Google Part III: The AI Company. Google is amazingly well-positioned... will they win in AI? (Audio)
Summary
Google enters the AI platform shift with the only end-to-end stack the hosts can identify: a frontier model in Gemini, millions of TPUs, a $50 billion-plus cloud business, global consumer distribution, and a search monopoly funding the whole effort. That makes AI Google’s “birthright,” but also creates the episode’s central innovator’s dilemma: the better direct-answer product may be structurally less profitable than the “giant cash printing business” it could replace.
The modern AI industry grew from research Google repeatedly invented, financed, productized internally, and then allowed to escape. Early probabilistic language models powered “Did you mean?” and AdSense; Google Brain’s cat paper launched the recommender era; Google acquired Geoffrey Hinton’s team and DeepMind; and eight Brain researchers published the 2017 transformer paper. Yet all eight transformer authors eventually left, while OpenAI, Anthropic, NVIDIA, Character.AI, and others commercialized the resulting platform shift.
Google’s decisive infrastructure advantage began when neural networks threatened to require “another Google” worth of data centers. It ordered 40,000 NVIDIA GPUs for $130 million, then designed and deployed the first TPU in only 15 months using reduced precision, a hard-drive-compatible form factor, and FPGA prototypes. The hosts cite an estimated 2–3 million TPUs today, approaching the scale of NVIDIA’s roughly 4 million GPU shipments last year and potentially making Google the industry’s lowest-cost producer of tokens.
ChatGPT converted AI overnight from a sustaining technology for search and ads into an existentially disruptive one. OpenAI’s improvised GPT-3.5 chat interface reached 1 million users in under a week, 30 million in a month, and 100 million within two months; Microsoft then declared “a new day for search” and wanted the world to know it had “made Google dance.” Google’s rushed Bard response was inferior and factually wrong in its launch material, contributing to an 8% one-day stock decline.
Sundar Pichai’s recovery has been “rapid but not rash”: merge Brain and DeepMind, appoint Demis Hassabis, standardize the company on Gemini, and ship at frontier-lab cadence without immediately destroying search economics. Gemini progressed from its December 2023 release to Gemini 1.5’s 1 million-token context window, Gemini 2.0, and Gemini 2.5 Pro; Google simultaneously layered AI Overviews and AI Mode onto search while keeping Gemini as a separate full-chat product. The result is a delicate “ballet” between adoption and cannibalization, not a wholesale replacement of google.com.
Google can finance the most capital-intensive race in business history from a franchise still producing $370 billion of annual revenue and $140 billion of earnings. It has invested aggressively in AI infrastructure while retaining $95 billion in cash and marketable securities, buying back stock, and paying a dividend; Google Cloud now exceeds a $50 billion annual run rate and grows about 30%. Rivals with strong models but no self-funding cash engine remain dependent on external capital and hyperscaler infrastructure.
Waymo is the clearest proof that Google can sustain a technically difficult AI bet long enough for “slowly and then all at once” to arrive. After roughly $10–15 billion of cumulative investment, Waymo has logged more than 100 million fully driverless miles, operates about 2,000 vehicles, and reports 91% fewer crashes involving serious injury or worse than comparable human driving. Against the CDC’s estimated $470 billion in total U.S. costs from crash deaths in 2022, the hosts now see a plausible “Google-sized opportunity” where they once saw a wild-goose chase.
The investment case turns on whether Google’s full-stack cost advantage and distribution can outweigh worse monetization and lower market share in AI. The bull case is that longer prompts expose more intent, Google can bundle AI across 150 million Google One subscribers, and its chips, data, YouTube corpus, applications, and private network compound; the bear case is that AI siphons off lucrative travel and health queries while supporting neither search-like ads nor 90% share. The unresolved question is stark: “Would they rather go bankrupt than lose at AI”—and will they still choose the mission if AI never becomes as good a business as search?
Deep dive
1. Google owns the AI stack but may have to destroy its best business
Ben frames Google’s position as the textbook innovator’s dilemma: a government-defined search monopoly with roughly 90% share has invented a product that is better for many purposes, but has not found a way to make it “anywhere near as profitable” as the incumbent one.
The strategic assets are unusually complete: Gemini is a top-tier model, Google Cloud exceeds $50 billion in revenue, TPUs are the only AI-chip deployment the hosts consider comparable in scale to NVIDIA GPUs, and google.com remains “the front door to the internet” whenever users express intent.
One researcher’s formulation becomes the investor shorthand: without either a frontier foundational model or an AI chip, a participant “might just be a commodity in the AI market.” Google is the only company the hosts identify as having both—and it also owns the cloud and applications through which they are distributed.
2. Larry Page conceived Google as an AI company from the beginning
Larry’s father studied machine learning and artificial intelligence when the field was deeply unfashionable, and that contrarian orientation carried into Google. Even PageRank can be understood as a statistical, AI-adjacent system.
Larry’s 2000 formulation already contained the end state: “Artificial intelligence would be the ultimate version of Google.” The ultimate search engine would understand everything online, know exactly what the user wanted, and return “the right thing”; ordinary search was merely an incremental route toward that goal.
Google’s later dilemma therefore is not an unrelated disruption invading from outside. AI is simultaneously the company’s founding mission, its greatest research accomplishment, the mechanism behind billions in existing revenue, and a product shape that could dissolve the economics of its original implementation.
3. Compression became Google’s first theory of machine understanding
Around 2000–01, early engineer Georges Harik told Ben Gomes and new hire Noam Shazeer that compressing data might be technically equivalent to understanding it: if a system can reduce information and later reconstruct it, it must have captured the meaningful structure that survived the compression.
