The Efficiency Paradox: Why Algorithmic Progress Fuels, Not Cuts, Compute Spending
In recent months, a familiar narrative has echoed through tech circles: as AI models become more efficient, the need for compute will shrink. The success of DeepSeek's V3 and R1 models, achieving performance comparable to far larger predecessors on a fraction of the hardware, seemed to vindicate this view. But the data tells a different story. Algorithmic progress has instead coincided with a rapid surge in compute investment, contradicting the notion that efficiency alone will reduce spending. The reason is both simple and profound: when efficiency unlocks new capabilities, the market does what it always does — it expands demand to fill the newly available space.
The Performance Effect: Bigger Models, Better Results
The most direct evidence comes from the empirical record. Epoch AI's analysis shows that algorithmic progress has accompanied a rapidly increasing rate of investment in computers used for training and deploying machine learning models. This flatly contradicts the idea that methodological improvements will necessarily lower total spend. The performance effect lies at the heart of this counterintuitive outcome. Algorithmic progress makes larger training runs more lucrative because it increases the anticipated performance of a model trained at any given compute level. When a breakthrough like DeepSeek's V3 arrives — achieving Llama 3 405B-level performance using roughly one-tenth the compute — it doesn't dissuade labs from scaling up. Rather, it raises the ceiling of what's possible, creating powerful incentives to invest in even bigger runs. Neural scaling laws continue to predict performance gains from increased compute, and DeepSeek's own innovations in multi-head latent attention, auxiliary-loss-free load balancing, and shared experts are expected to improve performance at unprecedented scales. Unlike model distillation, which primarily grants access, these innovations carry a strong performance effect, making large-scale scaling more attractive than ever.
The Substitution Effect: AI Within Reach
The second engine is the substitution effect. Algorithmic progress typically enables more people to afford AI services at higher quality than before. When AI can substitute well for other production inputs — above all, human labor, the overall effect is to spur higher spending on AI deployment, despite decreased per-unit costs. The data bears this out: the amount of compute needed to achieve a given level of pre-training performance for large language models has been shrinking on average by a factor of approximately three each year. If AI reaches the point where it can substitute for human workers across a wide range of industries, compute spending could rise to roughly match the labor share of GDP, currently about 60% in the U.S., versus roughly 0.6% of GDP spent on all computing hardware today. That would represent a 100-fold increase.
The Jevons Paradox and Its Limits
The analogy to Jevons paradox, the historical pattern whereby efficiency gains lower resource costs and trigger a rebound effect that increases total consumption, is tempting but ultimately insufficient to explain what's happening in AI. Most empirical studies suggest that in today's energy markets, the rebound effect is typically not strong enough to cause total consumption to increase as efficiency improves. More importantly, there's no economic law stating that greater efficiency in obtaining or using a resource must lead to increased consumption. The strength of the rebound effect varies widely by market. What makes AI distinctive is not merely its efficiency trajectory but its capacity to automate labor on a massive scale. Historical computing platforms like personal computers and smartphones followed a pattern of initial surge followed by saturation as markets filled. AI may diverge from this pattern precisely because it can become a general-purpose worker substitute rather than a saturated consumer product.
AI as a Unique Computing Product
The case for AI's unusual economic trajectory rests on its potential to automate remote work and, ultimately, to substitute for human labor across nearly all tasks. AI labs are focused on unlocking capabilities that could open the door to enormous economic value: multimodal agents that act in digital environments over long horizons, general-purpose robots that perform physical tasks, systems that function as genuine AI workers. Such developments would represent a fundamental departure from conventional computing. Merely creating autonomous digital workers could more than double GDP under conservative assumptions. If AGI, AI capable of substituting for human workers across all labor tasks, arrives, credible analyses indicate GDP could grow by more than 30% per year. In that scenario, businesses would have a strong financial incentive to increase compute spending to a level comparable to current wage expenditure, driving compute spending toward the labor share of GDP.
The Saturation Question
Of course, none of this is inevitable. If the AI paradigm plateaus at today's capability level, if deep learning hits a wall and AI development stalls at the level of current chatbots, then the efficiency-driven spending decline would likely play out. Market saturation is a real possibility for any technology. But the past decade of AI development offers scant support for such a near-term plateau. Scaling laws have repeatedly demonstrated that increased compute budgets lead to entirely new qualitative capabilities. The historical rate of compute scaling can likely be sustained until at least 2030, provided investors fund the necessary infrastructure. Algorithmic progress, by enabling frontier model capabilities sooner, may be the key factor that secures continued funding for compute scaling.
Conclusion
The expectation that compute spending will decline as algorithms improve makes sense only if one believes AI is already a finished product, its primary challenge being mere access expansion. Under that view, algorithmic progress would mainly make chatbot production cheaper, and saturated markets would naturally see total compute spending decline. But if one believes, as the evidence strongly suggests, that the next major step in AI development is creating systems that can automate many forms of human labor, then the potential for compute spending to rise is far greater. The past decade's clear trend supports this: as AI capabilities expanded, so did demand for compute, even as efficiency improvements lowered costs. If further algorithmic progress, combined with compute scaling, continues to unlock new and useful AI capabilities, there is strong reason to expect compute spending to grow rather than shrink.
The efficiency paradox is real, but its direction is opposite to what many assume. Algorithmic progress does not reduce the need for compute; it激活 it, by making the previously unaffordable suddenly within reach, and the previously impossible suddenly achievable. The question is not whether AI will drive compute spending higher, but how high that spending will climb, and whether the economic benefits it unleashes will prove worthy of the investment.
Matthew Barnett, Epoch AI, Gradient Updates, February 2025
Citation: Barnett, M. (2025). "Algorithmic progress likely spurs more spending on compute, not less." Published online at epoch.ai. Retrieved from https://epoch.ai/gradient-updates/algorithmic-progress-likely-spurs-more-spending-on-compute-not-less
