Generative AI and Analog Computing: The End of Moore's Law?

What happened?

In recent weeks, two developments have shaken the tech landscape. On one hand, the Fable model has managed to write a 'megakernel' for GPU that outperforms all previous solutions, achieving an 18.71X speedup over optimized PyTorch. This achievement, validated on the KernelBench-Mega benchmark, demonstrates that AI can automate fundamental R&D tasks in chips. According to the Fable team, the AI-written kernel executes a single cooperative call per token, while other systems require between 4 and 14 separate calls. This not only reduces latency but also optimizes memory and bandwidth usage. In comparison, Claude Opus 4.8 achieved 14.4X, GLM-5.2 reached 11.14X, and GPT 5.5 obtained 4.34X, all using Triton instead of CUDA. The gap is significant and suggests Fable has found a more efficient kernel fusion strategy.

On the other hand, the Remote Labor Index (RLI) shows that the success rate of AI systems in online freelance projects has risen from 2.5% in October 2025 to 16.1% in July 2026, with models like Fable 5 reaching 16.1%. This index, published by Import AI, measures AI's ability to complete complex design, programming, and analysis tasks on platforms like Upwork and Freelancer. The growth is exponential: in just nine months, the rate multiplied by six. If the trend continues, AI is expected to complete over 50% of high-complexity freelance projects by the end of 2027.

In parallel, interest in analog computing is resurging, which uses continuous signals instead of discrete ones to perform calculations. Companies like Analog Inference and Mythic are developing chips that promise up to 1000 times lower energy consumption than digital ones for certain AI workloads. For example, the Mythic M1076 chip can perform neural network inferences with only 25 watts, while an equivalent GPU consumes over 250 watts. Moreover, analog computing is not limited by Moore's Law in the same sense as digital, as its performance depends more on analog precision than transistor miniaturization.

Why is it important?

AI's ability to write GPU kernels represents a step toward recursive self-improvement (RSI) in R&D. If systems can improve their own hardware and software, we could enter a cycle of exponential improvement. The Fable kernel is a concrete example: AI generated CUDA code that surpasses optimized human implementations, and that code could be used to accelerate training of future AI models. This type of positive feedback is the basis of RSI, a concept that was theoretical until now but is beginning to have empirical evidence.

Furthermore, the automation of high-value freelance tasks (design, programming, analysis) threatens to redefine the global labor market, especially in economies dependent on remote work. According to the RLI, AI already directly competes with freelancers in areas like web development, graphic design, and data analysis. If the success rate continues to grow at the current pace, millions of workers could be displaced in the next five years. However, it could also create new opportunities in supervision, verification, and tasks requiring human creativity.

Analog computing, meanwhile, offers a way to bypass the physical limits of Moore's Law. At a time when transistor miniaturization is slowing down (3nm nodes are already in production and 2nm nodes are expected by 2025), analog chips could maintain the pace of progress in efficiency and performance for AI. Companies like IBM and Intel are also researching neuromorphic chips that combine analog and digital principles, indicating that the industry is seeking alternatives to traditional scaling.

Consequences for businesses and users

• Tech companies: must invest in AI infrastructure for R&D or risk falling behind. Analog chip startups could attract massive investment. For example, Analog Inference received $50 million in Series B funding in 2025, and Mythic already has agreements with IoT device manufacturers. Major companies like NVIDIA and AMD are also exploring hybrid analog-digital architectures.
• Freelance workers: competition with AI intensifies. Specialization in tasks requiring human creativity or oversight will be key. According to the RLI, the most vulnerable tasks are those with high levels of standardization, such as creating web templates or basic data analysis. In contrast, tasks involving contextual judgment or personal interaction remain difficult for AI.
• End users: cheaper and faster products and services, but possible market concentration in a few companies that own the technology. If AI automates R&D, companies controlling the most advanced models could dominate entire sectors. On the other hand, analog computing could democratize AI by reducing energy costs, allowing low-power devices to run complex models.

What should readers know?

First, that R&D automation is not science fiction: there are already benchmarks measuring it, like KernelBench-Mega, and the results are impressive. Second, analog computing is not new, but its resurgence is tied to the demand for efficiency in AI. From the 1940s to the 1960s, analog computers like ENIAC were used for physical simulations but were replaced by digital ones due to precision. Now, error tolerance in neural networks makes analog precision sufficient, and energy efficiency becomes critical. Third, these changes imply urgent political and economic decisions about labor regulation and intellectual property. For example, who owns the rights to an AI-written kernel? How are revenues from automated work taxed? Countries like the European Union are already discussing robot taxes, but the speed of change may outpace legislation.

“AI is learning to design its own brains. Analog computing could be the substrate enabling the next generation of intelligent systems.” — TheVortiq

Historical context

Analog computing dominated the early decades of computing (1940s-1960s) but was displaced by digital due to precision and scalability. Now, the rise of AI, where exact precision is not always necessary, revitalizes the analog approach. For example, in image recognition, a 1% error in neural network weights barely affects final accuracy but drastically reduces energy consumption. Meanwhile, the automation of cognitive tasks recalls the Industrial Revolution, but this time it affects intellectual work. In the 1990s, factory automation eliminated millions of manufacturing jobs but created new service positions. Today, AI threatens office and creative jobs, and the transition could be faster. According to a 2023 McKinsey study, up to 30% of work tasks could be automated by 2030, and the RLI suggests that estimate may be conservative.

In the realm of GPU kernels, progress has been gradual. The first AI-written kernels appeared in 2023 with models like GPT-4, but only achieved 2-3X speedups. In 2024, Claude Opus 3 reached 8X, and now Fable exceeds 18X. This evolution in just three years shows an acceleration that could lead to 100X speedups by 2027. If that happens, the impact on AI performance would be massive: training models like GPT-5 could be reduced from months to days, and inference on mobile devices would be much more efficient.

In summary, we are witnessing a convergence of three trends: AI automating its own development, analog computing breaking physical barriers, and a labor market transforming at unprecedented speed. The coming years will be critical in defining how society adapts to these changes.