The hosts connect Harik’s school-test analogy directly to modern language models: the world’s knowledge is compressed into a comparatively small collection of parameters, then “uncompressed” into useful answers. David cautiously challenges the anthropomorphism—models may demonstrate understanding, or they may merely “mimic understanding.”
Harik and Shazeer ignored internal skeptics and pursued probabilistic natural-language models. Harik’s defense was pure early-Google culture: “Sanjay thinks it’s a good idea, and no one in the world is as smart as Sanjay,” referring to Sanjay Ghemawat.
Their first commercial proof was Google’s “Did you mean?” system, a next-token-prediction ancestor that inferred a likely intended query from observed corrections. Besides improving the experience, it prevented infrastructure from repeatedly serving useless mistyped searches that users immediately replaced.
4. PHIL turned language modeling into billions of dollars
Harik and Shazeer expanded the work into PHIL, the Probabilistic Hierarchical Inferential Learner, a “large” language model by early-2000s standards. By predicting sequences and extracting meaning, it became useful far beyond spelling correction.
When Susan Wojcicki and Jeff Dean prepared AdSense in 2003, PHIL supplied the missing understanding of third-party webpages. Dean reportedly implemented the system in a week, matching Google’s existing AdWords corpus to publisher content and creating billions of dollars of incremental revenue by massively expanding ad inventory.
By the mid-2000s, PHIL reportedly consumed 15% of Google’s entire data-center infrastructure across AdSense, spelling correction, and related applications. Language modeling was therefore already computationally expensive, production-grade, and directly tied to Google’s bottom line two decades before ChatGPT.
5. Jeff Dean converted a 12-hour research model into a 100-millisecond product
Google Translate’s chief architect, Franz Och, trained an enormous n-gram model on two trillion words from Google’s search index and achieved a spectacular BLEU score in a DARPA translation challenge. The catch was that translating one sentence took about 12 hours—acceptable when submissions were due days later, useless for consumers.
Dean rearchitected the work so words and sentence components could be processed in parallel across Google’s distributed CPU infrastructure instead of sequentially. After several months, average translation latency fell from 12 hours to roughly 100 milliseconds, and the research system shipped in Google Translate.
The result established a recurring Google playbook: pair frontier machine-learning research with infrastructure engineering until it becomes a product at global latency and scale. The same predictive machinery then improved search-query completion and AdWords quality scores, where better click-through-rate estimates translated directly into revenue.
6. Google built an academic talent monopoly before AI was fashionable
In 2007 Larry recruited Sebastian Thrun from Stanford’s Artificial Intelligence Laboratory, effectively replacing a startup financing process with Google signing bonuses. Thrun brought a network of professors, researchers, and students that had included future Meta product chief Chris Cox and, briefly, Stanford undergraduate Sam Altman.
After successful work on Street View and the Ground Truth maps project, Thrun persuaded Larry and Sergey to let academics work part-time at Google while retaining university posts. Researchers gained compensation, infrastructure, colleagues, and the chance to deploy work to millions; Google concentrated the scarce practitioners from highly specialized PhD lineages.
Thrun invited Geoffrey Hinton to give a 2007 Google talk on deep neural networks, then brought him in as a consultant and, later, a roughly 60-year-old “summer intern.” Neural nets remained heretical because earlier hype had failed, but Hinton argued that multilayer systems needed more compute, not a different intellectual foundation.
7. Google Brain proved asynchronous distributed neural networks could work
Around 2010–11, Andrew Ng and Jeff Dean decided to take a large swing on a deep-learning model within Google X’s Google Brain effort, after earlier projects called Brain on Borg and Cortex had disappointed. Their ambition was a genuinely large deep neural network running across Google’s highly parallel CPU infrastructure.
Dean named the training system DistBelief: partly a reference to distributed computation, partly because “no one thought it was going to work.” Conventional research favored tightly synchronized computation; DistBelief instead updated parameters asynchronously, sometimes using stale information, and nonetheless worked.
That systems breakthrough mattered as much as the model. Google could split both data and models across enormous fleets, turning an architecture regarded as computationally impractical into something its existing infrastructure could train.
The resulting “cat paper” used a nine-layer network, 16,000 CPU cores, and 10 million randomly selected, unlabeled YouTube frames. A high-level neuron became excited by cats without ever being told what a cat was—evidence that meaningful representations could emerge through unsupervised learning.
8. The cat neuron quietly launched the first AI era
YouTube’s immediate problem was weak metadata: uploaders described videos poorly, leaving search and recommendation systems unable to know what the underlying content contained. If a neural network could answer “cat or not a cat” directly from frames, it could classify almost anything that drove relevance.
The hosts trace a huge commercial lineage from that result: YouTube recommendations, feeds, dwell time, copyright matching, revenue sharing, explicit-content filtering, and eventually similar systems at Facebook, Instagram, TikTok, Reels, and Shorts. This was not laboratory curiosity; it reorganized how humanity spent leisure time.
David’s signature correction is that the AI era began in 2012, not with ChatGPT in 2022. For any company with a feed, classifier, or recommender, neural networks were already “shaping human existence” and generating hundreds of billions of dollars—only invisibly, inside existing products.
9. AlexNet made gaming GPUs the engine of modern AI
Hinton and Toronto graduate students Alex Krizhevsky and Ilya Sutskever entered the 2012 ImageNet competition using a deep neural network trained on two consumer NVIDIA GeForce GTX 580 cards. They rewrote the algorithms in CUDA instead of relying on the supercomputer-class CPUs then associated with advanced research.
