Google Invented the Technology Behind ChatGPT. Now It Must Defend Its $400 Billion Empire.

Google invented the transformer, the technology that helped make ChatGPT possible. It employed many of the researchers who subsequently built OpenAI, Anthropic, and competing AI laboratories. It acquired DeepMind for approximately $550 million, developed its own AI chips, and built one of the world's largest computing infrastructures.

Yet when ChatGPT reached 100 million users in roughly two months, Google found itself defending a business it had spent decades making extraordinarily profitable.

The company had a problem few competitors could appreciate. Its scientists had helped create a potentially better way to answer questions, but replacing conventional search could threaten billions of dollars in advertising revenue.

The story also contains some extraordinary transactions. Google reportedly spent $2.7 billion to bring back a researcher who had helped invent the transformer. Elon Musk's frustration with Google's acquisition of DeepMind contributed to the creation of OpenAI. Microsoft subsequently financed OpenAI's challenge to Google's search business.

Meanwhile, Google was quietly building other advantages: proprietary chips, a rapidly expanding cloud business, and Waymo, its autonomous-driving company.

In a 2025 episode of Acquired, hosts Ben Gilbert and David Rosenthal reconstructed these developments through interviews, historical reporting, and conversations with researchers and executives, including Sundar Pichai, Demis Hassabis, Jeff Dean, Sebastian Thrun, and Greg Corrado.

Their central question was not whether Google understood artificial intelligence. Its history showed that it did.

The question was whether Google could transform the business that made it powerful without destroying the economics that funded its success.

1. Google Invented the Transformer, Then Lost Its Inventors

The 2017 paper that changed artificial intelligence

On June 12, 2017, eight researchers associated with Google published a paper titled Attention Is All You Need.

Its authors were Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin.

The paper introduced the transformer, an architecture that became foundational to modern large language models.

Before transformers, many advanced language-processing systems relied on recurrent neural networks. These systems processed language sequentially, making it difficult to efficiently train extremely large models. They also struggled to preserve information from earlier parts of long passages.

Long short-term memory networks, known as LSTMs, improved the situation by retaining selected information across longer sequences. Google had used them to improve translation and reported substantial reductions in translation errors during 2016.

But their sequential nature remained a problem.

Jakob Uszkoreit explored attention mechanisms that could determine which parts of a passage were relevant to interpreting another part. Instead of proceeding through language primarily one word at a time, a model could examine relationships across the sequence.

This approach offered two advantages: more effective use of context and greater parallelization of computations.

The initial implementation did not immediately outperform Google's existing translation models. Noam Shazeer joined the project, rewrote important parts of the implementation, and helped produce substantially better results.

The researchers discovered that the architecture performed well as they increased its scale.

In its original paper, the team reported strong machine-translation results while reducing training requirements. The English-to-French experiment, for example, reached a state-of-the-art result after approximately 3.5 days of training on eight GPUs.

That was important because it connected model quality to computational efficiency.

A language architecture capable of benefiting from larger datasets and greater computing power could eventually support applications much more ambitious than machine translation.

The paper also anticipated applications beyond text, including images, audio, and other information formats.

By 2025, the paper had reportedly accumulated more than 173,000 academic citations. Its influence extended into the models developed by OpenAI, Anthropic, Google, and numerous other organizations.

Google did use the transformer, but not as ChatGPT would

The popular version of this history suggests Google invented the transformer and simply ignored it.

That is incomplete.

Google incorporated transformer research into products that millions of people already used.

In October 2019, Google announced that it was introducing BERT into Search, improving its ability to interpret relationships between words and understand complicated queries.

The company later introduced MUM, another model designed to improve information understanding across languages and formats.

These were commercially useful applications. Better search results encouraged users to continue searching, while improvements in advertising relevance could produce financial returns.

Google's research was therefore supporting its existing business.

OpenAI ultimately pursued something different: it built products around the language model itself.

This distinction matters. Google was not necessarily missing the scientific significance of transformers. It was using them to strengthen a highly profitable business whose replacement offered uncertain financial benefits.

That made a radical change to search less attractive than it was for competitors without Google's advertising revenue to protect.

Every transformer-paper author eventually departed

All eight authors of Attention Is All You Need ultimately left Google.

Their departures illustrated a wider problem: a company could finance fundamental research, publish important discoveries, and still lose the people best positioned to commercialize them.

Those researchers joined or established organizations working on different AI opportunities. Their discoveries circulated throughout the industry, helping competitors improve products that eventually challenged Google itself.

The most striking example was Noam Shazeer.

Shazeer had worked on language models at Google since the early 2000s. After helping develop the transformer, he participated in Google's conversational AI research.

He eventually left in 2021 with Daniel De Freitas to establish Character.AI.

Their departure followed internal disagreements over releasing Google's chatbot technology. Google had concerns about safety, fairness, and the risks of publicly deploying conversational models.

Character.AI developed a consumer chatbot business and attracted substantial investment, but building frontier AI systems was expensive.

In August 2024, Google negotiated an arrangement with Character.AI. The agreement licensed its language-model technology and brought Shazeer, De Freitas, and other researchers back to Google.

The financial details are particularly revealing.

According to Alphabet's 2024 annual filing, Google paid Character.AI $2.7 billion in cash and canceled convertible instruments as part of the transaction. The company received a non-exclusive technology license and hired selected Character.AI employees.

The filing confirms the $2.7 billion payment. It does not establish that this entire amount was compensation for Shazeer personally.

That distinction is important because the transaction was widely described as Google spending billions to bring back one researcher. His return was clearly strategically important, but the agreement also involved intellectual-property rights and other employees.

Shazeer subsequently became a technical leader of Gemini alongside Jeff Dean.

The sequence captures one of Google's unusual AI challenges: it had helped invent the technology, lost central researchers, and then spent billions to recover some of the talent and capabilities it had previously possessed.

2. Google Was Building AI Long Before ChatGPT

Larry Page's original ambition

Google's interest in artificial intelligence began long before the transformer.

Larry Page's father was a computer science professor who studied machine learning and artificial intelligence at the University of Michigan when the field attracted relatively little academic enthusiasm.

That background influenced Page's understanding of computing.

The PageRank algorithm, which helped establish Google's search engine, evaluated relationships among webpages to estimate their importance. Although it was not a modern neural network, it represented a statistical approach to organizing information.

Page was interested in going considerably further.

In 2000, he described artificial intelligence as the ultimate form of Google's search engine: a system capable of understanding information, recognizing exactly what people wanted, and providing the appropriate result.

This was closely connected to Google's mission of organizing the world's information and making it universally accessible and useful.

The company therefore considered AI fundamental to its purpose, not merely an additional product category.

A conversation about compression leads to language models

Around 2000 or 2001, Google engineers Georges Harik, Ben Gomes, and Noam Shazeer discussed an unusual idea during lunch.

Harik, one of Google's earliest employees, had studied machine learning at the University of Michigan.

He suggested that compressing information might have something to do with understanding it.

The argument was that a system capable of reducing information into a compact representation and later reconstructing its meaning must have learned something about the underlying material.

The episode connects this concept to modern language models, which encode patterns from enormous datasets into comparatively compact numerical parameters.

There is an important qualification: accurately reproducing patterns does not, by itself, settle the philosophical question of whether a machine genuinely understands what it produces.

Shazeer nevertheless found the idea compelling.

He and Harik began developing probabilistic language models that estimated which words were likely to follow a given sequence.

The work resembled the broad next-token prediction principle later used by large language models.

Not everyone at Google considered this a worthwhile investment of engineering talent. Harik and Shazeer faced skepticism from colleagues.

Support from Sanjay Ghemawat, Jeff Dean's highly respected engineering partner, helped defend the research.

Google's relatively unusual management environment also mattered. Around 2001, Page had removed engineering managers, leaving researchers substantial freedom to pursue unconventional projects.

How language models became an advertising business

Shazeer's work helped produce Google's familiar spelling-correction feature, Did you mean.

The benefit went beyond convenience.

When users submitted misspelled queries, Google's infrastructure still had to process them. If those users immediately corrected their searches, the company had spent computing resources answering the wrong question.

Predicting the intended query improved search quality and reduced wasted processing.

Harik and Shazeer subsequently built a larger language model called PHIL, short for Probabilistic Hierarchical Inferential Learner.

In 2003, Susan Wojcicki and Jeff Dean needed a way to understand the contents of third-party webpages for Google AdSense.

Google already had advertisers purchasing search advertisements through AdWords. If the company could identify the subject of an independent webpage, it could display relevant advertisements on that page.

PHIL provided the necessary language-processing capabilities.

According to the episode, Dean used the model to implement the core AdSense system in approximately one week.

AdSense expanded Google's advertising inventory far beyond its own search results, creating a new source of billions of dollars in revenue.

By the mid-2000s, PHIL reportedly consumed approximately 15% of Google's data-center infrastructure.

Modern generative AI was still decades away from public adoption, but Google already understood that predicting and interpreting language could have enormous commercial value.

Translating a sentence: from 12 hours to 100 milliseconds

Google Translate offered another early demonstration.

Chief architect Franz Och, a machine-learning researcher from Germany, developed a large statistical translation model that performed exceptionally well in a DARPA translation competition around 2006–2007.

The system had reportedly been trained using approximately two trillion words from Google's search index.

Its quality was impressive, but it took around 12 hours to translate a sentence.

That was acceptable for a competition in which researchers received material at the beginning of a week and returned translations several days later. It was useless for a consumer expecting an immediate answer.

Jeff Dean worked with the translation team to restructure its computations, allowing more work to happen simultaneously across Google's distributed infrastructure.

The episode describes translation time falling from approximately 12 hours to 100 milliseconds.

The improvement demonstrated an important principle that would recur throughout Google's AI history: an algorithm could become commercially valuable only when the company found a way to operate it efficiently.

Google subsequently applied language prediction to search suggestions, advertising quality scores, and other products.

Sebastian Thrun brings academic AI into Google

In April 2007, Page recruited Sebastian Thrun, who led Stanford's Artificial Intelligence Laboratory.

Thrun had been preparing a startup with colleagues and had received proposed financing terms from Benchmark and Sequoia.

Page instead offered to bring the team into Google through signing bonuses, creating an arrangement resembling an acquisition before the startup had formally launched.

Thrun encouraged Google to recruit leading professors without requiring them to give up their university appointments.

Researchers could retain academic relationships while receiving compensation, stock, and access to Google's computing infrastructure.

This made Google particularly attractive in a field where advanced expertise was concentrated among a relatively small number of academics and doctoral researchers.

