Google Part III: The AI Company

I almost signed a lease on a so‑called studio with thin ceilings where you could hear the neighbor’s calls, which would’ve made him our accidental third co‑host. Let’s do a real show instead. Welcome to the Fall 2025 season of Acquired, where we break down great companies and their playbooks. Here’s the setup: imagine you run a wildly profitable, near‑monopoly business, then your own lab invents something that makes your core product look dated and you publish the recipe. Startups sprint to ship it. You can’t yet make the new thing as profitable as the old thing, so do you disrupt yourself or defend the cash machine? That’s Google right now. They birthed the transformer in 2017, power Gemini on their own cloud, and ship TPUs at scale. They still own the internet’s front‑door text box. So what’s the move? Today, the story of Google, the AI company.

Monopoly as in the government’s definition, yes. And one researcher told us if you do not have a frontier model or an AI chip, you’re likely a commodity; Google uniquely has both.

Quick housekeeping: join the email list and Slack at acquired.fm, we’re celebrating ten years with an open Zoom on October twentieth, 2025, and check out ACQ2 for our chat with Shopify’s Tobi Lütke on how AI changed his life. This is not investment advice.

Rewind a decade: nearly every big name in modern AI passed through Google—Hinton, Sutskever, Krizhevsky, Dario Amodei, Karpathy, Andrew Ng, Sebastian Thrun, Noam Shazeer, and the DeepMind founders—while Yann LeCun was the notable exception at Facebook.

It’s hard to trace any major lab’s origin without hitting Google along the way.

Larry Page always saw Google as an AI outfit, influenced by his father’s machine‑learning background. PageRank itself is statistical, and Larry’s early vision was a search engine that truly understands intent. Around 2000, Georges Harik told Ben Gomes and new hire Noam Shazeer over lunch that compression implies understanding, sparking a language‑model push.

That idea foreshadows today’s large language models as compressed, learnable representations of the world’s text, which can appear to reason even if they’re pattern learners at heart.

Noam and Georges went all‑in, with Sanjay Ghemawat’s blessing, and delivered Google’s “did you mean” correction, then a bigger model nicknamed Phil that powered AdSense page understanding. Jeff Dean wired it up fast, and the revenue impact was massive. By the mid‑two thousands Phil ate roughly fifteen percent of Google’s compute, but it proved language models fit the mission.

Cue the Jeff Dean legends, because he keeps showing up whenever something impossible needs to ship.

In 2007, Google Translate’s chief architect Franz Och won a DARPA contest with a giant n‑gram model that took twelve hours per sentence. Jeff parallelized the algorithm across Google’s CPUs and brought it to roughly a tenth of a second, making it the first big model to hit production at scale.

Those same modeling chops boost autocomplete and ad quality predictions, which flow straight into the bottom line.

That year Sebastian Thrun joined, launched Street View, and led Ground Truth to rebuild maps from imagery and dozens of data sources with a big human‑in‑the‑loop effort. He then recruited academics into Google part‑time and brought Jeff Hinton to campus; deep nets were fringe but Moore’s Law made them testable. Google X formed, then Google Brain with Andrew Ng, Greg Corrado, and Jeff Dean.

Dean’s DistBelief defied the era’s sync‑on‑one‑machine wisdom by training asynchronously across lots of CPUs. The “cat paper” followed: a nine‑layer net trained on unlabeled YouTube frames learned a cat concept with no labels, a TGIF moment that changed minds inside the company.

That unlocked YouTube’s content understanding and recommendations, plus safety and rights tools, and then spread to Facebook, Instagram, and TikTok. For feeds, the AI era quietly started around 2012.

Also in 2012, Hinton, Krizhevsky, and Ilya Sutskever ported their model to two Nvidia gaming GPUs bought off the shelf, used CUDA, and crushed ImageNet error. It validated deep nets and set Nvidia on its current trajectory.