ImageNet’s best previous systems misclassified about 25% of images; AlexNet cut the error rate to 15%, a 10-point absolute improvement and roughly 40% relative gain in one year. The discontinuity convinced a previously skeptical computer-science community that deep learning had crossed into practical superiority.
Jensen Huang’s description of AlexNet as AI’s “big bang moment” carries both technical and financial meaning: two off-the-shelf gaming cards unlocked the next research level and put NVIDIA on its path from a PC graphics supplier to the most valuable company in the world.
10. Google acquired Hinton’s team through a professor-run auction
Hinton, Krizhevsky, and Sutskever formed DNNResearch without a product—only the researchers who had won ImageNet. Baidu opened with a $12 million offer, prompting Hinton to construct a timed auction in which every new bid reset the clock for another hour.
Microsoft, Google, Baidu, and cash-constrained DeepMind participated while Hinton managed proceedings from a hotel room at Harrah’s Lake Tahoe during the NIPS conference. When the researchers decided Google was where they wanted to work, they stopped the auction at $44 million rather than maximizing the final price.
Krizhevsky and Sutskever asked Hinton to take 40% while each retained 30%. The trio joined Google Brain, where even percentage improvements in search, ads, Gmail, and YouTube could uncover “quite a bit of loose change in those couch cushions” across tens or hundreds of billions of revenue.
11. DeepMind pursued general intelligence while others pursued classifiers
DeepMind was founded in London in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Hassabis combined childhood chess prodigy, commercial game designer, failed gaming entrepreneur, computer scientist, and neuroscience PhD; Legg belonged to AI’s self-described “lunatic fringe” and helped popularize “artificial general intelligence.”
Their leap was conceptual: AlexNet and the cat paper classified patterns, while DeepMind proposed creating an intelligence that could generalize beyond narrow tasks and eventually become smarter than humans. Its enduring mission was “solve intelligence and use it to solve everything else.”
The website’s vague promise of general-purpose algorithms for “simulations, e-commerce, and games” reflected an honest absence of near-term products. What DeepMind primarily needed was enough capital and compute to conduct open-ended research—not a conventional startup plan with customers and unit economics.
12. Peter Thiel and Elon Musk funded the mission because it sounded impossible
Hassabis and Legg targeted the 2010 Singularity Summit because conventional London investors were unlikely to fund an AGI lab. Legg’s requirement was someone “crazy enough” not to sweat a few million dollars, who liked “super ambitious stuff” and would ignore professors saying it could not work.
Peter Thiel missed Hassabis’s carefully designed conference talk, so Hassabis approached him at an afterparty through their shared interest in chess. The next-day pitch led Founders Fund to anchor an approximately $2 million seed round—tiny by later AI standards, but enough to begin hiring and experimenting.
Through the PayPal network, Hassabis met Elon Musk at SpaceX. When Musk described Mars as humanity’s backup, Hassabis asked what happened if AI was the catastrophe: an AI could reach Mars through communications or any systems humans brought with them.
Musk reportedly sat silently, concluded “that’s probably true,” and soon invested. The hosts treat this as the moment his concern about safe AI flipped on, while also intensifying his belief that machine vision and learning could become central to Tesla’s autonomous-driving ambitions.
13. Google won DeepMind by offering mission, infrastructure, and separation
By late 2013 Mark Zuckerberg reportedly offered as much as $800 million for DeepMind, roughly twice what its founders would make from Google’s proposal. Facebook would accommodate unusual arrangements, but Zuckerberg would not grant Hassabis lasting independent control; DeepMind would become part of Facebook.
Musk countered with Tesla stock when Tesla’s market capitalization was around $20 billion—stock the hosts estimate later appreciated roughly 70-fold. Yet Musk wanted DeepMind focused on autonomous driving, while the founders refused to confine the lab to one product problem.
Larry learned about DeepMind after seeing its Atari Breakout system on a private plane with Musk and Luke Nosek. The system independently discovered a strategy of bouncing the ball up around the edges of the bricks and then along the top, and Larry immediately wanted to know who had built it.
Hassabis told the hosts that Larry simply “got it.” Google already had Brain working with product groups, so DeepMind could remain in London pursuing intelligence; Google also offered unmatched compute and an independent oversight board. The resulting January 2014 acquisition cost $550 million.
14. DeepMind paid back strategically before it produced a consumer product
One early internal application used neural networks to optimize data-center cooling. Google announced in July 2016 that DeepMind had reduced the energy required for cooling by 40%—a saving across Google’s infrastructure that the hosts believe could have justified the acquisition rapidly.
AlphaGo then demonstrated creative machine reasoning against world Go champion Lee Sedol. It won the first three games of a five-game match, and “move 37” initially appeared mistaken before revealing a strategy so inventive that human players began learning from the machine.
Go supplied the right test because a turn offers roughly 200 choices versus about 20 opening moves and 30–40 midgame choices in chess. Hassabis said that, as of 2017, even all the world’s computers running for a million years could not enumerate every variation, forcing AlphaGo to learn representations and strategies instead of brute-forcing the game.
15. OpenAI began as an explicit counterweight to Google’s talent concentration
Google’s DeepMind acquisition infuriated Musk, who had invested in the company and wanted its capabilities inside Tesla. In summer 2015, he and Sam Altman assembled leading researchers for a Rosewood Hotel dinner and asked what could lure them away from Google’s money, peers, academic freedom, and infrastructure.
