in the following figure: !Placeholder for visual described as: Line graph of AI System Performance on MMLU Benchmark Test, 2019-2024

Figure 5: AI System Performance on MMLU Benchmark Test, 2019-2024. This line graph shows AI models surpassing the human baseline on the Massive Multitask Language Understanding benchmark. Data per Stanford HAI. (Source PDF, p. 41)

This chart shows AI models surpassing the 89.8% human baseline on the MMLU benchmark in 2024, reaching 92.3%.

Competition and the Open-Source Factor:

The AI landscape is fiercely competitive, with tech giants, well-funded startups, and a burgeoning open-source community all vying for market share and technological leadership.

• Closed-Source Models: Companies like OpenAI, Anthropic, and Google primarily develop proprietary, closed-source models, often leading in raw performance and user experience for consumer applications. These models dominate consumer MAUs (p. 263).

• Open-Source Models: A resurgence of powerful open-source models (e.g., Meta's Llama series, models from DeepSeek) is providing viable, lower-cost alternatives. These are gaining traction among developers and are fueling innovation in areas like sovereign AI and local language models (p. 262). While closed models still lead in compute investment, the performance gap is closing, particularly with China emerging as a leader in open-source model releases (p. 265-266). The number of AI models available on platforms like Hugging Face has grown by +33x from 2022 to 2024 (p. 270), indicating the scale of open-source activity.

This dynamic creates a "flywheel of developer-led infrastructure growth" (p. 146), where more accessible models lead to more AI-native apps and tools, further accelerating adoption.

Section 5: AI's Expanding Dominion: Reshaping Work and the Physical World

AI's influence extends far beyond software and digital services; it is increasingly reshaping the nature of work and making significant inroads into the physical world.

Transformation of Work:

AI is acting as both an augmenter and an automator of human labor.

• Productivity Gains: AI tools are demonstrating tangible productivity improvements across various professions. For example, Stanford HAI research indicates a +14% increase in hourly chats per customer support agent when using AI (p. 332).
• Job Market Evolution: The demand for AI-specific skills is surging. The evolution of the workforce is clearly indicated by trends in job postings; the following figure highlights the significant growth in demand for AI-specific IT roles: !Placeholder for visual described as: Line graph comparing indexed change in USA AI vs. Non-AI IT Job Postings, 1/18-4/25

Figure 6: Indexed Change in USA AI versus Non-AI IT Job Postings, January 2018 - April 2025. This line graph shows AI job postings (blue line) increasing significantly while Non-AI IT job postings (red line) show a slight decline. Data per University of Maryland & LinkUp. (Source PDF, p. 332)

This shows a +448% increase in AI job postings over 7 years, while non-AI IT jobs declined by -9%. Companies are actively seeking talent for "Generative AI" roles (p. 335).

• New Work Paradigms: The report suggests a future where humans increasingly work alongside AI, potentially even teaching and refining AI systems (RLHF - Reinforcement Learning from Human Feedback) (p. 325).

AI in the Physical World:

Intelligence is becoming kinetic, embedded in vehicles, machinery, and infrastructure (p. 301).

• Autonomous Systems: Self-driving technology is maturing, with Tesla reporting a ~100x increase in fully self-driven miles over 33 months (p. 302) and Waymo capturing up to 27% of San Francisco rideshare gross bookings (p. 7, 303).
• Industrial & Specialized Robotics: AI is enhancing robotics in manufacturing, logistics (p. 6), and even agriculture, with companies like Carbon Robotics using AI-powered laser weeding to reduce herbicide use (p. 307). China's installed base of industrial robots now surpasses that of the rest of the world combined (p. 290).
• Defense and Exploration: AI is being integrated into defense systems for autonomous operations (e.g., Anduril, p. 305) and is accelerating resource discovery in mining (e.g., KoBold Metals, p. 306).

This "broader shift" implies a world where AI turns capital assets into software endpoints, moving intelligence from dashboards into real-world action (p. 301).

Section 6: The New Geopolitical Chessboard: AI and the Global Balance of Power

The rapid ascent of AI is not just a technological or economic phenomenon; it is a critical factor reshaping the global geopolitical landscape. As BOND's report underscores, "AI leadership could beget geopolitical leadership" (p. 8), turning AI development into a new frontier for international competition and strategic positioning, particularly between the United States and China.

