The 80% R&D Rule: How Open AI Ecosystems Actually Save Money

There is a pervasive illusion in the tech world regarding open-source artificial intelligence: many assume that open models automatically mean cheaper, off-the-shelf solutions for businesses. In reality, if you want a simple plug-and-play AI solution, using a closed, hosted API is often more cost-effective due to massive economies of scale. So, where does the financial magic of open-source AI actually happen? The answer lies not in deployment, but in development.

To understand this, we have to look at how AI compute is actually spent. Recent data from AI research organizations, including Ai2’s documentation of Olmo 3 and Epoch AI’s analysis of frontier labs, reveals a surprising ratio. Roughly 80% of the computational power required to build a leading model is burned during research and development—trial, error, and experimentation. The final, massive training run that produces the finished model accounts for only a fraction of the total cost.

This economic reality fundamentally changes the purpose of an open AI ecosystem. Traditional open-source software (OSS) thrives on Linus’s Law: "given enough eyeballs, all bugs are shallow." Users globally help fix code and add features, driving down costs. But you cannot easily crowdsource a bug fix in a neural network's weights. In open AI, the original developer bears almost the entire financial burden of creation. The true superpower of open-sourcing a model is that it drastically reduces future R&D costs for the entire developer ecosystem.

The Chinese AI landscape offers a compelling case study of this dynamic in action. Rather than operating in silos, many leading Chinese labs lean heavily into open models and publish incredibly thorough technical reports. By sharing their exact infrastructure setups, training recipes, and—crucially—their failed experiments, these labs effectively de-risk ideas for their peers. This intentional knowledge sharing prevents competitors from "double-spending" millions of dollars on redundant compute to learn the exact same lessons.

However, this compounding effect is currently facing a bottleneck. A common trope in the industry today is for companies to take open-source tools and fork them into proprietary, internal-only versions to gain a competitive edge. For instance, despite the rapid advancement of AI, there is still no fully open, at-scale recipe for the reinforcement learning training of Mixture of Experts (MoE) models. As foundational tools become less accessible, the open ecosystem loses its momentum.

Building a frontier model today is an intricate art of integrating hardware, data, and infrastructure at a breakneck pace. As long as R&D remains the dominant cost, the survival of open AI may depend on moving beyond just releasing model weights. It may require forming shared open-model consortiums—pooling resources so that the entire ecosystem can compound its knowledge and compete with the deep pockets of closed labs.

Key Points

• Approximately 80% of the compute used for frontier AI models goes toward R&D and experimentation, not the final training run.
• Unlike traditional open-source software, open AI relies on reducing future development costs rather than crowdsourcing direct fixes.
• The Chinese AI ecosystem accelerates innovation by publishing thorough technical reports, preventing peers from wasting compute on redundant research.
• The practice of forking open tools into closed, proprietary versions threatens the collaborative compounding effect of open AI.

Why It Matters

Recognizing that the bulk of AI costs lies in R&D rather than final training helps demystify the economics of the AI race, highlighting why collaborative knowledge-sharing is essential for open-source AI to survive against well-funded closed labs.

Sources:

• How open model ecosystems compound — Interconnects (Nathan Lambert)