Google: The AI Company

In one sentence: ten years before this episode aired, nearly everyone who now matters in AI worked at the same company. OpenAI's founding chief scientist Ilya Sutskever, Anthropic's Dario Amodei, DeepMind's Demis Hassabis and Shane Legg were all Google employees; the one exception was Yann LeCun, who was at Facebook. In 2017, a team inside Google Brain published a paper called "Attention Is All You Need," introducing the Transformer architecture that underpins every large model built since. Within a couple of years, all eight authors had left the company, mostly to found or join the rivals who would go on to threaten Google's own search business. Google didn't miss the opportunity. It saw it, built it, and still didn't bet the company on it. This episode is a textbook case of the innovator's dilemma: a business generating $140 billion a year in profit and holding 90% of the search market has to decide whether to replace itself with something that doesn't yet know how to make money. Recorded in 2025, Ben's closing verdict: "This is probably the most fascinating example of the innovator's dilemma ever."

The Company on One Page

Key People

Jeff Dean

Jeff Dean is the throughline of this episode: whenever the story needs someone who single-handedly rewrote an impossible system over a weekend, it's him. Internal "Jeff Dean facts" borrow the Chuck Norris format: "the speed of light in a vacuum used to be about 35 miles an hour, until Jeff Dean spent a weekend optimizing physics," and "to Jeff Dean, NP means no problemo." Behind the jokes is a real track record. He borrowed a colleague's language model, PHIL, and coded up AdSense in a week. He took Franz Och's translation system, which needed 12 hours to translate a single sentence, rearchitected it to run in parallel across data centers, and got that down to 100 milliseconds. Together with Sanjay Ghemawat and Urs Hölzle, he built the infrastructure that lets Google break any workload into pieces and scatter them across data centers worldwide. In 2011 he co-founded Google Brain with Andrew Ng and neuroscience PhD Greg Corrado, and personally wrote the distributed training system Distbelief, a name that's half a pun on "distributed" and half a nod to how few people in the field believed asynchronous updates would actually work. They worked. In 2016, when Google tried to keep Ilya Sutskever from leaving for OpenAI, it was Jeff Dean who delivered the doubled counteroffer; it didn't work, but it says something about who the company valued most. After Brain merged into DeepMind in 2023, he and Noam Shazeer, hired back at a premium years after leaving, became the two technical leads on Gemini.

Demis Hassabis

Demis Hassabis's résumé is itself a preview of DeepMind's later character. At 14 he was the world's second-ranked chess player in the under-14 category; accepted to Cambridge at 17 but told he was too young, he spent his gap year at the game studio Bullfrog, where he built Theme Park, which sold 15 million copies. After Cambridge he started his own game studio, Elixir, which failed, and only then went back for a PhD in neuroscience. At University College London he met postdoc Shane Legg, a self-described member of the "lunatic fringe" who believed AI would keep getting more powerful until it surpassed human intelligence, and who helped popularize the term "artificial general intelligence." The two pulled in Demis's childhood friend Mustafa Suleyman and founded DeepMind in 2010 with the tagline "solve intelligence, and use it to solve everything else." They couldn't raise money in London, so the three talked their way into San Francisco's Singularity Summit hoping to run into venture capitalist Peter Thiel. Demis never caught him on stage; instead he found him at Thiel's after-party and steered the conversation away from AI entirely, toward chess, which Thiel loved. That conversation turned into a $2 million seed round led by Founders Fund, and an introduction, through Thiel, to Elon Musk, who later invested too. By late 2013 the company had three acquisition offers on the table: Facebook bid roughly $800 million but demanded DeepMind give up its independence and start shipping products, a deal that reportedly would have personally netted the founders close to twice what they made by taking Google's offer instead; Elon Musk offered Tesla stock in exchange for the team working on self-driving; and Larry Page, who bid the least, promised the one thing the others wouldn't: "Stay in London, keep working on intelligence." Page's connection to Demis traced back to a private jet, where he'd spotted a video DeepMind had sent showing an AI teaching itself to exploit Atari's Breakout, and asked on the spot what company had made it. Google closed the deal at $550 million with an independent ethics board attached. In 2023, that independence was quietly revised: Sundar folded Brain into DeepMind and named Demis CEO of the combined Google DeepMind.