The USA-China AI Rivalry:

Global competition, especially concerning USA and China tech developments, is described as "acute" (p. 3). This rivalry extends across multiple dimensions of AI:

• Model Development: While US companies initially led in frontier model innovation (e.g., OpenAI's GPT series, Google's Gemini), China is rapidly advancing, particularly in open-source model releases and achieving competitive performance benchmarks (p. 262, 265-266, 283-285). Epoch AI data shows China outpacing the rest of the world (excluding the USA) in cumulative large-scale AI system releases by 2024 (p. 282).
• Hardware and Semiconductors: Control over the hardware that powers AI (CPUs, TPUs, AI accelerators) is a key strategic battleground. The US has implemented export controls to limit China's access to advanced semiconductor technology, while China is intensifying efforts to develop its domestic chip industry (e.g., Huawei's advanced AI chip clusters, p. 273, 288). Taiwan's TSMC remains a critical player, producing the majority of the world's most advanced semiconductors (p. 273, 280).
• National Strategies and Investment: Both nations view AI as crucial for economic growth and national security. The US relies heavily on private sector innovation and CapEx, while China employs state-backed coordination, national infrastructure projects, and a strong push in specific sectors like robotics, where its installed base is now higher than the rest of the world combined (p. 6, 290).
• Data Governance and Influence: The differing approaches to data privacy and governance also play a role. The report alludes to concerns that nations leading in AI could leverage it to project influence, potentially forcing data sharing or developing cyber capabilities (p. 272).

Broader Geopolitical Implications:

The AI race has implications beyond the two superpowers:

• Technological Sovereignty: Nations are increasingly recognizing the need for "sovereign AI" capabilities (p. 8, 262, 78) to reduce dependence on foreign technology and ensure national interests are protected. This involves developing local language models, domestic infrastructure, and fostering local talent.
• Shifting Economic Power: Leadership in AI is expected to translate into significant economic advantages. The global public market capitalization landscape is already reflecting this, with AI-centric companies like NVIDIA seeing their valuations soar. In May 2025, 83% (25 of 30) of the world's most valuable public companies were US-based, a significant increase from 53% (16 of 30) in 1995. China is a new entrant on this list with 2 companies, alongside other new geographic entrants like Taiwan and Germany (p. 275-277).
• Global Standards and Ethics: The competition also extends to shaping global norms, standards, and ethical guidelines for AI development and deployment.

The BOND report concludes that in this environment, "innovation is not just a business advantage; it is national posture" (p. 339). The race for AI supremacy is well underway, with profound implications for the future global order.

Critical Analysis

The BOND May 2025 AI Trends report provides a compelling, data-rich overview of the AI revolution. Its primary strength lies in its comprehensive aggregation of diverse data points to illustrate the sheer velocity and multifaceted nature of AI's current trajectory.

Strengths:

• Data-Driven Narrative: The report excels at using quantitative data (user growth, CapEx, revenue, performance benchmarks) to substantiate its claims about the unprecedented nature of AI development.
• Breadth of Coverage: It effectively connects disparate trends across user adoption, technological advancement, economic shifts, physical world integration, workforce evolution, and geopolitical dynamics.
• Identification of Key Tensions: The report clearly highlights critical tensions, such as high training costs versus falling inference costs, the promise of AI versus monetization challenges, and the closed-source versus open-source debate.

Limitations:

• Future Projections: While data-grounded, some forward-looking statements (e.g., regarding profitability, market dominance) are inherently speculative in such a rapidly evolving field. The report acknowledges this, stating "Only time will tell" (p. 155, 186) regarding certain outcomes.
• Depth vs. Breadth: Given its scope, the report offers a high-level view. Deeper dives into specific technical architectures or nuanced socio-economic impacts are beyond its purview.
• Data Aggregation Nuances: As with any report aggregating data from multiple sources, there can be slight variations in definitions or methodologies (e.g., "active users"). The report often notes these where applicable.

Practical Implementation Considerations:

• Infrastructure Strain: The massive CapEx and energy consumption (p. 126-129) associated with AI, particularly for data centers, pose significant logistical, environmental, and economic challenges. Power availability is becoming a critical bottleneck (p. 120).
• Talent Development: The rapid evolution of AI necessitates a workforce equipped with new skills. The surge in AI-related job postings (p. 333) highlights an ongoing need for talent in AI development, deployment, and management.
• Ethical and Regulatory Frameworks: The speed of AI development is outpa_cing regulatory and ethical frameworks. Issues of bias, misinformation, job displacement, and security (p. 51) require careful and proactive consideration. The Bletchley Declaration on AI Safety (p. 30) is an early step in this direction.

Scalability and Performance:

The report shows that while AI model performance is rapidly improving (p. 42, 143) and inference is becoming more efficient, scaling these systems globally presents challenges. The infrastructure to support "tens of billions of units" (p. 13) using AI technology is a massive undertaking. The trend towards specialized hardware (ASICs, TPUs) alongside GPUs (p. 158) is a response to the need for both performance and efficiency at scale.

Potential Challenges and Trade-offs:

• Monetization vs. Open Access: The tension between commercializing AI (requiring investment returns) and the democratizing power of open-source models presents a fundamental trade-off.
• Innovation Speed vs. Safety/Control: The race for AI leadership might incentivize rapid deployment over cautious, safety-first approaches. Economic Benefits vs. Inequality: While AI promises productivity gains, its benefits may not be evenly distributed, potentially exacerbating existing inequalities if not managed proactively.