Google soon bought their startup, DNN Research, via a live auction at the Lake Tahoe conference. Baidu, Microsoft, and Google bid, with DeepMind briefly in the mix before dropping out; the final price was forty‑four million, with Hinton receiving a larger share at the team’s urging. They joined Brain and the ROI was immediate.

Then came DeepMind. In 2014, Google acquired the secretive London lab founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Demis was a chess prodigy turned game designer turned neuroscience PhD; Shane helped popularize the idea of AGI. Their north star was to build general intelligence before worrying about products.

They raised a seed from Founders Fund after a Singularity Summit pitch, and Peter Thiel introduced Demis to Elon Musk. Demis pushed Elon to consider that AI risks would follow us to Mars, which flipped the bit for Elon, who invested and soon chased AI talent for Tesla.

Mark Zuckerberg, meanwhile, hired Yann LeCun to start FAIR and tried to buy DeepMind for as much as eight hundred million, but would not grant the governance independence Demis wanted. Elon offered Tesla stock. Larry Page learned about DeepMind on a flight while watching its Atari Breakout demo and clicked with Demis, offering the best cultural and technical fit.

Google already had Brain to ship product wins and the compute to fuel frontier research. Alan Eustace flew to London with Jeff Hinton strapped into a custom takeoff harness, and the deal closed at five hundred fifty million with an independent ethics board that later included Reid Hoffman. DeepMind quickly cut data center cooling energy by approximately forty percent and then shocked the Go world when AlphaGo won with creative play, including the famous Move 37.

The ripple effects were immediate. Elon was furious and, with Sam Altman, convened a 2015 Rosewood dinner to entice researchers out of Google. Most declined; the combination of pay, peers, and infrastructure at Google was unmatched.

OpenAI’s pitch was irresistible to some: do open research for humanity, publish freely, no product grind. Ilya Sutskever took the leap despite a rich counter from Google, a handful of top folks followed, and a one billion dollar pledge only turned into about one hundred thirty million collected, which covered high salaries while they chased DeepMind-style projects like Dota, Universe, and even a robot hand solving a Rubik’s Cube.

Back at Google, Alex Krizhevsky quietly stuck a single GPU in a closet, then Jeff Dean and John Giannandrea pushed for real scale: forty thousand Nvidia GPUs for roughly one hundred thirty million dollars, greenlit by Larry Page. That spend juiced Google’s products and hinted to Nvidia that neural nets were a business, but it also exposed a looming bill that led Google to build its own chip.

Google’s TPU bet was all about efficiency: lower numerical precision, a quick FPGA bridge, and a clever hard-drive form factor so data centers could slot them in fast. AlphaGo ran on early TPUs, and today Google operates millions of them, a parallel universe to Nvidia’s footprint that most people forget exists.

They also shipped TensorFlow so models could run across TPUs, GPUs, or CPUs, and Brain kept pushing the state of the art while driving real revenue wins.

Language at Google moved from recurrent nets and LSTMs, which kept context but didn’t parallelize well, to a new idea: give the model attention over the whole text at once. A researcher prototyped it, Noam Shazeer rewrote the code, it scaled beautifully, and the transformer was born.

Google used transformers in products like BERT and improved search quality, but treated it as a feature, not a platform shift. They let the paper, Attention Is All You Need, go public, which was wonderful for the world and paved the way for rivals as its authors soon left to start or join AI startups.

OpenAI, meanwhile, hit internal turmoil when Elon Musk demanded a merger with Tesla or he’d walk, then exited in early 2018 with his money. Necessity sharpened focus, and by mid-2018 OpenAI published GPT-1: pretrain on broad internet text, then fine-tune for tasks.

Reid Hoffman brokered a meeting with Satya Nadella in Sun Valley that led to a one billion dollar Microsoft deal with Azure credits and an exclusive license, housed in a capped-profit LP controlled by the nonprofit. Azure’s scale made Microsoft the perfect partner.

