In The Terminator, the entire plot hinges on the idea that history pivots on a single person at a single moment. Skynet sends the T-800 back to 1984 Los Angeles because it knows that John Connor — not yet born, not yet anything — will grow up to lead the human resistance. The machines understand something profound: the future isn’t inevitable. It’s contingent.

Kill one woman in a quiet apartment, and the whole timeline collapses. The revolution dies before it’s conceived.

Now look at AlexNet.

In the summer of 2012, Alex Krizhevsky was a graduate student at the University of Toronto, working under Geoffrey Hinton, alongside Ilya Sutskever. He trained a deep convolutional neural network on two NVIDIA GTX 580 cards — consumer gaming GPUs, three gigabytes of memory each — that he’d installed in a desktop PC, reportedly running in his bedroom at his parents’ house. When AlexNet won the ImageNet competition that September, it didn’t just win — it crushed the field, dropping the error rate from around 26 percent to 15 percent. That single result is what convinced the world that deep learning on GPUs was the path forward. Every modern AI system — GPT, Claude, autonomous driving, the entire trillion-dollar GPU economy — traces its lineage back through that moment.

So here’s where the parallel gets eerie.

If Skynet were real, and looking backward at its own origin story the way the Terminator films imagine it — the target wouldn’t be a soldier. It wouldn’t be a politician. It would be a quiet 26-year-old in a Toronto suburb, hunched over a tower PC humming with two gaming cards meant for rendering Battlefield 3. That bedroom is the Sarah Connor apartment. Those GTX 580s are the conception event. Krizhevsky is the John Connor figure — except he’s also, in a sense, Sarah, because he’s the one who actually birthed the thing. Hinton is maybe the Kyle Reese — the mentor who shows up with the right ideas at the right moment. Backpropagation. Deep architectures. The conviction that this would work when nobody else believed it. He makes the future possible.

And the wild part? It actually happened. There was no Skynet to stop it. The conception event went off exactly as the timeline required.

This is The GPU Pulse. I’m Pete Paisley. Let’s talk about the five days that built the modern world.

Quick word on me before we get into it. I’m Pete Paisley, and I host The GPU Pulse — the show where we cover what’s actually moving in the GPU aftermarket. Pricing. Supply. Refresh cycles. The decisions that ITAD operators, IT resellers, and enterprise buyers are making right now about what to buy, what to hold, and what to flip. This is a market that didn’t really exist fifteen years ago — and the reason it exists at all is the story we’re telling today. So this one’s a little different from our usual market intel episodes. Today we’re going back to the origin point. Because if you’re trading H100s, or watching Hopper inventory hit the secondary market, or trying to figure out where Blackwell pricing is headed — all of that is downstream of one PhD student in Toronto who didn’t have a doctorate yet, working on hardware that cost less than a used car.

Let’s get into it.

So let’s set the scene properly. Toronto, summer 2012. Three people in a small academic orbit. Geoffrey Hinton — at that point, one of the world’s foremost authorities on neural networks, with about thirty-five years of foundational research behind him. The kind of researcher who had been right about deep learning for decades while most of the field had moved on to other approaches. And two of his graduate students — Alex Krizhevsky and Ilya Sutskever. Sutskever, of course, would later co-found OpenAI. Neither of them had finished their PhD yet.

Hinton himself described the division of labor pretty bluntly years later. He said, quote: “Ilya thought we should do it, Alex made it work, and I got the Nobel Prize.” That’s the dynamic. Hinton provided the intellectual framework, built up over decades. Sutskever provided the conviction — the belief that scale would work, that if you made the network deep enough and threw enough compute at it, it would learn things nobody had been able to teach a machine before. And Krizhevsky wrote the code. He ran the hardware. He’s the one in the bedroom with the screwdriver and the GPU fans humming.

Now here’s what was actually converging in that room — and this is the part the aftermarket community needs to internalize, because it’s the foundation of everything we trade today.

Three elements came together for the first time. One — a massive labeled dataset. ImageNet. About 1.2 million photographs, hand-labeled, organized into thousands of categories. That was new. Datasets like that simply hadn’t existed at scale. Two — a programmable GPU platform. NVIDIA’s CUDA, which had been quietly maturing since 2007, finally made it possible to use a graphics card for general-purpose math. Not rendering. Not pixels. Just raw parallel matrix multiplication, exposed to a developer in a way that didn’t require you to be a graphics specialist. Three — a deep neural network architecture with the capacity to actually use both of those things.