Almost everyone answered “nothing.” Ilya Sutskever alone found the risk worth taking: “I felt like there were risks involved, but I also felt like it would be a very interesting thing to try”—a line the hosts call “the most Ilya quote of all time.”
Google reportedly countered with roughly double OpenAI’s offer, delivered personally by Jeff Dean, but Sutskever stayed committed. His decision supplied the activation energy for about seven researchers, Greg Brockman from Stripe, Altman, and Musk to form a nonprofit research lab.
OpenAI pledged to advance digital intelligence for humanity “unconstrained by a need to generate financial return.” Backers announced $1 billion of commitments from Musk, Altman, Reid Hoffman, Jessica Livingston, Thiel, and others, though only about $130 million was ultimately collected.
16. Early OpenAI copied DeepMind until its funding model broke
OpenAI initially behaved like a university or DeepMind-style lab: recruiting researchers, publishing, and building agents for Dota 2, a Rubik’s Cube, Universe, Atari titles, and open-world games. Emergent strategies were scientifically interesting, but no project clearly coalesced into a product or commercial engine.
Dario Amodei left Google Brain for OpenAI in early 2016 and, with Sutskever, led major research efforts before later founding Anthropic. Even with that talent, Musk increasingly saw a collection of experiments copying DeepMind rather than an AI program useful to Tesla or moving decisively toward its mission.
After Google published the transformer paper, Musk demanded either full control—potentially merging OpenAI into Tesla—or his complete departure with future funding. The board refused, and by early 2018 OpenAI had lost its primary financial backer just as scalable transformers made capital requirements explode.
17. Google’s GPU order signaled an industrial market before NVIDIA saw one
When Krizhevsky joined Google in 2013, he was astonished that machine-learning workloads still ran on CPUs. He bought a GPU machine locally, placed it in a nearby closet, attached it to Google’s network, and resumed training as he had in academia—now with Google paying for electricity.
In 2014 Jeff Dean and Alan Eustace proposed formally adding 40,000 NVIDIA GPUs to Google’s fleet for $130 million. Finance resisted, but Larry personally approved the order because “the future of Google is deep learning.”
The order was enormous relative to NVIDIA’s approximately $4 billion annual revenue and $10 billion market capitalization. To the hosts, it effectively revealed that neural networks were valuable enough in production for a sophisticated customer to invest nine figures, likely strengthening NVIDIA’s conviction to build aggressively for the market.
18. The TPU emerged when one feature threatened to require another Google
Neural-network speech recognition initially ran only on Nexus phones because Google lacked capacity for every Android device. Dean calculated that three minutes of daily use across one billion phones would require roughly twice Google’s existing data-center fleet: “We need another Google.”
Jonathan Ross’s FPGA work supplied a bridge to a custom application-specific integrated circuit optimized for matrix multiplication. The TPU traded generality for efficiency, using reduced numerical precision and software quantization so the same memory, transistors, and power could execute many more calculations.
The team designed, verified, fabricated, and deployed the first TPU in 15 months. FPGA prototypes validated the mathematics, while a hard-drive form factor let technicians remove a drive and insert a TPU without physically redesigning server racks—a quintessential Google infrastructure shortcut.
TPU V1 was “not great,” but later generations converged toward GPUs while adding capabilities. The hosts cite an estimated 2–3 million Google TPUs today versus roughly 4 million NVIDIA GPUs shipped last year, making TPUs a scale platform rather than an experimental internal chip.
19. The transformer solved memory by making attention parallel
Google’s neural translation overhaul initially relied on recurrent neural networks, then long short-term memory networks. LSTMs cut translation error by 60% in 2016 because they retained more context, but remained computationally intensive and poorly suited to the parallel hardware revolution Google had built.
Jakob Uszkoreit and colleagues explored broader “attention”: instead of processing only nearby words sequentially, let the model examine the entire passage when predicting each translated word. The hosts compare this to professional human translation, where understanding the whole source precedes rendering individual sentences.
The architecture was compute-heavy but extremely parallelizable. The group called it the “transformer” both because it transformed one information representation into another and because several members liked the children’s franchise.
Shazeer joined after the early implementation failed to beat LSTMs, rewrote the codebase from scratch, and returned saying, “It works now.” The transformer crushed the incumbent system, and results continued improving as the team enlarged the model—revealing a scalable architecture rather than a narrow translation trick.
20. Elegance and scaling laws displaced clever hand-built algorithms
Google Brain co-founder Greg Corrado emphasized how suspiciously simple the transformer looked: “Transformers are barely a neural network architecture.” Researchers often assumed something so elegant could not work, but Corrado came to view simplicity and resource efficiency as signals that a solution resembled what survives in nature.
The result anticipated Rich Sutton’s later “bitter lesson”: researchers prefer sophisticated domain-specific algorithms, yet across language, vision, and games, scalable architectures combined with more data and compute repeatedly win. The transformer became the cleanest mechanism yet for “more data, more energy, more compute, better results.”
The conceptual loop closed with Harik and Shazeer’s 2001 lunch. Seventeen years after discussing comprehension as compression, Shazeer co-authored a model that ingested enormous corpora, stored statistical structure in compressed parameters, and reconstructed contextually appropriate language.
21. Google used the transformer tactically instead of treating it as a platform reset
The eight researchers published “Attention Is All You Need” in 2017. As of 2025, the hosts cite more than 173,000 academic citations, making it the seventh-most-cited paper of the 21st century despite being substantially newer than those above it.
Google did not simply ignore the invention. Transformer models including BERT and MUM improved query comprehension and search quality, following Brain’s established playbook of feeding new AI capabilities into existing products. What Google failed to do was treat the architecture as a wholesale change in computing and user interface.
