AI 1.08 — The Silicon Beneath
Big Orgs and Compute and AI-zation of Everything
Seventy Years of Waiting:
Turing asked if machines could think in 1950. McCarthy coined the term AI in ’56. Rosenblatt built his perceptron in ’58. Hinton and his students vindicated neural networks some thirty years after Minsky had tried to bury them. The ideas were always there. The math was always there. The people who believed in it, stubborn and brilliant, were always there. And yet for most of those seventy years, the field didn’t see mainstream consumption. It dazzled in the research labs but stumbled in the real world. So what changed?
Three things have improved in continuous iteration over the last seventy year and these three have more overlap than we realize.
Firstly, there was Computation. Gordon Moore, co-founder of Intel, had observed in 1965 that the number of transistors on a chip was doubling roughly every two years. This observation, which became Moore’s Law, remarkably held true for decades. The regular CPU, the brains inside your computer, was never designed for the kind of massively parallel number-crunching that neural networks demand. CPUs were built for versatility, not volume. But what Neural Networks really needed was something that could do thousands of small, repetitive calculations all at once. Something, as it turned out, that had been sitting inside every teenager’s gaming PC. A Graphics card.
Secondly, there was Data. By the mid-2010s, the internet had produced more text, images and video than all of human civilisation had generated in its entire prior history. The shift from symbolic AI to statistical AI that we explored in the last article only worked if you had enough data for the statistics to be meaningful. By 2015, meaningful was an understatement, the internet was drowning in data and AI researchers were learning to make the best use of them.
Thirdly, there was Funding. When Geoffrey Hinton’s students crushed ImageNet in 2012, it was not just an academic milestone. It was a flare gun. The venture capitalists saw it. The corporate R&D departments saw it. The governments saw it. Deep learning was finally working well enough for funding. It was giving results to real problems. Between 2013 and 2023, global private investment in AI grew from roughly $3 billion to over $90 billion annually. Unlike the dot-com bubble, where money chased promises, now, money was chasing results it had already seen.
Jensen’s Bet:
Born in Tainan, Taiwan in 1963, Jensen Huang moved to the United States as a child. He studied electrical engineering at Oregon State and got his master’s at Stanford. In 1993, at the age of 30, he co-founded NVIDIA with Chris Malachowsky and Curtis Priem. For almost two decades, NVIDIA was a graphics card company. It made the chips that rendered the explosions in your video games, the textures on your virtual race tracks. Not exactly the sort of company you’d expect to end up at the centre of a global technological upheaval.
But Jensen made a bet. In 2006, NVIDIA released CUDA, a platform that allowed developers to use its GPUs for general-purpose computation, not just rendering pixels. It was an unusual move. GPUs were built for graphics. Why would anyone want to use them for anything else? For years, barely anyone did. CUDA was a solution looking for a problem and Jensen, wearing his trademark leather jacket, stood behind it anyways.
Then deep learning happened. Researchers discovered that the very thing a GPU was good at, doing thousands of simple mathematical operations simultaneously and was exactly what training a neural network required. The matrix multiplications. The gradient calculations. The forward and backward passes. All of it mapped beautifully onto GPU architecture. NVIDIA had, almost by accident, built the engine that AI needed. And because of CUDA, it was the only company which had the software ecosystem as well to take advantage of it.
NVIDIA’s market capitalisation, which had hovered around $10 billion for most of the 2010s, crossed $1 trillion in May 2023. By early 2024, it had surpassed $2 trillion. Briefly, Jensen Huang’s company was the most valuable on the planet. Ahead of Apple. Ahead of Microsoft. Its flagship AI chip, the A100, became the de facto standard for training large language models. Its successor, the H100, was so coveted that companies were placing orders years in advance. H100s were being resold at two to three times their list price. There is something poetic about the fact that the same chip architecture that once made video game water look realistic was now powering machines that could write poetry.
Others tried to catch up. Lisa Su, the Taiwanese-American engineer who had turned AMD around from near-bankruptcy, launched the MI300X accelerator. Purposefully built for AI. Backed by a competitive software stack. It was a credible challenge. Major cloud providers began offering AMD-based instances. But Jensen’s decade-long head start with CUDA, the optimised libraries, the developer community, these were impossible replicate overnight. Jeff Dean at Google, one of the most accomplished engineers in computing history, had been quietly building Google’s own silicon since 2013. The Tensor Processing Unit, first deployed internally in 2015, was custom-designed for the very matrix operations that neural networks require. By 2024, Google was on its fifth generation of TPUs and used them to train Gemini. Amazon followed Google, built its own chips, Trainium and Inferentia, to provide AWS customers with cost-effective options for AI research, training, and inference.
