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

2025-10-06 - source - Read full transcript

Ben Gilbert (host)David Rosenthal (host)

Key insights

Google invented nearly every technology that powers the current AI boom, then failed to productize the biggest one, the Transformer, for five years.

Eight Google Brain researchers published 'Attention Is All You Need' in 2017, but all eight had left Google within a couple of years to join AI startups. Google used Transformers internally (BERT, search ranking) but never restructured its business around them until ChatGPT forced its hand in December 2022.

innovators-dilemma

The Google antitrust remedy in 2025 was effectively softened by the AI race, which arguably saved Google from a much harsher breakup outcome.

A federal judge ruled Google an illegal monopolist in search but declined to order a Chrome spinoff or end payments to Apple, citing competitive pressure from AI startups funded by tens of billions of venture dollars. Ben and David note the irony that the researcher exodus (Ilya Sutskever leaving to found OpenAI) that once threatened Google's AI lead may now be the reason Google avoided a structural breakup.

innovators-dilemma

Google is the only company today that owns all four pillars of the AI stack: a frontier model, a chip, a hyperscale cloud, and consumer-scale applications.

NVIDIA has chips only; Meta has applications only; Microsoft and Amazon have cloud plus partial applications; OpenAI and Anthropic have models only and depend on external capital and external clouds. Google alone is self-funding across all four, which the hosts argue is a structural advantage no other AI player currently has.

full-stack-ai-advantage

Google's TPU cost structure gives it a meaningful and possibly durable low-cost-producer advantage in an era where compute cost is unusually central to competitiveness.

NVIDIA GPU buyers pay roughly 75-80% gross margin (a 4-5x markup); Google's TPU partner Broadcom takes roughly 50% margin (a 2x markup). Since chips and their depreciation are over half the cost of running an AI data center, this margin gap materially lowers Google's cost per token versus everyone else, an argument sourced from Gavin Baker of Atreides Management.

scaling-laws-and-compute-economics

AlexNet's 2012 ImageNet win, achieved on two off-the-shelf $500 NVIDIA gaming GPUs, is the true 'big bang' moment that set NVIDIA on its path and proved GPUs beat CPUs for neural nets.

The Toronto team of Geoff Hinton, Alex Krizhevsky, and Ilya Sutskever cut the ImageNet error rate from 25% to 15% by rewriting their algorithms in CUDA and training on consumer graphics cards instead of supercomputer CPUs, a result Jensen Huang has repeatedly called AI's founding moment.

scaling-laws-and-compute-economics

Elon Musk's anger at losing DeepMind to Google, and later his ultimatum to take over OpenAI, both directly shaped the modern AI landscape.

Elon was an early DeepMind investor and wanted it for Tesla's self-driving effort; when Larry Page outbid him (via a personal rapport with Demis Hassabis), Elon co-founded OpenAI in 2015 partly out of frustration. In 2017-2018 he demanded control of OpenAI or threatened to walk with his funding; the board refused, he left, and the funding gap arguably forced OpenAI to pivot hard toward the Transformer-based GPT line to survive.

talent-exodus-and-founding-myths

ChatGPT's launch was accidental, not a planned consumer product strategy, and its speed shocked even OpenAI.

OpenAI built a simple chat wrapper around GPT-3.5 in about a week, partly to preempt a rumored Anthropic chat launch. It hit 1 million users within a week, 30 million by year end, and 100 million within two months, the fastest consumer product ever to do so, catching OpenAI's own infrastructure and business model (it expected a B2B licensing business) off guard.

talent-exodus-and-founding-myths

Google had a working ChatGPT-like product internally years before ChatGPT existed but chose not to ship it, for both safety and business-model reasons.

Noam Shazeer built an unsafe raw chatbot called Meena around 2019-2020 that lacked RLHF and would go off the rails (reportedly naming people who 'should die'). Google also faced a structural conflict: giving direct chat answers instead of ten blue links would cannibalize ad revenue and risked legal exposure over disintermediating publishers, both non-starters before the market forced the issue.

innovators-dilemma

Sundar Pichai's December 2022 'code red' response followed the textbook incumbent playbook for a disruptive innovation, not a sustaining one, and the hosts credit Google with executing it unusually well.