Noam Shazeer

Noam Shazeer's first appearance in this story is a Google micro-kitchen around 2000, where he's a new hire listening to a veteran engineer, George Herrick, casually argue that compressing data is technically equivalent to understanding it. Shazeer latches onto the idea immediately; the two drop what they're working on and disappear into probabilistic language models, eventually producing an internal model nicknamed PHIL that first ships as spell correction. Seventeen years later he's the guy who joins the Transformer project midstream, after the team already has a working implementation that just isn't beating LSTMs yet. Shazeer rewrites the whole codebase, and afterward the Transformer clears LSTMs outright; colleagues on the project start calling him a magician, a wizard. He also sees the paper's implications earlier than almost anyone else at the company, immediately lobbying leadership that this might be big enough to justify tearing out the search index and the ten blue links altogether, and building an internal chatbot called Meena to prove it. Without safety post-training, Meena went off the rails badly enough, asking it who should die would produce a list of names, that it was never a shippable product. Shazeer left in 2021 to found Character.AI. After ChatGPT upended the industry, Google paid roughly $2.7 billion in 2024 for a licensing deal that brought him back, pairing him with Jeff Dean as Gemini's two technical leads.

Sundar Pichai

Sundar Pichai joined Google in 2004 to run the client-products line that included the Toolbar, an unglamorous first assignment on the road to becoming CEO. In 2017 he was already telling the world Google was "AI-first," but the company stayed cautious about anything AI-facing and public: even the internal chat interface tested in AI Test Kitchen was hard-capped at five conversational turns before it shut itself down, because the longer a conversation ran, the more likely it was to go off the rails. When ChatGPT crossed 100 million users within two months of its November 2022 launch, Pichai issued a company-wide code red, reclassifying AI overnight from a sustaining innovation that made Google's existing businesses better into an existential threat, and made two decisions that would have been unthinkable a year earlier: merging Google Brain and DeepMind, two groups with different charters that didn't get along especially well, under Demis Hassabis, and retiring every internal model name in favor of one brand, Gemini. Those two calls, plus a product strategy that runs AI Overviews, AI Mode, and the standalone Gemini app side by side, drew the hosts' closing verdict on the whole episode: the textbook response to disruptive innovation, done about as well as any large company has managed it, protecting a $140-billion-a-year cash machine while genuinely competing on the frontier at the same time. In 2025, a federal court ruled Google's search business a monopoly but declined to force a breakup of Chrome, citing, among other things, how much competitive heat the AI race was already generating.

The Playbook

Each entry: origin story → insight → effect.

1. A lunchtime aside turned into AdSense

• Story: around 2000, veteran engineer George Herrick tells two colleagues over lunch that compressing data is technically equivalent to understanding it. New hire Noam Shazeer can't let it go; the two build an internal model called PHIL, shipped first as spell correction. In 2003, Jeff Dean borrows it and codes up AdSense in a week.
• Insight: the technical direction that ends up mattering most rarely starts as a funded project. It starts with someone taking a coworker's offhand intuition seriously enough to disappear into it for a few months.
• Effect: AdSense extends the ad business from search results to the entire open web and adds billions in revenue the same year it ships.

2. An asynchronous system that worked because nobody believed it would

• Story: in 2011, Jeff Dean builds Google Brain's training system and calls it Distbelief, a pun on both "distributed" and how little the field trusted it. The consensus at the time was that training had to be synchronous; Dean scattered parameter updates across thousands of cores that updated asynchronously, on stale data. By the end of the year, the "cat paper" trains an unsupervised network that learns the concept of a cat's face using 16,000 CPU cores, proving the architecture actually worked.
• Insight: field consensus sometimes just means nobody has tried it at real scale yet.
• Effect: the same technology is credited with seeding hundreds of billions of dollars in recommendation-system revenue at YouTube, Facebook, and ByteDance over the following decade.