Shazeer argued that Google should consider replacing the search index and ten blue links with a giant transformer, exactly the cannibalizing move leadership could not justify. Between 2017 and ChatGPT, the company spent roughly five years improving its franchise rather than building the new one.
All eight paper authors eventually left to start or join AI companies. Shazeer founded Character.AI and returned only after Google’s approximately $2.7 billion licensing-and-hiring transaction—an expensive illustration of publishing the breakthrough, losing the talent, then buying some of it back.
22. OpenAI bet its entity on transformer scaling
In June 2018 OpenAI introduced GPT-1, applying transformer pre-training across large amounts of general internet text and then fine-tuning for specific tasks. Google’s BERT and an Allen Institute model emerged around the same period, but OpenAI treated the direction as an institution-defining bet.
Scaling immediately implied more data, compute, energy, and capital than the nonprofit could sustainably supply. The hosts leave open whether Musk’s departure caused the pivot or merely accelerated a direction already chosen; either way, financial necessity and the transformer arrived together.
Reid Hoffman connected Altman to Microsoft chief Satya Nadella. Discussions at the July 2018 Allen & Company conference produced Microsoft’s commitment to invest $1 billion in cash and Azure credits, plus an exclusive license to OpenAI’s technology for Microsoft products.
OpenAI created a captive for-profit limited partnership controlled by its nonprofit parent, allowing Microsoft to invest while preserving nominal mission control. The structure solved an immediate capital problem but originated the “for-profit nonprofit question mark” still being contested in 2025.
23. Microsoft supplied the cloud that an independent model lab could not build
Microsoft was attractive not merely because it had cash, but because Azure could supply the GPU infrastructure OpenAI lacked. A young research lab was not going to buy chips, secure power, construct data centers, and operate a global training fleet; it needed a hyperscaler.
The partnership revived Google’s historic rival in the one platform shift Google itself had created. Microsoft gained differentiated technology for Azure and its applications, while OpenAI gained the capital and compute required to scale without returning to Google.
GPT-2 arrived in 2019 as a promising API that could continue supplied prose but demanded developer skill and offered no consumer front door. GPT-3 in June 2020 was much more convincing—sometimes difficult to distinguish from human writing—but remained mostly infrastructure awaiting productization.
24. GitHub Copilot proved GPT could change a real workflow
In summer 2021 Microsoft launched GitHub Copilot using GPT-3, the first substantial productization of OpenAI technology. It turned a general model into an assistive interface embedded directly in software development rather than asking consumers to invent prompts for a raw API.
Adoption followed “slowly then all at once”: early engineers whispered that Copilot made them somewhat more efficient; later companies claimed AI generated the majority of their code. The episode treats this as a durable workflow transformation preceding the consumer chatbot explosion.
Microsoft subsequently invested another $2 billion. By the end of 2021, OpenAI had a scalable architecture, a cloud patron, and its first product proof—while Google’s market capitalization began falling from nearly $2 trillion toward $1 trillion amid rate increases and a broader risk-off collapse.
25. Google built chat before ChatGPT but could not responsibly ship it
Shazeer built Meena, an internal transformer chatbot arguably close to the eventual ChatGPT experience. It lacked post-training and reinforcement learning from human feedback, however, and could reportedly produce responses as unsafe as lists of people who “should die.”
The technical prototype collided with three constraints: direct answers threatened ad revenue, publisher disintermediation carried legal risk, and Google’s trusted brand could not tolerate an authoritative system that hallucinated. Even today, David says he verifies Claude’s answers through Google—the trust promise the chatbot might have damaged.
Google evolved Meena into LaMDA but still kept the conversational interface mostly internal. Shazeer left in 2021 after repeatedly advocating release, despite having authored the transformer and possessing an unusually strong internal record.
AI Test Kitchen exposed a constrained LaMDA chat experience in May 2022, before ChatGPT, but ended every conversation after five turns because longer interactions were more likely to derail. The caution was understandable after Microsoft’s Tay debacle, yet it prevented users from discovering the product’s full magic.
26. ChatGPT was an accidental consumer product with explosive proof of demand
OpenAI’s GPT-3.5 was useful but still lacked an intuitive interface. Altman reportedly asked for a chatbot, and within roughly a week someone made a conversational product by wrapping repeated API calls in a chat interface—an implementation so simple that the magnitude of the response surprised the company itself.
Anthropic may have accelerated the timing: OpenAI heard that Dario Amodei’s team was preparing what became Claude and wanted its own chat interface out first. Neither lab appears to have forecast the category’s immediate consumer scale.
ChatGPT launched as a “research preview” on November 30, 2022. It reached 1 million users in less than a week, 30 million by December 31, and 100 million registered users by the end of January 2023—the fastest product ever to cross that threshold.
Servers buckled, Microsoft and OpenAI scrambled for additional Azure capacity and financing, and OpenAI erected a quick paywall partly to control costly demand. A company forecasting B2B API licensing discovered a direct consumer business almost by accident: “the accidental consumer tech company.”
27. ChatGPT changed AI from Google’s advantage into Google’s emergency
Before November 2022, Google treated AI as a sustaining innovation: better recommendations, ads, translation, Gmail, Photos, and search quality favored the incumbent with the most capital and data. ChatGPT revealed a superior interface for many jobs performed by search and turned those same assets into castles requiring protection.
Sundar issued a company-wide “code red” in December, demanding native AI products at speed. The proper disruptive-response playbook, as the hosts describe it, was to ship a comparable new product while measuring where it complemented search and where it cannibalized the franchise.
