Frontier AI Compute Thresholds: The 10^25 and 10^26 FLOP Standards Explained

Global regulators have anchored their oversight of artificial intelligence to a highly technical metric: the floating-point operation, or FLOP. As the European Union activates its sweeping AI enforcement powers in 2026, the debate over whether 10^25 or 10^26 operations is the correct threshold defines the frontier of AI policy.[14]

A FLOP is a single arithmetic calculation, like adding or multiplying two decimal numbers. When an artificial intelligence model is trained, it performs these operations trillions of times over across massive clusters of specialized graphics processing units.[3]

Counting these operations provides a quantifiable, objective measure of a model's scale before it is ever deployed. Because training compute has historically correlated tightly with a model's capabilities, governments use it as an initial filter to identify systems that warrant regulatory scrutiny.[3]

The division between jurisdictions is stark. The European Union presumes systemic risk at 10^25 FLOPs, capturing several current-generation models. Meanwhile, California, New York, and South Korea have set their legislative thresholds an order of magnitude higher at 10^26 FLOPs, a scale that targets the next generation of frontier development.[7][12]

The Mathematics of a FLOP

Estimating the training compute of a dense transformer model relies on a standard industry heuristic. The usual rule of thumb multiplies the number of parameters by the number of training tokens, and then multiplies that product by six.[3]

The multiplier of six accounts for the operations required during the training process. Specifically, it represents roughly two operations per parameter per token in the forward pass, and about four operations in the backward pass.[3]

For mixture-of-experts models, the calculation shifts slightly. Analysts count only the parameters that are actively engaged for each token, rather than the total parameter count of the entire model, to accurately reflect the computational load.[3]

When developers do not disclose their parameter counts or training data sizes, outsiders must estimate compute from benchmark scores or hardware utilization. These outside estimates can be highly speculative, often missing the mark by a factor of two or more.[3]

To avoid this guesswork, modern artificial intelligence legislation places the reporting burden directly on the developers. By requiring companies to track and disclose their own compute metrics, regulators bypass the need for external estimation and ensure accurate classification.[3][4]

The precision of the operations also matters. Most frontier laboratories train their models using lower-precision formats, such as 8-bit floating-point, which allows hardware to perform calculations faster and more efficiently than traditional formats.[10]

When regulators define a FLOP, they typically include both integer and floating-point operations, regardless of the precision used. This broad definition prevents developers from circumventing thresholds simply by switching to a different mathematical format during the training run.[6][8]

The European Union's 10^25 Standard

The European Union AI Act establishes a two-tier framework for general-purpose AI models. While all models face baseline transparency requirements, Article 51 creates a rebuttable presumption of systemic risk for any model trained with a cumulative compute exceeding 10^25 FLOPs.[4]

This threshold is not an outright ban, but an administrative trigger for closer scrutiny. Providers of models that cross the 10^25 line must notify the European Commission within two weeks and adhere to strict obligations under Article 55 of the Act.[4][5]

These obligations include conducting state-of-the-art model evaluations, assessing and mitigating systemic risks at the Union level, and ensuring the cybersecurity of the model and its physical infrastructure. Providers must also track and report any serious incidents.[4][5]

The 10^25 threshold captures several models that exist today. Epoch AI counted 33 publicly disclosed models that it estimated at or above 10^25 FLOPs as of its June 2025 update, beginning with GPT-4 in March 2023.[3]

The Digital Omnibus on AI, which entered into force in July 2026, maintained the 10^25 FLOPs presumption without alteration. The European AI Office now enforces these rules, requiring full compliance from systemic-risk models regardless of whether they use open-source licenses.[5]

The European framework also addresses the modification of existing models. If a downstream developer fine-tunes a general-purpose model in a way that significantly alters its capabilities, they become the new provider under the law and must reassess the systemic risk.[4]

The Commission presumes a significant change occurs when the compute used for modification exceeds one-third of the original model's training compute. This ensures that heavily modified systems do not escape the 10^25 FLOPs scrutiny simply because they were built on an older foundation.[4]

The United States and the 10^26 Line

In the United States, the federal government initially set its sights an order of magnitude higher. President Biden's Executive Order 14110, signed in October 2023, established a 10^26 FLOPs threshold for dual-use foundation models, triggering mandatory reporting to the Department of Commerce.[9]

The Bureau of Industry and Security proposed rules in September 2024 to formalize this requirement. Covered American persons were expected to provide quarterly reports on any AI model training run using more than 10^26 computational operations, sharing safety testing results and cybersecurity measures.[6]

However, the federal landscape shifted dramatically in 2025. Executive Order 14148 revoked the original Biden directive on January 20, 2025, eliminating the federal 10^26 reporting rule and leaving a vacuum that state legislatures quickly moved to fill.[3]

California enacted Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, in September 2025. The law applies to foundation models trained using more than 10^26 FLOPs, requiring developers to publish safety frameworks and report critical safety incidents to the state.[11]

New York followed suit with the RAISE Act, which also adopted the 10^26 FLOPs threshold to define a frontier model. Under the amended law, developers must submit transparency reports before deploying new models and conduct applicability assessments based on their compute scale.[8]

