Google Part III: The AI Company. Google is amazingly well-positioned... will they win in AI? (audio)

Why does it matter? Google invented the future of AI and almost lost it

Google is the company that invented the transformer, built the first large neural networks, hired nearly every AI researcher on earth — and then watched startups use its own published research to threaten its $140 billion profit machine. This episode traces how Google got here, why it nearly missed the wave it created, and whether its extraordinary asset base is enough to win.

• Google is the only company with all four pillars needed to win in AI: a frontier model, custom chips, a hyperscale cloud, and self-sustaining profits
• The AI era didn't start with ChatGPT in 2022 — it started with a cat paper in 2012 and quietly generated hundreds of billions in revenue through YouTube and Facebook feeds
• OpenAI exists specifically because Google monopolized AI talent, and the founding dinner at the Rosewood Hotel almost ended with everyone saying no
• Waymo's 91% crash reduction and Lyft-beating San Francisco revenue cost Google less than one month of profits

The cat paper started an invisible AI revolution worth hundreds of billions

In 2012, a nine-layer neural network trained on 16,000 CPU cores learned to recognize cats from unlabeled YouTube frames without ever being told what a cat was. That single result — the cat paper — is the actual start of the modern AI era, not ChatGPT.

Sundar Pichai told Ben and David that seeing the cat paper cross his desk was one of the defining moments of Google's story. The reason is what happened next: the same technology that found a 'cat neuron' firing in an unsupervised model could go inside videos and understand what they were about — enabling YouTube to build a recommendation feed that actually worked. Facebook borrowed the technique and built Instagram's feed. ByteDance took it further with TikTok. The feeds that now define how most humans spend their leisure time trace directly back to that paper.

As David puts it: 'The AI era started in 2012.' For the next decade, AI was already reshaping human existence and driving hundreds of billions in revenue — it was just happening inside algorithmic feeds, not in chatbots. 'It's not that much of a leap to say that the cat paper led to probably hundreds of billions of dollars of revenue generated by Google and Facebook and by ByteDance over the next decade.' The public AI narrative is roughly ten years behind the actual business impact.

The technical unlock was Jeff Dean's distributed system called Disbelief, which proved neural networks could run asynchronously on stale data across thousands of machines — against everything the research community believed was necessary. AlexNet the same year confirmed the GPU insight that would eventually make Nvidia the most valuable company in the world.

Google published the transformer, watched all eight authors walk out the door, and handed competitors the weapon to threaten its core business

In 2017, eight researchers on the Google Brain team published a paper titled 'Attention Is All You Need.' It has since been cited over 173,000 times — the seventh most-cited paper of the 21st century. Every frontier AI model running today is built on it.

Google allowed the publication out of genuine open-science values. 'In perhaps one of the greatest decisions ever for value to humanity and maybe one of the worst corporate decisions ever for Google, Google allows this group of eight researchers to publish the paper.' Within a couple of years, all eight authors had left Google to either start or join AI startups. Noam Shazeer — the engineer who rewrote the transformer codebase from scratch and made it actually work, the one teammates described as 'a magician' — founded Character AI. Google eventually paid roughly $2.7 billion to get him back.

What Google did with the transformer afterward is often misunderstood. 'It is a false narrative out there that Google did nothing with the transformer after the paper was published. They actually did a lot. What they didn't do was treat it as a wholesale technology platform change.' BERT went into search. Language models improved ad quality scores. But the move that mattered — building a consumer chatbot — was blocked by a $140 billion profit machine that couldn't afford to cannibalize itself.

'2017 begins the five-year period of Google not sufficiently seizing the opportunity that they had created with the transformer.' The Innovator's Dilemma, playing out in real time, at the company that literally invented the dilemma's subject matter.

OpenAI was founded because Google had monopolized AI talent — and one defection unraveled the whole fortress

The dinner at the Rosewood Hotel on Sand Hill Road in summer 2015 was supposed to lure Google's AI researchers into a new nonprofit lab. It almost completely failed.