Sundar merged the rival Google Brain and DeepMind teams into one unit under Demis Hassabis (in violation of DeepMind's original acquisition independence terms), standardized on a single model line (Gemini) instead of multiple parallel efforts, and shipped an early Gemini within about six months, while carefully avoiding cannibalizing core Search revenue with AI Overviews and AI Mode rather than a wholesale replacement.

innovators-dilemma

Google's own AI monetization is unproven, in stark contrast to how fast it monetized Search in 1998-2000.

Google reportedly earns roughly $400/user/year from ads; a $20/month AI subscription tier is a hard sell to more than a thin slice of that base. Search's clear ad-based value capture mechanism (AdWords) emerged within two years of launch; AI's business model, ad-supported or subscription, is still undetermined more than two and a half years after ChatGPT's launch.

ai-monetization-uncertainty

AI queries may be more monetizable per-query than search, in theory, because of longer, higher-intent prompts, but no one has built that ad product yet.

Bill Gross observed to Ben that AI chat queries average 20+ words versus 2-3 words for web search, implying far more precise user intent and potentially much higher ad rates if a workable ad unit is ever built into chat interfaces; this remains speculative and unbuilt as of the recording.

ai-monetization-uncertainty

Waymo could become a Google-sized standalone business, built for a fraction of what foundation model labs are spending.

Using CDC data (roughly $470B in annual US crash costs) and Waymo's reported 91% reduction in serious-injury-or-worse crashes versus human drivers, the hosts estimate over $400B/year in addressable cost savings, larger than Google's entire current revenue, achieved so far on only about $10-15B of cumulative Alphabet investment, roughly one year of Uber's profit.

full-stack-ai-advantage

Applying the Seven Powers framework to Google's AI products shows a narrower moat than Google enjoyed in Search.

Scale economies (amortizing training costs across roughly a quadrillion tokens served per month by mid-2025), branding, and cornered-resource distribution (Search as the default front door) are strong. But switching costs and network economies are largely absent in today's AI products, and Google has no counter-positioning or process-power advantage versus OpenAI, Anthropic, and other labs, a much thinner set of structural advantages than Search had at its peak.

scaling-laws-and-compute-economics

Books referenced

In the Plex - Steven Levy - Primary source for all three Google episodes; contains the 2001 micro-kitchen story where George Herrick, Ben Gomes, and Noam Shazeer first theorized that compressing data is equivalent to understanding it.
Supremacy: AI, ChatGPT, and the Race That Will Change the World - Parmy Olson - Main source for the DeepMind founding story, the Demis/Shane/Mustafa origin, and the competing acquisition offers from Facebook, Tesla, and Google.
Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World - Cade Metz - Source for the DNN Research auction story (Geoff Hinton auctioning his company from a Lake Tahoe hotel room) and other AI history details.

Media referenced

The Bitter Lesson - paper - Rich Sutton's 2019 essay argued that scalable architectures fed more data and compute beat hand-crafted algorithms; framed as the philosophy the Transformer proved a few years early.
AlphaGo (documentary) - movie - Google-produced documentary about DeepMind's AlphaGo beating world Go champion Lee Sedol; Ben highly recommends watching it in full.
Nova: The Great Robot Race - show - 20-year-old PBS/Nova documentary on the 2005 DARPA Grand Challenge that Ben watched to research the Stanford self-driving team's origin story, on a tip from Bret Taylor.
Jeff Dean and Noam Shazeer episode - podcast - The DoorDash podcast episode where Jeff Dean and Noam Shazeer tell the story of parallelizing Google Translate's n-gram model from a 12-hour to a 100-millisecond translation time.
The Great AI Awakening - article - David's 2016 New York Times Magazine piece on Google Translate's switch to neural networks, which he says left him with his 'jaw on the floor.'
The YouTube Tip of the Google Spear - article - Ben Thompson/Stratechery piece published the week of recording, arguing Google's video and UGC advantage (training data plus shoppable video) is an underrated AI bull case.
Glue Guys podcast - podcast - Ben's carve-out; hosted by Ravi Gupta (Sequoia), Shane Battier, and Alex Smith. Ben and David recently guested on it.
Wright Thompson interview on Glue Guys - podcast - Specific episode David carves out as excellent despite low listen counts.
Stepchange podcast - podcast - New podcast by Ben Eidelson; its episode 3 on the history of data centers is recommended by Ben Gilbert.
F1 (2025 film) - movie - Ben's other carve-out; recommends seeing it in theaters even for non-F1 fans.