GPT-2 arrived as a promising API without a consumer front door, GPT-3 pushed quality much higher, and then GitHub Copilot landed in 2021 as the first major product built on it. Microsoft added another two billion dollars, and momentum kept building.

Markets turned in 2022 and Google looked slow, then OpenAI wrapped a simple chat UI around GPT-3.5 and accidentally launched the next big consumer product. Servers buckled, a quick paywall tempered demand, rumors of Anthropic’s chat sped timing, and it rocketed to one hundred million users in about two months.

Inside Google, Noam Shazeer had already built Meena and then LaMDA, but without RLHF and with big safety and legal risks, leadership kept them behind glass with short, nerfed conversations. The business model also resisted replacing links with direct answers.

Microsoft seized the moment with a new Bing powered by OpenAI and bragged it would make Google dance, which triggered a code red. Google rushed Bard out on LaMDA, botched the launch with a factual error, took an eight percent stock hit, swapped in PaLM, and still trailed GPT-4.

Sundar Pichai then made two big calls: merge Brain and DeepMind under Demis Hassabis, and unify around a single model named Gemini for everything inside and outside Google. Mustafa Suleyman later moved from Inflection to Microsoft, the Gemini brand subsumed Bard, scaling-law logic favored one giant model, and Sergey Brin returned to help.

There’s a parallel AI triumph in Alphabet: Waymo. The roots go back to the 2005 DARPA Grand Challenge, where Sebastian Thrun’s Stanford team won with off-the-shelf sensors and machine learning to fuse camera color and laser depth, a software-first approach that carried into Google X’s Project Chauffeur.

Larry Page set the “Larry 1000” gauntlet of hard routes, the team hit it quickly, and then spent years grinding through edge cases. Waymo spun out, aimed for robotaxis over driver assist, and launched paid driverless rides in Phoenix and then San Francisco, where it’s become a go-to option.

The experience feels different and safer, with kid seats, dogs, and private calls all feeling natural. Waymo reports approximately ninety-one percent fewer serious injuries than human drivers on comparable streets and has surpassed one hundred million miles without a safety driver.

Operations are heavy—charging, cleaning, sensor upkeep—but funding is deep and scale is improving. If it meaningfully reduces the roughly four hundred seventy billion dollars in annual U.S. crash costs, the upside is enormous.

Back to Google’s core AI push: Sundar’s 2023 mandate to unify teams and rally around Gemini set the course, with Jeff Dean and Oriol Vinyals joining DeepMind to build it and Sergey back on badge, all-in on one model for the company.

If Jeff Dean is on it, I’m sold.

Jeff and Noam are now co‑leading Gemini, which Google unveiled at I/O in May 2023 alongside AI Overviews in search, and they shifted into true AI speed after ChatGPT.

Running inference across a meaningful slice of Google search is staggering scale.

Gemini launched multimodal from the start, hit public preview by December 2023, expanded to a one‑million‑token context in early 2024, and rolled out the 2.0 and 2.5 waves plus a switchable AI mode on google.com in 2025.

They’re even auto‑opting some users into AI mode to gauge reactions, which is the golden goose experiment.

On top of that came NotebookLM and new video and image tools like VO and Genie with those real‑time world‑building demos.

If that demo is real, pairing it with a headset would feel like living inside a generated world.

Google now claims roughly four hundred fifty million monthly Gemini users, including lightweight clients, which is a rocket ride from zero.

I buy the momentum, though I wonder how they’re counting usage given how some rivals stretched similar stats.

Either way, it is a huge ramp.

And they’re doing it while revenue hits records and the core business keeps humming.

Quick sidebar: a judge found Google monopolized search but stopped short of heavy remedies, in part pointing to AI competition.

I think that’s flimsy since the AI challengers burn cash and are not self‑funding yet, and it is funny to imagine the domino chain where an ex‑Googler helps found a rival that inadvertently spares Google a breakup.