Stanford's laboratory also connected several future technology leaders. Chris Cox, later Meta's chief product officer, had conducted undergraduate research there. Sam Altman worked there before leaving Stanford to found Loopt, part of Y Combinator's first batch in 2005.

Thrun contributed to Google's Ground Truth mapping project.

Google Maps had relied heavily on commercial geographic data supplied by companies such as Tele Atlas and Navteq. Because competitors could purchase similar information, those datasets offered limited differentiation.

Ground Truth combined Street View imagery, census information, and roughly 40 additional sources to construct more accurate maps.

The work also required substantial human labor. Approximately 1,000 workers in India reportedly helped resolve discrepancies and construct geographic information.

It was an early example of Google's willingness to combine computing, proprietary data, and large-scale operations, rather than expecting algorithms to solve everything independently.

Geoffrey Hinton and the revival of neural networks

In December 2007, Thrun invited University of Toronto professor Geoffrey Hinton to present his research at Google.

Neural networks were still viewed skeptically in much of computer science.

Earlier generations of researchers had promised capabilities they could not deliver. Limited computing power made training deep networks especially difficult.

Hinton and researchers such as Yann LeCun argued that networks containing multiple computational layers could learn increasingly sophisticated patterns if supplied with sufficient data and computing resources.

Hinton developed a relationship with Google, initially consulting and later participating in an unusual internship arrangement around 2011–2012, despite being approximately 60 years old.

Meanwhile, Google X emerged around late 2009 as an organization for ambitious research projects.

Thrun's work in connecting academic research with Google's computing capabilities contributed to its development.

Google Brain became one of its early major projects.

In 2011, Jeff Dean, Andrew Ng, and Greg Corrado founded Google Brain to investigate whether extremely large neural networks could learn useful representations from data.

The team combined Dean's distributed-computing expertise, Ng's academic machine-learning experience, and Corrado's neuroscience background.

Previous internal efforts, including Brains on Borg and Cortex, had encountered difficulties.

Dean developed a system called DistBelief that allowed training calculations to operate asynchronously across many processors.

Some processors worked with slightly outdated model parameters rather than waiting for every other machine to synchronize.

Researchers initially worried that this would produce unreliable training.

Instead, the technique allowed Google to train much larger models using its existing computing infrastructure.

The Cat Paper and YouTube's commercial transformation

One of Google Brain's earliest achievements became known as the Cat Paper.

Formally titled Building High-level Features Using Large Scale Unsupervised Learning, the research demonstrated that a large neural network could develop representations of recognizable objects without being supplied with human-labeled examples.

The system used millions of YouTube video frames, a nine-layer neural network, and approximately 16,000 CPU cores across a large computing cluster.

One of the learned representations responded strongly to images of cats.

The significance was not that recognizing cats was itself an extraordinary commercial product. It was evidence that neural networks could discover useful categories from raw information.

Google possessed an enormous quantity of video material, but users often uploaded videos with inaccurate or incomplete titles and descriptions.

Understanding the actual contents of videos could improve search, recommendations, and content classification.

That helped YouTube recommend material more effectively and identify content relevant to copyright management, advertising, and safety.

The hosts regard approximately 2012 as the beginning of the commercial AI era because these systems were already influencing what people watched and how advertising was delivered.

Similar recommendation architectures became central to Facebook, Instagram, TikTok, and other platforms.

This earlier generation of AI was less visible than ChatGPT, but it changed how billions of people encountered information and entertainment.

AlexNet changes the importance of Nvidia

Another major breakthrough occurred in 2012.

Stanford researcher Fei-Fei Li had assembled ImageNet, a dataset containing approximately 14 million labeled images.

At the time, leading image-recognition systems still made mistakes on roughly one-quarter of benchmark images.

Geoffrey Hinton and University of Toronto researchers Alex Krizhevsky and Ilya Sutskever developed AlexNet, reducing the error rate from approximately 25% to 15%.

That represented roughly a 40% relative reduction in errors.

The improvement changed attitudes toward deep learning.

AlexNet also demonstrated an unexpected computing advantage: it was trained using two comparatively inexpensive Nvidia GeForce GTX 580 graphics cards.

Nvidia's CUDA platform allowed the researchers to use graphics processors for the mathematical operations required by neural networks.

The result helped transform GPUs from primarily graphics hardware into essential AI infrastructure.

Hinton, Krizhevsky, and Sutskever subsequently established DNN Research. The company had no established commercial product; its primary assets were the researchers themselves.

Baidu initially offered approximately $12 million to acquire the company.

Hinton organized an auction while attending a conference near Lake Tahoe, attracting Google, Microsoft, Baidu, and DeepMind.

DeepMind withdrew because it could not compete financially. Facebook did not participate.

The bidding reportedly reached approximately $44 million, at which point the researchers chose Google without necessarily extracting the maximum possible offer.

The proceeds were divided 40% to Hinton and 30% each to Krizhevsky and Sutskever.

The younger researchers reportedly supported giving Hinton the larger share.

Their arrival substantially strengthened Google Brain and helped accelerate the incorporation of neural networks into Google's commercial products.

Google would soon become one of Nvidia's most important early AI customers. It would also begin developing its own chips to avoid depending entirely on expensive general-purpose hardware.

3. DeepMind: The $550 Million Acquisition That Helped Create Google's Rivals

Three founders pursuing general intelligence

DeepMind was founded in London in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman.

Hassabis was a childhood chess prodigy who entered the video-game industry as a teenager. He worked at Bullfrog Productions, helping develop Theme Park, a commercially successful simulation game reportedly selling approximately 15 million copies.

He later studied computer science at Cambridge, established another games company called Elixir, and pursued a neuroscience doctorate at University College London.

His interests moved increasingly toward the nature of intelligence itself.

Legg, also associated with University College London, researched machine intelligence and was an early advocate of artificial general intelligence, or AGI.

Suleyman was Hassabis's childhood friend and had pursued entrepreneurial projects after leaving Oxford.

The founders wanted to create systems capable of learning many different kinds of tasks rather than designing specialized programs for individual problems.

Their mission was to solve intelligence and then use intelligence to address other challenges.

This presented an obvious financing difficulty.

DeepMind had no proven consumer product. Its research objectives were speculative, and meaningful progress could eventually require enormous computing resources.

How a chess conversation helped secure financing

In June 2010, Hassabis and Legg traveled to San Francisco for the Singularity Summit.

They hoped to meet wealthy investors comfortable financing unconventional technological ambitions.

Their principal target was Peter Thiel, the PayPal co-founder, early Facebook investor, and Founders Fund partner.

Hassabis arranged to speak at the event partly to attract Thiel's attention.

Unfortunately, Thiel did not attend his presentation.

Hassabis subsequently encountered him at a gathering at Thiel's home. Knowing Thiel enjoyed chess, he began their conversation with the game rather than an immediate investment pitch.

His competitive chess background provided a natural connection.

Thiel invited Hassabis to present DeepMind to Founders Fund the following day.

The meeting succeeded, and Founders Fund led an initial financing round of approximately $2 million.

Thiel's network also introduced Hassabis to Elon Musk.

The SpaceX conversation that influenced Musk's AI concerns

Musk invited Hassabis to visit SpaceX.

During the visit, the two discussed existential risks and Musk's ambition to establish a human civilization on Mars.

Musk regarded Mars as a possible backup if a catastrophe threatened life on Earth.

Hassabis raised an unusual objection: a dangerous superintelligent AI might not be contained by physical distance. If Mars maintained communication links with Earth, a sufficiently capable AI could potentially reach it.

According to the account presented by Acquired, the conversation influenced Musk's thinking about AI risk.

Musk subsequently invested in DeepMind.

He also recognized that improved machine vision could benefit Tesla. Neural networks capable of understanding images might eventually help autonomous vehicles interpret roads, obstacles, and traffic.

Musk began recruiting researchers with relevant machine-learning expertise.

His investment in DeepMind therefore connected two interests: concern about advanced AI and the practical requirements of developing autonomous-driving technology.

The game-playing demonstration that reached Larry Page

DeepMind initially concentrated on research rather than products.

One demonstration involved an AI system learning to play Atari's Breakout.

Through training, the system discovered that directing the ball behind the rows of bricks allowed it to clear them more efficiently, without being explicitly taught that strategy.

The hosts recount an unusual introduction between the company and Larry Page.

Musk and investor Luke Nosek were reportedly watching a demonstration of DeepMind's achievement while traveling on a private aircraft.

Page was aboard the same flight.

He noticed the demonstration and became interested in the company.

This account comes from the episode's research and has not been independently established here as a documented transaction record. Its importance is that Page's interest emerged from the company's progress toward learning systems, not an established commercial product.

Facebook reportedly offered more money than Google

By late 2013, major technology companies had recognized the value of acquiring scarce AI research talent.

Mark Zuckerberg wanted to strengthen Facebook's machine-learning capabilities, particularly as neural-network recommendations became increasingly important to social-media engagement.

Facebook reportedly offered as much as $800 million for DeepMind.

Musk also proposed acquiring the company using Tesla shares. Tesla's market capitalization was approximately $20 billion at the time.

An acquisition by Tesla could have created substantial financial returns for DeepMind's investors if they retained shares through Tesla's later appreciation.

But DeepMind's founders had another priority.

They wanted to preserve their mission of developing general intelligence.

Facebook was unwilling to provide the degree of independence they sought. Tesla wanted researchers focused on autonomous vehicles, which risked redirecting DeepMind toward one company's commercial requirements.

Google offered a different arrangement.

Larry Page understood DeepMind's long-term research ambitions. Google already possessed Brain, an AI organization applying machine learning to existing products.

That meant DeepMind could remain focused on fundamental research without immediately becoming responsible for improving Google Search, Gmail, or advertising.

Google also possessed the computing infrastructure required for increasingly ambitious experiments.

Hassabis reportedly felt that Page understood what DeepMind was trying to accomplish.

Google's acquisition and unusual research protections

Google's pursuit of DeepMind formed part of a broader effort to recruit leading AI researchers.

Page had instructed executives to invest aggressively in deep learning. At a strategy gathering on an undisclosed South Pacific island, he reportedly told them:

“Let's really go big.”

Engineering executive Alan Eustace took that direction seriously.

After acquiring DNN Research, Google began pursuing additional research talent.

In December 2013, Eustace organized a trip to London to assess DeepMind.

Geoffrey Hinton needed to participate, but a back condition prevented him from sitting comfortably during long flights.

According to the episode, Google arranged a private aircraft fitted with an unusual restraint system so Hinton could travel lying down.

