Feature
For decades, Moore’s Law — the idea that microchip transistor counts double every two years as costs drop — has driven semiconductor progress. But as AI chip development accelerates, experts are questioning its relevance.
Specialized chips such as graphics processing units (GPUs), tensor processing units (TPUs), neural processing units (NPUs) and custom accelerators are pushing traditional scaling models to their limits and driving innovations beyond them. Some believe artificial intelligence (AI) hardware has already outgrown Moore’s Law, while others argue that future breakthroughs will depend more on architectural and algorithmic innovation than transistor density.
Table of Contents
- Why Moore’s Law Is Important
- Understanding Moore’s Law in the Context of AI Chips
- Has AI Hardware Surpassed Moore’s Law?
- Parallel Processing Breakthroughs
- The Post-Moore’s Law Era: What It Means for AI Chip Development
- Impact on AI Development and Deployment
- The Future of AI Hardware
Why Moore’s Law Is Important
Moore’s Law, introduced by Gordon Moore in 1965, predicted that microchip transistor counts would double every two years, boosting computing power while lowering costs. This principle drove decades of technological breakthroughs, from personal computers to massive data centers. But AI’s rise has strained this model. Modern AI workloads are too complex for traditional scaling to keep up, shifting the focus toward new approaches like energy-efficient designs, specialized chips and architectural innovations. Today’s AI chips are optimized for tasks beyond what Moore’s Law originally envisioned.
This shift raises fundamental questions: Is Moore’s Law still applicable in the AI era, or has it been rendered obsolete by the demands of modern AI applications?
Some argue that AI chip development has outpaced the traditional principles of Moore’s Law, relying instead on breakthroughs in hardware design, parallel computing and software optimization. Others contend that while traditional transistor scaling may be slowing, advancements in chip packaging, materials science and new computing concepts continue to push performance improvements — while Nvidia outright says they are outpacing Moore’s Law.
The increasing complexity of AI workloads is outpacing traditional transistor scaling, prompting innovations that go beyond Moore’s Law. Adam Yong, founder of AgilityWriter, said, "Moore’s Law has been a cornerstone of semiconductor progress for decades, but it is becoming less relevant in the context of AI chip development. AI workloads are highly specialized and usually demand chips that are not just faster but tailored to specific tasks."
He added that specialized architectures like GPUs, TPUs and custom accelerators now offer much better performance for AI tasks than simply scaling up transistor density. Plus, AI workloads have outgrown traditional scaling, shifting the focus to specialized architectures and algorithmic innovations that optimize AI performance.
Related Article: Meet the Startups Taking on Big Tech With Smarter AI Chips
Understanding Moore’s Law in the Context of AI Chips
Moore’s Law has long served as the benchmark for general-purpose computing, driving the exponential growth of processing power in central processing units (CPUs). Traditionally, the doubling of transistor density every two years meant that CPUs could deliver faster computations while becoming more power-efficient and cost-effective. This steady progression fueled advancements across industries. However, the demands of AI have introduced new challenges that general-purpose CPUs struggle to effectively address.
Modern AI tasks like real-time language translation or image recognition require vast computational power, processing data simultaneously rather than sequentially. Traditional CPUs struggle to handle this load, which is why specialized chips such as GPUs and TPUs have become essential to training AI models and making accurate real-time predictions.
The demands of AI workloads also place significant stress on traditional chip designs, leading to challenges in power delivery and heat dissipation. Mike Frank, senior scientist at Vaire Computing, told VKTR that "Regardless of transistor density, chip performance is increasingly limited by power delivery and cooling constraints. Power is currently being addressed through aggressive (e.g., forced liquid) cooling, but there are limits to this approach." He explained that while cooling techniques such as forced liquid cooling help manage heat, the physical constraints of power dissipation still limit AI chip performance.
Although general-purpose CPUs have long driven performance gains under Moore’s Law, AI workloads demand specialized solutions. Nathan Brunner, CEO at Boterview, said, "AI, especially deep learning, thrives on parallel computations. When training a neural network, each neuron in a layer performs computations at the same time as others. These operations require special architectures like GPUs and TPUs. It is now a gold rush — chip manufacturers are racing to build hardware that can handle the immense demands of AI applications."
Unlike traditional computing tasks that focus on sequential processing, AI-specific processing demands emphasize simultaneous operations on large datasets. AI chips are optimized for large-scale computations, reducing delays (latency) during tasks like neural network training and AI applications such as image recognition and language translation. These specialized chips take advantage of architectural innovations such as tensor cores, on-chip memory and advanced interconnects to maximize efficiency and performance.