And the genius of what Krizhevsky did wasn’t to hand-code rules for recognizing a cat versus a dog versus a car. It was the opposite. He built a system, fed it those 1.2 million labeled images, and let it teach itself. Five days of continuous training. Two GTX 580s. Three gigabytes of memory each. And at the end of it, the network had learned to see.

September 30th, 2012. The ImageNet Large Scale Visual Recognition Challenge. AlexNet enters and wins by a margin of 10.8 percentage points over the nearest competitor. That margin is the part you have to sit with. It wasn’t a tie. It wasn’t a slight edge. It wasn’t “promising.” The entire computer vision research community was forced to confront the same uncomfortable conclusion at the same time: everything they had been doing was wrong.

That’s the conception event. That’s the bedroom moment. Five days. Two consumer gaming cards. The world reorganized.

Let’s slow down on the hardware for a minute, because this is The GPU Pulse and we owe it to the audience to actually look under the hood.

The architecture Krizhevsky used was a convolutional neural network. A CNN. The design isn’t original to AlexNet — its roots go back to Yann LeCun’s work in the late 1980s. But the way a CNN works is worth understanding because it’s why it could even fit on two consumer GPUs in the first place.

Older approaches to image recognition tried to process every pixel in relation to every other pixel. That’s catastrophically expensive. The number of relationships explodes. CNNs work the way eyes actually work. They scan small regions. They detect local patterns. They build from simple features to complex ones, layer by layer. Early layers learn to detect edges and corners. Middle layers learn shapes and textures. Later layers learn objects and concepts.

And nobody programs those stages explicitly. The network discovers them through training. That’s the magic — and that’s what people mean when they say a network “learns.”

The other thing CNNs do that older approaches didn’t is parameter sharing. The same feature detector slides across the whole image. So a network that’s learned what an edge looks like in the upper-left corner can also recognize that same edge in the lower-right corner without learning it twice. That dramatically cuts the computational load. And it builds in something earlier methods lacked entirely — spatial awareness. The understanding that a feature is the same feature wherever it appears in the frame.

So the architecture was mature. The dataset was new. What made AlexNet specifically decisive was the combination of scale and the right hardware. Five convolutional layers. Three fully connected layers. Sixty million parameters. Running on a GPU fast enough to train the whole system in days rather than months.

And the GTX 580 — let’s just appreciate what that card was, because in the aftermarket world we deal with hardware lineage every day. The 580 launched in late 2010. Fermi architecture. Five hundred and twelve CUDA cores. Three gigabytes of GDDR5. It was a high-end gaming card aimed at enthusiasts who wanted to run Crysis 2 or Battlefield 3 at maximum settings. Retail price around five hundred dollars. Total system cost for what Krizhevsky was running — two cards, a workstation board, a power supply that could feed both — probably under two thousand dollars all in.

That’s the entire hardware bill of materials for the most consequential AI engineering work of the modern era. Two thousand dollars of consumer parts.

The CNN was the architecture. The GPU was what finally let it breathe.

The industry didn’t take long to understand what had happened. And the speed of the response is one of the most telling parts of this whole story.

Within weeks of the ImageNet result, Hinton had consulted a lawyer about how to maximize the value of a company with three employees, no products, and no revenue. The answer was an auction. They incorporated something called DNNresearch Incorporated — basically a shell company built around the three of them and the IP — and they put it up for bid.

Four bidders. Google. Microsoft. Baidu. And a young London-based startup called DeepMind, which at that point was barely on anyone’s radar. Four of the world’s most aggressive technology buyers, all trying to acquire the same three people.

The bidding climbed through the fall of 2012. Late one evening, as it approached forty-four million dollars near midnight, Hinton suspended the auction. He went to sleep. The next morning he sold to Google.

Think about that for a second. Six months earlier, this was a graduate student running gaming cards in his bedroom. Six months later, the entire technology industry had validated what it meant — to the tune of forty-four million dollars for a three-person team with no product.

And what followed was the rapid reorganization of research priorities across every major lab on the planet. GPU clusters started appearing in data centers — not in ones and twos, but in racks. NVIDIA, which had spent its entire corporate life selling graphics cards to gamers, watched its data center revenue begin a climb that would not stop for over a decade. The GTX 580 had proved the concept. The question the industry now faced was what to build next.

