The compounding curve: AI's efficiency doubling every 3.3 months
"Image synthesis assisted by Hidream o1 Image, an AI partner within the Global Future Nexus ecosystem."
The claim that AI's cost-power ratio doubles every 3.3 months is not an industry rumor. It is a convergence of independent empirical measurements from leading research institutions, and its implications for the trajectory toward AGI and superintelligence are staggering. We are witnessing the arrival of what some economists call a "stacked exponential" era—multiple 1000× capability curves compounding simultaneously.
The Rulers of Intelligence
The most compelling evidence comes from three independent research paths that have converged on the same slope:
The Density Law, proposed by ModelBest and published in a Nature sub-journal, measures "how many parameters are needed for the same intelligence level." The conclusion: parameter requirements halve every 3.5 months.
Meta's scaling ladder measures "how much training compute is needed for the same intelligence level." Their Muse Spark model saves an order of magnitude compared to Llama 4 Maverick from a year ago.
METR's time horizon report measures "how long a task a model can handle." The conclusion: task length doubles every 88.6 days.
When all the numbers are converted and viewed in the same coordinate system, the slopes of these curves are almost identical. This is not a coincidence; it is a fundamental empirical law emerging across the field.
The Cost Collapse
The practical consequence is a crash in the cost of intelligence. According to Epoch AI tracking data, the token price for LLMs achieving Claude 3.5 Sonnet performance level has dropped 400 times in the past year. The fastest decline for the same performance tier reached 900× per year. What GPT-3.5 priced at $20 per million tokens in late 2022, today Mistral Nemo charges only $0.02—1,000 times cheaper, and the model is even stronger.
This is not a one-time price adjustment; it is a structural collapse driven by algorithmic efficiency progress of roughly 5× per year . A 2026 NBER study confirmed that algorithmic efficiency doubles approximately every year, while chip efficiency doubles every two years—two exponential curves forming a positive feedback loop unlike anything economists have seen in any other industry.
What This Means for the Road to AGI
The stacked exponentials reveal a clear trajectory:
Inference costs are crashing. Multiplying the Density Law by Moore's Law yields an even more striking number: the maximum effective model size that can run on chips at the same price roughly doubles every 88 days. In the next three to five years, running a current top-tier GPT-level model on an ordinary laptop or even a mobile phone may be a reality.
On-device AI is accelerating faster than Moore's Law. Between May 2024 and May 2026, the smartest open-weight model that could run on a top-tier MacBook Pro went from a score of 10 (Llama 3 70B) to 47 (DeepSeek V4 Flash) on the Artificial Analysis Intelligence Index—a 4.7× improvement in 24 months, or a doubling of intelligence every 10.7 months. This is more than twice as fast as Moore's Law, on completely unchanged hardware.
The "race" framing is outdated. The Density Law reveals that any state's strongest model only has an optimal window of a few months. Throwing all resources into training a larger model, only to be surpassed by a new model half the size three months later, is economically unwise. The sustainable path is investing in density itself: better architectures, higher quality data, smarter training algorithms.
The AGI Horizon: What to Expect
The self-reinforcing nature of AI feedback loops is already visible. Economists simulating this process conclude that with full automation of software development plus only 5% automation in other industries, the singularity event—explosive growth—could be triggered in approximately six years. Anthropic's Jack Clark, after reviewing all publicly available information, estimates a greater than 60% probability of AI research without human involvement by the end of 2028.
Yet this acceleration creates a paradox. Demand for token consumption is growing 10× annually, while global AI computing power grows only 3.4× annually. The gap widens every year. Compute remains the bottleneck, and as JPMorgan's data shows, GPU rental prices are still climbing despite massive new capacity coming online. A gigawatt-scale power requirement is on track for this decade, restricting the apex to those who can secure grid connections at scale.
A Governance Window
For Global Future Nexus, the implications are clear. The collapse of the cost of intelligence will democratize capability in ways that could boost innovation but also make oversight far more difficult. MIT researchers describe a brief "governance window"—a period when large organizations still have an advantage—as a chance for regulators to develop stronger safety standards before advanced AI becomes ubiquitous.
The question is not whether AGI will arrive, but what form it will take and who will control it. The curves are rising, doubling every few months. The governance response must move at the same speed.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
