Google is the originator of the AI revolution—and now faces the innovator’s dilemma of its own making. The 2017 Transformer paper, published by Google Brain researchers, is the foundational architecture behind every major AI system today, including ChatGPT, Claude, and Gemini. Google not only invented the core technology but also possesses the most complete stack of AI assets in the world: a frontier model (Gemini), custom AI chips (TPUs), a major cloud platform (Google Cloud at $50B+ revenue), and the dominant distribution channel to users (Google Search, with ~90% market share). Yet the company must navigate the tension between protecting its enormously profitable search-and-ads business and aggressively pursuing a future where AI could displace that business.
Google’s Deep AI Roots
Larry Page always defined Google as an AI company, long before it was fashionable. In 2000, two years after founding, he said: “Artificial intelligence would be the ultimate version of Google.” His father was a machine learning PhD from the University of Michigan, and Page saw PageRank itself as a statistical/AI method.
In 2001, engineers Jeff Dean, Gnome Shazir, and George Heric began exploring language models based on the theory that compressing data is equivalent to understanding it. This led to “Phil” (Probabilistic Hierarchical Inferential Learner), a language model that by the mid-2000s consumed 15% of Google’s data center infrastructure and powered AdSense’s ability to understand third-party web pages—unlocking billions in new ad revenue.
Jeff Dean rearchitected Google Translate in 2007, parallelizing a language model that previously took 12 hours per sentence down to 100 milliseconds, demonstrating Google’s unmatched ability to scale distributed workloads across data centers.
The Talent Concentration and the Birth of Google Brain
By the mid-2010s, virtually every major AI researcher worked at Google, including Ilia Sutskever (later OpenAI chief scientist), Dario Amodei (later Anthropic founder), Andrej Karpathy (later Tesla AI lead), and the DeepMind founders. The analogy: it was as if IBM had hired every person who knew how to program at the dawn of computing.
Sebastian Thrun joined Google from Stanford in 2007, leading Street View and “Ground Truth” (rebuilding Google Maps from scratch using Street View imagery). He then proposed bringing AI academics to Google part-time, which led to recruiting Geoff Hinton—the “godfather of neural networks”—from the University of Toronto.
Google Brain launched in 2011 as the second project inside Google X, founded by Andrew Ng, Jeff Dean, and Greg Corrado. They built “Disbelief,” a distributed training system that ran asynchronously across CPU cores—counterintuitively working despite using stale data.
The “Cat Paper” (2012) was a landmark: a nine-layer neural network trained on 10 million unlabeled YouTube frames across 16,000 CPU cores learned to recognize cats without being told what a cat was. This directly enabled YouTube’s recommendation engine, content filtering, and copyright detection—transforming YouTube into the internet’s largest media property and setting the template for AI-driven feeds at Facebook, Instagram, and TikTok.
AlexNet, GPUs, and the DeepMind Acquisition
AlexNet (2012) was the “big bang moment for AI”: Geoff Hinton’s students Alex Krizhevsky and Ilia Sutskever trained a deep neural network on two off-the-shelf Nvidia GeForce GTX 580 gaming cards, achieving a 40% improvement in the ImageNet competition. This proved GPUs—not CPUs—were the hardware path forward for AI and set Nvidia on its trajectory to becoming the world’s most valuable company.
Google acquired the three-person DNN Research (Hinton, Krizhevsky, Sutskever) for $44 million in a hotel-room auction at Lake Tahoe, beating out Baidu, Microsoft, and DeepMind. Hinton took 40%, the students 30% each.
Google acquired DeepMind in 2014 for ~$500 million after a competitive process involving Facebook (which offered up to $800M) and Elon Musk (who offered Tesla stock). DeepMind’s founders chose Google because Larry Page shared their mission, Google Brain already handled product integration (leaving DeepMind free to pursue pure research), and Google offered unmatched compute infrastructure. An independent ethics board was established as a condition.
DeepMind’s AlphaGo defeated world champion Lee Sedol in 2016, including the famous “Move 37” that no human would have played. The match ran on four TPUs in Google Cloud.
TPUs: Google’s Custom AI Chip Advantage
Google built its first Tensor Processing Unit (TPU) in just 15 months (2015–2016) as a “hair on fire” response to the realization that rolling out speech recognition to all Android phones would require doubling Google’s data center footprint. The TPU used reduced-precision arithmetic (rounding numbers) to dramatically increase calculations per second for matrix multiplication.