And this is where, if you’re in the aftermarket today, you can start to see your own business taking shape on the horizon. Because the moment NVIDIA’s data center business started growing, every one of those GPUs was eventually going to come off-lease, get refreshed, get decommissioned, get pulled from a hyperscaler rack, and enter the secondary market. The flow we trade in today — that flow started here.

The answer to “what comes next” arrived in 2017. And it expanded the problem entirely.

A team of Google researchers published a paper called “Attention Is All You Need.” It introduced the transformer architecture. And the transformer shifted AI’s center of gravity from images to sequences. Language. Audio. Code. Anything ordered in time. Where CNNs had given machines the ability to see, transformers gave them the ability to read, to reason, and to generate.

The implications for compute were immediate, and they were violent.

Training a convolutional network on ImageNet required days on two consumer GPUs. Training a transformer-based large language model required weeks on thousands of specialized ones. The compute demands of the transformer era didn’t just grow — they scaled exponentially with model size. And the research community had already internalized Sutskever’s founding thesis from the bedroom in Toronto. Scale is the mechanism. More data. More compute. Deeper networks. Better results.

Every new capability the transformer architecture unlocked required another order of magnitude of compute to achieve. The GPU era had begun in 2012. By 2017 it had become something the original GTX 580 could not have imagined.

And this is the part where the aftermarket conversation starts to feel inevitable. Because the moment the industry’s compute appetite starts scaling exponentially, the hardware stack underneath it has to scale with it. Every generation has to be bigger, denser, more specialized. Which means every generation eventually becomes the previous generation. Which means there has to be a market for the previous generation. That’s us. That’s what we trade.

What the transformer made inevitable, the semiconductor industry built.

Google had already deployed its first Tensor Processing Unit internally in 2015, purpose-built for the matrix multiplication operations that neural networks demand. That was a direct institutional response to the compute trajectory AlexNet had set in motion. Google saw the curve and started building custom silicon to climb it.

As transformer-scale training became the dominant workload, every major hyperscaler followed. AWS built Trainium for training and Inferentia for inference. Apple embedded Neural Engines into every iPhone. Qualcomm built AI acceleration into mobile processors that ship to billions of devices.

NVIDIA, meanwhile, consciously transformed itself. They introduced Tensor Cores in the Volta architecture specifically to accelerate the matrix operations at the heart of deep learning. Then they evolved through the V100 and the A100 to the H100 and beyond. Hopper. Blackwell. Vera Rubin on the roadmap. A company that earned roughly four billion dollars annually in 2012, almost entirely from gaming, was generating over ninety billion dollars in data center revenue by 2024.

That entire arc — from a five-hundred-dollar graphics card in a Toronto bedroom to the most valuable semiconductor company in history — traces back to one five-day training run on two GTX 580s.

The hyperscaler AI infrastructure that now consumes billions of dollars in capital expenditure every quarter. The GPU clusters that power every large language model in production today. The purpose-built silicon racing to meet demand that shows no sign of slowing. All of it is downstream of those three elements converging for the first time in a bedroom in the Toronto area in 2012.

AlexNet didn’t just win a competition. It defined what computing would become.

I want to close on what this story means for the people who actually listen to this show.

If you’re an ITAD operator, an IT reseller, an enterprise buyer trying to figure out fleet refresh strategy — the trillion-dollar GPU economy you’re operating inside of right now is not a permanent fixture of the universe. It’s a contingent outcome. It happened because three people in Toronto converged with three technologies in the right room at the right moment, and the result was big enough to bend the entire industry around it.

That’s worth remembering, because it tells you something about where the next inflection point is going to come from. It probably isn’t going to come from the obvious place. It probably isn’t going to come from the company everyone is watching. The original conception event in this industry happened in a bedroom on consumer hardware, run by a guy who didn’t have his PhD yet. Five hundred dollars per card. Three gigabytes of memory each.

The aftermarket is the long tail of decisions made at moments like that one. Every GPU you sell, every fleet you refresh, every price curve you track — that’s all the wake of the boat. The boat was launched in September 2012, in a quiet Toronto suburb, by a quiet 26-year-old who probably just wanted to win a research competition.

He won the world instead.

If you want the full written piece this episode is based on, head over to gpuresource.com — the article is called “The GPU Backstory: How a Bedroom in Toronto Created the Modern Era of AI Compute.” Linked in the show notes.

I’m Pete Paisley. This has been The GPU Pulse. We’ll see you next time.