AI sector faces $700B capex surge while inference costs drop 99.7% amid bubble burst

Market volatility has intensified debates regarding the existence of an artificial intelligence bubble, with conflicting narratives emerging from top financial and technological leaders. Bridgewater Associates founder Ray Dalio characterizes the current market level as relatively high, signaling a bubble, while NVIDIA CEO Jensen Huang asserts that computing power demand is only beginning to explode. Both perspectives hold validity; the presence of a bubble is inevitable when society confronts disruptive advanced productive forces, serving as a necessary tribute rather than a purely pejorative phenomenon. This dynamic mirrors the dawn of the internet era, where speculative excess paved the way for foundational infrastructure that would eventually empower nearly every sector.

The comparison to the 2000 internet bubble is frequent, recalling a period where the Nasdaq crashed nearly 78% and over $5 trillion in wealth evaporated. During that era, companies like MicroStrategy plummeted 62% in a single day due to accounting scandals, while Pets.com and Webvan went bankrupt immediately.

However, the physical infrastructure built during that frenzy, such as global submarine cables and dense wavelength division multiplexing networks laid by telecom giants like WorldCom, became the low-cost breeding ground for future supergiants like Netflix and Zoom. Amazon's stock fell from $107 in 1999 to $7 in 2001, a drop exceeding 90%, yet it survived because its core logic of restructuring retail aligned with long-term technological trajectories, illustrating Amara's Law where short-term impacts are overestimated while long-term impacts are severely underestimated.

In the current 2026 landscape, the AI bubble appears even more pronounced in terms of capital deployment. The five major cloud service providers—Amazon, Google, Meta, Microsoft, and Oracle—are projected to reach $690 billion in capital expenditures by 2026, with total AI infrastructure investment expected to hit $5.3 trillion by 2030. Of this massive outlay, only approximately 25% is allocated to GPUs, while the remaining 75% funds physical infrastructure including liquid cooling systems, power transmission, network switches, and optical modules. In stark contrast, leading pure-play AI companies such as OpenAI, Anthropic, Cohere, and Mistral are expected to generate no more than $40 billion in total revenue by 2026. Data compiled by Woofun AI highlights this severe asymmetry, where nearly $700 billion is invested in the foundational layer against a few hundred billion in application-layer returns, raising questions about sustainability.

However, a critical economic shift is occurring that defies traditional linear thinking regarding cost and demand. When OpenAI released GPT-4 in March 2023, the mixed cost per million tokens input was approximately $30. By April 2025, optimized model architectures and improved inference computing power drove this price down to between $0.1 and $0.15 per million tokens. ' and TokenCost data, AI inference costs have dropped by over 99.7% in the past two years. Despite this precipitous decline in unit costs, corporate AI cloud spending tripled between 2024 and 2025. This phenomenon validates the Jevons Paradox, where technological progress improves efficiency, leading not to reduced consumption but to an exponential increase in demand as marginal costs approach zero.

As the cost of intelligence collapses, AI is transitioning from simple text summarization to an era of intelligent agents and multimodal enhanced retrieval. Companies are now deploying AI agents to autonomously execute thousands of tasks, write code, scan millions of legal contracts, and simulate biological experiments, unlocking vast long-tail demand previously constrained by expense. Woofun AI notes that this shift is comparable to the divergence between NVIDIA in 2026 and Cisco in 2000; while their ecological niches as hardware enablers are similar, their underlying financial health and the depth of their integration into the economy differ significantly. The market is rapidly sobering up, recognizing that algorithmic optimization lowers adoption thresholds, thereby driving total computing power consumption upward rather than downward.

The market is currently undergoing a profound self-purification process, marking the eve of the 'valley of disillusionment' on the Gartner technology maturity curve. Venture capital sentiment has shifted dramatically; startups that previously raised funds with dozens of pages of PPTs wrapped around OpenAI APIs are now dying in large numbers as the tide recedes. This surface-level bust is accompanied by three deeper structural evolutions. First, value is shifting from capital expenditure (CapEx) to operational expenditure (OpEx). While hardware vendors like NVIDIA and TSMC currently reap the benefits, excess profits will gradually migrate to AI-native companies that use low-cost tokens to solve vertical industry pain points. Second, valuation multiples are compressing as performance digests high valuations through a 'time-for-space' approach, provided revenue growth keeps pace with capital expenditure depreciation.

Real-world applications are already demonstrating the efficacy of this transition. Global automotive and chip manufacturing giants have shortened the R&D to mass production cycle of new products by 35% and improved overall line equipment efficiency by 18% through end-to-end AI twin technology. In the financial sector, quantitative trading, risk control, and credit assessment are increasingly dominated by multimodal agents by 2026, with AI processing macro expectations at microsecond timestamps and engaging in micro-level asset pricing. In knowledge-intensive fields like law, medicine, and auditing, AI has evolved from a junior assistant to a partner-level expert. Woofun AI analysis suggests that with ChatGPT, Gemini, and Claude surpassing 1 billion active users, the technology is effectively substituting daily high-intensity intellectual labor for a significant portion of the global workforce.

Schumpeter's concept of 'creative destruction' remains the governing principle of this technological lifecycle. The capital market's impatience to convert $700 billion in infrastructure investment into immediate application-side profits will inevitably trigger a brutal reshuffle, eliminating speculative shell companies while preserving those with genuine technological foundations. Post-reshuffle, the resulting cheap and massive computing centers, coupled with highly optimized algorithms, will serve various industries at extremely low prices. Just as the post-2000 era ushered in a digital age where the internet became indispensable, the current trajectory points irreversibly toward an intelligent era where all industries are vertically integrated and empowered by AI. Amidst the clamor of a bursting bubble, the underlying productive potential remains robust and un-inflated.