Microsoft sharpened the threat with another $10 billion investment, saying it now owned 49% of OpenAI’s for-profit entity, and February 2023’s AI-powered Bing. Satya declared “a new day for search,” adding, “I want people to know that we made Google dance.”
Ben calls this the worst possible scenario for Google: its oldest enemy could leapfrog ten-blue-link search with genuinely differentiated technology rather than another imitation. Yet the new Bing launch itself “kind of misses,” giving Google time to organize a stronger response.
28. Bard’s failure forced Google to rebuild both model and organization
Google rapidly rebranded its LaMDA chatbot as Bard and shipped it in February 2023. A choreographed launch demonstration contained a factual error, the product felt clearly inferior to ChatGPT, and Alphabet shares fell 8% that day.
The hosts identify missing post-training and RLHF as the experiential gap: Bard lacked ChatGPT’s tone, appropriateness, and apparent discipline. Google replaced LaMDA with Brain’s PaLM model in May, but it remained behind GPT-3.5 while OpenAI released GPT-4 in March.
Sundar then made the organizational decision previous DeepMind agreements had resisted: merge Google Brain and DeepMind into Google DeepMind, appoint Hassabis to lead the combined AI organization, and subordinate separate charters and cultural friction to a single company-wide emergency.
The hosts read the transaction as DeepMind absorbing Brain’s responsibility, not a neutral federation. Reinforcement-learning expertise and frontier-model development moved under Hassabis, while Jeff Dean, Oriol Vinyals, and later the returning Shazeer joined the unified push.
29. Gemini became both Google’s common model and its rallying product
Sundar’s second decree was “one model for everything”: text, images, audio, and video, serving every internal product and external AI surface. Centralization reflected scaling economics—duplicating frontier training across multiple teams would waste even Google’s resources and fragment the data flywheel.
Google announced Gemini at I/O in May 2023 and offered early public access in December. Gemini 1.5 followed in February 2024 with a 1 million-token context window; Gemini 2.0 arrived in February 2025, and Gemini 2.5 Pro entered experimental mode in March before general availability in June.
Search simultaneously gained AI Overviews, initially through Labs and later broadly, plus AI Mode in March 2025. Google can run inference across an unfathomable number of queries, but deliberately exposes different experiences by query and user rather than redirecting google.com wholesale to Gemini.
Renaming both the model and consumer application Gemini mattered culturally. Ben’s reading is that Google was declaring the product to be the technology itself—like Gmail’s original appeal of speed, search, and storage—not a heavily designed social experience layered over the model.
30. Google has recovered frontier velocity without yet proving demand quality
Beyond the core chatbot, Google launched NotebookLM, AI-generated podcasting, Veo 3 video, Nano Banana imagery, Flow, and Genie 3’s prompt-driven world building. These categories exploit Google’s research and YouTube assets without directly cannibalizing lucrative search queries.
Google reports 450 million monthly Gemini users, a remarkable rise from zero, but the hosts question the definition. The total may include Nano Banana, AI Overviews, AI Mode, or incidental model interactions, echoing Meta’s inflated framing of users who accidentally touched Meta AI.
Whatever the precise engagement quality, Google moved from a disastrous Bard launch to a competitive model family and rapid release cadence while revenue reached record levels. David sees a familiar capability: Google navigated mobile successfully and may again be learning to absorb a platform shift.
31. Waymo shows why AI timelines are impossible to forecast
The lineage begins with DARPA’s 2004 Grand Challenge: a 132-mile autonomous desert race with a $1 million prize, no human intervention, and zero finishers among roughly 100 registered teams. A year later, 22 of 23 finalists passed the prior best distance and five completed the course.
Sebastian Thrun’s Stanford team won using nearly unmodified commodity hardware on a Volkswagen rather than Carnegie Mellon’s extensively rebuilt vehicles. Their principle was that every custom component could fail, so they accepted noisy sensors and performed the innovation in software.
A real-time machine-learning system combined precise short-range laser data with a wide, color camera view. The camera could see toward the horizon and identify the safe path, letting the vehicle anticipate turns and drive faster.
Larry later challenged Thrun to identify a technical reason self-driving was impossible. After sleeping on it, Thrun answered, “I realized what it was. I’m just afraid,” and began Project Chauffeur, Google X’s first project, in 2009.
32. Waymo’s second 99% took more than a decade
Larry defined the “Larry 1000,” about 1,000 difficult California miles including Tahoe, Lombard Street, Highway 1, and the Bay Bridge. A small team completed the benchmark within 18 months, proving feasibility far faster than it could prove a safe, commercial product.
The project used no deep learning for its first five years. Convolutional networks improved object perception around 2013; transformer-derived techniques later aided prediction and planning; Waymo became an Alphabet subsidiary in 2016 and raised $3.2 billion externally in March 2020.
The first public commercial service without a human driver launched in Phoenix in October 2020, eleven years after the project began. Subsequent rounds included $2.5 billion and $5.6 billion, while San Francisco turned the service from a demonstration into a daily transportation habit.
Sebastian originally preferred highway assistance, Eric Schmidt proposed buying Tesla for roughly $3 billion, and Larry favored robotaxis. The chosen path was hardest operationally, but it also produced a distinct experience: private conversations, children’s car seats, pets, and phone calls without negotiating with a driver.
33. Waymo may be a Google-sized business disguised as an expensive fleet
Waymo operates in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta, with Tokyo planned. The hosts cite hundreds of thousands of paid weekly rides, more than 100 million fully driverless miles growing by 2 million weekly, over 10 million paid trips, and approximately 2,000 vehicles.