Companies

Google / Alphabet - Central subject: the company that invented the Transformer and DeepMind's parent, now racing to defend Search while building Gemini, TPUs, and Google Cloud.
OpenAI - Founded in 2015 by researchers poached from Google (notably Ilya Sutskever) with Elon Musk and Sam Altman backing; launched ChatGPT in November 2022, triggering Google's 'code red.'
DeepMind - London AI lab founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman in 2010; acquired by Google in 2014 for $550M after a bidding war with Facebook, Tesla, and Baidu.
Anthropic - Founded by Dario Amodei after leaving OpenAI; rumored plans to launch a chat interface reportedly pushed OpenAI to rush ChatGPT out the door first.
Microsoft - Invested $1B in OpenAI in 2018 (cash plus Azure credits), then $10B more in 2023; launched the AI-powered Bing in Feb 2023 that Satya said was meant to 'make Google dance.'
NVIDIA - Benefited from AlexNet's GPU breakthrough in 2012; Google's 40,000-GPU, $130M order in 2014 (on a $4B NVIDIA revenue base) is cited as an early tell of NVIDIA's future scale.
Meta / Facebook - Made an up-to-$800M offer for DeepMind that Demis rejected; later built FAIR under Yann LeCun; discussed as 'application only' in today's AI stack, no frontier model or cloud.
Waymo - Alphabet's self-driving subsidiary, born from the 2009 Project Chauffeur/Larry 1000 challenge; now doing more gross bookings than Lyft in San Francisco with roughly $10-15B invested to date.
Tesla - Elon Musk tried to buy DeepMind with Tesla stock and separately considered having Google buy Tesla for $3B; AlexNet's success helped convince Elon to push AI into Tesla's self-driving efforts.
Broadcom - Google's TPU hardware partner, handling chip production and TSMC interface; estimated ~50% margin versus NVIDIA's ~80%, a structural cost advantage for Google's token production.
Character.AI - Chatbot startup founded by Noam Shazeer after he left Google in 2021; Google paid roughly $2.7B in a 2024 licensing/hiring deal to bring Noam back to lead Gemini development.
Inflection AI - Founded by DeepMind co-founder Mustafa Suleyman and Reid Hoffman; later acqui-hired by Microsoft, with Mustafa becoming Microsoft's head of AI.

Techniques and frameworks

The Transformer / attention mechanism - 2017 Google Brain paper 'Attention Is All You Need' (173,000+ citations); lets a model attend to an entire text corpus at once rather than just nearby words, and parallelizes far better than LSTMs.
RLHF (reinforcement learning from human feedback) - The missing ingredient in Google's early chatbots (Meena, LaMDA/Bard) that let ChatGPT tune tone, appropriateness, and correctness; Bard's 2023 launch lacked comparable polish.
TPU quantization - Google's Tensor Processing Unit uses reduced numerical precision (e.g., rounding 4586.8272 to 4586.8) to fit far more calculations per chip, trading precision for throughput and enabling custom ASICs built in 15 months.
Distbelief's asynchronous distributed training - Jeff Dean's system trained neural nets across thousands of CPU cores without synchronizing parameters in real time, defying the field's synchronous-training consensus and still working.
Seven Powers framework - Hamilton Helmer's framework (scale economies, network economies, counter-positioning, switching costs, branding, cornered resource, process power) used to assess Google's AI moat: scale economies, branding, and cornered resource (Search distribution) are strong; counter-positioning and process power are largely absent.
Bull/bear/quintessence analysis - Acquired's standard closing framework, applied here to weigh Google's AI upside (full-stack assets, self-funding, distribution) against the downside (weaker AI monetization, market share erosion, loss of underdog goodwill).

