🎥 Watch the full episode here

📆 Published: October 6th 2025 🕒

Estimated Reading Time: 8 mins. Time saved: 320 mins! 🔥

🤖 Google's $4 Trillion Mistake? How the Inventor of Modern AI Got Caught Napping

The Acquired team are back with another mega episode as part of their complete Google Teardown. For those unfamiliar, Ben Gilbert and David Rosenthal produce some of the most rigorous company deep-dives in tech—think multi-hour explorations that go far beyond surface-level analysis.

While this episode isn't product-specific per se, it's an essential case study for anyone trying to understand one of the biggest shifts happening in product and tech right now: how organizational structure, incentives, and culture can trump technical superiority. Don't have time for a 4hr+ podcast? Don't worry, I've got you!

The greatest irony in technology history might be unfolding right now. Google invented the Transformer—the foundational architecture powering ChatGPT, Claude, and every frontier AI model reshaping our world. They assembled the densest concentration of AI talent ever seen. They built custom silicon, deployed it at planetary scale, and generated hundreds of billions in AI-driven revenue years before "AI" became a boardroom buzzword.

And yet, when ChatGPT launched in November 2022, Google was caught completely flat-footed.

This is the story of how the company that created modern artificial intelligence now faces an existential innovator's dilemma: protect $140 billion in annual search profits, or cannibalise the core business to win the AI era they invented.

🧬 The Original Sin: When Compression Became Intelligence

The seeds of Google's AI dominance—and its current predicament—were planted in a micro kitchen in 2000. George Heric, one of Google's first ten employees with a machine learning PhD from Michigan, casually mentioned to colleagues that compressing data is technically equivalent to understanding it. The logic was elegant: if you can shrink information, store it, and perfectly recreate it later, you must understand what it means.

Gnome Shazir, a new hire, latched onto this idea. Despite widespread internal skepticism ("a large number of people thought it was a really bad thing"), they built Phil—the Probabilistic Hierarchical Inferential Learner. This early language model powered Google's "Did you mean?" feature, then became the engine behind AdSense's content matching. By the mid-2000s, Phil consumed 15% of Google's entire data centre infrastructure.

The business impact was immediate and massive. AdSense generated billions overnight by matching Google's ad corpus to third-party web pages. Language models weren't just working—they were printing money.

Key Takeaways:

• Google's AI journey began with language models in 2001, two decades before ChatGPT

• Early LMs like Phil drove core revenue products (AdSense, search quality) from day one

• The compression-as-understanding thesis foreshadowed modern LLMs compressing world knowledge into parameter weights

⚡ The Infrastructure Advantage: Jeff Dean's Parallel Universe

When Franz Och's translation model won DARPA's 2006 challenge, there was one problem: it took 12 hours to translate a single sentence. The model was trained on two trillion words and designed for a competition where you had Monday to Friday to submit results—not for a production system serving millions.

Enter Jeff Dean, Google's legendary infrastructure architect. Dean rearchitected the algorithm to run in parallel across Google's distributed infrastructure, reducing translation time from 12 hours to 100 milliseconds. They shipped it in Google Translate immediately.

This moment crystallised Google's structural advantage: the ability to take cutting-edge research and deploy it at planetary scale. While academia published papers, Google had Jeff Dean, distributed systems expertise, and data centres spanning continents. The first large language model in production wasn't at a research lab—it was powering Google products in 2007.

By 2011, when Andrew Ng and Jeff Dean launched Google Brain, they built DistBelief—a system that ran neural networks asynchronously across thousands of machines. Conventional wisdom said this couldn't work; you needed synchronous, tightly-coupled compute. DistBelief proved otherwise, enabling the famous "cat paper" that trained a nine-layer neural network on 16,000 CPU cores to recognise cats in YouTube videos without labels.

Key Takeaways:

• Google's infrastructure moat enabled production deployment of AI years before competitors

• The cat paper (2011) proved unsupervised learning at scale, unlocking YouTube recommendations and hundreds of billions in revenue

• DistBelief's asynchronous architecture was heretical—and it worked

🧠 The Talent Vacuum: How Google Became the AI Ivy League

By 2014, Google had assembled an unprecedented roster: Ilya Sutskever, Jeff Hinton, Alex Krizhevsky (the AlexNet team), Demis Hassabis, Shane Legg, Mustafa Suleyman (DeepMind), Dario Amodei, Andrej Karpathy, Andrew Ng, Sebastian Thrun, and Gnome Shazir. The only notable AI researcher not at Google was Yann LeCun at Facebook.