Snapshot: about three hundred seventy billion in revenue and one hundred forty billion in profit over the past year, and the market cap just crossed three trillion.

They hold roughly ninety five billion in cash while pouring money into AI data centers, doing buybacks, and even paying a dividend.

So the core is a cash geyser while they sprint on AI.

Gemini revenue is opaque, but Google One has more than one hundred fifty million subscribers, most on low tiers, with AI perks kicking in at the twenty‑dollar plan and the base growing about fifty percent year over year.

A strong bundle could make paid AI reach real scale.

Okay, tell us the cloud story.

Cloud started with an opinionated App Engine in 2008, then shifted to infrastructure with Compute Engine in 2012, but early on they lacked enterprise go‑to‑market and kept crown‑jewel tech to themselves.

Kubernetes made multicloud real, Thomas Kurian rebuilt enterprise sales, AI demand plus TPU supply did the rest, and now it is a profitable, more than fifty‑billion run‑rate business growing about thirty percent.

Cloud is the delivery channel for AI, and Google is rare in having an application, a frontier model, its own chips, and a hyperscale cloud.

That stack lets them run their own workloads, fill data centers, and push TPUs, maybe even into new clouds, to seed a broader ecosystem.

Let’s do bull and bear.

Bull case: Google still owns the default entry points, Gemini is competitive, and they are the only self‑funded model maker with private fiber, YouTube’s video moat, and abundant TPUs that avoid the Nvidia premium.

Chip economics matter: if Nvidia keeps around eighty percent margins while Google’s partner sits closer to fifty, that gap compounds when chips drive most of AI’s total cost.

That could make Google the low‑cost producer of tokens in a category that may settle near fifty percent gross margins.

They also hold unmatched personal data across Gmail, Maps, Docs, and more to build tailored assistants others cannot.

On monetization, chat carries richer intent than web search, so ads could price higher once the product shape lands and more time shifts into AI.

And Waymo could become a massive AI business on its own.

Bear case: value capture lags value creation, search monetizes at hundreds of dollars per user per year while a similar paid AI ARPU is unlikely, and ads in chat have not clicked yet.

The product is not obviously superior to rivals, share will be split, and high‑value search tasks like travel and health may shift to AI, draining premium queries.

Being the incumbent flips the vibe; startups get the goodwill Google once had, which makes every move a bit harder.

On seven powers for AI: huge scale economies, a trusted brand, and a cornered resource in distribution; switching costs are low today, network effects are thin, and challengers are counter‑positioning.

This is the fiercest innovator’s dilemma yet, and leadership has to choose how hard to lean into AI if it dents the search fountain; so far Sundar and team are threading the needle impressively.

Agreed; unifying research, focusing on one model, and shipping fast without panic is real execution, and the next decade will test it.

That wraps our Google series for now. Carve‑outs and a fun announcement: the NFL called us, and we’re hosting their Innovation Summit the Friday before the Super Bowl in San Francisco with a stream for everyone.

We’ll share details as February approaches; it’s going to be a blast.

It will be an incredible lead‑in to Sunday.

Carve‑out: I finally saw the F1 movie in theaters; it is beautiful and worth the big screen if you can.

Mine is the Glue Guys podcast; the Wright Thompson episode is a gem more people should hear.

Family gaming update: I bought a Steam Deck, my daughter got curious, learned to play Sea of Stars on vacation, and now we mostly play together; a Switch for co‑op may be next.

Huge thanks to the many researchers and authors who informed this series, including Stephen Levy, Parmy Olson, and Kate Met, plus folks across Google, TPUs, Waymo, investors, and friends who helped us stress‑test the economics.

Also check out StepChange on the history of data centers, our ACQ2 conversation with Toby Lutke, and come hang with us in Slack.

And join our tenth‑anniversary open Zoom on October twentieth at 4 p.m. Pacific time; details are in the notes. We’ll see you next time.