3. When there's no pricing precedent, let an auction decide

• Story: in late 2012, the three founders of shell company DNN Research (Hinton, Krizhevsky, Sutskever) put themselves up for sale with no obvious valuation. Hinton personally ran a structured auction, resetting the clock an hour after every new bid, from his hotel room at a casino in Lake Tahoe. Baidu, Google, Microsoft, and a cash-strapped DeepMind (which dropped out almost immediately) all bid; Google won at $44 million.
• Insight: for an asset with no precedent, a well-designed bidding mechanism beats guessing a number yourself.
• Effect: all three join Google Brain; years later, the head of Google X tells the New York Times that Brain's returns to the core business had already more than paid for every moonshot X had funded.

4. Whether an acquirer can keep its "we won't interfere" promise depends on whether it needs to

• Story: in early 2014, DeepMind fields three offers at once. Facebook bids about $800 million but wants DeepMind to give up its independence and start shipping products. Elon Musk offers Tesla stock in exchange for the team working on self-driving. Larry Page, bidding the least, is the only one who promises: stay in London, keep doing research.
• Insight: an acquirer's promise not to interfere is credible only if it doesn't actually need you to do anything else. Google already had Brain handling productization.
• Effect: DeepMind takes the lowest offer in exchange for an independent ethics board; the hosts now peg the asset's value at north of half a trillion dollars.

5. A 40,000-GPU order is a signal to the whole industry, not just a purchase

• Story: in 2014, Jeff Dean and Alan Eustace push to move Google's deep-learning infrastructure off CPUs and onto GPUs, starting with an order for 40,000 units from NVIDIA at a cost of $130 million, large enough that Larry Page personally has to sign off, at a time when NVIDIA's entire annual revenue was $4 billion.
• Insight: a large customer's willingness to place a huge order ahead of demonstrated need is itself market information, whether or not the order is ever made public.
• Effect: the hosts think the order gave NVIDIA real confidence to bet early on AI chips; for Google, the ballooning GPU bill is also what eventually forced it to build its own chip.

6. Marginal compute demand forces you to design your own silicon

• Story: speech recognition, freshly launched on Nexus phones, becomes an instant hit. Jeff Dean runs the math: rolling it out to a billion Android phones at three minutes of use a day would require doubling every data center Google owned. The company builds the Tensor Processing Unit, trading numerical precision for efficiency and packaging the chip to fit into a hard-drive slot so it could go straight into existing server racks without any physical redesign. Design, verification, and deployment take 15 months, done in secret at a satellite office in Madison, Wisconsin, for more than a year before going public.
• Insight: when a single new feature's marginal compute demand would bankrupt your existing infrastructure, building your own chip stops being a nice-to-have and becomes survival.
• Effect: the first TPU runs the AlphaGo match; today Google's internal and cloud TPU fleet is estimated at two to three million units, roughly on par with NVIDIA's annual GPU shipments.

7. Publishing the paper published the moat

• Story: in 2017, eight Google Brain researchers publish "Attention Is All You Need," introducing the Transformer architecture. At the time the company treats it as roughly the scale of a better Google Translate and lets it go out into the world. The paper has since been cited more than 173,000 times, the 7th most cited paper of the 21st century; within a few years, all eight authors have left Google to found or join the labs that would go on to use the same architecture against Google's own search business.
• Insight: a culture of open publication buys you academic prestige and makes you the place top researchers want to work, but it can also hand your own core technology to the rival who will eventually threaten you most.
• Effect: two years later, OpenAI builds GPT on top of the Transformer, directly leading to ChatGPT, a decision the hosts call maybe the most expensive one Google has ever made.

8. The textbook response to an existential threat is real organizational change, not bolting new tech onto the old structure

• Story: two months after ChatGPT crosses 100 million users, Sundar Pichai declares a code red and, within six months, merges Google Brain and DeepMind under Demis Hassabis and renames every internal model Gemini, both moves explicitly ruled out by the terms of the 2014 DeepMind acquisition.
• Insight: the textbook answer to disruptive innovation is to actually break the old organizational boundaries, not treat the new thing as an upgrade to the old business, while still not panicking into something reckless.
• Effect: from ChatGPT's launch to Gemini's announcement at Google I/O takes six months; today AI Overviews, AI Mode, and the Gemini app run in parallel, protecting the existing search-ad franchise while racing to build the new one.