The arrangements demonstrate the lengths to which Google was prepared to go to evaluate scarce research talent.

In January 2014, Google acquired DeepMind for approximately $550 million.

The transaction included protections intended to preserve DeepMind's research independence and provide oversight of its mission.

Reid Hoffman later participated in the organization's oversight structure.

This arrangement mattered because DeepMind had declined other potential buyers partly over disagreements about control.

The hosts speculate that DeepMind could eventually be worth approximately $500 billion as an independent company.

That is their hypothetical valuation, not a market-tested value. DeepMind has not been separately listed, and the economic value attributable to its research cannot be cleanly separated from Google's infrastructure, products, and distribution.

Nevertheless, the acquisition was strategically important.

It gave Google another leading AI laboratory while keeping an ambitious research program away from competing technology companies.

DeepMind was not merely an expensive science project

DeepMind's work soon produced measurable commercial benefits.

In July 2016, Google announced that DeepMind researchers and data-center engineers had developed a system capable of reducing energy consumption for cooling by up to 40%.

The system analyzed operating conditions such as temperature, power usage, and equipment settings, then modeled how changes could affect efficiency.

An important clarification appears in Google's original technical explanation: the 40% reduction referred to cooling energy, not total data-center electricity consumption.

Google reported that the improvement translated to approximately a 15% reduction in overall power-usage-effectiveness overhead.

The difference matters. A reduction in cooling energy is valuable, but it must not be mistaken for a 40% reduction in every category of data-center energy expenditure.

DeepMind subsequently produced one of the most widely recognized AI demonstrations in history.

In March 2016, AlphaGo defeated champion Go player Lee Sedol by four games to one.

Go had long been regarded as especially difficult for computers because of its enormous number of possible moves and strategic combinations.

AlphaGo combined neural networks, reinforcement learning, and search techniques.

Its famous Move 37 in the second game initially surprised expert observers. The move later proved strategically effective, and human players began studying ideas introduced by the system.

DeepMind continued investigating challenging environments, including StarCraft, where systems must operate in real time under uncertainty.

The company had demonstrated that techniques developed through game-playing research could produce capabilities that experts did not expect conventional programming to achieve.

Its methods also became relevant to wider scientific research.

In a March 2026 retrospective, Hassabis connected AlphaGo's technical legacy with DeepMind's continuing work on scientific problems and general-purpose intelligence.

The commercial and research achievements helped justify Google's acquisition. They also increased concern among outsiders that one company might control an unusually large concentration of advanced AI expertise.

4. How Google's DeepMind Acquisition Helped Produce OpenAI

Musk and Page move from cooperation to disagreement

Musk had invested in DeepMind and maintained a close relationship with Larry Page.

Google's acquisition changed their interests.

Musk reportedly became unhappy that DeepMind was controlled by an organization he considered a competitor for AI researchers and, potentially, autonomous-driving technology.

His concerns also extended to the concentration of advanced intelligence research within a single large company.

In summer 2015, Musk and Sam Altman organized a dinner at the Rosewood Hotel on Sand Hill Road.

They wanted to persuade leading AI researchers to leave Google and help create an independent laboratory.

This was a difficult proposition.

Google employed many of the world's most accomplished researchers. They received substantial compensation, access to specialized colleagues, academic flexibility, and computing resources that a new organization could not easily reproduce.

Most of the invited researchers were reluctant to leave.

One important exception was Ilya Sutskever.

Ilya Sutskever becomes the decisive recruit

Sutskever had helped develop AlexNet and subsequently joined Google through its DNN Research acquisition.

His willingness to establish an independent laboratory made the proposed organization more credible.

According to Acquired, Google offered approximately twice his existing compensation to remain, with the counteroffer reportedly communicated through Jeff Dean.

Sutskever declined.

His decision made it easier to attract other researchers.

Approximately seven technical researchers formed the initial group, accompanied by Altman, Musk, and Greg Brockman, who had previously worked at Stripe.

OpenAI was announced in December 2015.

Its original statement described a nonprofit research organization intended to advance artificial intelligence for broad human benefit without being constrained by financial-return requirements.

That mission differentiated it from conventional venture-backed companies.

OpenAI also positioned itself as an alternative to private laboratories controlled by Google and Facebook.

Researchers were promised substantial freedom to publish discoveries and pursue fundamental scientific questions.

The difference between $1 billion pledged and $130 million received

OpenAI announced approximately $1 billion in financial commitments from prominent supporters.

They included Musk, Altman, Reid Hoffman, Peter Thiel, Jessica Livingston, and other technology investors and organizations.

The headline created the impression of an exceptionally well-financed research institution.

But a commitment was not necessarily an immediate contribution.

TechCrunch's examination of OpenAI's nonprofit tax filings found that total contributions received from the organization's founding through 2021 were approximately $133.2 million.

This is consistent with the episode's description of roughly $130 million ultimately collected for the original nonprofit effort.

The distinction is substantial, although OpenAI's early research was much less expensive than its later large-model development.

Initial expenditures concentrated on researcher salaries and relatively modest computing needs.

The capital received was sufficient to support several years of experimentation.

OpenAI initially resembled DeepMind

The organization did not begin with a clearly defined consumer chatbot strategy.

It investigated reinforcement learning, game-playing agents, robotic manipulation, and general-purpose learning environments.

DeepMind was working on games such as StarCraft. OpenAI pursued Dota 2, another complex competitive environment requiring coordination and real-time decision-making.

OpenAI also developed Universe, which allowed agents to interact with various software and game environments.

Separate robotics experiments explored manipulation tasks, including solving a Rubik's Cube.

Dario Amodei joined from Google Brain in early 2016 and became an important research leader alongside Sutskever.

He would later leave OpenAI and establish Anthropic, reinforcing the recurring movement of researchers between major laboratories.

At this stage, OpenAI was pursuing several promising ideas rather than a single dominant technological architecture.

Musk leaves after a disagreement over direction and control

Musk became dissatisfied with OpenAI's progress.

According to the episode, he believed the organization had spent too much effort pursuing research similar to DeepMind's without establishing a decisive advantage.

He also wanted research that could become relevant to Tesla.

Musk reportedly proposed taking control of OpenAI and potentially connecting its work more directly to Tesla.

OpenAI's leadership rejected that arrangement.

Musk left the organization by early 2018 and withdrew his continuing financial support.

Around the same period, OpenAI was becoming increasingly interested in transformers and large-scale language modeling.

The hosts question whether Musk's departure accelerated that change.

The evidence presented does not establish a direct causal relationship. Both developments, however, made OpenAI's future financing more challenging.

Training increasingly large models would require vastly more computing power than the nonprofit had initially needed.

GPT becomes OpenAI's central research direction

OpenAI released the first GPT research in June 2018.

GPT stood for Generatively Pre-trained Transformer.

The idea was to train a model on large quantities of general text, allowing it to learn reusable language representations before adapting to specific tasks.

This differed from developing a separate specialized model for every application.

Google's BERT research and related work at the Allen Institute showed that multiple organizations were exploring transformer-based approaches.

OpenAI became particularly committed to increasing model size, training data, and computing resources.

GPT-2 arrived in 2019.

It could generate plausible passages, continue supplied text, and perform a range of language tasks. But access and usability remained limited for ordinary consumers.

GPT-3, released in June 2020, substantially improved the quality and flexibility of generated text.

Its capabilities attracted developers and investors, but OpenAI still lacked the polished, broadly accessible conversational product that would later define the company.

The immediate commercial opportunity appeared to be licensing models through APIs and integrating them into other software.

One of the first significant examples would come from Microsoft's software-development ecosystem.

Reid Hoffman connects Altman with Microsoft

Reid Hoffman occupied an unusual position within the developing AI industry.

He had participated in DeepMind's oversight arrangements, supported OpenAI, and joined Microsoft's board after LinkedIn was acquired.

He was therefore connected to organizations whose research and commercial interests were beginning to overlap.

Hoffman helped introduce Sam Altman to Microsoft CEO Satya Nadella.

The hosts discuss conversations surrounding the Allen & Company gathering in Sun Valley, Idaho, in July 2018.

The relationship eventually produced Microsoft's initial $1 billion investment in OpenAI in 2019.

Microsoft supplied financial resources and access to Azure computing infrastructure. In return, it obtained opportunities to commercialize OpenAI technology through its products and services.

OpenAI could not independently construct the computing facilities needed for increasingly ambitious models.

Microsoft, meanwhile, gained access to a research organization developing technology with the potential to challenge Google's most important consumer product.

A nonprofit creates a commercial investment structure

Accepting substantial outside investment created a governance problem.

OpenAI had been established as a nonprofit organization without a conventional obligation to maximize financial returns.

To accommodate financing, it created a controlled commercial entity, OpenAI LP.

This structure attempted to allow investors to receive returns while keeping ultimate governance connected to the nonprofit mission.

The arrangement introduced tensions between research objectives, financial incentives, and the authority of the nonprofit board.

Those tensions became more visible in later years.

In November 2023, Altman was removed as CEO and subsequently reinstated after employees threatened to leave.

Microsoft emerged as a potential destination for departing researchers, demonstrating how important the partnership had become to OpenAI's continued operations.

The events also illustrated a structural weakness: OpenAI depended heavily on Microsoft's capital and infrastructure despite maintaining a distinct governance system.

Microsoft finds a commercial use for GPT

Microsoft introduced GitHub Copilot using OpenAI technology in 2021.

Copilot helped developers generate and complete code within their existing programming workflows.

Its early adoption demonstrated that language models could produce immediate economic value in software development.

Microsoft subsequently committed additional financing, described in the episode as approximately $2 billion around 2021.

Its relationship with OpenAI now served several objectives: selling Azure computing services, improving Microsoft software, and developing alternatives to Google's search and productivity products.

The partnership was particularly consequential because Microsoft and Google had competed for decades across browsers, operating systems, enterprise software, and internet services.

OpenAI provided Microsoft with a new opportunity to challenge Google through technology originating substantially from Google's own research ecosystem.

The wider migration of AI talent

Other researchers continued moving between institutions.

Mustafa Suleyman had co-founded DeepMind, later shifted into Google-related AI responsibilities, and subsequently established Inflection AI with Reid Hoffman.

After OpenAI's 2023 leadership crisis, Microsoft entered an acquisition-like arrangement with Inflection in 2024 and brought Suleyman into the company to lead its AI operations.

These relationships illustrate how a comparatively small group of researchers, founders, executives, and investors repeatedly appeared across Google, DeepMind, OpenAI, Anthropic, Microsoft, and other organizations.

