A Brief History of Technology: From the Vacuum Tube to the End of Moore's Law

For most of the last century, progress in technology meant progress in hardware. That era is ending. Understanding how we got here, and why the curve is bending, matters if you want to understand where AI is taking us next.

The horse that started it all

Silicon Valley's origin story is, oddly, about horse breeding. In 1876, Leland Stanford began applying a data-driven method to training racehorses: identify the fastest foals as early as possible, then pour resources into them rather than waiting for a horse to mature naturally. He called it the Palo Alto system. It made him a fortune, and after his son's early death, it became the founding logic of Stanford University, built to apply the same principle to people. That instinct, spot potential early, bet everything on it, would resurface again and again as the actual engine of Silicon Valley's technical progress, from venture capital to hiring.

Hardware's long climb (1900s–1980s)

The technical history that follows is a chain of hardware breakthroughs, each solving the previous one's limitations:

• The triode (1910s): Lee de Forest, working at the Federal Telegraph Company near Stanford, added a third electrode to the vacuum diode, creating the first electronic amplifier and switch. It underpinned radio, television, and the earliest computers, but vacuum tubes were bulky, fragile, hot, and unreliable.
• The klystron (1937): A solution to the triode's high-frequency limits, developed at Stanford and critical to WWII radar. This cemented a funding relationship between Stanford and the Pentagon that would shape the region for decades, what's sometimes called "military Keynesianism": using defense spending, not civilian infrastructure, as the engine of growth and innovation.
• The transistor (1947): William Shockley, John Bardeen, and Walter Brattain at Bell Labs replaced the vacuum tube with a solid-state semiconductor switch. It was smaller, cooler, and, crucially, could be miniaturized indefinitely.
• The integrated circuit (1958–59): Jack Kilby (Texas Instruments) and Robert Noyce (Fairchild Semiconductor) independently figured out how to etch multiple transistors onto a single piece of silicon. This is the birth of the modern chip.
• The microprocessor (1971): Intel, founded by Fairchild alumni Noyce and Gordon Moore, put an entire computer's logic on one chip.

Each step shrank the switch. Each shrinkage meant more switches per chip, more computing power per dollar, and, per Moore's observation, a doubling of transistor density roughly every one to two years. In 1961, a chip held four transistors. By 2020, a single phone processor held nearly 12 billion.

From vertical integration to the fabless model (1980s–2000s)

By the mid-1980s, Japanese manufacturers were out-competing American firms on the exact memory chips the US had invented, controlling around 70 percent of the global DRAM market. Intel's response, abandoning memory chips entirely to bet everything on microprocessors, was a survival move that reshaped the industry.

It also forced a deeper structural shift. Chip design and chip manufacturing had always been fused: if you wanted to make a chip, you needed hundreds of millions of dollars to build the factory (the "fab") yourself. New chip-design software, developed by Carver Mead and Lynn Conway in the late 1970s, separated those two steps. An engineer could design a chip on a computer and send the file anywhere to be manufactured. Morris Chang built an entire company, TSMC, around this insight in 1987: a foundry that only manufactures other companies' chip designs and never competes with its own. This "fabless" model is why today a company like Nvidia or Apple can design a cutting-edge chip without owning a single factory, and why a single company, TSMC, has become one of the most geopolitically important firms on Earth.

The Dutch company at the center of it all

TSMC's fabs, and every other leading-edge fab in the world, depend on a single machine that only one company can build: the extreme ultraviolet (EUV) lithography system made by ASML, based in Veldhoven, the Netherlands. EUV machines use light with a wavelength of just 13.5 nanometers to etch the impossibly fine patterns that make modern chips possible, and no competitor, American, Japanese, or otherwise, has managed to replicate the technology after decades of trying. ASML holds what amounts to a 100 percent monopoly on the tool the entire advanced chip industry depends on.

ASML spun out of the Dutch electronics giant Philips in 1984 and survived its early years partly on Dutch and European subsidies, and its EUV breakthrough came out of a decades-long, internationally shared research effort, most of the foundational physics was actually done in US national laboratories (Livermore, Sandia, Berkeley) and at Bell Labs, alongside European public-private partnerships and collaboration with the Belgian research institute IMEC. ASML was the company that took that shared research and, through a high-risk, decades-long bet that its Japanese rivals ultimately weren't willing to make, turned it into the only commercially viable EUV machine on Earth. It's a useful corrective to the Silicon Valley "lone genius founder" mythology running through the rest of this history: Europe's most valuable technology company exists because of patient, publicly supported, collaborative research, not a garage and a monopoly-obsessed venture capitalist.

The economic payoff for the Netherlands has been substantial. Each EUV machine costs well over $150 million and takes months to ship and assemble, and ASML is Europe's largest technology company by market capitalization. It is by far the country's biggest private R&D spender, spending more than the next several companies on the Dutch R&D list combined, and the Brainport Eindhoven region built around it now accounts for more than a tenth of the Netherlands' entire GDP. That concentration also creates real exposure: economists have explicitly compared the Netherlands' dependence on ASML to Finland's earlier dependence on Nokia, a single dominant firm that a country's innovation base and export earnings become structurally tied to.

The wall we've now hit

That's the part of the story that's now running out of road. Transistors are approaching atomic dimensions. You cannot indefinitely shrink a switch that is already only a few atoms wide, quantum effects and heat dissipation become fundamental physical limits, not engineering ones. Moore's Law was never a law of physics, it was an economic and engineering observation that held for six decades because each generation of shrinkage kept paying off. That payoff is diminishing. Building a leading-edge fab now costs upward of $20 billion, and the lithography machines required to make these chips are produced by essentially one company on Earth (ASML). Hardware progress hasn't stopped, but its exponential curve is flattening.

Where the exponential curve moved: software

This is the pivot worth naming clearly. For a century, the compounding gains in technology came from hardware, smaller transistors, denser chips, cheaper computing. That compounding is what made computers, the internet, and smartphones possible on a predictable, ever-accelerating timetable.

The current wave of AI is, in large part, a story about that exponential curve migrating from hardware to software. The physical substrate, chips, is no longer where the fastest gains are being made; it's a comparatively fixed, if extremely expensive, foundation. The gains are increasingly coming from what runs on top of it: model architectures, training techniques, and the scale and quality of data, layers of software improvement that don't require physically manufacturing anything smaller than before.

This is a genuinely different kind of progress than the twentieth century experienced, and it's worth being precise about the difference rather than assuming AI is simply "more Moore's Law." Hardware progress was constrained by physics and capital intensity: shrinking a transistor is a manufacturing problem with a manufacturing cost. Software progress is constrained by different things: data availability, algorithmic insight, and, increasingly, energy and compute at inference time. Whether software gains can compound as reliably and for as long as transistor density did for sixty years is, honestly, an open empirical question, not a settled one. Betting the next decade's economy on a software equivalent of Moore's Law is a real possibility, but treating it as guaranteed is the same kind of hype that inflated and then burst the dot-com bubble in 2000.