The TPU was designed to fit into existing hard drive slots in server racks, avoiding physical rearchitecture. The project was developed in secret at a Google satellite office in Madison, Wisconsin, and wasn’t announced until Google I/O.
Google now has an estimated 2–3 million TPUs, compared to Nvidia’s ~4 million GPUs shipped last year. TPUs give Google a major cost advantage: while Nvidia charges ~80% gross margins (a ~5x markup), Google’s chip partner Broadcom charges ~50% margins (~2x markup), and chips represent over half the total cost of running an AI data center.
The Transformer Paper and Google’s Strategic Miss
“Attention Is All You Need” (2017) was published by eight Google Brain researchers, including Gnome Shazir. It introduced the Transformer architecture, which replaced sequential LSTM processing with parallelizable attention mechanisms—allowing models to consider all context simultaneously. The paper has been cited over 173,000 times, making it the 7th most-cited paper of the 21st century.
Google did build on the Transformer internally (BERT, MUM, and other models improved search quality), but did not treat it as a wholesale platform shift. All eight authors eventually left Google, including Gnome Shazir (who founded Character AI, which Google later reacquired for ~$2.7 billion).
The Transformer gave OpenAI its opening: after Elon Musk left OpenAI in 2018 (following a power struggle with Sam Altman), OpenAI pivoted to the Transformer architecture, released GPT-1 and GPT-2, and then partnered with Microsoft, which invested $1 billion in July 2018 at the Sun Valley Conference.
ChatGPT and Google’s Code Red
ChatGPT launched November 30, 2022, reaching 1 million users in under a week and 100 million by January 2023—the fastest product in history to do so. OpenAI didn’t plan for this; they threw up a paywall over a weekend to manage demand, having expected their business to be API licensing, not consumer chat.
Google had working chatbots years before ChatGPT: Gnome Shazir built “Mina” (and later “Lambda”) internally by ~2020–2021, but they lacked RLHF (reinforcement learning with human feedback) for safety and tone. Google’s AI Test Kitchen limited Lambda conversations to five turns to prevent the model from going off the rails.
Sundar Pichai issued a “Code Red” in December 2022, reclassifying AI from a sustaining innovation to a disruptive, existential threat. Google’s response was initially poor: Bard (launched February 2023) gave a factual error in its launch video, and the stock dropped 8%. The underlying Lambda model was clearly behind GPT-3.5 and then GPT-4.
Unifying Under Gemini and the Modern Era
In 2023, Sundar made two transformative decisions: (1) merge Google Brain and DeepMind into a single entity, Google DeepMind, led by Demis Hassabis; and (2) standardize all of Google on one model, Gemini, built jointly by Jeff Dean, Oriol Vinyals, and eventually Gnome Shazir.
Gemini launched in December 2023 (just six months after announcement at Google I.O.), with Gemini 1.5 (February 2024) introducing a 1 million token context window, and Gemini 2.5 Pro (March 2025) shipping at a pace comparable to Nvidia’s product cadence. Google now reports 450 million monthly Gemini users.
Google is running AI inference at massive scale: from 10 trillion tokens processed in April 2024 to nearly 1 quadrillion by mid-2025—a ~100x increase in 14 months. AI Overviews now appear on a large subset of Google searches.
Waymo: Google’s Other AI Moonshot
Sebastian Thrun’s Stanford team won the 2005 DARPA Grand Challenge using commodity sensors and machine learning (rather than custom hardware), pioneering the software-first approach to autonomy.
Project Chauffeur (later Waymo) launched inside Google X in 2009 after Larry Page challenged Thrun to find a technical reason it was impossible—Thrun realized there wasn’t one, he was just afraid. The team completed the “Larry 1000” (1,000 miles of difficult California roads) within 18 months.
Waymo took 11 years to reach commercial launch (October 2020 in Phoenix) because of the long tail of edge cases in self-driving. It didn’t use deep learning at all for its first 5+ years; convolutional neural nets were added in 2013–2014, and Transformer-based planning came in 2017.
Today Waymo operates in five US cities (Phoenix, San Francisco, LA, Austin, Atlanta), with hundreds of thousands of paid rides weekly, over 100 million autonomous miles driven, and a new data-driven study showing 91% fewer serious-injury-or-worse crashes versus human drivers. It’s expanding to Tokyo and more US cities. Total investment has been $10–15 billion—modest relative to the opportunity and small compared to Google’s monthly profits.
Google Cloud: The Strategic Distribution Layer
Google Cloud reached $50B+ annual revenue run rate (growing 30% YoY), making it the fastest-growing major cloud. It only became profitable in 2023 after hiring former Oracle president Thomas Kurian in 2018, who built a 10,000-person go-to-market organization and pivoted to enterprise needs.