The institutions competed fiercely, but their histories were closely connected.

5. ChatGPT Exposes Google's Product Problem

Google already had conversational AI

Before ChatGPT launched, Google had developed models capable of producing sophisticated conversational responses.

Noam Shazeer worked on a chatbot called Meena.

Its capabilities suggested that transformer-based systems might support an alternative to traditional information retrieval.

Shazeer advocated a much more ambitious approach to Google's consumer products, including replacing parts of conventional search with model-driven responses.

Google was cautious.

Early chatbots could produce inappropriate material, incorrect information, or responses inconsistent with Google's expectations for safety and reliability.

Unlike a startup releasing a research demonstration, Google was responsible for services used by billions of people.

The company's reputation, commercial relationships, and regulatory obligations made those risks significant.

There was also the question of revenue.

A conventional search engine could show several results, advertisements, and links to external websites.

A chatbot might deliver one answer directly.

That could be a better experience for the user while providing fewer opportunities to earn advertising revenue.

It also complicated Google's relationship with publishers whose content helped make search useful. Direct answers could reduce the need to visit the original websites.

LaMDA and the limits Google imposed

Google continued developing conversational models through LaMDA.

In May 2022, it introduced selected AI demonstrations through AI Test Kitchen.

The company restricted conversations to approximately five exchanges, reducing the risk that prolonged interactions would produce unsafe or inappropriate content.

There were precedents for caution.

Microsoft's earlier Tay chatbot had generated offensive material after interacting with users on Twitter, becoming an example of how quickly an unrestricted public AI product could damage its creator's reputation.

Google had declared itself an AI-first company in 2017, but its practical strategy remained largely focused on embedding machine learning into established products.

That approach had previously generated enormous financial returns.

The company also lacked clear evidence that consumers wanted a general-purpose chatbot as their primary interface to online information.

This changed almost overnight.

ChatGPT turns an API into a consumer phenomenon

In late 2022, OpenAI created a simple conversational interface around GPT-3.5.

According to the episode, the initial product was assembled in approximately one week.

Researchers and developers had already been using language models for years.

ChatGPT made the capabilities accessible to ordinary users without requiring programming knowledge, complicated integrations, or extensive prompt preparation.

The hosts suggest that competition with Anthropic's developing Claude chatbot may have influenced OpenAI's timing.

That remains a possible contributing factor rather than an independently established explanation for the entire launch decision.

ChatGPT became publicly available on November 30, 2022.

It reportedly attracted approximately one million users within its first week and 30 million by the end of December.

By January 2023, it had achieved a scale that surprised even its creators.

A February 2023 Reuters report, citing UBS and Similarweb estimates, placed ChatGPT at approximately 100 million monthly active users in January.

This independently reported figure is slightly different from the episode's characterization of 100 million registered users. Monthly active users and registrations are not interchangeable measurements.

Nevertheless, the rate of adoption was extraordinary.

OpenAI's commercial forecasts had concentrated on APIs and business licensing. It was suddenly operating one of the fastest-growing consumer applications in history.

Demand created infrastructure problems.

OpenAI needed more Azure computing capacity, and the cost of serving users rose rapidly.

Paid access offered a way to recover some of those costs while controlling demand.

In February 2023, OpenAI introduced a $20 monthly ChatGPT subscription.

An accessible interface had transformed a research model into a consumer business.

Microsoft tries to challenge Google Search

Microsoft moved quickly.

In January 2023, it announced another multibillion-dollar investment in OpenAI, commonly described as approximately $10 billion.

The episode discusses reporting that Microsoft could obtain a substantial economic interest in OpenAI's commercial operations, potentially reaching 49% under certain arrangements.

An economic interest, however, is not equivalent to outright ownership or control of the nonprofit-governed organization.

In February, Microsoft introduced an AI-powered Bing.

Satya Nadella presented the product as an opportunity to change the search experience itself rather than merely improve conventional search rankings.

Microsoft wanted to challenge Google's position as the principal entry point to the internet.

The hosts consider the initial Bing launch commercially disappointing relative to Microsoft's ambitions.

But that did not eliminate the threat.

Microsoft had demonstrated a willingness to compete through a different search interface, supported by models and infrastructure that had not existed during earlier search-engine battles.

Sundar Pichai declares Code Red

In December 2022, Sundar Pichai initiated an internal Code Red in response to ChatGPT's success.

The challenge was organizational as much as technical.

Google had spent years using AI to improve existing businesses.

It now needed to release AI-native products quickly enough to compete with startups that were not constrained by Google's existing revenue model.

The company accelerated development of Bard, a public chatbot initially based on LaMDA.

Its launch in early 2023 was troubled.

Bard lacked some of ChatGPT's conversational polish, and the underlying model was perceived as weaker.

A prepared promotional demonstration also included a factually incorrect answer.

That was particularly embarrassing because the presentation had been recorded in advance.

Alphabet's shares fell approximately 8% following the announcement, intensifying investor concerns that Google had fallen behind commercially.

Google replaced Bard's original LaMDA foundation with PaLM in May 2023.

The change improved the product, but OpenAI had already released GPT-4 in March, increasing the competitive gap perceived by the hosts.

Pichai responded with two major organizational decisions.

Google Brain and DeepMind become one organization

In April 2023, Google announced the combination of Google Brain and DeepMind.

The new organization became Google DeepMind, led by Demis Hassabis.

This reversed an important element of DeepMind's original acquisition arrangement.

When DeepMind joined Google in 2014, its founders had secured independence to pursue general intelligence without being absorbed into Google's ordinary product-development operations.

That separation had now become difficult to justify.

Google Brain emphasized practical research and production systems connected to Google's services.

DeepMind had concentrated more heavily on fundamental research and reinforcement learning.

Maintaining competing laboratories and model-development programs could fragment talent, data, and computing resources at precisely the moment Google needed to move faster.

The merger brought their capabilities under one leadership structure.

It also gave Hassabis responsibility for developing models intended to serve Google's broader commercial needs.

One model family: Gemini

Pichai's second major decision was to standardize Google's model strategy around Gemini.

Instead of allowing different teams to maintain unrelated foundation models for different products, Google would invest in a central model family and use it across the company.

That strategy had two components.

Existing product teams would increasingly use Gemini where AI capabilities were required.

At the same time, they were expected to identify additional opportunities to incorporate AI into their services.

The hosts characterized the arrangement as involving both pressure to adopt the shared technology and encouragement to discover new uses for it.

Jeff Dean, Oriol Vinyals, Hassabis, and other researchers became central to the effort.

Sergey Brin also returned to active technical involvement.

After Google's Character.AI agreement brought Shazeer back, he joined the Gemini leadership effort alongside Dean.

The arrangement reunited researchers who had participated in Google's language-model work across several decades.

Gemini becomes a product, not merely a model

Google moved quickly after the organizational changes.

It announced plans for Gemini at Google I/O in May 2023 and introduced early public versions in December.

The company increasingly developed Gemini as a multimodal model family, intended to interpret and generate information across text, images, audio, video, and other formats.

In February 2024, Gemini 1.5 introduced a context window supporting up to one million tokens.

That allowed applications involving far more information than many competing models could process in one interaction at the time.

Google subsequently retired Bard as its principal consumer brand and adopted Gemini for the standalone assistant.

The company also accelerated model releases during 2025, including Gemini 2.0 and successive versions of Gemini 2.5 Pro.

The strategy was no longer to treat conversational AI as a limited demonstration.

Google intended Gemini to become a foundational technology across its consumer and enterprise products.

Why Google separated AI Overviews, AI Mode, and Gemini

Rather than immediately replacing conventional search with a chatbot, Google developed several different AI experiences.

AI Overviews introduced generated answers within ordinary search results. This allowed Google to improve selected searches while retaining the traditional interface.

AI Mode provided a more conversational search experience. It was introduced during 2025, with Google testing ways to expose additional users to the format.

The Gemini app operated as a standalone AI assistant, competing more directly with ChatGPT and Claude.

These offerings served different purposes.

AI Overviews helped Google introduce AI to existing search users without requiring them to adopt an entirely new product.

AI Mode allowed experimentation with deeper changes to information retrieval.

The Gemini app gave Google a place to compete directly in the conversational AI category.

Google could adjust the balance among these products as usage, advertising demand, and operating costs became clearer.

This was particularly important because every additional AI-generated response required computing resources. Introducing generative inference at Google's search volume was an enormous infrastructure undertaking.

NotebookLM, video generation, and interactive worlds

Google's AI strategy also extended beyond search.

NotebookLM became a notable product by allowing users to work with documents and generate conversational audio discussions based on source material.

The hosts saw it as a practical example of AI creating a new kind of information experience rather than simply answering conventional search queries.

Google also developed generative media products.

Veo 3 demonstrated advanced video-generation capabilities. Nano Banana became associated with image generation and editing. Genie 3 demonstrated the possibility of generating interactive environments from prompts.

The hosts were particularly impressed by demonstrations of environments that could develop dynamically as users explored them.

They also acknowledged that they had not personally verified every demonstrated capability.

These products presented commercial opportunities that did not necessarily require sacrificing traditional search revenue.

YouTube could become particularly important.

Google already possessed extensive experience handling video, distributing content, identifying visual material, and operating advertising systems.

AI-generated video could increase the amount of material creators produce, while improved video understanding could support discovery, accessibility, advertising, and commerce.

Gemini's 2025 progress and the much larger 2026 numbers

By the time of the episode, Google reported approximately 450 million monthly Gemini users.

The hosts questioned precisely what was included in the figure. Deliberate usage of a standalone chatbot differs from AI interactions automatically embedded in other products.

That distinction remains important when comparing competitors.

Subsequent disclosures show substantial growth.

In a June 2026 investor presentation, Google reported that AI Overviews had more than 2.5 billion monthly users, AI Mode had exceeded one billion monthly users, and the Gemini app had surpassed 900 million monthly users.

Alphabet's July 22, 2026 earnings release subsequently placed the Gemini app at 950 million monthly active users.

The company also reported that Gemini models were processing approximately 22 billion API tokens per minute.

These figures are company-reported and represent different usage measures. They cannot simply be added together to calculate unique AI users.

They nevertheless show that Google's AI response had moved well beyond its troubled Bard launch.

The harder question remains whether those interactions can generate the same economic value as the search activity they may eventually replace.

6. Google's Infrastructure Advantage: It Makes Models, Chips, and Cloud Services

Nvidia helped create the industry Google now competes in

Google's AI infrastructure history contains another tension.