Elon Musk and Sam Altman gathered the most important AI researchers of the era and asked what it would take to leave Google. The answer, going around the table, was essentially: nothing. 'We're getting paid way more money than we ever imagined. Many of us get to keep our academic positions and affiliations — and we get to hang out here with each other.' Jeff Dean was there. The infrastructure was the best in the world. There was no reason to leave.

Except one person was intrigued. 'The trouble was so many of the people most qualified to solve these problems were already working for Google. And no one at the dinner was quite sure that these thinkers could be lured into a new startup even if Musk and Altman were behind it.' But Ilya Sutskever was at least open: 'I felt like there were risks involved, but I also felt like it would be a very interesting thing to try.'

Google came back with a counter offer — reportedly double — delivered personally by Jeff Dean. Ilya said no. That single decision broke the logjam. 'OpenAI gets founded in 2015 with the goal of, hey, let's shake all this talent out of Google and level the playing field — and Google just accelerates.' The moat was a people moat, not a structural one. The moment mission alignment cracked, the intellectual capital walked.

DeepMind sold to Google at half Facebook's price because Larry Page understood what Demis was trying to do

Facebook was willing to pay up to $800 million for DeepMind. Google paid $550 million. The founders left roughly twice as much money on the table by choosing Google — and they did it on purpose.

'The crazy thing is this kinship between Larry and Demis is I think the reason why the deal gets done at Google. Once the two of them get together, they are like peas in a pod.' Demis Hassabis told Ben and David directly: he 'just felt like Larry got it.' Larry had always viewed Google as an AI company. Demis didn't even want to build products until DeepMind reached AGI. Facebook's Mark Zuckerberg wouldn't grant the independent oversight board and separate mission structure that the DeepMind founders required. Google — which already had Brain working on product applications — could credibly promise to leave DeepMind alone.

The acquisition was physically closed by Alan Eustace flying to London on a charter jet with Jeff Hinton laying in a custom harness rig on the floor because his back condition prevented him from sitting. That is the level of urgency.

The result: 'It's worth what $500 billion today. This is as good as Instagram or YouTube in terms of greatest acquisitions of all time.' A 2014 deal struck at $550 million, won on mission fit over price, is now arguably the single best acquisition in tech history by return multiple. The data center cooling application alone — a 40% energy reduction announced just two years after acquisition — arguably paid back the purchase price quickly.

Google's secret chip program built in 15 months in Madison, Wisconsin may be the most important structural advantage in AI

After rolling out speech recognition on Nexus phones, Jeff Dean did the math: if Google extended it to all billion Android phones and people used it three minutes a day, they would need twice as many data centers as Google currently operated — just for that one feature. 'There's a great quote where Jeff goes to Holtzel and goes: we need another Google.'

The alternative was building a custom chip. A Google engineer named Jonathan Ross had been spending his 20% time on FPGAs. The team took that work and built the Tensor Processing Unit — designed specifically for matrix multiplication, useless for anything else, and dramatically more efficient than GPUs for neural network inference. The whole thing — designed, verified, fabricated, and deployed into data centers — took 15 months. It was built not in Mountain View but in a satellite office in Madison, Wisconsin. The team fit the TPU into the form factor of a hard drive so it could slot into existing server racks without physical rearchitecture. They kept it secret for over a year.

Today Google has an estimated 2 to 3 million TPUs. Nvidia shipped roughly 4 million GPUs last year. Nobody talks about this as a two-horse race in AI chips, but it is.

The economics matter enormously. Nvidia carries roughly 80% gross margins — a ~5x markup on chip cost. Google's chip partner Broadcom reportedly runs ~50% margins — a ~2x markup. Since chips represent over half the cost of running an AI data center, 'Google being definitively the low-cost provider of tokens because they operate all their own infrastructure and because they have access to low markup hardware — it actually makes a giant difference and might mean that they are the winner in producing tokens for the world.'

ChatGPT was an accident — and the accidental interface layer captured more value than years of model development

By late 2022, OpenAI had GPT-3.5. It was impressive. It had no front door.