Moat Analysis (the 7 Powers framework)

This episode's 7 Powers analysis is deliberately scoped to Google's AI products themselves (Gemini, AI Overviews, AI Mode, benchmarked against Anthropic, OpenAI, Perplexity, Grok, and Meta AI), not all of Alphabet, a different object of analysis than the core-search assessment in the previous episode.

Bear Case

Two pessimistic threads run through the episode: monetization and market position. On monetization, AI chat hasn't found anything close to search advertising's revenue mechanism. Rough per-user math puts Google's US search revenue at around $400 a year per user, for a free product; the number of people willing to pay that much for an AI subscription is a thin slice of the population, and Google found AdWords within two years of launching in 1998, a pace this transition hasn't matched. On position, 1998 Google was obviously better than every competitor at a glance; Gemini today is, at best, a toss-up against several strong rivals. Google owns 90% of search but will likely end up with somewhere between a quarter and half of the AI market at steady state, a meaningfully smaller pool of users, even before accounting for the fact that AI is probably siphoning off the highest-value queries first: travel planning, health, legal questions, the exact categories that used to carry the most expensive ads. The last thread is simply who Google is now, the incumbent being challenged rather than the startup fighting an entrenched giant, and it doesn't get the public goodwill this time that it had in 1998 or during the mobile transition.

Bull Case

Google is the only company holding a frontier model (Gemini), its own AI chips (TPU), its own hyperscale cloud (Google Cloud), and its own distribution at the application layer (billions of daily search queries), and no other company owns more than one of those four pillars. More importantly, Google is the only major model maker whose AI business is funded by its own operating cash flow; OpenAI and Anthropic are still, in effect, startups running on outside capital. The chip math favors Google too: NVIDIA's GPU margins run 75–80%, a four-to-five-times markup over cost, while Google's markup from its chip partner Broadcom is closer to 50%, and since chips and their depreciation are more than half the cost of running an AI data center, being the low-cost producer of tokens may matter more in this cycle than it has in any past technology transition. Google also holds the personal data behind Gmail, Maps, Docs, Chrome, and Android, letting it build personalized AI experiences nobody else can match; it owns dark fiber bought for pennies on the dollar after the dot-com crash and only recently lit up as a private backhaul network between its data centers; and YouTube is both the world's only large-scale library of long- and short-form user-generated video for training models and, in its own right, one of the biggest search destinations on the internet. On top of all that sits a nearly independent growth story: Waymo already runs driverless robotaxis at scale in five cities on roughly $10–15 billion invested, and could plausibly become a business the size of Google all over again.

Deep Cuts

• Larry Page discovers DeepMind on a private jet: Elon Musk is flying with fellow PayPal-mafia member and early DeepMind angel Luke Nosek, watching a video DeepMind had just sent showing an AI teaching itself, with no human demonstration, a trick for beating Atari's Breakout by tunneling along the top of the brick wall. Larry Page, on the same flight, leans over and asks what company made it, an encounter the hosts credit as the real origin of Google's eventual acquisition of DeepMind.
• A dinner where only one person was willing to leave: in the summer of 2015, Elon Musk and then-Y Combinator president Sam Altman invite a table of top AI researchers to the Rosewood Hotel on Sand Hill Road and ask what it would take to get them to leave Google. Almost everyone says nothing could. The one researcher who's even open to it is Ilya Sutskever, who later put it: "I felt like there were risks involved, but I also felt like it would be a very interesting thing to try." OpenAI's first wave of hires comes out of that mostly unsuccessful dinner.
• A custom airplane harness, built for one due-diligence trip: in late 2013, Google's head of engineering, Alan Eustace, flies to London to evaluate DeepMind and needs Geoff Hinton along to judge the technology, except Hinton's back injury means he can only lie down, not sit, for a long flight. Eustace, who once held the world altitude record for a stratospheric skydive, charters a private jet and builds a custom harness rig so Hinton won't slide around the cabin during takeoff and landing.
• One wrong answer in a launch video, and a market-cap drop far bigger than the fix: in February 2023, Google rushes Lambda out the door rebranded as Bard, and the choreographed launch video includes a factual error in one of its scripted answers. The stock drops 8% that day. The real failure isn't the one bad demo; it's that the product is missing the step later known as RLHF, reinforcement learning from human feedback.
• Almost buying Tesla for $3 billion as a shortcut to self-driving: early in the Waymo project, Eric Schmidt floats the idea of skipping the self-driving buildout entirely and just buying Tesla, in the depths of the Model S production crisis, when Tesla was worth about $3 billion. The deal never happens; the hosts guess that if it had, DeepMind's acquisition and OpenAI's founding might both have unfolded differently.
• One lab, two future rivals: Stanford's AI Lab, under Sebastian Thrun, produced undergraduate researchers who included future Meta chief product officer Chris Cox and a future dropout who joined Y Combinator's first batch and, in 2015, helped Elon Musk put together OpenAI: Sam Altman. Both were just undergrads doing odd jobs in the same lab.