Moore’s Law focused on transistor scaling, but today’s AI chip development depends on energy efficiency, scalability and real-time data processing. The shift from general-purpose to specialized AI hardware signifies a departure from the traditional scaling model, prompting the industry to explore new approaches such as 3D chip stacking, heterogeneous computing, and domain-specific architectures to meet the growing computational demands of AI-driven applications.
Has AI Hardware Surpassed Moore’s Law?
While traditional transistor scaling remains an important factor, AI chip development has embraced new design principles that extend beyond simple density increases. The demand for faster, more efficient AI processing has led to advancements that challenge the relevance of Moore’s Law as the primary driver of progress.
AI hardware is leaving Moore’s Law behind through innovations in parallel processing and distributed computing. Traditional CPUs process tasks one at a time, but AI demands large-scale parallelism. Specialized chips like GPUs and TPUs use thousands of cores to run neural network calculations all at once. Distributed AI frameworks further scale performance by splitting tasks across multiple devices or cloud systems, bypassing the limits of single-chip transistor density.
Nvidia is one example of a company claiming to have exceeded the constraints of Moore’s Law. Nvidia CEO Jensen Huang argued that Nvidia’s AI chips are advancing at a faster pace due to the company’s ability to innovate across the entire computing stack, including architecture, chips, systems and algorithms. Nvidia’s latest data center chip (GB200 NVL72) is 30 to 40 times faster at running AI inference workloads than its previous best-selling chip. Huang said this kind of performance leap is possible because Nvidia can optimize performance holistically rather than relying solely on transistor density improvements. This holistic approach purportedly allows Nvidia’s AI hardware to deliver exponential performance improvements while reducing the cost of inference over time.
Another significant innovation is 3D stacking and chiplet architectures. As Frank noted, traditional 2D chip layouts are reaching physical limits in terms of heat dissipation and energy efficiency, prompting engineers to explore three-dimensional designs that stack multiple layers of silicon to improve performance without increasing footprint. Chiplet-based designs, where modular processing units are combined to optimize specific tasks, have become increasingly popular. This approach allows manufacturers to scale performance by interconnecting specialized components, improving flexibility and reducing production costs.
Related Article: Supercomputing for AI: A Look at Today’s Most Powerful Systems
Parallel Processing Breakthroughs
AI chip development has embraced parallel processing and new architectures to push beyond traditional scaling.
"Although industry roadmaps aim to continue increasing device densities for a number of years," Frank explained, "we are now in an era of diminishing returns in terms of the effect on compute efficiency and performance, due to increasing quantum effects which limit transistor size and performance, and fundamental thermodynamic limits on transistor switching energies." Physical constraints on transistor scaling, he said, are driving the search for alternative approaches like specialized architectures and parallel processing.
With traditional transistor scaling hitting its limits, custom architectures and emerging technologies are now driving AI hardware innovation. Frank emphasized that "specialized designs and optimized algorithms are increasingly important today, but soon, all the low-hanging fruit in that area will have been plucked as well. Further progress will require fundamental advances through new computing paradigms." While specialized architectures have fueled recent progress, Frank explained, continued breakthroughs will require adopting new paradigms beyond current technological limits.
AI chips face a power challenge: they need to handle enormous workloads without overheating or consuming too much energy. Companies such as Google and Intel are focusing on custom chip designs that maximize efficiency, helping reduce both power use and costs.
For instance, Google's TPU v5 is designed to deliver high performance with reduced power consumption, enabling faster training of machine learning (ML) models at a lower operational cost. Similarly, Intel's Gaudi AI accelerators emphasize energy-efficient deep learning capabilities, making AI workloads more sustainable.
These AI chips represent the cutting edge of AI hardware innovation and a departure from the traditional Moore’s Law scaling model, relying instead on architectural breakthroughs, software optimizations and tailored design approaches.
Webinar
Sep
25
CMS Briefing: A Live Look at What’s Next in AI-Driven Platforms
Learn how leading organizations are using AI‑driven tools to publish faster, personalize smarter and stay secure.
Register
Webinar
On demand
Ready or Not: How Data-First Organizations Are Unlocking Agentforce Potential
Learn how to cut through the noise, activate Agentforce and build a Salesforce AI strategy that actually delivers.
Watch Now
Webinar
On demand
AI in Customer Service: Faster Resolutions, Happier Customers
Don’t let rising demand burn out your team. See how to build a smarter, more resilient support org.
Watch Now
Webinar
On demand
From Hype to High-Impact CX Strategies That Actually Scale
Turn buzzworthy AI and outsourcing trends into measurable CX wins with fresh 2025 data.
Watch Now
Webinar
On demand
Insights to Action Rethinking the Contact Center for Real Business Impact
Join our exclusive webinar to hear CX executives share their innovative strategies for transforming service delivery.