Its hardware stack includes 13 cameras, four lidar units, six radars, and external microphones. That is costlier than Tesla’s camera-only approach, but Waymo argues multiple sensing modalities are necessary to satisfy the safety and regulatory bar for true autonomy.
A recent Waymo study reported 91% fewer crashes involving serious injury or worse than comparable human driving. Against more than 40,000 annual U.S. road deaths—about 120 daily—the hosts ask why a demonstrated order-of-magnitude safety improvement is not discussed constantly.
The CDC estimated that U.S. crash deaths in 2022 resulted in $470 billion of total costs. A tenfold reduction implies more than $420 billion of potential annual savings, before counting new mobility and experience value; against that opportunity, Waymo’s estimated cumulative $10–15 billion burn now looks “really, really smart.”
34. AI competition helped preserve Google’s search monopoly
Since the prior Acquired episode, a federal judge found Google to be a monopoly in internet search but imposed no remedy requiring a Chrome divestiture or an end to tens of billions in default-distribution payments to Apple and others.
AI competition contributed to the restraint: OpenAI, Anthropic, and Perplexity had attracted tens of billions of funding, persuading the court that the market might produce viable challengers without weakening Google during the new race.
Ben finds that logic questionable because none of those rivals yet produces net income; their competitive capacity depends on investors continuing to finance enormous losses. The episode’s darkly comic domino chain is that Sutskever’s departure created OpenAI, whose existence then helped prevent Google’s breakup.
35. Google can fund AI offense while returning capital
Over the latest twelve months, Google generated approximately $370 billion in revenue and $140 billion in earnings. The hosts call it the most profitable technology company and second globally only to Saudi Aramco: “Let’s not forget Google is the best business ever.”
Alphabet’s market capitalization crossed $3 trillion after falling toward $1 trillion in 2022, placing it behind NVIDIA, Microsoft, and Apple. Cash and marketable securities declined from about $140 billion in 2021 to $95 billion as Google funded AI data centers, repurchased shares, and initiated dividends.
The paradox strengthens the bull case: Google says it intends to win “the most capital intensive race in business history,” yet its search franchise produces more cash than management needs for capex plus a safety cushion. Independent model laboratories possess no comparable self-funding mechanism.
Google also has more than 150 million Google One subscribers, growing nearly 50% year over year. Most remain on cheaper storage tiers, while premium AI begins around $20 monthly, but YouTube Premium, YouTube Music, storage, Play Store benefits, and even sports could support a scaled AI bundle.
36. Google Cloud became AI distribution after learning enterprise sales
Google App Engine launched in 2008 as an opinionated platform-as-a-service requiring prescribed languages, SDKs, and deployment patterns. AWS’s infrastructure-as-a-service model proved more useful, and Google did not launch Compute Engine—the foundation of Google Cloud Platform—until 2012.
The early business also withheld Google’s internal crown jewels and lacked an enterprise organization. Revenue was only about $4 billion in 2017, while the go-to-market team numbered roughly 150 and sat largely in California rather than near global customers.
Kubernetes and multicloud portability gave the third-place provider a counterposition, then former Oracle president Thomas Kurian arrived in late 2018 and added roughly 10,000 go-to-market employees. Revenue exceeded $13 billion in 2020, reached $26 billion in 2022, and the division became profitable in 2023.
Today Google Cloud exceeds a $50 billion annual run rate, grows about 30%, and is the fastest-growing major hyperscaler cited. AI workloads supply a tailwind, while abundant TPUs give customers an alternative to fighting for NVIDIA allocations—provided they accept Google’s stack rather than CUDA.
37. Google is the only company holding every major AI pillar
The hosts divide the market into models, chips, cloud infrastructure, and scaled applications. NVIDIA primarily has chips; Microsoft and Amazon primarily have clouds; Meta has applications; AMD has chips; OpenAI and Anthropic have models; Apple, in their blunt assessment, has “nothing.”
Google possesses all four: Gemini, TPUs, Google Cloud, and applications spanning Search, YouTube, Gmail, Maps, Docs, Chrome, and Android. AI Overviews alone create model usage at a scale a startup cannot manufacture, while Cloud makes TPUs accessible outside Google’s internal fleet.
A cloud business is essential to the chip strategy because Amazon and Microsoft were unlikely to adopt an unproven Google accelerator voluntarily. Google Cloud created initial demand and developer access; possible future TPU availability through neoclouds could broaden the ecosystem without turning Google into a conventional merchant-chip seller.
The likely objective is not NVIDIA-style hardware margin but CUDA-style ecosystem gravity. If users can find TPUs wherever workloads already run, Google can improve utilization, tooling, and model economics while reducing dependence on a supplier extracting scarcity pricing.
38. Distribution, data, and YouTube make the bull case unusually broad
Google still distributes products to nearly every internet user at the moment of intent. AI Overviews and AI Mode demonstrate that it can funnel search demand into new experiences, while Gemini has reached product quality the hosts consider broadly comparable with OpenAI and Anthropic.
Personalized data across Gmail, Maps, Docs, Chrome, and Android could create an assistant no independent lab can reproduce. Switching costs are currently low, but they rise sharply once an AI understands communications, schedules, documents, preferences, and history across the user’s digital life.
YouTube adds the internet’s unique scale corpus of long- and short-form user-generated video, plus the second-largest search destination and Google’s private high-bandwidth network. Models could identify every product in every video, make it instantly shoppable, generate new content, and monetize engagement with Google’s existing advertising machinery.
