Why Moore’s Law Still Won’t Die
In the fall of 2025, engineers in Chandler, Arizona did something the semiconductor industry had waited nearly a decade to see: they shipped chips built on a 1.8-nanometer-class manufacturing process before their longtime rival TSMC could get there first. The company was Intel, a firm many analysts had already written off. The achievement briefly reopened a debate that technologists have been having for twenty years and never quite finished: is Moore’s Law, the founding assumption of the digital age, still functioning — or are we watching its last surviving reflexes?
Gordon Moore, a co-founder of Intel, made a simple observation in 1965. The number of components that could be packed onto an integrated circuit was roughly doubling every year, and he expected that trend to continue. He later revised the pace to every two years. What started as an editorial aside in a trade magazine became the closest thing the technology industry has to a law of physics, even though it was never one. For six decades, it set expectations for how fast computers would get faster, how quickly prices would fall, and how soon today’s impossible idea would become tomorrow’s product.
The question of whether that trend is still alive is not a nostalgic one. It shapes how much AI companies can compute, how governments plan industrial policy, and how much every smartphone, car, and data center will cost. To answer it honestly, the industry’s own experts don’t even agree with each other, and understanding why reveals more about the technology than a simple yes or no ever could.
The Bet That Became an Industry
Moore’s original insight was not a law of nature. It was an economic argument disguised as an engineering one. As transistors shrank, more of them could fit on a single sliver of silicon, and because a factory could produce that silicon in bulk, the cost per transistor kept falling even as performance rose. Miniaturization and affordability moved together, and that pairing is what made the prediction so powerful. A company that could shrink its transistors faster than its rivals could sell more computing power for less money, year after year.
For the semiconductor industry, this turned into something close to a shared discipline. Companies structured their entire roadmaps — new factories, new lithography tools, new materials — around hitting the next node on schedule. Missing a generation meant falling behind competitors who could offer smaller, cheaper, faster chips. The doubling became self-reinforcing: engineers didn’t just observe the trend, they organized their careers around meeting it.
Why the Easy Gains Ran Out
For most of Moore’s Law’s history, shrinking a transistor was mostly a matter of refining existing techniques. Chipmakers used increasingly precise lithography to draw finer lines on silicon wafers, and each new generation of tools allowed features to shrink further. That approach worked beautifully until transistors became small enough that ordinary physics started interfering with the plan.
Below roughly 20 nanometers, electrons no longer stayed obediently inside their intended channels. A transistor that is supposed to be either fully on or fully off begins to leak current even when switched off, wasting power and generating heat that a chip cannot dissipate fast enough. The cost of developing and manufacturing cutting-edge microchips has skyrocketed, and next-generation fabrication technologies such as extreme ultraviolet lithography require highly specialized equipment that makes it increasingly difficult for companies to achieve the economies of scale that once drove Moore’s Law. Where “smaller” once reliably meant “cheaper,” that relationship has weakened considerably.
Materials themselves became part of the constraint. Silicon, the traditional material used in semiconductor manufacturing, is reaching its limits in terms of performance and scalability, and as transistor dimensions shrink, the electrical properties of silicon are no longer sufficient to maintain the desired levels of performance. This is the origin of nearly every “Moore’s Law is dead” headline of the past decade: not a single dramatic wall, but a slow accumulation of physical penalties that each new generation had to fight harder to overcome.
The Architecture That Bought More Time
Rather than abandon the trend, engineers changed what they were scaling. The most significant recent shift has been a move away from the transistor design that powered computing for over a decade, called FinFET, toward a new geometry known as gate-all-around, or GAA. In a FinFET transistor, the gate wraps around three sides of a thin silicon fin. In a GAA design, the gate surrounds the channel on all four sides, giving engineers far tighter control over when current flows and when it doesn’t. Intel calls its version RibbonFET; TSMC calls its version Nanosheet GAA. Both are attempts to solve the same leakage problem that plain shrinking could no longer fix.
A second innovation addressed a different bottleneck: getting power into the chip without interfering with its signals. Traditionally, both power and data wiring sit on the same side of a chip, competing for the same limited space. Intel’s answer, called PowerVia, moves power delivery to the back of the wafer, physically separating it from the signal wiring on the front. By moving power routing to the back of the wafer, Intel has decoupled power and signal wires, drastically reducing electrical interference and voltage droop, allowing chips to achieve higher clock speeds at lower power. This single change, more than any transistor shrink, is why Intel’s newest chips can be meaningfully faster without simply running hotter.
These innovations arrived just as two companies converged on the same finish line. Intel’s mid-2025 volume production of its 18A process allowed the American chipmaker to leapfrog the industry’s traditional leader, TSMC, in the race to deploy the next generation of transistor architecture, marking the first time in nearly a decade that the process lead was not a foregone conclusion in favor of TSMC. Weeks later, TSMC answered. TSMC’s N2 process entered early production in 2025 and was expected to expand its production capacity toward broader commercial application in 2026.
