The rapid progress and adoption of large language models in recent years have sparked extensive discussions about how artificial intelligence (AI) will shape the future of our economy. Central to these discussions are questions about whether AI will substantially accelerate economic growth, how much investment AI will attract, the timing and scale of these investments, and the speed at which automation will transform labor markets.

To advance this discussion, we introduce the Growth and AI Transition Endogenous (GATE) model. GATE brings together concepts from machine learning and economic growth theory to illustrate the key dynamics of AI development, task automation and their downstream macroeconomic effects. It draws heavily on scaling laws—empirical regularities relating compute scaling to performance for both training and inference—and semi-endogenous growth—a theory that explains economic growth as a result of R&D efforts that generate scientific advances. You can find a technical description of the model in our whitepaper.

Alongside our paper, we are releasing an interactive model. This tool lets you simulate a variety of scenarios for AI development and economic growth. Users can change model parameters and visualize the resulting path of key economic and AI development variables. The tool makes explicit some of the key assumptions and parameters that shape the trajectory of AI automation, which we hope will lead to a more transparent debate about the economic impacts of AI.

Explore GATE

A brief introduction to GATE

GATE synthesizes many years of our research into algorithmic progress, compute-centric views of AI development, returns to R&D investment, and estimates of the fundamental limits of CMOS processor efficiency. Our previous work consistently identified compute scaling and algorithmic efficiency as central drivers of AI advancement. We’ve placed these insights at the core of GATE: increasing compute resources and ongoing algorithmic breakthroughs directly enhance AI capabilities, enabling the automation of a broader set of economic tasks. This automation accelerates economic growth, a portion of which is reinvested into further AI development, creating a powerful feedback loop between compute, innovation, and growth.

The GATE model has three core components:

Compute. Investments in computing infrastructure grow the available computing resources for developing and deploying AI systems. Moreover, investments in hardware and software R&D result in more efficient hardware and AI algorithms.

Automation. Part of the installed computing capacity is dedicated to AI training. In keeping with the findings of the scaling literature, models trained on more compute are capable of automating more tasks. Moreover, as with reasoning models, more tasks can be automated by increasing the amount of compute used during inference. The model combines both training and inference scaling to determine the feasibility and cost of progressively automating more tasks.

Production. The compute reserved for inference is used to run large quantities of “digital workers” that perform automatable tasks. This digital workforce is combined with human labour, capital and natural resources to produce economic output, part of which is then reinvested into capital and AI development. The high-level decisions of allocating investment and compute are handled by a social planner that maximizes the lifetime utility of a representative household.¹

Each GATE module incorporates realistic features to generate plausible scenarios of AI-driven economic growth. For instance, the AI development module captures diminishing returns on R&D investments and adjustment costs associated with rapidly scaling up computing infrastructure. Users can explore a variety of scenarios by adjusting key assumptions—such as the pace of hardware efficiency improvements or the magnitude of algorithmic breakthroughs—through the interactive model.

Additionally, GATE provides optional components to examine the impact of investor uncertainty, which can limit large upfront investments, and R&D externalities, where benefits of innovation spill over to competitors and the broader economy. These components help illustrate how real-world frictions might influence the trajectory of AI automation.

Read our model documentation for an accessible overview, or dive into the whitepaper for a detailed technical exposition.

Preliminary insights

This first version of the model helps us explore important questions about AI’s economic impact, though we’re still working to understand how robust its findings are. Let’s look at some of the most interesting things we’ve learned so far while testing GATE’s behavior—keeping in mind these are early results that need more investigation. We plan to release a follow-up technical paper exploring the GATE’s predictions in more detail in the near future.

Enormous AI investments. GATE predicts that the investment in global compute supply may exceed 10% of world GDP, an approximately 50-fold increase over current levels. This investment precedes AI value generation, as economic agents anticipate vast potential returns from eventual task automation. Despite generating little initial economic output, substantial resources are reallocated from conventional capital to compute-related infrastructure (fabs, data centers, etc.) due to the immense incentives to accelerate AI automation timelines.

Large scale-up of compute. Even under conservative assumptions—such as modest returns to hardware and software R&D—the model consistently projects that the global economy can marshal enough effective compute to automate most tasks within two decades. Note that we assume that any task ultimately becomes automatable once sufficient compute and innovations accumulate.

Massive economic growth. The GATE model predicts that AI automation leads to significantly accelerated economic growth, with rates elevated by 2-20 times compared to the recent historical average of ~3% per year during the period of automation. As automation progresses, growth rates steadily increase, with particularly dramatic acceleration at higher levels of automation. These dramatic economic gains occur despite production bottlenecks created by complementarities between tasks in the economy, and they materialize over roughly a decade as AI systems progressively automate a larger fraction of economic tasks.

It is important to note that while we believe GATE is a significant step forward in our understanding of AI and its economic implications, the model still has significant limitations. For example, we don’t model the advance of non-AI technologies, which would lead to even faster economic growth. Conversely, we also assume that disembodied AIs can perfectly substitute human labor across all tasks, when in practice they will likely require additional equipment such as robots. This results in some unrealistic predictions; for example, with our default choice of parameters the model immediately assigns 20% of the world economic output (nearly $20T/year) to AI development. A detailed discussion of the model limitations can be found in our whitepaper.

Because of these limitations, we don’t recommend interpreting the model outcomes as precise quantitative predictions. Instead, these results are best treated as a qualitative description of some key dynamics that we are likely to see play out in AI. Our upcoming work on AI modeling will examine how robust these dynamics are.

Conclusion and next steps

The GATE model provides a compute-centric framework to systematically explore the potential trajectory and macroeconomic impacts of AI-driven automation.

While this approach significantly advances our understanding of interactions between AI investment, technology improvements, and labor automation, several important limitations remain.

In the coming months, we plan to engage researchers and other GATE users to prioritize and address the more important limitations of the model. We also plan to release a detailed technical analysis of the model dynamics and predictions. We invite researchers to engage and provide their feedback — you can contact us at info@epoch.ai.

We have already begun this process, and would like to thank Ben Golub and Ryan Greenblatt for their extensive reviews. We would also like to thank Chad Jones, Tyler Cowen, and Anton Korinek for their engagement and feedback.

Notes

1. High level decisions to allocate investment and compute in the model are solved via gradient descent. The allocation of inference compute per task, human labor, capital and land is solved analytically. ↩