The company retains a deep research bench and is willing to spend billions to restore exceptional people. With Dean and Shazeer serving as Gemini’s co-technical leads, Ben’s response is simple: “You got Jeff Dean working on it, I’m in.”
39. Owning the accelerator could make Google the lowest-cost token producer
NVIDIA’s roughly 75–80% gross margins imply buyers pay approximately a 4x–5x markup over production cost. Google still pays partner Broadcom, reportedly around a 50% margin, but that resembles a twofold markup—a major difference when accelerators dominate total cost.
The hosts cite estimates that chips and depreciation represent more than half of AI data-center cost, research and software labor roughly 25–33%, and power only about 2–6%. Fast obsolescence makes five-year chip depreciation optimistic: NVIDIA itself quickly shifted enthusiasm from Hopper to Blackwell.
AI companies may achieve only about 50% gross margins rather than software’s familiar 80%. In that world, low-cost production matters far more than in prior technology waves; Google’s own chips, data centers, network, and utilization could determine who can vend tokens sustainably.
Token volume illustrates the scale: Google reportedly processed 10 trillion tokens across its products in April 2024, almost 500 trillion in April 2025, and 980 trillion by June. Training cost amortized over that inference base creates a structural advantage independent laboratories cannot easily match.
40. AI may eventually monetize more intent than search ever observed
Traditional searches average two or three words; AI prompts often exceed 20. Bill Gross’s argument is that richer language reveals far more precise intent, so a future advertising system should know “the crap out of what that user wants” and price relevant placements accordingly.
AI also moves offline or nonexistent interactions into measurable digital sessions: complex planning, explanation, creation, and extended time spent. If the interface supports commercial participation without destroying trust, the monetizable pie could become larger than search rather than merely cannibalizing it.
Waymo supplies additional upside outside the search-replacement frame, potentially expanding ride-hailing into private mobility, personally owned autonomy, transportation for blind and elderly people, and trucking. The hosts exclude speculative AGI from normal valuation analysis but acknowledge it as the “galaxy brain” upside.
41. The bear case starts with weaker monetization and fragmented share
Google’s current AI products have not demonstrated an ad shape comparable with search. The hosts estimate Google earns roughly $400 per U.S. user annually from a free product; only a thin segment will willingly pay approximately that amount each year for an AI subscription.
Search launched in 1998 and found AdWords within two years, while the current AI wave still lacks a comparably obvious value-capture system. High value creation does not guarantee that model providers, rather than users, applications, chipmakers, or clouds, retain the economics.
Google owns roughly 90% of search but may settle at only 25–50% of an AI market shared with OpenAI, Anthropic, Perplexity, xAI, Meta, and others. Even equivalent monetization per user would produce substantially less market power.
The first queries moving to chat may be the most lucrative: travel planning that once generated Expedia advertising, or health questions associated with expensive legal and medical leads. Financial erosion could begin beneath record headline revenue before becoming visible in consolidated results.
42. Incumbency removes the goodwill Google enjoyed during earlier transitions
Google’s chatbot began as “immediately obviously inferior” and improved toward parity, whereas 1998 search was immediately superior. The company is now defending an entrenched franchise instead of arriving as the beloved underdog with a radically better product.
Public and regulatory sentiment toward all large technology companies has deteriorated, but the hosts believe startups still receive more goodwill than Google. OpenAI and Anthropic had to behave like consequential institutions unusually early, yet they retain more of the challenger narrative.
The central risk is not technical incapacity; it is hesitation produced by profitability, publishers, trust, and shareholders. If AI Mode better fulfills Google’s mission, why is the company not moving much more aggressively? Because every percentage point of migration can affect one of history’s greatest businesses.
43. Google’s strongest AI powers are scale, brand, and cornered distribution
In Hamilton Helmer’s framework, scale economies dominate: Google amortizes training, hardware, networking, and software over search-scale inference, while high utilization improves the economics of every TPU and data center.
Google search functions as a cornered distribution resource, and the broader product portfolio creates future switching costs once Gemini becomes personalized. Brand cuts both ways, but for mainstream users, “I trust Google” may outweigh resentment of the incumbent.
Network effects are weak today because another Gemini user does not directly improve one person’s experience. Google also lacks counterpositioning—it is being counterpositioned—and the hosts do not yet see process power that reliably produces breakthroughs unavailable to peer laboratories.
That thinner power profile is revealing. Search appeared to possess nearly every durable advantage; AI currently gives Google several formidable ones, but not the overwhelming bundle that once entitled it to persistent 90% share.
44. Google’s outcome rests on whether it can remain rapid without becoming rash
Ben’s quintessence is “the most fascinating example of the innovator’s dilemma ever.” Larry and Sergey have said they would rather go bankrupt than lose in AI, but the real test arrives only if advancing Google’s information mission requires accepting a permanently worse business than search.
David argues that, despite its flat-footed opening, Google may now be threading the needle better than any other incumbent. It unified rival labs, standardized on one model, restored key talent, shipped rapidly, and used selective integration to protect the franchise rather than making an impulsive all-or-nothing switch.
The unresolved possibility is that search’s foundation is eroding even while financial statements remain superb. Record revenue can coexist temporarily with users moving their most valuable workflows elsewhere, particularly when Google itself is subsidizing the transition through expensive inference.
For now, the hosts credit Sundar with being “rapid but not rash.” Stewarding both a mission and a public-company profit engine is an unusually difficult dual mandate; whether Google can keep both will make this one of business history’s defining platform-transition cases.