Frontier AI

Frontier AI describes the small set of the most advanced, general-purpose AI systems — typically large language models and their multimodal descendants — developed by a handful of well-funded laboratories at the edge of current capability. Because no single metric captures "most advanced", regulators and developers define the frontier mainly through training compute thresholds, with 10^25 and 10^26 floating-point operations (FLOP) the most widely used reference points.^[1][2]

Background

The term "frontier model" entered policy vocabulary as generative AI advanced rapidly after 2022. Policymakers wanted a way to impose heightened obligations only on the most capable systems rather than on every piece of software using AI. The concept also carried an epistemic argument: the capabilities of each new frontier generation are hard to predict in advance, so transparency and evaluation obligations should attach to systems that push into that unknown territory.^[1]

In the United States, Executive Order 14110 (October 2023) used a placeholder threshold of 10^26 FLOP to flag models posing national security risks; it was revoked in January 2025 and replaced by an order emphasising American competitiveness. California's SB 1047 (vetoed in September 2024) proposed training-cost and compute thresholds; its successor, the Transparency in Frontier Artificial Intelligence Act (SB 53), was signed on 29 September 2025 and requires large developers to publish safety frameworks and incident reports for models trained above 10^26 FLOP.^[2][3] In the European Union, the AI Act treats general-purpose AI models trained above 10^25 FLOP as having systemic risk, triggering documentation, evaluation and serious-incident reporting duties.^[1]

Diplomacy moved in parallel: the AI Safety Summit at Bletchley Park (2023) was followed by Seoul (2024) and Paris (2025), while industry created the Frontier Model Forum in 2023 as a self-governance venue. Developers also publish responsible scaling policies that tie deployment decisions to internal capability thresholds.^[3]

Key Concepts

Three properties define the frontier in practice. Scale — frontier training runs consume thousands of accelerators for weeks and cost hundreds of millions of dollars. Generality — the resulting models perform many unrelated tasks from one pre-trained checkpoint. Emerging risk concentration — the same concentration means a few companies' choices shape what the public, governments and businesses everywhere can access.

Compute is the preferred regulatory yardstick because it is measurable and verifiable before deployment, but critics note it is an imperfect proxy: algorithmic efficiency improves, fine-tuning can add capability below thresholds, and different model families spend compute differently. The California report on frontier AI policy documents these scoping trade-offs directly.^[1]

Applications and Impact

Frontier models are consumed mostly as hosted APIs, which means their influence far exceeds the handful of builders: they power search assistants, coding tools, call centres and scientific workflows across industries, and each generation resets expectations for what "automation" can do. Their build-out has also become an infrastructure phenomenon — gigawatt-scale data centre projects, long-term power contracts and a global demand spike for accelerators and high-bandwidth memory.^[3]

>Key Takeaways

• Frontier AI means the most capable general-purpose models, defined in policy mainly by compute thresholds of 10^25–10^26 FLOP.
• The EU AI Act, California's SB 53 and earlier US executive orders each drew thresholds for heightened obligations.
• A handful of laboratories control frontier development; the models reach everyone else through APIs and cloud services.
• Compute is an imperfect but verifiable proxy for capability, and regulators deliberately keep thresholds updateable.
• Malaysia's priorities are governed access: PDPA-safe procurement, framework alignment, and shared compute infrastructure.

See Also

• EU AI Act
• AI Safety
• Bletchley Declaration
• Sovereign AI
• National AI Office Malaysia
• AI Data Centres in Malaysia

References