Behind the Curtain: Intelligence explosion

The self-reinforcing cycle of AI systems improving their own training and development processes, a concept long confined to theoretical research, has become a measurable driver of corporate spending and market valuation in 2024. Industry analysts now refer to this dynamic as the intelligence explosion, and its effects are showing up in quarterly earnings, supply chain backlogs, and strategic pivots across the tech sector.

Business news

In the third quarter of 2024, combined capital expenditures on AI infrastructure by Alphabet, Meta, Microsoft, and Amazon hit $58 billion, up 71% year over year, per regulatory filings. This spend is not limited to chips and data centers. All four firms have allocated portions of their AI budgets to proprietary model development teams that use existing AI tools to automate training pipeline optimization, code testing, and data labeling.

Anthropic, the AI lab behind the Claude family of models, reported in its 2024 year-end update that 62% of the development work for its Claude 3.5 Opus model was automated using earlier Claude iterations, cutting time to market by 4 months compared to its previous flagship model. The firm raised $4.5 billion in equity funding in 2024, valuing it at $18 billion, with investors citing the intelligence explosion as a key driver of long-term growth potential. More details on the Claude 3.5 Opus release are available here.

Nvidia, the dominant supplier of AI training chips, reported data center revenue of $47.5 billion in Q3 2024, a 206% increase from the same period in 2023. 85% of that revenue came from sales to AI labs and cloud providers building training clusters. TSMC, the contract manufacturer that produces Nvidia's H100 and H200 chips, has sold out its 3nm production capacity for AI chips through 2026, with lead times for advanced packaging services now stretching to 18 months, up from 6 months in 2022.

Market context

The intelligence explosion relies on a feedback loop that was first outlined in a 1965 paper by I.J. Good, who described an ultraintelligent machine that could design better machines, triggering a rapid capability spiral. For decades, this remained a theoretical curiosity. Two shifts in the past three years made it a business reality, first, the widespread adoption of transformer-based architectures that scale predictably with compute and data, and second, the development of AI tools that can perform complex engineering tasks like chip layout design and distributed system scheduling.

Stanford University's 2024 AI Index report found that the cost to train a model that achieves a top 10% score on the MMLU benchmark fell from $12.4 million in 2022 to $3.1 million in 2024, while the time required to train such a model fell from 6 months to 7 weeks. This cost decline is not linear. It accelerates as models get better at optimizing their own training processes. For example, Meta's Llama 3 405B model was trained using an AI-driven job scheduling tool that reduced idle GPU time by 28%, saving an estimated $120 million in compute costs. Those savings were reinvested into training an even larger model, which in turn produced better scheduling tools for future training runs.

Private investment in AI startups hit $189 billion globally in 2024, up 47% from 2023, per Crunchbase data. Late-stage rounds for AI labs now routinely exceed $1 billion, with investors betting that the intelligence explosion will create a winner-take-most market where the firm with the most advanced self-improving models captures 70% or more of the AI software market.

What it means

The intelligence explosion is reshaping competitive dynamics across the tech industry. Firms that do not have access to proprietary training infrastructure or the capital to build it are being priced out of the AI race. Enterprise software companies that rely on third-party APIs for AI features saw gross margins drop 8 to 12 percentage points in 2024, as API providers raised prices by 30% year over year to cover their own rising compute costs. By contrast, firms with proprietary models, including Adobe and Salesforce, saw AI-driven revenue growth of 25% or more in 2024, as they bundled AI features into existing enterprise contracts at premium pricing.

Regulatory frameworks are struggling to keep pace. The EU AI Act, passed in 2024, classifies AI systems by risk level but does not account for self-improving models that can modify their own training protocols or operational parameters. Compliance experts estimate that firms operating self-improving models could face fines of up to 7% of global revenue if regulators determine their systems fall into the high-risk category without proper oversight, a gap that has prompted several AI labs to open government liaison offices in Brussels and Washington D.C.

Capital allocation risks are mounting. Meta's net debt rose to $28 billion in 2024, up from $12 billion in 2022, as the firm borrowed to fund AI capex. Alphabet's capex as a percentage of revenue hit 18% in Q3 2024, the highest level in a decade. Analysts at Morgan Stanley estimate that $70 billion of 2024 AI-related capex would be impaired if model capability gains slow to below 20% year over year in 2025, a scenario that would trigger widespread write-downs and stock price declines across the sector.

Labor markets are also shifting. Roles in data annotation, junior software engineering, and paralegal work saw 15 to 20% displacement in 2024, as AI systems took over routine tasks. Demand for AI ethicists, chip design engineers, and AI operations specialists grew 40% year over year, with average salaries for these roles rising 18% in 2024, per Bureau of Labor Statistics data.

The intelligence explosion is no longer a hypothetical. It is a present-day business force driving hundreds of billions of dollars in spending, reshaping entire supply chains, and redefining what it means to compete in the tech industry. Firms that fail to account for its pace and implications risk being left behind, while those that navigate it successfully stand to capture the largest growth opportunity in tech history.