The People and Companies Racing to Prove a Point
The debate over whether Moore’s Law survives is not merely academic. It has become a public argument between the two men most responsible for the chips inside modern computers. In 2022, Nvidia’s chief executive, Jensen Huang, declared that Moore’s Law was no longer valid. Intel’s leadership pushed back, insisting the trend remained intact. Neither claim was simply wrong; they were describing different things.
Nvidia’s argument rests on the fact that the old style of scaling — smaller transistors, same design, automatic speed gains — has essentially stopped delivering the returns it once did. Huang has argued that the era of geometrical scaling is over, and instead focused on accelerated computing, where Nvidia achieves performance improvements by optimizing the entire stack, from specialized GPU architectures and high-bandwidth memory to the software that runs on them. Under this view, the meaningful gains left in computing come from redesigning entire systems, not from making individual transistors smaller.
Intel’s argument rests on the transistor itself. Its executives point to the atomic-scale precision now achievable in production, and to a public roadmap that has, for the first time in years, been met on schedule. Intel has executed a “five nodes in four years” roadmap culminating in its 18A process, with its Panther Lake processors shipping to hardware partners ahead of a January 2026 retail launch.
Both companies are, in a sense, correct — and both have a financial interest in being seen as correct. Intel sells transistor manufacturing; emphasizing that transistors can still shrink protects its core business. Nvidia sells system-level acceleration; emphasizing that transistor scaling has stalled makes its own strategy look indispensable.
Where the Bottleneck Has Actually Moved
The more useful question is not whether transistors can still get smaller — they can, at extraordinary cost — but whether that shrinking still delivers the kind of broad, cheap, general-purpose performance gain that Moore originally described. Here the evidence is more mixed.
Manufacturing at the leading edge has become so expensive that only a small handful of companies can afford to participate at all. TSMC’s N2 process officially entered volume production in late 2025, promising up to 15 percent performance gains or substantial power reductions versus previous nodes, but much of that capacity was effectively sold out through 2026, with major customers including Apple, Nvidia, Qualcomm, and AMD locking in large shares of the initial output. This scarcity is itself a sign of how the economics have changed. Cutting-edge chip capacity is no longer a commodity that scales down to consumer prices the way it once did; it is a rationed resource that a few of the world’s largest technology companies compete to secure years in advance.
At the same time, demand for that capacity has never been higher, driven overwhelmingly by artificial intelligence. AI accelerators require enormous amounts of silicon per chip, which has pulled manufacturing capacity away from ordinary consumer devices and toward a small set of extremely well-funded buyers. The shrinking transistor has not disappeared; it has become a strategic asset controlled by a handful of governments and corporations, rather than a broadly distributed engine of falling consumer prices.
What Popular Discussion Gets Wrong
A common misconception treats Moore’s Law as a strict, unbroken doubling that either holds perfectly or has failed completely. In reality, the pace of transistor doubling has been slowing for well over a decade, long before recent headlines declared its death. The more accurate description is not a light switch but a dimmer: fewer companies can keep pace, each generation takes longer and costs more to achieve, and the raw transistor count matters less on its own than how those transistors are arranged, powered, and connected to memory.
Another misunderstanding is treating “Moore’s Law” and “faster computers” as identical. Many of the largest recent leaps in usable performance, particularly for AI workloads, have come not from smaller transistors but from specialized chip designs, faster memory connections, and software tuned to specific hardware. This is the shift Nvidia describes as accelerated computing, and it explains why a computer’s real-world speed can still be climbing sharply even in years when transistor density gains slow to a crawl.
Why This Still Matters
The stakes extend well beyond semiconductor executives arguing past each other in interviews. Every industry that depends on cheap, ever-improving computing — from AI research to medical imaging to climate modeling — has built its expectations on the assumption that computing power keeps getting less expensive relative to what it can do. If that assumption breaks down, or slows sharply, the cost of the next generation of AI systems will not fall the way earlier generations of software did. It will instead be gated by how much manufacturing capacity a handful of factories can build, and how much a handful of buyers can afford to pay for it.
This also reshapes geopolitics. Advanced chip manufacturing is now concentrated in a small number of facilities, overwhelmingly in Taiwan, with the United States, South Korea, and increasingly Japan racing to build domestic capacity of their own. A production milestone that would once have been a technical footnote — which company reaches a given node first — now carries the weight of national industrial strategy, export controls, and supply chain security.
Conclusion
Moore’s Law was never really a law of physics. It was a forecast about human ingenuity: that engineers would keep finding ways to pack more computing power into the same space, at a falling cost, for as long as anyone needed them to. That forecast has not failed so much as it has changed shape. Transistors are still shrinking, chipmakers are still hitting milestones once considered close to impossible, and the underlying ambition Gordon Moore described in 1965 is still driving billions of dollars in investment every year.
What has genuinely ended is the old assumption that this progress would arrive automatically, cheaply, and evenly across the industry. Today’s gains come from architectural reinvention as much as raw miniaturization, and they are increasingly reserved for the companies that can afford to buy them. Moore’s Law, in other words, is not dead. It has simply stopped being free.