The chip wars were not just a business story. It was a signal that AI had foundationally become a hardware problem. And the people who controlled the hardware, controlled the AI war.
And nearly all of that hardware, NVIDIA’s GPUs, AMD’s accelerators, Apple’s Neural Engine, was fabricated by one company at one location: TSMC in Taiwan. Over 90% of the world’s most advanced chips passed through its foundries. The concentration was staggering and it was too many eggs in the same basket. So, the CHIPS Act poured $52 billion into reshoring semiconductor manufacturing. Export controls tightened around China. The silicon beneath AI was not just expensive. It was fragile.
The Cloud and the Death of the Garage Genius:
Here is where the story becomes a bit uncomfortable for certain romantic ideals held about technology. The garage startup. The lone genius. Two college kids with a dream and a computer. Training GPT-4, as we discussed, reportedly cost over $100 million and required 25,000 GPUs running for months. Training Gemini Ultra used tens of thousands of custom TPUs across multiple data centres. These are not numbers that a brilliant postdoc at a smaller university or a self-taught programmer can contemplate.
But it wasn’t completely over as we have the cloud, a beautiful abstraction layer built by some of the world’s largest companies, became the equaliser. It still costed money, but there was hope beneath superficial despair.
Andy Jassy built Amazon Web Services from a side project selling Amazon’s excess server capacity. It became the backbone of the modern internet. By the 2020s, a researcher in São Paulo could rent the same GPU cluster that a Fortune 500 company in Seattle was using. Pay by the hour. Shut it down when done. That brilliant postdoc somewhere in the middle east could spin up an instance, train a model overnight and have results by morning for the cost of a few hundred dollars. The democratization of technology wasn’t lost afterall, you might need some form of research funding to dream big, but you could still dream.
Satya Nadella had reinvented Microsoft from a stagnating Windows company into a cloud-first behemoth. Then he made perhaps on of the most consequential business decision of the AI era by investing into OpenAI. Microsoft Azure was then made the exclusive cloud infrastructure behind ChatGPT. Every query you’ve ever typed into ChatGPT. Every essay it’s written. Every line of code it’s debugged. All of it ran on Azure. Nadella didn’t just bet on AI. He made sure the AI ran using Microsoft. Thomas Kurian at Google Cloud offered access to Google’s TPUs and tightly integrated AI services. Together, AWS, Azure and GCP controlled over 65% of the global cloud infrastructure market.
This raises a question worth sitting with. If the infrastructure required to build frontier AI is controlled by a handful of companies, and the cost runs into hundreds of millions, who actually gets to build the future?
Adam Selipsky, then CEO of AWS, in a 2023 conversation with The Verge, framed it as democratization. The cloud, he argued, was levelling the playing field. “You don’t need to build your own data centre anymore.” And he was not entirely wrong. Arthur Mensch and his co-founders at Mistral AI trained competitive models only using rented cloud infrastructure. University research groups around the world were doing the same. But the counterargument is hard to ignore: the playing field might be level, but the field itself is owned by someone else. The usage is not cheap, and it is getting more expensive as demand outpaces supply.
The cloud providers knew exactly what they were. They were inflating an ever increasing Golden bubble. And the Golden bubble showed no signs of reaching it’s bursting point. Yet.
The AI-zation of Everything:
While Jensen and Satya and Andy were building the plumbing, something else had happened at street level. Quietly, and then all at once, AI had seeped into everything we touch. Not as a feature we chose, but as a layer we never noticed.
Our phones, the one in our pockets right now, runs more AI operations per second than the world’s fastest supercomputer could manage in the year 2000. The portrait-mode blur. The night-mode clarity. The magic eraser that removes random strangers from a holiday photo. All of it is powered by neural networks. When Apple’s Johnny Srouji and his team designed the Neural Engine into the A11 Bionic chip in 2017, it processed 600 billion operations per second. By the A17 Pro in 2023, that number had climbed to 35 trillion.
The algorithm chose for you before you opened your Netflix app. The model predicts. You stay. That’s the product. So does your car. Not the massive self-driving promises. The lane-keeping assist. The adaptive cruise control. The automatic emergency braking that comes standard in most vehicles sold after 2020. Amnon Shashua, a computer science professor at the Hebrew University of Jerusalem, founded Mobileye in 1999. His vision was simple: put computer vision into vehicles. Intel acquired the company for $15.3 billion in 2017. By 2024, Mobileye’s technology was in over 150 million vehicles across more than 50 automakers. Every time your car gently nudges you back into your lane, that’s Shashua’s work.
Get Vishnu Kumar V H’s stories in your inbox
Join Medium for free to get updates from this writer.