Google Cloud is the only way external developers can access TPUs, giving Google a chip ecosystem play. Kubernetes (open-sourced by Google in 2018) became a multi-cloud strategy pillar, counter-positioning against AWS and Azure.
Google is the only AI company with all four pillars: frontier model, custom chips, hyperscale cloud, and massive consumer distribution. Every other major AI player has at most one or two of these.
The Bull Case
Distribution: Google Search remains the front door to the internet for ~90% of users. AI Overviews and AI Mode give Google the ability to funnel users into AI experiences while protecting the core franchise.
Self-sustaining funding: Google generates $140B in annual earnings—more than any tech company except Saudi Aramco. Unlike OpenAI, Anthropic, or Perplexity, Google’s AI efforts are funded by a cash-generating machine so large it returns excess capital to shareholders via buybacks and dividends even while investing heavily in AI capex.
Infrastructure moat: Google owns its own private dark fiber backbone (bought for pennies after the 2000 dot-com crash), connecting all data centers. No one else has this.
Cost advantage in chips: With TPUs at ~2x markup vs. Nvidia GPUs at ~5x markup, and chips being over half the cost of AI data centers, Google is the lowest-cost token producer in the world.
Multimodal data assets: YouTube is the only large-scale source of user-generated video (long and short form), enabling training for video AI models like VO3 and Genie 3 (a real-time world builder). Ben Thompson has argued the future internet is video, not text, and Google owns that future.
Personalized AI: Google has personalized data from Gmail, Maps, Docs, Chrome, and Android that no competitor can match, enabling deeply personalized AI products.
Waymo optionality: Waymo alone could become a Google-sized business through accident reduction savings ($420B+ annually in the US) plus ride-share market expansion and new use cases (elderly, blind, trucking).
The Bear Case
Value capture is unproven: Google makes ~$400 per user per year from search ads. It’s unclear anyone will pay $400/year for AI. The subscription model (150M Google One subscribers) is mostly at $5–10/month tiers; premium AI is $20/month and a small fraction.
AI doesn’t lend itself to ads yet: High-value search queries (travel, health, legal) are migrating to AI chat, but there’s no clear ad model for chat interfaces. These were among the most monetizable search queries.
No longer the obviously superior product: Google Search in 1998 was instantly better than AltaVista. In AI, Google was initially behind (Bard vs. ChatGPT) and is now arguably on par with several competitors—not dominant.
Market share will be lower: Google owns ~90% of search. In AI, steady-state market share is likely 25–50%, meaning even with equal per-user monetization, total revenue would be lower.
Incumbent disadvantage: The ecosystem no longer roots for Google the way it did in the startup era. Big tech is broadly out of favor, and startups have the “hearts and minds” narrative.
The innovator’s dilemma is real: Larry Page and Sergey Brin have said they’d “rather go bankrupt than lose at AI.” But if AI is less profitable than search, the tension between mission and shareholder returns becomes acute. Google is threading a delicate needle—protecting the core franchise while building a new one—and it’s unclear how long both can coexist.
Power Analysis (Scoped to AI Products)
Scale economies: The dominant power. Google amortizes model training costs across quadrillions of inference tokens—50x more tokens processed in one year than the prior year. No competitor comes close.
Cornered resource: Google Search’s ~90% market share and the data from Google’s product ecosystem (Gmail, Maps, YouTube, Android) are unmatched.
Branding: Net positive—most people trust Google more than “who knows AI companies,” though the advantage is smaller than in Google’s early days.
Switching costs: Currently low for AI products, but likely to increase as AI integrates with personal data (calendar, email, documents).
Network economies, counter-positioning, process power: Largely absent in AI today. Google is being counter-positioned, has no clear process advantage in producing breakthroughs, and AI products don’t yet exhibit network effects.
Quintessence
This is the most fascinating case study of the innovator’s dilemma in history. Google invented the technology, hired all the talent, built the chips, and has the best product—yet must choose between two futures: fulfilling its mission of organizing the world’s information (which demands aggressive AI adoption) or preserving the most profitable business in tech history. Sundar Pichai and Google’s leadership are threading this needle remarkably well—unifying teams, shipping products rapidly but not rashly, and protecting the core franchise while building a new one. But the tension is unresolved, and the next decade will determine whether Google’s unparalleled assets translate into dominance or whether the innovator’s dilemma claims its most consequential victim.