This wasn't luck. It was strategy. Sebastian Thrun pioneered the model: bring AI professors in part-time, let them keep academic posts, pay them well, and give them Google-scale problems. When Jeff Hinton gave a 2007 tech talk at Google on deep learning, it catalysed everything. Google hired him as a "summer intern" in 2011—at age 60—to work around employment policies.

The $44 million acquisition of DNN Research (Hinton, Sutskever, Krizhevsky) in 2012 was run as an auction from a hotel room at a casino in Lake Tahoe. Four bidders: Baidu, Microsoft, Google, and a scrappy London startup called DeepMind that had to drop out because they had no money. The researchers chose Google even though Baidu bid higher. They wanted the infrastructure, the talent density, and the mission.

Then came DeepMind. Google paid $550 million in 2014 for a company with no products, vague website copy about "simulations, e-commerce, and games," and a mission to "solve intelligence." Facebook offered $800 million. Elon Musk offered Tesla stock (which would've been worth ~$40 billion today). DeepMind chose Google because Larry Page got it. He didn't need them to build products—Google Brain was already doing that. DeepMind could stay in London, publish research, and chase AGI.

Key Takeaways:

• Google's talent strategy: hire the best, pay well, provide infrastructure, and let them research

• The DeepMind acquisition ($550M) may be worth $500B+ today—rivalling Instagram and YouTube as greatest acquisitions ever

• By 2015, leaving Google for AI research seemed irrational—until OpenAI changed the game

💎 The Hardware Bet: TPUs and the $130 Million Gamble

When Alex Krizhevsky arrived at Google in 2013, he was shocked: everything ran on CPUs. He bought a GPU from a local electronics store, stuck it in a closet, and started training models. It worked so well that by spring 2014, Jeff Dean and John Giannandrea planned to order 40,000 Nvidia GPUs for $130 million.

Finance wanted to kill it. Larry Page personally approved it. "The future of Google is deep learning."

But Google didn't stop there. When they rolled out speech recognition on Nexus phones, Jeff Dean calculated that supporting it across all Android devices would require doubling Google's entire data centre footprint. The alternative: build custom chips.

Enter the Tensor Processing Unit (TPU). Designed, verified, built, and deployed in 15 months, the TPU used reduced computational precision and fit into the form factor of a hard drive so it could slot into existing server racks. The project was done in Madison, Wisconsin, kept secret for a year, and deployed in time for the AlphaGo match in 2016.

Today, Google has an estimated 2-3 million TPUs. For context, Nvidia shipped ~4 million GPUs last year. Google operates an almost-Nvidia-scale chip operation internally—a fact often overlooked in "Nvidia dominance" narratives.

Key Takeaways:

• Google's $130M GPU order in 2014 signalled to Nvidia that enterprise AI was real

• TPUs gave Google cost control, performance optimisation, and independence from Nvidia

• Custom silicon is a moat: if you don't have a frontier model or AI chips, you're a commodity

🔮 The Transformer: Eight Researchers, One Paper, Infinite Consequences

In 2017, eight Google Brain researchers published "Attention Is All You Need." The paper introduced the Transformer architecture, replacing recurrent neural networks and LSTMs with a ‘parallelisable’ attention mechanism that could process entire text sequences at once.

The insight was elegant: instead of processing text sequentially, give the model attention to the entire corpus. This mirrored how human translators work—read the whole passage, understand context, then translate. It was computationally expensive but infinitely parallelisable, perfect for Google's infrastructure.

Gnome Shazir, the same engineer who built Phil in 2001, joined the project and rewrote the codebase. Suddenly, it worked—and it scaled beautifully. The bigger the model, the better the results.

Google's response? "Cool, this is the next iteration of our language model work." They built BERT and integrated Transformers into search quality. Meaningful improvements, but incremental.

Meanwhile, the rest of the world saw something else: a technology platform shift. OpenAI, Anthropic, and every AI lab built on Transformers. ChatGPT, Claude, Gemini—all Transformer-based. Google invented the architecture powering the entire AI revolution and treated it as a product feature, not a paradigm shift.

Key Takeaways:

• The Transformer's elegance was its simplicity: minimal architecture, maximum scalability

• Google integrated Transformers into products (BERT, search) but didn't recognise the platform shift

• Publishing the paper openly was scientifically noble—and strategically catastrophic

🚪 The OpenAI Insurgency: When the Talent Walked Out

Summer 2015. Rosewood Hotel, Sand Hill Road. Elon Musk and Sam Altman host a dinner for AI researchers. The pitch: leave Google, start a nonprofit AI lab, publish openly, work for humanity instead of ad revenue.