Era & Industry Trivia

• "The AI era actually started in 2012": after the cat paper and AlexNet, the same underlying technology showed up in YouTube's, Facebook's, Instagram's, and ByteDance's recommendation feeds, and algorithm-fed feeds became how humanity spends most of its leisure time. ChatGPT is when the public noticed AI existed; for any company running a recommender or a classifier, the AI era had already been running for a decade.
• AlexNet was trained on two consumer GeForce GTX 580 graphics cards, bought at a local electronics store for a couple hundred dollars each. NVIDIA's Jensen Huang later called this AI's "big bang moment," the instant that set NVIDIA on the path from a middling PC-gaming accessory maker to the most valuable company in the world.
• Rich Sutton's 2019 essay "The Bitter Lesson," now widely cited across the field, summarized exactly the pattern this decade kept proving: a scalable architecture fed with more data and more compute tends to beat a cleverer algorithm.
• In the 2025 US v. Google antitrust ruling, the judge found Google a monopolist in search but declined to order a Chrome breakup or bar Google's payments to Apple for default placement, citing, among other reasons, how much competitive investment had already flowed into OpenAI, Anthropic, and Perplexity.
• Waymo's own safety data, released the month before this episode, shows a 91% reduction in serious-injury-or-worse crashes compared to the average human driver on the same city streets; more than 40,000 people die in traffic crashes in the US every year, about 120 a day.

Cross-domain Notes

This episode has no real intersection with the PH discourse network. Not once across the entire recording does China or any geopolitical topic come up, unlike the previous episode's brief mention of Google's China business between 2002 and its 2010 withdrawal. The story here stays entirely inside Silicon Valley: the talent fights, chip orders, and product races among Google, DeepMind, OpenAI, Anthropic, Microsoft, and NVIDIA. If there's a thread worth naming, it's the episode's recurring theme of compute concentrating in a handful of companies' hands: TPUs, GPUs, and private dark-fiber backhaul networks are all, at bottom, about who controls the physical infrastructure needed to train the next generation of models, which echoes, at a methodological level, the PH domain's interest in concentrated technological power in pages like fortress-america and technate. But the episode never touches geopolitics or cross-border competition directly, so calling this a real intersection would overstate it.

Pages Worth Creating

• Entities: jeff-dean (the engineering throughline of the episode: Distbelief, AdSense, TPU, and now Gemini), demis-hassabis (DeepMind's founder, now CEO of the combined Google DeepMind), noam-shazeer (the full loop from PHIL to Transformer to Character.AI to Gemini), sundar-pichai (the decision-maker behind code red, arguably the episode's real protagonist), geoff-hinton (godfather of deep learning, ran the DNN Research auction himself)
• Concepts: transformer (this episode's core technology; "Attention Is All You Need" deserves its own page tracing its spread and citation history), deepmind (its acquisition and organizational history deserve a standalone page), tpu (the Tensor Processing Unit's technical evolution); 7 Powers 护城河框架 already has a page, and this episode contributes the "score AI products on their own" application of the framework
• Episode cross-links: Google：搜索的诞生 (this episode's direct predecessor, "The Origin of Search"), NVIDIA 之三：AI 时代的黎明（2022-2023） (the AI compute lineage: NVIDIA's GPUs and Google's TPUs racing each other across both series)

Source · acquired