The company helped establish Nvidia's importance in AI while simultaneously developing a strategy to reduce dependence on Nvidia hardware.

When Alex Krizhevsky arrived at Google following the DNN Research acquisition, he was surprised by the company's reliance on CPUs.

Google's infrastructure engineers preferred relatively standardized fleets of machines. GPUs introduced new hardware requirements and operational complexity.

Krizhevsky reportedly purchased a GPU machine himself, installed it near his office, and demonstrated that the technology could accelerate neural-network training.

Jeff Dean and engineering executive John Giannandrea subsequently developed plans to introduce GPUs throughout Google's data centers.

In spring 2014, Larry Page approved a purchase of approximately 40,000 Nvidia GPUs for $130 million, despite concerns about the expenditure.

Nvidia's annual revenue was then approximately $4 billion, with a market capitalization near $10 billion.

Google's order was therefore meaningful evidence that neural-network computing could become a substantial commercial market.

The GPUs supported wider deployment of neural networks in Google Photos, speech recognition, search, Gmail, and advertising.

Improved advertising predictions helped generate financial returns from the investment.

But these successful applications created another problem.

Speech recognition reveals an expensive future

Dean calculated what might happen if neural-network speech recognition were deployed across approximately one billion Android phones.

Even if each phone used the technology for only a few minutes daily, the additional processing could require Google to roughly double its data-center capacity.

This was economically unattractive.

Google could continue purchasing general-purpose computing hardware, but expanding AI into every major product would create enormous costs.

Its engineers began pursuing specialized chips optimized for neural-network calculations.

Jonathan Ross had already investigated programmable hardware through Google's internal experimental-project arrangements.

That work contributed to a formal effort to develop an application-specific integrated circuit.

The result was the Tensor Processing Unit, or TPU.

Why TPUs matter

A TPU is designed to perform the mathematical operations used by machine-learning models efficiently.

In particular, neural networks rely heavily on matrix multiplication.

General-purpose processors and GPUs can perform these operations, but their flexibility also consumes hardware resources.

Google's engineers designed TPUs to prioritize useful neural-network operations over broader computing versatility.

Reduced numerical precision became an important feature.

Neural networks often do not require every calculation to use the same precision expected in traditional scientific computing.

By accepting carefully controlled rounding, engineers could perform more calculations using less power and hardware.

Techniques known as quantization helped compensate for the reduced precision.

Google used programmable FPGAs to test the proposed architecture before committing to specialized manufactured chips.

The first TPU was designed to fit into existing server infrastructure, occupying approximately the space of a hard drive.

This allowed deployment without completely rebuilding Google's data centers.

Much of the engineering work involved Google's Madison, Wisconsin operation, demonstrating that important contributions came from outside its Mountain View headquarters.

The initial TPU was designed, verified, manufactured, and deployed in approximately 15 months.

Google's 2017 technical disclosure confirms the development period and explains why the project became urgent.

Google reported that its first TPU achieved approximately 15–30 times the performance, and 30–80 times the performance per watt, of certain contemporary CPUs and GPUs on the workloads examined.

Those figures describe a particular hardware generation and workload comparison. They should not be treated as universal performance ratios for all modern AI chips.

From an internal secret to millions of chips

Google initially kept TPU development confidential.

The company had little reason to publicize an advantage that might allow it to deploy AI more efficiently than competitors.

Early TPUs were used for inference, meaning the operation of trained models rather than the complete training process.

AlphaGo demonstrated part of their potential. The system used four TPUs during its landmark match against Lee Sedol.

Google continued developing successive generations, eventually incorporating increasingly sophisticated capabilities comparable to those found in advanced GPUs.

The hosts estimate that by 2025, Google operated approximately two to three million TPUs.

This was presented as an industry estimate rather than a confirmed public inventory.

For comparison, the hosts discussed estimates of approximately four million Nvidia GPUs shipped during the preceding year.

The figures are not directly equivalent: one concerns an installed population of proprietary accelerators, while the other concerns annual shipments across Nvidia customers.

Even with that limitation, they illustrate the considerable scale of Google's internal chip operation.

TensorFlow makes AI easier to deploy

Hardware was only one part of the system.

Google Brain also developed TensorFlow, a machine-learning framework that replaced the earlier DistBelief system.

TensorFlow made it easier to build and operate neural networks across CPUs, GPUs, and TPUs.

Researchers could use a more consistent software environment rather than rebuilding models for every hardware architecture.

By making TensorFlow available beyond Google, the company also influenced the wider machine-learning ecosystem.

This became strategically relevant as external customers began adopting Google's cloud services and specialized AI hardware.

7. Google Cloud's Turnaround Gives Google Another AI Business

Why Google initially struggled against AWS

Google possessed advanced data centers long before Google Cloud became a major enterprise business.

Its first important commercial cloud offering, App Engine, arrived in 2008.

App Engine simplified application deployment, but required developers to follow relatively restrictive programming and infrastructure arrangements.

Amazon Web Services pursued a more flexible approach.

AWS allowed customers to purchase computing capacity, storage, databases, and other infrastructure components without requiring them to adopt one prescribed application architecture.

That flexibility appealed to developers and enterprises.

Google eventually introduced Compute Engine in 2012, providing infrastructure closer to the model customers increasingly preferred.

But product architecture was not Google's only obstacle.

Google lacked the enterprise-sales organization required to compete effectively for large corporate customers.

Its traditional businesses had grown through consumer adoption and self-service advertising.

Cloud infrastructure required a different commercial relationship: direct sales, technical assistance, support commitments, and long-term customer management.

Google had also been reluctant to commercialize some of its most sophisticated internal systems.

Technologies such as Bigtable, Borg, and DistBelief had initially been developed primarily for Google's own requirements.

Kubernetes and multicloud change the strategy

Google increasingly embraced software portability and Kubernetes, the container-orchestration system that became important for deploying applications across computing environments.

This supported a multicloud strategy.

Rather than insisting that customers move everything into Google Cloud, Google could help them operate systems across Google, Microsoft, Amazon, and other infrastructure providers.

For a provider trailing AWS and Microsoft Azure, compatibility offered an important commercial advantage.

Google was not always trying to replace every existing supplier. It could win individual workloads by offering useful technology and competitive economics.

Thomas Kurian builds an enterprise organization

In late 2018, Google appointed former Oracle president Thomas Kurian to lead its cloud business.

Kurian brought experience selling complex technology systems to large enterprises.

He expanded Google's sales and go-to-market organization substantially.

The hosts describe an earlier commercial operation of approximately 150 employees growing by thousands, including a broader expansion involving roughly 10,000 positions.

Google also developed geographically distributed sales teams rather than concentrating its commercial relationships in California.

The financial results demonstrate the change.

Google Cloud generated approximately $4 billion in revenue in 2017, about $13 billion in 2020, and approximately $26 billion in 2022.

It achieved profitability in 2023.

By the time of Acquired's 2025 episode, it had exceeded a $50 billion annualized revenue run rate.

Cloud had become a major business rather than an expensive adjunct to Google's consumer operations.

The latest filings show a much larger company

Alphabet's 2025 annual report provides independently inspectable financial figures.

Google Cloud generated $58.7 billion in revenue in 2025, compared with $43.2 billion in 2024.

Operating income increased from approximately $6.1 billion to $13.9 billion.

This is more informative than an annualized revenue run rate because it reports the complete year's realized revenue and operating income.

The expansion continued in 2026.

For the quarter ended June 30, 2026, Google Cloud reported $24.8 billion in revenue, an increase of 82% compared with the same quarter a year earlier.

The scale of growth reflects strong demand for cloud infrastructure and AI services, although cloud revenue includes products beyond AI.

Google Cloud is now strategically valuable in two different ways.

It earns revenue by selling computing services to external customers.

It also operates infrastructure connected to Google's own frontier-model development, creating opportunities to share engineering capabilities and reduce costs.

Google's chips become a commercial alternative

External customers can access TPUs through Google Cloud.

This creates an alternative to purchasing or renting Nvidia-based computing capacity.

Customers willing to develop software compatible with Google's hardware can potentially benefit from its specialized architecture.

Google also benefits because a broader customer base increases utilization and encourages software development around its chips.

The hosts discuss rumors that independent AI computing providers, sometimes called neoclouds, might eventually distribute TPU capacity.

They do not identify a confirmed specific arrangement in the episode.

The underlying strategic possibility is clear: wider TPU availability could create a larger developer ecosystem, making Google's hardware useful beyond its own data centers.

That would strengthen an advantage Nvidia already possesses through CUDA.

The economics of owning the computing stack

The hosts believe hardware costs may become unusually important in the AI industry.

Traditional software businesses frequently operate with high gross margins because distributing another copy of a software product costs relatively little.

Generative AI requires substantial computing resources every time a model processes a request.

The hosts suggest gross margins around 50% may be realistic for some AI businesses, compared with approximately 80% for especially profitable conventional software companies.

They also discuss Nvidia's gross margins, estimated around 75–80%, as evidence that purchasing external AI hardware can be expensive.

Google's proprietary chips may reduce the amount of supplier margin embedded in its computing costs.

However, the savings are not equivalent to manufacturing chips for free.

Google relies on partners, including Broadcom, for aspects of TPU development and production. The hosts discuss an unverified estimate that some of these arrangements may involve supplier margins around 50%.

The episode also offers rough estimates of AI data-center operating costs: more than half associated with chips and depreciation, approximately one-quarter to one-third with researchers and engineers, and a much smaller share directly attributable to electricity.

These percentages are not universal accounting figures. Different companies allocate costs differently, and electricity also creates infrastructure constraints not captured by its direct operating expense.

Hardware depreciation matters because expensive AI accelerators may lose economic value rapidly as newer generations arrive.

A company operating millions of chips must continually balance replacement costs against performance and efficiency improvements.

Google has another advantage: its established private networking infrastructure.

The company acquired inexpensive unused fiber capacity following the early-2000s telecommunications downturn and subsequently developed extensive data-center connectivity.

That network supports distributed computing as well as services such as YouTube.

Google's extraordinary token-processing scale

High usage allows Google to spread the fixed cost of model training over more AI interactions.

The hosts describe Google's reported token-processing volume growing from approximately 10 trillion tokens in April 2024 to almost 500 trillion in April 2025.

That represents an increase of roughly fifty times in one year.

By June 2025, the episode reports approximately 980 trillion tokens processed, with the possibility of exceeding one quadrillion in a subsequent month.

These figures were company-reported or estimated at the time, not independently measured by the hosts.