Summary

Ben Gilbert and David Rosenthal close their three-part Google series by arguing that Google's AI story is the most striking case of the innovator's dilemma they have ever covered. Google invented nearly every ingredient of the modern AI stack, from George Herrick and Noam Shazeer's 2001 lunch-table theory that "compression equals understanding," to Google Brain's cat-recognition paper, to the acquisition of DeepMind, to the 2017 Transformer paper itself, yet by the mid-2010s had employed almost every major AI researcher in the world (Ilya Sutskever, Geoff Hinton, Demis Hassabis, Dario Amodei, and more) without turning that concentration of talent and technology into a shipped consumer AI product. The episode traces how that talent gradually scattered to OpenAI, Anthropic, and Tesla, driven substantially by Elon Musk's frustration after losing DeepMind to Larry Page's personal rapport with Demis Hassabis, and how OpenAI's accidental hit product, ChatGPT, blindsided Google in November 2022 despite Google having built comparable (if unsafe and unshipped) internal chatbots years earlier.

The hosts spend significant time on the infrastructure story that made all of this possible: Jeff Dean's Distbelief system defying consensus by training asynchronously across CPU cores, the 40,000-GPU order that gave NVIDIA an early signal of its future scale, and the TPU's 15-month emergency build cycle to keep Google's data center footprint from doubling. A long detour covers Waymo's two-decade path from a DARPA Grand Challenge team to a business the hosts argue could rival Google's own revenue, given how large the cost savings from reduced traffic fatalities could be relative to Waymo's roughly $10-15 billion in cumulative investment.

The core analytical payoff is Sundar Pichai's "code red" response to ChatGPT: merging the historically rival Google Brain and DeepMind teams, standardizing on a single Gemini model line, and shipping fast (an early Gemini within about six months) while deliberately avoiding cannibalization of Search's ad revenue with AI Overviews and AI Mode rather than a wholesale replacement. Both hosts credit this as an unusually disciplined incumbent response, closer to the mobile transition Google navigated well than to a Kodak-style failure.

The bull case for Google rests on its unique position as the only company holding all four AI pillars simultaneously (model, chip, cloud, and application) with self-sustaining funding, versus rivals who typically hold only one or two and depend on external capital. Google's TPU cost advantage over NVIDIA GPU buyers (roughly 50% versus 75-80% supplier gross margin) is framed as an unusually important edge in an era where, unlike prior tech cycles, being the low-cost producer of tokens may actually matter for who wins. The bear case centers on unproven AI monetization: Google earns roughly $400 per user per year from ads today, a number far short of what a paid AI subscription at scale would require, and applying the Seven Powers framework shows Google's AI moat (scale economies, branding, cornered-resource distribution) is real but thinner than the near-total moat it built in Search, with meaningful market share now split among several credible competitors for the first time in Google's history.

Notable Quotes

"Somebody put it to me in research that if you don't have a foundational frontier model or you don't have an AI chip, you might just be a commodity in the AI market. And Google is the only company that has both." - David Rosenthal

"The value creation is there in spades. The value capture mechanism is still TBD. Google's old value capture mechanism is one of the best in history. That's the issue at hand." - Ben Gilbert

"Once ChatGPT comes out, on a dime overnight, AI shifts from being a sustaining innovation to a disruptive innovation. It is now an existential threat." - David Rosenthal

"My quintessence when I boil it all down is just that this is the most fascinating example of the innovator's dilemma ever." - Ben Gilbert

"Is AI a good business to be in compared to Search? Search is a great business to be in. So far AI is not, but in the abstract... it should be." - Ben Gilbert