Sam Altman said, roughly: can someone just make a chat? Within a week internally, someone did. 'They also just throw up a paywall randomly because they thought that the business was going to be an API business.' The paywall wasn't a product decision — it was a cost-dampening measure because servers were tipping over. 'To say that OpenAI had any idea what was coming would also be completely false. They did not get that this would be the next big consumer product when they launched it.'

On November 30, 2022, Sam Altman tweeted: 'Today we launched ChatGPT. Try talking with it here.' Within a week: 1 million users. By December 31: 30 million. By end of January 2023: 100 million registered users — the fastest product in history to that milestone. Ben Thompson's framing nailed it: OpenAI is 'the accidental consumer tech company.'

Meanwhile, Google had Noam Shazeer's internal chatbot — first called Meena, then Lambda — since the late 2010s. It existed. It wasn't shipped, partly for safety reasons (ask it who should die and it would generate names), partly because replacing ten blue links with a chatbot directly threatened the mechanism that generated $140 billion in annual profit. The marginal product decision — 'let's make a chat box' — turned out to be worth more in consumer mindshare than years of model development. Whoever makes transformative technology accessible first captures the narrative, regardless of who built the better model underneath.

Waymo burned less than one month of Google's profits to become a real business with a Google-sized market opportunity

For most of its existence Waymo looked like the canonical moonshot boondoggle — 15 years, no product, burning billions. The numbers tell a different story.

Waymo has now driven over 100 million miles with no human behind the wheel, is adding 2 million miles per week, runs hundreds of thousands of paid rides weekly across five cities, and is reportedly doing more gross bookings than Lyft in San Francisco as of January 2025. A study released by Waymo shows 91% fewer crashes with serious injuries or worse compared to the average human driver, controlled apples-to-apples. Over a million motor vehicle fatalities occur globally every year; 40,000 deaths per year in the US alone — 120 every day. The CDC puts total US crash costs at $470 billion annually. A 10x reduction in serious crashes implies over $420 billion per year in avoided costs — more than Google's entire annual revenue.

The total capital burned to reach this point: somewhere between $10 and $15 billion. 'That's one year of Uber's profits.' More relevantly: 'That cost of $10 to 15 billion is the profits that Google made last month.'

What looked like a distraction was funded by a rounding error of Google's cash generation. The question for long-duration bets is never 'will this ever work?' — it's whether the parent has enough runway to let the compounding happen. Google's search machine made Waymo possible.

The race ahead is really a cost-per-token race — and only one company has all four pillars to run it indefinitely

The standard framing of the AI race — OpenAI versus Google, frontier model benchmark scores, consumer chatbot market share — misses the actual structural question: who can fund billion-dollar training runs without a VC check?

Google is the only model maker with self-sustaining funding. Every other frontier lab is effectively a startup dependent on external capital. 'Google is the only model maker who has self-sustaining funding' — and they're generating so much excess cash they've started paying dividends and buying back stock on top of record AI capex. 'They're giving extra dollars back to shareholders for fun.'

The four pillars required to win in AI infrastructure — frontier model, custom chips, hyperscale cloud, self-sustaining profits — belong exclusively to Google as a complete set. 'There is no other company that has I think more than one.' Meta has applications. Amazon has cloud. Microsoft has cloud and a model partnership. Nvidia has chips. OpenAI and Anthropic have models and nothing else. Google has all four.

The AI era will ultimately be decided not by who has the best benchmark score in any given quarter, but by who can sustain the capital intensity of the arms race across a decade. That question has a clear current answer — and it traces back to a lunch conversation in a Google micro-kitchen in 2001, a cat paper in 2012, a transformer published in 2017, and a search monopoly that funds all of it. The innovator's dilemma cuts both ways: Google's greatest threat is also its greatest asset.

---

Topics: Google, artificial intelligence, Waymo, DeepMind, OpenAI, transformer, TPU, Google Cloud, innovator's dilemma, AI strategy, search, Gemini, competitive strategy, tech history, acquisitions