Watch Now
Webinar
On demand
Accelerating Healthcare Ops with AI
Sign up to see how your peers are putting AI to work — and making it stick.
Watch Now
The Post-Moore’s Law Era: What It Means for AI Chip Development
In the post-Moore’s Law era, AI chip development is undergoing major changes driven by the limitations of traditional transistor scaling. As AI workloads become more complex and data-intensive, new design paradigms are emerging to sustain performance gains and efficiency improvements without relying solely on increasing transistor density.
Instead of relying on a one-size-fits-all approach, AI hardware is evolving to include specialized processing units (GPUs, TPUs and NPUs) alongside general-purpose CPUs. This approach enables AI systems to allocate specific tasks to the most suitable hardware, optimizing performance for ML inference, training and edge computing. Domain-specific architectures, tailored to tasks like natural language processing or computer vision, also enhance efficiency by reducing computational overhead and latency.
Integration of AI algorithms directly into hardware is another crucial shift. Traditionally, AI models relied on software optimization to achieve performance improvements, but the increasing complexity of deep learning models has required tighter integration between hardware and software. AI accelerators now feature built-in support for neural network operations, enabling faster and more efficient processing without the need for extensive software tuning. This trend is particularly evident in edge AI devices, where power efficiency and real-time inference capabilities are critical.
The future of AI chip development will rely on diverse computing frameworks that complement each other. Frank told VKTR that his company's approach "offers a near-term prospect to fundamentally change the paradigm for implementing general digital computing. Meanwhile, thermodynamic computing offers utility for randomized computations, and photonic computing provides short-term solutions for increasing communication bandwidth. They each may have their place in heterogeneous computing systems of the future." He added that emerging paradigms like reversible computing, thermodynamic models and photonic solutions will likely coexist in such systems to address different AI workloads.
Emerging technologies like quantum computing and neuromorphic computing could reshape AI chip development. Quantum computing promises to solve complex problems exponentially faster than today’s systems, while neuromorphic chips mimic how the human brain works, delivering AI performance with lower energy use. Although neither is ready to replace current AI hardware, they’re being heavily researched and could shape future breakthroughs.
With the diminishing returns of transistor scaling, emerging technologies like photonic computing and quantum computing are opening new doors for AI hardware. "Photonic computing offers potential short-term solutions for increasing communication bandwidth, while thermodynamic computing could improve the efficiency of specialized AI tasks," suggested Frank. "Meanwhile, reversible computing presents a path for sustained improvements in system efficiency and performance."
Related Article: Why AI Companies Are Creating World Models
Impact on AI Development and Deployment
The performance of AI chips plays a critical role in shaping the development and deployment of AI applications. As AI models grow in complexity, the speed and efficiency of hardware directly impact training and inference processes, driving innovation while posing significant challenges in terms of cost, scalability and optimization.
AI training, which involves processing huge datasets to fine-tune models, requires immense computational power. High-performance AI chips enable faster processing, reducing the time needed to train large-scale models such as GPT and Gemini. Brunner said that the cost of training foundation models is rapidly falling due to improvements in AI hardware, allowing more companies to independently build and deploy their own models. "I believe we will see more open-source models and companies developing their own foundation models from scratch instead of refining existing ones."
Despite advancements in chip performance, the financial aspect of AI adoption remains a critical concern for businesses. The development and deployment of AI models require substantial investment in specialized hardware, cloud infrastructure and energy consumption. High-performance chips, while offering unparalleled processing power, come with significant capital and operational costs. Frank pointed out that novel hardware architectures haven’t yet achieved widespread adoption, but with conventional chip technologies nearing their limits, new players in the AI chip market may gain prominence.
In the face of hardware limitations, software optimization has emerged as a crucial factor in maximizing AI chip performance. Yong emphasized the importance of combining hardware and software optimizations to improve efficiency. "AI chip development has outgrown the traditional scaling model that Moore’s Law predicted. It has now focused on architectural and algorithmic improvements to boost performance."
As conventional chip designs reach their limits, new players and architectures may disrupt the dominance of incumbents such as Nvidia. "So far," said Frank, "novel AI hardware offerings have not really broken through in terms of widespread market adoption. However, as conventional CMOS technology approaches its limits, contenders in the area of specialized AI chips might gain increased prominence." He pointed out that while Nvidia currently dominates AI chip development, emerging technologies could shift market dynamics as traditional designs face limitations.
The Future of AI Hardware
As AI chip development moves beyond the traditional approach of Moore's Law, the industry faces a drastic shift where progress depends on factors beyond transistor density. Innovations in specialized architectures, heterogeneous computing and 3D chip stacking are driving performance improvements, while energy efficiency and cost considerations remain critical. The future of AI hardware will require balancing these advancements with challenges in manufacturing complexity and supply chain constraints.