Beyond California and New York, the 10^26 FLOPs threshold has become a de facto standard for American state-level legislation. It provides a clear, bright line that exempts small startups and academic researchers from the heavy compliance burdens placed on massive corporate laboratories.[11]

The threshold also applies to the hardware clusters themselves. Federal and state proposals frequently include reporting requirements for computing clusters capable of performing 10^20 operations per second, ensuring oversight of the physical infrastructure required to reach the 10^26 training mark.[6][9]

Global Adoption and the Asian Market

The 10^26 FLOPs standard has also gained traction internationally. South Korea's AI Basic Act, which took full effect on January 22, 2026, imposes extra duties on advanced AI models trained with a cumulative compute of at least 10^26 FLOPs.[12][13]

South Korea's framework is notable for its explicit extraterritorial reach. The law applies to any foreign business whose AI activities affect users in the Korean market, meaning global developers must comply if their 10^26 models are utilized by Korean subsidiaries or consumers.[12]

By setting the bar at 10^26 FLOPs, South Korea aligned its advanced-AI obligations with the American state-level approach rather than the European Union's 10^25 standard. This higher threshold ensures that only a small handful of global frontier-model developers face the most stringent requirements.[12]

The AI Basic Act in South Korea also introduces a grace period for its most severe penalties. While the law is in force, the government is deferring direct fines through at least January 2027, allowing companies time to build compliance programs for their 10^26 models.[12][13]

This grace period reflects a broader Asian strategy of balancing safety with industrial promotion. South Korea's framework includes heavy investments in data infrastructure and talent, ensuring that the 10^26 threshold does not stifle the domestic innovation ecosystem it seeks to build.[12]

Other Asian nations have taken different paths. Japan's AI Promotion Act of 2025 relies on guidance and voluntary frameworks rather than strict compute thresholds, while China continues to enforce targeted, multi-regulator rules that focus on specific applications rather than raw computational scale.[13]

The Financial Cost of Compute

The difference between 10^25 and 10^26 FLOPs is not merely a zero on a page; it represents a massive leap in financial investment. Training an AI model requires vast clusters of specialized hardware running continuously for months, consuming enormous amounts of electricity.[10]

Industry analysts estimate that training a model to the European Union's 10^25 FLOPs threshold costs between $7 million and $10 million in compute resources. This price point is accessible to well-funded startups and established tech companies, capturing a broad swath of the industry.[10]

In contrast, reaching the 10^26 FLOPs threshold requires an estimated $70 million to $100 million in training compute alone. This financial barrier restricts the development of 10^26 models to a select group of heavily capitalized frontier laboratories and major cloud providers.[10]

California's legislative history reflects this financial reality. Earlier proposals explicitly paired the 10^26 FLOPs threshold with a requirement that the model cost at least $100 million to train, ensuring the law only targeted the most resource-intensive projects in the world.[11]

As hardware efficiency improves and chip costs fluctuate, the financial proxy for compute may shift. However, the order-of-magnitude difference ensures that 10^26 models will remain the exclusive domain of frontier developers for the foreseeable future.[10][11]

Why the Compute Proxy is Breaking Down

Despite its widespread adoption in policy, the raw sum of pre-training compute is becoming a progressively worse proxy for model capabilities. When the EU AI Act was drafted, pre-training accounted for up to 99 percent of total training compute, making it a reliable metric.[2]

Modern AI development now leans heavily on post-training techniques. Methods like distillation, synthetic data generation, and reasoning post-training consume significant computational resources and drive massive capability gains, yet they are often unaccounted for in simple pre-training thresholds.[2]

Reasoning training compute has grown roughly ten times every three to five months, far outpacing the annual growth in pre-training compute. If this trend continues, post-training could soon constitute the majority of total training compute in frontier models.[2]

This shift creates a regulatory blind spot. A future model could be trained with a fraction of the pre-training compute specified by existing thresholds, yet achieve capabilities that far exceed those the threshold was intended to capture, effectively bypassing the regulations.[2]

Furthermore, algorithmic efficiency continues to improve. Researchers are constantly discovering new architectures and training methods that extract more performance from fewer FLOPs, loosening the historical correlation between raw computational scale and the resulting intelligence of the system.[2]

The Future of Compute Thresholds

Regulators are beginning to acknowledge the limitations of static compute thresholds. California's SB 53 explicitly includes compute used for subsequent fine-tuning, reinforcement learning, or other material modifications in its 10^26 FLOPs calculation, attempting to capture post-training enhancements.[11]

The European Union AI Act includes provisions to update its 10^25 FLOPs threshold in light of technological advances. The European Commission can also designate a model as posing systemic risk based on its capabilities or impact, providing an alternative route to classification.[4]

Policy experts have proposed alternative frameworks, such as thresholds that move dynamically with the frontier. This could involve setting the regulatory line at a specific percentile of the most compute-intensive models, rather than a fixed numerical value that degrades over time.[3]

Other proposals suggest keeping compute as an initial, inexpensive filter, but relying on rigorous capability evaluations to determine the actual risk profile of a model. This hybrid approach pairs the objectivity of FLOP counting with the nuance of empirical testing.[3]

For now, compute thresholds persist because they are cheap to apply, hard to dispute, and measurable before a model is ever deployed. They remain the foundational metric of global AI governance, defining the regulatory reality for the world's most advanced laboratories.[3][14]