More recent disclosures use different measures.

In July 2026, Alphabet reported that Gemini models processed approximately 22 billion API tokens per minute.

API throughput is not directly comparable with every token processed across embedded Google products, so the different figures should not be treated as one continuous series.

They nevertheless demonstrate how large Google's AI infrastructure has become.

The strategic advantage is the combination of enormous existing demand and substantial control over the technology needed to serve it.

Google can train models, manufacture specialized chip designs through partners, operate data centers, distribute services through Google Cloud, and deliver AI through products already used by billions.

Few competitors possess all those capabilities within one corporate group.

8. Google's Financial Strength Is Also Its Biggest Obstacle

The profits that make disruption difficult

Google's business presents a classic innovator's dilemma.

It controls approximately 90% of traditional search and earns substantial margins from advertising.

Its position is strengthened by consumer habits, distribution arrangements, brand recognition, and the integration of search into browsers and devices.

A competing AI startup can build a product that changes how people retrieve information without sacrificing an established search-advertising business.

Google cannot make the same decision without considering the financial consequences.

The problem is particularly important because Google Search is closely connected to commercial intent.

People search for products, travel, services, businesses, and information that influences purchasing decisions.

Advertisers pay to appear when those intentions are expressed.

If a conversational assistant supplies a complete answer without requiring traditional search results, some of those advertising opportunities may disappear.

Google must therefore consider two different outcomes: providing a better experience and earning sufficient revenue from it.

They are not necessarily the same thing.

What Google's financial statements actually show

The episode describes Alphabet approaching approximately $370 billion in annual revenue and around $140 billion in earnings during 2025.

Those figures reflected the financial scale being discussed at the time.

The company's subsequently published full-year accounts provide a clearer reference.

According to Alphabet's 2025 Form 10-K, the company recorded:

These are full-year reported figures, not estimated annualized rates.

They explain why Google has the resources to finance AI research on a scale that would be difficult for an independent laboratory to match.

Alphabet can invest heavily in models, chips, and data centers while continuing to fund other businesses, repurchase shares, and pay shareholder dividends.

Its market capitalization reflected this financial strength.

After reaching approximately $2 trillion in late 2021, Alphabet's valuation fell substantially during the technology-market downturn of 2022, at one stage approaching $1 trillion.

By September 2025, its market capitalization had reportedly reached approximately $3 trillion.

The company's share-price recovery occurred despite continued concerns about AI competition.

AI infrastructure is becoming increasingly capital-intensive

Google's financial capacity does not eliminate the cost of the AI transition.

Alphabet spent approximately $91.4 billion on capital expenditure during 2025, primarily connected to technical infrastructure.

Spending accelerated further in 2026.

Following its July 2026 earnings announcement, Reuters reported that Alphabet had increased its expected 2026 capital expenditure to approximately $195–205 billion.

The forecast reflected demand for AI computing capacity that continued to exceed available supply.

The company's second-quarter 2026 revenue reached $119.8 billion, a 24% year-over-year increase.

Google Search and related revenue grew 17%, while Google Cloud revenue rose 82%.

These results complicate the argument that AI had already caused a visible collapse in Google's search business.

As of those financial disclosures, search revenue was still growing.

However, current growth does not establish what search advertising will look like after a larger share of information retrieval moves into conversational AI.

It is also important to distinguish business profitability from accounting gains. Alphabet's second-quarter 2026 results included approximately $98 billion in net other income, primarily from unrealized gains on equity securities. Those gains should not be mistaken for ordinary operating profit from search, advertising, or cloud services.

Why replacing search advertising is difficult

The hosts estimate that Google generates approximately $400 in annual search revenue per American user.

The figure is a rough calculation, not a reported per-user accounting measure.

Its significance is that a free service can generate substantial value from advertising.

A consumer subscription must recover that revenue directly from the customer.

At $20 monthly, a subscription generates $240 annually before operating costs.

Even if a consumer considers an AI assistant useful, willingness to pay may be lower than the value advertisers generate through that person's commercial search activity.

The host's estimate therefore illustrates why Google's choice is more complicated than simply charging everyone for Gemini.

Google could develop a widely adopted AI product while capturing less revenue per user.

Alternatively, it could lose high-value commercial searches to competitors while retaining lower-value informational searches.

Travel planning is an especially relevant example.

People planning trips often research destinations, accommodation, flights, and local activities.

These searches support valuable advertising markets.

If a conversational assistant performs much of that planning directly, the traditional series of searches and paid placements may become less important.

Health and legal information present similar changes in user behavior, although their advertising economics and safety considerations differ.

AI could create a better advertising opportunity

The hosts do not assume that AI necessarily destroys advertising.

Conversational prompts can provide more detailed information about what users want.

They compare conventional search queries averaging roughly two or three words with AI prompts that may contain 20 words or more.

A detailed conversation could reveal preferences, constraints, budgets, and buying intentions more clearly than a short search query.

In principle, this could support more relevant advertising.

AI could also help users complete tasks they previously abandoned because the required research or coordination was too difficult.

That might create new commercial activity rather than merely redistribute existing searches.

But the opportunity is not equivalent to a proven business model.

By the time of the episode, conversational AI had not demonstrated an advertising system capable of consistently reproducing Google's traditional search economics at comparable scale.

The distinction between potential advertising relevance and actual advertising revenue remains central to evaluating the company's strategy.

Subscriptions provide another route

The hosts reconsider an assumption that consumer AI would necessarily need advertising.

Discussion with former YouTube executive Shishir Mehrotra encouraged them to examine subscription bundles more seriously.

Google already possessed a substantial subscription business through Google One.

During 2025, it reported more than 150 million paying Google One subscribers, with annual subscriber growth approaching 50%.

However, the majority were believed to use relatively inexpensive plans, while premium AI capabilities were associated with higher-priced subscriptions.

The total number of Google One subscribers cannot be used as a measure of how many people pay specifically for Gemini.

Google could nevertheless integrate AI into an existing collection of paid services.

Storage, YouTube Premium, YouTube Music, and other consumer products offer potential bundling opportunities.

The hosts also mention services such as NFL Sunday Ticket when discussing the breadth of Google's subscription portfolio.

A bundle might encourage customers to pay more for a collection of services than they would for one standalone chatbot.

Google's existing subscriber relationships could therefore reduce the difficulty of introducing paid AI.

Personalization could become another advantage

Google possesses extensive consumer relationships through Search, Gmail, Maps, Docs, Chrome, Android, and other products.

With appropriate user permissions, AI systems could use information from these services to provide more relevant assistance.

An assistant connected to a person's calendar, communications, documents, and travel preferences could perform tasks that a disconnected chatbot would find harder.

Such integrations could also create switching costs.

At present, many consumers can move among Gemini, ChatGPT, Claude, and other assistants relatively easily.

If an assistant becomes deeply integrated into daily workflows, replacing it may become more difficult.

Enterprise customers already face stronger switching costs because AI systems can be connected to internal databases, compliance requirements, documents, and specialized business processes.

Personalization therefore offers Google both a product opportunity and a potential source of competitive protection.

It also raises questions about privacy, consent, and the appropriate use of personal information.

YouTube could become an AI commerce platform

YouTube represents another major opportunity.

It already combines search, video distribution, creator relationships, audience attention, and advertising.

The hosts discuss a future in which AI recognizes products appearing in videos and makes them directly purchasable.

A viewer watching a cooking video, for example, could identify kitchen equipment without requiring the creator to manually provide every product link.

Generative video tools could also lower production costs, allowing more people to create content.

That might increase engagement and advertising inventory.

These possibilities remain dependent on product development and commercial adoption, but they illustrate why Google's AI opportunity extends beyond answering text questions.

The company owns established distribution platforms where AI can become useful without necessarily replacing the core service.

9. Can Google Defend Its Competitive Advantages in AI?

Search dominance does not guarantee AI dominance

Google's approximately 90% search share is unusually difficult to reproduce.

Its AI position is different.

Users have credible alternatives from OpenAI, Anthropic, Microsoft, and other developers.

The hosts speculate that Google might eventually control only 25–50% of the AI market.

This is a scenario for considering potential business outcomes, not a verified market forecast.

It matters because even improved AI monetization might not fully compensate for a reduced share of consumer information retrieval.

The company could remain one of the leading AI providers while experiencing lower market concentration and reduced revenue per interaction.

Applying Hamilton Helmer's Seven Powers

The episode evaluates Google using Hamilton Helmer's Seven Powers framework.

The framework identifies several possible sources of durable competitive advantage: scale economies, network economies, counterpositioning, switching costs, branding, cornered resources, and process power.

Google's strongest apparent advantages are concentrated in a few categories.

Scale economies come from its data centers, TPUs, networking infrastructure, and enormous inference volume. Spreading fixed costs over more usage can reduce the average cost of serving AI requests.

Distribution comes from Google's position as the established entry point to the internet. Its relationships with consumers, browsers, and devices are extremely difficult for a new company to reproduce. The hosts treat this as a strategically controlled resource.

Branding provides recognition and trust. Many consumers know Google better than they know newer AI companies, even though being an incumbent can also create reputational disadvantages.

Other forms of competitive protection are less certain.

Switching costs remain limited for consumers

Consumers can often move between AI assistants without a major financial or technical burden.

They may prefer a particular interface, but another model can answer many of the same questions.

This differs from enterprise software, where changing providers can require migrating records, rebuilding integrations, retraining employees, and revising operational procedures.

Google could strengthen consumer switching costs through personalization, but this had not yet become an established advantage at the time of the episode.

AI assistants lack obvious direct network effects

A social network can become more useful when additional friends join.

General-purpose AI assistants do not yet display the same obvious direct relationship.

One additional Gemini user does not automatically make Gemini more useful to another user.

Greater usage may support product improvement, feedback, and economic scale, but those are not identical to direct network effects.

The distinction matters because Google's overwhelming distribution could be less protective in AI than it has been in conventional search.

Startups hold the counterpositioning advantage

Counterpositioning occurs when a challenger adopts a business strategy that an incumbent finds difficult to copy because doing so threatens its existing operations.

That description fits Google's AI dilemma.

OpenAI could build ChatGPT around direct answers without needing to protect a search-advertising franchise.

Google faced a much more expensive choice.

The challenger could pursue a potentially disruptive product because it had relatively little established revenue to lose.

Google was more vulnerable precisely because its existing business was so successful.

Research excellence is not permanent protection

Google's history of scientific breakthroughs does not guarantee continued technical leadership.

OpenAI, Anthropic, and other laboratories have also produced important research.