A study in Nature Medicine showed that radiologists working alongside AI made fewer diagnostic errors than either working alone. Farmers were using drone-mounted models to predict crop yields before harvest. Law firms were reviewing thousands of contracts in hours. Trading floors were running on algorithms that never slept. AI had not announced itself in these places. It had simply shown up and started working.
At some point, you have to stop and wonder: does everything need to be smart? Does your thermostat need reinforcement learning? Here’s what happened. Companies discovered that slapping “AI-powered” on a product raised the valuation, attracted the press and opened the investor calls. And once the incentive existed, the restraint disappeared. A lot of it was theater. But enough of it was real to keep the whole machine turning and most importantly, keep the costs “justified”.
So, where does this end?
A Golden bubble waiting to burst:
This is the question that hangs over all of it, and it still does not have a clean answer. In 2023 alone, global corporate investment in AI exceeded $150 billion. Satya Nadella put $13 billion into OpenAI. Google, Amazon and Salesforce collectively poured billions into Anthropic. Saudi Arabia’s Public Investment Fund and the UAE’s sovereign wealth funds were building “AI-focused data centres” in the desert, quite literally. Governments from Washington to Beijing to Brussels were earmarking billions in national AI strategies.
Why this much money, this fast, into something that most companies are still figuring out how to profit from?
The surface answer is fear. No CEO wants to be the executive who missed the AI wave, like many missed the internet in the ’90s. But the deeper answer is more interesting. AI, unlike most previous technology waves, is what economists call a general-purpose technology. The steam engine was one. Electricity was another. The internet and iPhones were the most recent. These technologies don’t solve one problem. They provide a fundamentally new way to approach nearly every problem. That is why the money is so broad. It’s not a bet on chatbots. It’s a bet on a new substrate for everything.
The hope is that AI will eventually generate enough productivity gains, enough new revenue streams, enough cost savings to justify the expenditure. The venture capital world calls this the J-curve. The question that keeps CFOs up at night is just how deep the bottom of that J goes before it curves back up.
And then a team of engineers in Hangzhou, China, threw a curve ball at everyone’s assumptions. Liang Wenfeng, a former quantitative fund manager, had founded DeepSeek and in January 2025 released its R1 model. It matched frontier models on several benchmarks at a reported fraction of the training cost. Jensen Huang’s NVIDIA lost nearly 17% of its stock price in a single day. Roughly $600 billion in market value. Wiped out. The largest single-day loss for any company in history.
The DeepSeek moment sent a shiver through Silicon Valley. Could clever engineering substitute for brute-force hardware? The answers were not clean. DeepSeek’s efficiency gains were real but so were the questions about how they achieved them under US chip export controls.
The money, in short, is following conviction more than proof. Whether that conviction is prescience or hubris is a question only time will answer.
Now What?
So here we stand. The infrastructure exists. The chips are being minted. The cloud providers are expanding data centres at a pace that would have seemed absurd a decade ago. AI is in your phone, your car and everything you can think of. The models keep getting bigger. The costs keep getting higher. The capabilities are getting more remarkable.
The infrastructure story was, in many ways, the easier story. Chips can be designed, data centres can be built, cloud instances can be spun up. The harder story from a professional and a human standpoint is what would happen when these systems walk into domains that we humans have always considered our own. The chess board. The artist’s studio. The programmer’s IDE.
Now that we have included AI in almost everything around us. It’s come to come to the time that in some of those domains, machines have already started winning outright. The implications are not exactly what you might expect.
Up Next: AI 1.09 — Towards Technological Singularity: Part 1: New World Order, New Rules
References and further reading:
1. NVIDIA CUDA: https://developer.nvidia.com/cuda-toolkit
2. In conversation with AWS CEO — The Verge (2023): https://www.theverge.com/23824200/ai-cloud-amazon-aws-adam-selipsky
3. AI Index Report 2024 — Stanford HAI: https://aiindex.stanford.edu/report/
4. AI-assisted Radiology — Nature Medicine (2020): https://www.nature.com/articles/s41591-020-01174-9
5. OpenAI Funding and Valuation — Reuters (2024): https://www.reuters.com/technology/openai-talks-raise-funding-100-billion-valuation-wsj-reports-2024-09-11/
6. DeepSeek and NVIDIA Market Impact: https://www.deepseek.com/
7. AMD MI300X and the AI Chip Race: https://aibusiness.com/verticals/amd-vs-nvidia-the-battle-for-ai-gpu-supremacy-begins
8. Google TPU Architecture: https://cloud.google.com/tpu
9. Mobileye and ADAS: https://www.mobileye.com/
10. Apple Neural Engine: https://machinelearning.apple.com/