Comparative Context

The BOND report implicitly and explicitly touches upon several areas where different technical approaches and strategies are in play.

1. Closed-Source vs. Open-Source AI Models:

Closed-Source (e.g., OpenAI's GPT-4, Anthropic's Claude):

• Approach: Developed and controlled by single entities, often with proprietary datasets and architectures. Access is typically via APIs or paid subscriptions.
• Strengths:Often lead in cutting-edge performance (though the gap is closing), generally offer more polished user interfaces, and may provide stronger security/support for enterprise. Dominate consumer MAUs (p. 263).
• Limitations: Lack of transparency ("opacity," p. 262), higher costs, potential for vendor lock-in.
• Use Cases: Preferred by many enterprises for robust, supported solutions and by consumers seeking ease-of-use.

Open-Source (e.g., Meta's Llama, various models on Hugging Face):

• Approach: Models whose weights and often training methodologies are publicly available, allowing for modification, fine-tuning, and local deployment.
• Strengths: Lower cost, greater transparency, fosters community innovation, enables customization and sovereign AI initiatives. Rapidly improving in performance (p. 265-266). China is showing strong leadership in large-scale open-source releases (p. 262).
• Limitations:Can be less polished, may require more technical expertise to deploy and manage, support can be community-driven rather than guaranteed.
• Use Cases: Favored by developers, researchers, startups, and nations aiming for technological independence.

Trade-offs: The choice involves balancing performance, cost, control, transparency, ease of use, and speed of innovation. The report suggests a "flywheel of developer-led infrastructure growth" (p. 146) driven by open options.

2. AI Monetization Strategies: Horizontal Platforms vs. Specialized Software:

Horizontal Platforms (e.g., Microsoft integrating Copilot across its suite, OpenAI's ChatGPT aiming for broad enterprise adoption):

• Approach: Building broad, general-purpose AI capabilities embedded across many functions or offered as a unified interface (p. 215).
• Strengths: Large addressable market, potential for strong network effects, ability to leverage existing distribution channels (e.g., Microsoft Office users, p. 228).
• Limitations: May lack the deep domain-specific expertise of specialized tools; competition from both specialized players and other horizontal platforms.
• Use Cases: General productivity, communication, basic research, coding assistance.

Specialized Software (e.g., Anysphere Cursor for code editing, Abridge for healthcare conversations, Harvey for legal workflows):

• Approach: Developing AI tools fine-tuned for specific industries or tasks, often leveraging proprietary industry data (p. 216, 232-244).
• Strengths: Deep domain expertise, can solve very specific high-value problems, potentially faster adoption within niche markets due to clear ROI. Rapid ARR growth is seen in several specialized AI companies (p. 234-244).
• Limitations: Smaller initial market per tool, may face integration challenges with broader enterprise systems.
• Use Cases: Industry-specific tasks like medical scribing, legal document analysis, specialized code generation, financial analysis.

Trade-offs: Horizontal platforms offer breadth and scale, while specialized software offers depth and tailored solutions. The report suggests a "convergence" rather than a winner-take-all scenario, where the ability to "abstract the right layer, own the interface, and capture the logic of work itself" will be key (p. 216).

3. National AI Strategies: USA vs. China:

USA:

• Approach: Primarily private-sector driven innovation, significant CapEx from tech giants, strong university research, and a focus on frontier model development. Government support through initiatives like the CHIPS Act (p. 273).
• Strengths: Leading in frontier model R&D, strong venture capital ecosystem, home to most leading AI companies by market cap (p. 275).Limitations: Potential for less centralized coordination compared to state-led approaches; debates around regulation and industrial policy.
• Limitations: Historically reliant on foreign semiconductor technology (though this is changing), questions around data privacy and international trust.

China:

• Approach: State-backed coordination, national infrastructure projects, strong focus on open-source model development and rapid deployment in strategic sectors like robotics and manufacturing. Growing domestic semiconductor industry (p. 272-273, 282-290).
• Strengths: Rapid scaling, strong government support, large domestic market, leadership in certain AI application areas (e.g., industrial robots, p. 290). Closing performance gaps with US models, sometimes with lower training costs (p. 286-287).
• Limitations: Historically reliant on foreign semiconductor technology (though this is changing), questions around data privacy and international trust.

Trade-offs: The US model leverages market dynamism and leading-edge research, while China's approach emphasizes strategic alignment and rapid national scaling. The competition is driving innovation but also creating geopolitical tensions and concerns about technological decoupling.

Source

This article is based on the findings and data presented in the following technical document:

Document Title: Trends – Artificial Intelligence (AI)

Author/Organization: BOND (Mary Meeker / Jay Simons / Daegwon Chae / Alexander Krey)

Publication Date: May 30, 2025

Document Type: Report

URL: https://www.bondcap.com/report/tai/#view/33