Future 5 November 2026 4 MIN READ
From One Transistor to a Quadrillion in a Data Centre
How transistor counts grew from one at Bell Labs in 1947 to 2,300 in 1971 and 1.2 quadrillion in an AI data-centre POD, with caveats on NVIDIA's figures.
THE SHORT ANSWER
The first transistor, made at Bell Labs in 1947, was a single device. Intel's 4004 of 1971 held 2,300. NVIDIA says its Rubin GPU has 336 billion, and rates a 1,152-GPU Vera Rubin POD at 1.2 quadrillion transistors. These are NVIDIA's figures, not independent counts.
On this page
One device, in 1947
On 16 December 1947, John Bardeen and Walter Brattain at Bell Labs made the first semiconductor amplifier. The Computer History Museum describes two closely spaced gold contacts pressed on a slab of high-purity germanium, with the voltage on one contact controlling the current through the other and amplifying the signal up to 100 times. Bell Labs announced the device, named the “transistor”, on 30 June 1948, and Bardeen, Brattain and William Shockley shared the 1956 Nobel Prize in Physics.
That was one transistor. The first computers built from them used few. A prototype at Manchester University in November 1953 used 92 point-contact transistors and 550 diodes.
Thousands on one chip, 1965 and 1971
In April 1965 Gordon Moore wrote in Electronics that, with unit cost falling as the number of components per circuit rises, “by 1975 economics may dictate squeezing as many as 65,000 components on a single silicon chip.” It was a statement about cost, and it counted components, not only transistors.
In November 1971 Intel released the 4004, which the Smithsonian describes as the first complete CPU on one chip and the first commercially available microprocessor. It held 2,300 transistors, used a 10-micrometre process and could execute about 92,000 instructions per second.
A GPU today
Hopper, NVIDIA’s GPU generation of the early 2020s, has 80 billion transistors, according to DatacenterDynamics. Blackwell, announced in 2024, has 208 billion, and NVIDIA’s documentation says it achieves that with two reticle-limited dies joined by a 10 terabyte-per-second link in one GPU. A reticle-limited die is as large as the manufacturing equipment can print in one go, so a bigger chip has to be made of several.
NVIDIA’s technical blog of 21 July 2026 says its Rubin GPU uses 336 billion transistors and up to 50 petaflops in a low-precision number format called NVFP4. That is about 146 million times the 4004’s 2,300, a ratio we worked out from the two figures.
A rack, then a POD
The data centre does not buy one chip at a time. NVIDIA’s Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs. Multiplying 72 by 336 billion gives about 24 trillion transistors in the GPUs of one rack, before counting the CPUs, switches and memory. That is our arithmetic.
NVIDIA’s March 2026 blog goes further. It says a Vera Rubin POD, which links five kinds of rack-scale system, holds 1,152 Rubin GPUs and 1.2 quadrillion transistors in total, and is rated at 60 exaflops. In the American counting used in NVIDIA’s text, a quadrillion is 1,000 trillion. The GPUs alone would come to about 0.39 quadrillion (1,152 times 336 billion), so the total includes the POD’s other chips; the headline figure is NVIDIA’s.
What the count does and does not say
A transistor count is not a measure of speed or usefulness. Every figure above for NVIDIA’s chips comes from NVIDIA or from reports of its announcements, as a specification or a design claim, and we have not found an independent count. Moore’s own sentence was about economics, and the later rise came partly from adding dies, boards and racks, not only from fitting more on one die.
That is also where the electricity goes. Our guide to data-centre power and water use shows what such racks add up to across the grid.
Our take: the surprise is not any one chip but how little of the growth is physics. A device that amplified a signal about a hundredfold became a unit counted in billions per package because it kept getting cheaper to make more of them, which is what Moore wrote in 1965. Today the count rises by joining dies and racks, and the bill for that arrives as power and cooling.
Questions and answers
When was the transistor invented?
On 16 December 1947 at Bell Labs, by John Bardeen and Walter Brattain, with William Shockley leading the group. They shared the 1956 Nobel Prize in Physics. The device used gold contacts on germanium and amplified a signal up to 100 times.
How many transistors did the first microprocessor have?
The Intel 4004, released in November 1971, had 2,300 transistors. The Smithsonian describes it as the first complete CPU on one chip and the first commercially available microprocessor.
How many transistors does an AI GPU have?
NVIDIA says its Rubin GPU has 336 billion and its Blackwell GPU 208 billion. Blackwell reaches that with two reticle-limited dies in one GPU. These are the maker's figures.
Is a transistor count the same as performance?
No. It says how many switches a chip holds, not how fast or useful it is. NVIDIA rates Rubin by operations a second in a chosen number format, and those ratings are the company's own.
Do you agree?
Take it to Chevlume: write your view, add your sources and let others answer. The closed alpha is by invitation.
Keep reading
What Quantum Computers Can and Cannot Do
Qubits, errors, advantage claims and the threat to encryption, with every result labelled as demonstrated, claimed by a company or projected.
FUTURE / DISCOVERYBelgium 2003: The 4,096 Extra Votes
One candidate, 4,096 phantom votes and a bit flip. What the official report concluded, and what later retellings added about cosmic rays.
FUTURE / DISCOVERYKnight Capital: 45 Minutes, $460 Million
A forgotten code path on one of eight servers cost a trading firm $460 million in 45 minutes. What the SEC found about deployment and controls.
Sources
- 1947: Invention of the Point-Contact Transistor · Computer History Museum · Accessed
- Inventing the Transistor · Computer History Museum · Accessed
- Cramming More Components onto Integrated Circuits (G. Moore), Electronics, 19 April 1965 · IEEE reprint · Accessed
- Intel 4004 Microprocessor · Smithsonian National Museum of American History · Accessed
- Nvidia announces next GPU generation, Blackwell · DatacenterDynamics · Accessed
- The NVIDIA Grace Blackwell Superchip · NVIDIA documentation · Accessed
- Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI (21 July 2026) · NVIDIA Technical Blog · Accessed
- NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer (16 March 2026) · NVIDIA Technical Blog · Accessed
Published
Discuss on X