The next major AI architecture may emerge from any of several institutions.

The hosts therefore question whether Google possesses sufficient process power to ensure that its research organization will consistently outperform competitors.

They conclude that Google's strongest demonstrated advantages remain scale, distribution, and branding.

Future personalization and enterprise integration might add stronger switching costs.

But Google's search-era advantages do not automatically transfer unchanged into AI.

10. Google's Search Monopoly Survives a Major Antitrust Test

The federal court's monopoly finding

Google's distribution advantage is not merely a matter of business strategy.

It has also been central to federal antitrust litigation.

In August 2024, U.S. District Judge Amit Mehta concluded that Google had illegally maintained monopoly power in general search services.

The court found that Google's distribution arrangements helped preserve its position in the market.

The U.S. Department of Justice's summary quotes the court's finding:

“Google is a monopolist, and it has acted as one to maintain its monopoly.”

The court's decision followed extensive litigation concerning default search placement, distribution agreements, and the barriers confronting potential competitors.

Google has disputed the legal conclusions and argues that consumers choose its search engine because of its quality.

The September 2025 remedies decision

The court's September 2, 2025 remedies decision stopped short of imposing some of the most aggressive structural changes sought by government lawyers.

Google was not required to sell Chrome or Android.

It could also continue certain search-distribution payments, including arrangements involving Apple, subject to restrictions.

However, the outcome was not simply unrestricted permission to continue its previous practices.

The court prohibited specified exclusive distribution arrangements and required Google to make certain search-index and user-interaction data available to competitors.

It also imposed obligations concerning search and search-advertising syndication.

The purpose was to make it easier for rival services to compete without immediately recreating every component of Google's infrastructure.

AI competition influences the legal outcome

The rapid development of AI competitors influenced the remedies debate.

OpenAI, Anthropic, Perplexity, and other services suggested that search could face meaningful competitive pressure from products outside the established search-engine market.

The hosts highlight an unusual irony.

Google had employed many of the researchers who helped establish competing AI laboratories.

Some of those competitors originated partly because researchers wanted alternatives to Google's control of advanced AI development.

Their emergence subsequently strengthened arguments that competition might develop without forcing Google to divest major assets.

That is a significant connection between Google's research history and its legal position.

It does not mean AI competition eliminates the court's monopoly finding.

Nor does it establish that the new companies can sustainably challenge Google.

Can venture-funded competition last?

The hosts question whether emerging AI laboratories provide durable competition when many depend heavily on outside financing.

Operating frontier models requires substantial capital.

A startup may offer an excellent product while continuing to spend more than it earns.

Google, by contrast, can finance research through mature profitable businesses.

If investment conditions become less favorable, some competitors could struggle while Google continues investing.

The hosts therefore distinguish between the existence of competitors and the existence of sustainable competitive market structures.

Google may face genuine technological competition without losing the financial strength that protects its existing position.

What changed after the episode

Google appealed the search decision in January 2026.

In its public explanation of the appeal, the company argued that the court had underestimated consumer choice and rapidly increasing competition.

It also sought to pause selected remedies involving data sharing and syndication, citing privacy and innovation concerns.

The Justice Department's case docket continued listing proceedings concerning implementation and enforcement during 2026, including filings in October.

This means the September 2025 remedies decision should not be presented as the final resolution of every legal issue.

Google also faced a separate antitrust case concerning advertising technology.

In September 2026, the Justice Department announced additional court-ordered remedies in that separate litigation, including behavioral requirements intended to support interoperability with competing advertising technology.

The cases concern different markets and should not be conflated.

Together, they demonstrate that Google's search distribution and advertising infrastructure remain subjects of significant legal scrutiny, even as the company expands its AI offerings.

11. Waymo: Google's Other Major AI Investment

Google's artificial-intelligence history is not limited to search, language models, and cloud computing.

One of its longest-running projects is now a commercial autonomous-transportation business.

Waymo began years before deep learning became mainstream. Its origins trace to government research competitions, Stanford engineering, and Larry Page's conviction that computers could eventually replace human drivers.

The project required more than a decade of development before it became a commercial driverless service.

By 2025, it had accumulated more than 100 million fully driverless miles and provided millions of paid rides.

The hosts increasingly viewed it as a potentially substantial business rather than merely an expensive experiment.

The DARPA competition that started the industry

In 2004, the Defense Advanced Research Projects Agency organized the Grand Challenge.

Autonomous vehicles were required to navigate approximately 132 miles through the Mojave Desert without human intervention.

The competition offered a $1 million prize, an unusual arrangement that required special congressional authorization.

Approximately 100 teams registered.

None completed the course.

The following year's competition offered $2 million, and technological progress was dramatic.

Of 23 finalists, 22 exceeded the previous year's best distance, while five vehicles completed the entire course.

Sebastian Thrun led Stanford's winning team.

The Stanford engineers took a distinctive approach.

Carnegie Mellon competitors emphasized extensive vehicle modifications and sophisticated mechanical systems for stabilizing sensors.

Stanford favored a relatively standard Volkswagen equipped with commercially available sensing hardware.

The team combined laser sensors with ordinary video cameras.

Lasers helped identify nearby road surfaces and obstacles with considerable precision. Cameras could observe areas farther ahead, although interpreting visual information was more difficult.

Machine learning connected the two.

The system identified safe terrain using laser measurements, learned what that terrain looked like through cameras, and used visual information to estimate the route beyond the immediate range of its laser measurements.

This allowed the vehicle to anticipate curves and operate at higher speeds.

Stanford's car eventually overtook a Carnegie Mellon competitor autonomously and won the race.

The achievement demonstrated that advanced software could compensate for some limitations of comparatively ordinary hardware.

Larry Page creates Project Chauffeur

Thrun later joined Google, where Page became interested in turning autonomous-driving research into a major initiative.

Thrun was initially reluctant.

A vehicle operating on public roads introduced safety risks far beyond those encountered during a controlled desert competition.

Page challenged him to identify a fundamental technical reason self-driving cars could not eventually work.

Thrun concluded that much of his hesitation concerned the risks of attempting the project rather than an established technological impossibility.

Project Chauffeur began in 2009, becoming an early major initiative within Google X.

Page established an ambitious benchmark called the Larry 1000.

It consisted of approximately 1,000 miles of challenging California driving routes, including roads around Lake Tahoe, San Francisco's Lombard Street, the Bay Bridge, and Highway 1 toward Los Angeles.

Google's small team completed the challenge within approximately 18 months.

That was encouraging, but the project remained far from a reliable commercial service.

A demonstration route could be completed successfully without proving that a vehicle could handle every unusual road condition, unexpected obstacle, weather event, or unpredictable human driver.

Solving those problems required many more years.

Assistance systems or fully autonomous vehicles?

Google considered different commercial directions.

Thrun favored developing a highway-driving assistance system resembling an advanced cruise-control product.

Page preferred a more ambitious idea: vehicles capable of transporting passengers without a human driver.

That distinction became fundamental to Waymo's eventual strategy.

A driver-assistance system assumes a human remains available to intervene.

A driverless robotaxi must operate safely without relying on continuous human supervision.

Google also considered other ways to enter the automotive industry.

According to the episode, Eric Schmidt proposed acquiring Tesla during a period when the company was struggling with Model S production.

The hosts discuss a possible acquisition price of approximately $3 billion.

No transaction occurred.

They speculate about how acquiring Tesla might have affected Google's relationships with Musk, DeepMind, and OpenAI.

Those possible alternative histories cannot be established, but the discussions demonstrate how closely the emerging autonomous-driving and AI industries were connected.

Neural networks change Waymo's capabilities

Google's original autonomous-driving project began before modern deep learning became central to the field.

For several years, its systems relied on other methods for perception, navigation, and decision-making.

Around 2013, neural networks began improving the vehicles' ability to recognize and interpret their surroundings.

Convolutional neural networks were particularly useful for identifying objects and understanding visual scenes.

Google's expanding GPU infrastructure supported the development and deployment of these systems.

Transformers and related model advances later became relevant to prediction, planning, and other autonomous-driving functions.

The connections show how research developed for one application could become valuable in another.

Technology originally associated with image classification, language modeling, and large-scale computing eventually helped vehicles interpret complex environments.

Waymo becomes a separate Alphabet company

In 2016, the project became Waymo, a distinct Alphabet subsidiary.

The separation reflected growing confidence that autonomous driving could become a commercial business.

Waymo subsequently raised substantial outside financing.

In March 2020, it announced a funding round of approximately $3.2 billion, with investors including Silver Lake, the Canada Pension Plan Investment Board, Mubadala, Andreessen Horowitz, and Alphabet.

Later fundraising included approximately $2.5 billion in a Series B round and $5.6 billion in a Series C round.

Alphabet remained the majority owner.

The outside investments provided capital while allowing Alphabet to retain significant exposure to the business.

Commercial rides begin in Phoenix

In October 2020, Waymo launched a public commercial driverless service in Phoenix.

Passengers could pay for transportation in vehicles without a human safety driver behind the steering wheel.

The launch occurred approximately 11 years after Project Chauffeur began.

That interval illustrates the difference between proving a technical concept and delivering a service that can operate reliably in everyday conditions.

Waymo subsequently expanded into San Francisco, where its driverless service became broadly accessible in June 2024.

By 2025, it operated in five American metropolitan markets:

Phoenix, San Francisco, Los Angeles, Austin, and Atlanta.

The company also discussed additional American expansion and an international launch in Tokyo.

Waymo did not manufacture its own automobiles.

It used vehicles supplied by automotive manufacturers, including Jaguar electric vehicles, while developing its autonomous-driving hardware and software.

This allowed it to concentrate on the driving technology and transportation service rather than building an entirely new automobile-manufacturing operation.

Waymo's scale by 2025

The hosts cite several indicators of commercial progress.

By 2025, Waymo had reportedly completed more than 10 million cumulative paid rides.

Its fully driverless fleet had accumulated over 100 million miles, with approximately two million additional miles being driven each week.

The episode describes a fleet of approximately 2,000 vehicles carrying out hundreds of thousands of paid rides weekly.

Waymo reportedly generated more gross bookings than Lyft in San Francisco during January 2025.

That comparison concerned one market and one period. It did not mean Waymo had overtaken Lyft's overall business.

Still, it provided evidence that consumers were willing to pay for autonomous rides at meaningful scale.

This represented a major change from the project's earlier research demonstrations.

Waymo was no longer proving only that an autonomous car could navigate a difficult route.

It was operating a transportation service used repeatedly by paying customers.