The response from most attendees: "Why would we leave? We're paid millions, we keep our academic posts, we have the best infrastructure and colleagues in the world."

Except one person was intrigued: Ilya Sutskever. When Ilya said yes—turning down a counter-offer from Jeff Dean personally—others followed. Seven researchers left to found OpenAI with a $1 billion pledge (only $130M actually collected initially).

For the first few years, OpenAI mimicked DeepMind: play games (Dota 2), do research, publish papers. No single big thing. Then they got access to the Transformer paper.

The irony is exquisite. Google published the Transformer openly, consistent with research norms. OpenAI took that architecture, scaled it relentlessly, and built GPT. By GPT-3, they had something magical. By ChatGPT, they had a consumer phenomenon that made "AI" a household term overnight.

Google's talent raid backfired. Instead of slowing Google down, OpenAI gave ex-Googlers the platform to out-execute their former employer.

Key Takeaways:

• OpenAI's founding was a direct response to Google's AI dominance and DeepMind acquisition

• The nonprofit model attracted talent who wanted mission over money—until the mission required billions

• Google's open research culture armed its competitors with the Transformer

🚨 Code Red: The ChatGPT Shock

November 30, 2022. ChatGPT launches. Within days, it's a cultural phenomenon. Within weeks, Google declares "code red."

The internal scramble was real. Google had LaMDA, a conversational AI that engineer Blake Lemoine famously (and incorrectly) claimed was sentient. They had BERT powering search. They had models. But they didn't have a consumer product that captured imagination.

Why? The innovator's dilemma in its purest form. Every AI-powered search interaction potentially cannibalises a high-margin search ad. Launching a ChatGPT competitor means risking the $140 billion cash cow. OpenAI, with no search revenue to protect, had no such constraint.

Google's response: unify DeepMind and Google Brain under the Gemini brand, accelerate product launches, and lean into the "we invented this" narrative. Gemini's initial launch was rocky (remember the historically inaccurate image generation controversy?), but subsequent iterations showed Google's technical chops remained intact.

The question isn't capability—it's strategy. Can Google disrupt itself before someone else does?

Key Takeaways:

• ChatGPT's launch was Google's "iPhone moment"—a product that redefined user expectations overnight

• Google's technical capabilities were never in doubt; organisational willingness to cannibalise search was

• The DeepMind/Brain merger signalled Google was taking the threat seriously

📊 The Bull and Bear Case: What Happens Next?

Bull Case: Google has everything needed to dominate AI. Frontier models (Gemini), custom chips (TPUs), cloud infrastructure at scale, the best talent bench, oceans of data, and $140B in annual profits to fund R&D. Search remains the front door to the internet. Waymo is operationally ahead in robotaxis. Google Cloud is positioned as the AI infrastructure play. The innovator's dilemma is real, but so is Google's execution muscle.

Bear Case: Organisational inertia, revenue protection instincts, and internal complexity will hamstring Google. They'll integrate AI into existing products but won't create the next platform. OpenAI, Anthropic, and startups unburdened by legacy revenue will move faster. Google invented the future and will watch others monetise it—just like Xerox PARC invented the GUI and Apple commercialised it.

The truth? Probably somewhere in between. Google will remain a massive AI player, but the question is whether they'll lead or merely participate in the era they created.

Key Takeaways:

• Google's assets (talent, infrastructure, capital) are unmatched, but culture and incentives matter more

• The innovator's dilemma isn't theoretical—it's Google's daily reality

• Winning requires creating autonomous teams insulated from search revenue concerns

⚖️ The Verdict: Inventing the Future Isn't Enough

Google's AI story is a an incredible case in both execution and missed opportunity. They built the talent, the infrastructure, the models, and the architecture that powers modern AI. They deployed it at scale years before "AI" was a buzzword, generating hundreds of billions in revenue.

And yet, when the consumer AI moment arrived, they were caught off guard.

The lesson for product leaders and execs: technical superiority doesn't guarantee market leadership. Organisational design, incentive structures, and willingness to disrupt yourself matter as much as R&D budgets. Google invented the Transformer, but OpenAI made it a household name.

The next chapter is unwritten. Google has the assets to win. The question is whether they have the courage to risk the present for the future—or whether protecting $140 billion in search profits will prove too tempting to resist.

The innovator's dilemma isn't a thought experiment. It's Google's reality. And the clock is ticking.

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know whether you liked this separate Pod Shot, or whether you’d rather stick with your usual programming….. 🚀👋

Alastair 🍽️.