Why Waymo uses more sensors than Tesla

Waymo's approach to autonomous driving differs substantially from Tesla's.

Waymo uses several types of sensors to gather overlapping information about its surroundings.

The hosts describe a system incorporating approximately 13 cameras, four lidar units, six radar units, and external microphones.

Cameras provide visual information.

Lidar uses laser measurements to estimate the positions and shapes of nearby objects.

Radar measures aspects of the surrounding environment and can provide useful information under conditions that challenge ordinary cameras.

Microphones can help detect sounds such as emergency-vehicle sirens.

The combination provides redundancy.

Waymo accepts the added hardware complexity and cost because it considers multiple sensing methods valuable for fully driverless operation.

Tesla has pursued a more camera-centered approach.

The distinction matters because Tesla's Full Self-Driving system, as discussed in the episode, still required human supervision.

Waymo's commercial service was designed to operate without a human driver supervising the vehicle.

The hosts emphasize that these were not equivalent products, even though both used advanced AI.

Rosenthal described using Tesla's driving assistance regularly while remaining responsible for monitoring the road.

His experience with Waymo was different because he could travel as a passenger without needing to perform that supervisory role.

Driverless transportation changes the passenger experience

The absence of a human driver produces practical differences beyond vehicle technology.

Passengers may feel more comfortable having private conversations.

Families can organize journeys without interacting with a stranger driving the vehicle.

Travelers may find it easier to bring pets or manage personal activities during a trip.

Rosenthal described using Waymo with his young daughter, including installing a child seat.

He considered the experience more convenient for certain family journeys than conventional ride-hailing.

These are personal observations rather than universal consumer findings.

They nevertheless illustrate why autonomous transportation might attract demand from people who do not regularly use Uber or Lyft.

A driverless service could expand the total number of trips people take, not merely transfer existing ride-hailing customers from one company to another.

Waymo's reported safety advantage

Safety is central to Waymo's commercial proposition.

At the time of the episode, the hosts cited data suggesting approximately 91% fewer crashes involving serious injuries or worse compared with human-driver benchmarks.

The comparison attempted to account for differences between Waymo's operating conditions and broader human-driving statistics.

Subsequent data provide a more recent view.

Waymo's September 24, 2026 safety update analyzed more than 270 million fully autonomous miles completed through June 2026.

The company reported that its vehicles were involved in approximately:

• 95% fewer crashes causing serious injury or worse than comparable human-driver benchmarks.
• 82% fewer crashes involving any reported injury.
• 93% fewer injury-causing crashes involving pedestrians.
• 86% fewer involving cyclists.
• 82% fewer involving motorcyclists.

These are Waymo's reported results, not independently verified guarantees of future performance.

The comparisons cover its operating areas and use human-driving benchmarks designed to account for the types of roads and conditions in which Waymo operates.

They should not be generalized automatically to highways, unfamiliar cities, severe weather, or every possible driving environment.

Waymo also reports that its comparison includes crashes regardless of fault.

Its published safety methodology and data provide more detail about the human-driver benchmarks, reporting adjustments, crash categories, and uncertainty estimates.

The updated results strengthen the evidence that Waymo has achieved substantially lower serious-injury crash rates than comparable human driving in its established operating environments.

What safer roads could be worth

The hosts discuss the considerable social and economic costs associated with road accidents.

They cite approximately 40,000 annual road deaths in the United States and a CDC estimate placing the annual economic costs associated with fatal crashes at approximately $470 billion in 2022.

Those costs include medical expenses and monetary estimates associated with lives lost.

The hosts offer a hypothetical calculation: a 90% reduction applied to $470 billion would represent more than $420 billion in annual societal savings.

This is illustrative, not a forecast of Waymo's financial results.

It also should not be confused with applying Waymo's measured reduction in serious-injury crashes to every category of American road death.

The actual benefits of autonomous transportation would depend on adoption, operating environments, vehicle performance, and the replacement of existing human driving.

More importantly, societal savings do not automatically become company revenue.

Waymo would need a sustainable business model that captures enough value to pay for its services.

Waymo's operating costs remain substantial

Operating autonomous vehicles involves many expenses.

Vehicles must be purchased or leased, equipped with sensors, maintained, cleaned, charged, insured, and managed.

Waymo must also fund software development, model training, inference, mapping, support operations, and hardware replacement.

The company has experimented with different commercial arrangements, including direct operations and partnerships with organizations experienced in managing vehicle fleets or ride-hailing services.

Profitability therefore depends on more than improving the underlying AI.

Waymo must reduce the cost of operating each vehicle while maintaining safety and service quality.

How large could Waymo become?

The hosts compare Waymo with two established industries.

They estimate the combined market capitalization of the global automotive industry at approximately $2.5 trillion including Tesla, or roughly $1.3 trillion excluding Tesla.

They also estimate the combined market capitalization of major ride-hailing companies at approximately $300 billion.

These comparisons provide a sense of possible economic scale, but they are imperfect.

Waymo does not primarily manufacture cars, so automobile-company valuations are not directly comparable.

Ride-hailing is a closer commercial reference, although autonomous transportation could eventually create additional markets.

Those markets might include privately owned autonomous vehicles, long-distance trucking, transportation for older people, and mobility for visually impaired passengers.

The most important uncertainty is whether autonomous operation can become sufficiently inexpensive and reliable to support widespread adoption.

The hosts estimate that Waymo's cumulative investment and losses had reached approximately $10–15 billion by 2025.

That would be a substantial amount for most companies.

For Alphabet, however, it was relatively modest compared with its annual earnings and the scale of the potential opportunity.

Rosenthal described changing his assessment of the project:

“I used to think this was like some wild goose chase. It now looks really, really smart.”

That is a personal investment assessment, not proof of Waymo's future profitability.

Its significance is the underlying change in evidence.

Waymo had moved from controlled demonstrations to a paid service, accumulated extensive driverless mileage, and published increasingly substantial safety data.

The company still faced considerable operating costs and expansion challenges, but it had established a commercial presence that did not exist during most of its earlier development.

For Alphabet, Waymo could become a significant business even if generative AI changes the economics of Google Search.

12. What the Acquired Episode Adds to Google's History

The episode's value comes partly from connecting events that are often discussed separately.

Google's language-model research helped improve advertising long before ChatGPT.

Its recruitment of university researchers contributed to Google Brain.

Google Brain's technical breakthroughs helped establish the modern AI industry.

Google's acquisition of DeepMind strengthened its research capabilities but also intensified concerns among outsiders about concentrating AI expertise.

Those concerns contributed to OpenAI's creation.

OpenAI subsequently obtained Microsoft's financing and infrastructure, developed GPT models, and launched the product that forced Google to reorganize.

The transformer researchers' departures created another unusual consequence: Google eventually paid billions to recover talent and technology associated with work it had originally supported.

Meanwhile, its investment in GPUs helped Nvidia establish AI demand, while the development of TPUs gave Google a potential cost advantage over other frontier-model companies.

And Waymo, which began before these developments became commercially fashionable, emerged as a separate transportation opportunity.

Reporting behind the episode

Gilbert and Rosenthal drew on discussions with executives, engineers, and researchers involved in these developments.

Their research included conversations with Sundar Pichai, Demis Hassabis, Jeff Dean, Sebastian Thrun, Greg Corrado, and others.

They also drew on books including Steven Levy's In the Plex, Cade Metz's Genius Makers, and Parmy Olson's Supremacy.

These sources helped reconstruct early recruiting decisions, scientific disagreements, investments, acquisitions, and internal organizational choices.

The interview material is particularly useful for understanding the relationships between people and institutions.

However, some anecdotes, acquisition discussions, estimates, and reported conversations depend on participant recollections or secondary reporting.

They should not all be treated as having the same evidentiary weight as a company's financial filing, a published research paper, or a court judgment.

The episode's closing recommendations and announcements

The hosts also discussed several subjects outside Google's corporate history.

Acquired announced an NFL innovation summit planned for February 2026 during Super Bowl week in San Francisco, including onstage programming and streaming.

Gilbert recommended the 2025 film F1, particularly its visual presentation and sound. He also discussed his family's experience with Travelpro luggage, praising its durability and practical design while citing approximately $416 for a large checked suitcase before discounts.

Rosenthal recommended the Glue Guys podcast, featuring Sequoia partner Ravi Gupta, former basketball player Shane Battier, and former NFL quarterback Alex Smith. He particularly highlighted an interview with sportswriter Wright Thompson.

He appreciated the podcast's willingness to value conversations even when relatively few people listened. One discussion attracted approximately 5,000 listeners, but its hosts considered the work worthwhile regardless of its limited initial audience.

Rosenthal also described introducing his young daughter to video games through a Steam Deck and Sea of Stars. Although she could not yet read, she gradually learned the controls and basic mechanics. The experience prompted him to consider buying a Nintendo Switch so they could play together.

These discussions were personal recommendations and announcements rather than evidence about Google's AI business.

13. Google's Next Challenge Is Economic, Not Simply Technical

Google's original ambition was to create a system that could understand information and provide people with what they needed.

Its research contributed directly to that ambition.

The company helped develop early language models, large-scale neural networks, the transformer, specialized AI chips, and increasingly capable consumer AI systems.

It also assembled an unusually comprehensive technology operation, combining models, cloud computing, proprietary hardware, consumer distribution, and substantial financial resources.

The concern that Google had fallen permanently behind after ChatGPT's launch became harder to support as Gemini improved, research operations were consolidated, and adoption increased.

Subsequent financial and product disclosures show that Google remained capable of expanding AI while continuing to grow its existing businesses.

But the innovator's dilemma has not disappeared.

Google's traditional search business produces revenue through a pattern of user behavior that AI may change.

The company can improve the quality of answers without necessarily increasing its advertising revenue.

It can expand AI usage while facing higher computing costs.

It can retain millions of paying subscribers while still needing to determine how much consumers will pay specifically for intelligent assistance.

And it can become one of the world's leading AI providers without maintaining the extraordinary market share it has enjoyed in search.

The financial consequences depend on questions that remained unresolved at the time of the episode and have not been fully answered by subsequent growth figures: how search behavior will change, how AI services will be monetized, whether proprietary computing will sustain a cost advantage, and how durable competition among frontier-model developers will become.

Google's position is unusual because it has helped create both sides of this transition.

It developed much of the technology enabling its competitors while building the infrastructure, products, and financial resources required to compete against them.

The central issue is no longer whether Google can build advanced artificial intelligence.

It is whether a company built around delivering links and advertisements can profitably become one that increasingly delivers answers, decisions, and completed work.