Google must turn AI invention into a new business
Google's problem is not that it missed artificial intelligence. It is that Google helped invent the future while operating the most profitable interface to the present.
In 2017, a team of Google researchers introduced the Transformer, the architecture that became foundational to modern large language models. Google subsequently produced BERT, LaMDA, PaLM, Gemini, custom AI chips, and some of the most consequential scientific systems in the field. Yet OpenAI, not Google, turned general-purpose conversation into the defining consumer AI product of the early generative era.
Jeff Dean's departure makes that contradiction harder to dismiss. Dean spent 27 years at Google, co-founded Google Brain, and became one of the company's most important technical leaders. His exit does not establish that Google's AI organization is collapsing. It does show how much more valuable autonomy has become to researchers who can raise capital and buy compute outside a large company.
Google can build frontier AI. The harder job is redesigning Search, advertising, and its own decision-making before an external AI interface becomes the primary gateway to commercial intent.
Google is not yet Kodak, and its current financial position makes near-term collapse implausible. Its dilemma still resembles Kodak's. Google must accept lower certainty, higher compute costs, and possible pressure on a growing profit engine if it wants to control the next interface. Treating AI only as a feature that protects Search would leave the layer above Search open to competitors. Advertising cash flow gives Google the means to fund the transition, if management is willing to use it that way.
Jeff Dean's departure is a signal, not a verdict
Jeff Dean joined Google in 1999, helped create systems that supported the company's early scale, co-founded Google Brain in 2011, and later served as Google's chief scientist. His career connects Google's original search infrastructure to its modern AI ambitions.
In August 2026, Dean left that role to form Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The company is intended to automate experimental loops in science and engineering. Alphabet participated in its financing and entered a long-term cloud and compute partnership, which makes the departure more nuanced than a clean break. Google may still benefit as investor and infrastructure provider.
The arrangement softens the rupture, but it does not remove the institutional message. Researchers with unusual status, resources, and internal access decided that a new organization was a better vehicle for their next work. Dean's departure also arrived near other high-profile exits. Noam Shazeer, a Transformer co-author and Gemini leader, left for OpenAI. John Jumper, the Nobel-winning AlphaFold co-creator, left Google DeepMind for Anthropic. Axios reported internal concerns about morale and model delays. Google disputed that morale caused model shortfalls and said AI attrition had improved year over year.
Public reporting identifies a small number of unusually important departures, not a mass exodus. Google says its broader hiring acceptance and retention remain strong. The concentration still matters. These are people who helped define Transformers, Gemini, and AlphaFold, and frontier teams depend heavily on trust and tacit knowledge. Alphabet preserved a commercial relationship with Dean through its investment in Discovery Loop, but it lost the employee relationship.
Dean's exit is an early-warning indicator. Repeated departures of technical leaders, longer release delays, or a widening gap between research and products would turn it into evidence of a broader problem. One departure cannot do that on its own.
Google invented the architecture, but not the defining product
The familiar claim that Google "invented the Transformer" is broadly correct but incomplete. Eight researchers working at Google authored Attention Is All You Need in 2017. The paper replaced recurrent processing with an attention-based architecture that trained more efficiently in parallel. That design became the base layer for much of the subsequent large-model boom.
Google continued to advance and deploy the technology:
- BERT, introduced in 2018, improved language understanding and was incorporated into Google Search.
- LaMDA, publicly demonstrated in 2021, showed open-ended conversational capabilities built on Transformers.
- PaLM extended scale and generality, while Google's TPUs supplied specialized training infrastructure.
- Google DeepMind later consolidated major research efforts and developed Gemini as a multimodal model family.
Google kept investing in AI. It excelled at research, infrastructure, and improvements inside existing products, but moved more slowly when the technology called for a new consumer interface.
ChatGPT's breakthrough in late 2022 was partly a product breakthrough. It placed a general-purpose model behind one simple conversational box, opened it to the public, and let usage reveal what the technology could become. Google had comparable ingredients but approached public release through a higher-liability lens. Bard was announced in February 2023 as an experiment powered initially by a lightweight version of LaMDA, after ChatGPT had already made conversational AI a global category.
During a platform transition, inventing the enabling technology does not guarantee ownership of the interface, the developer ecosystem, or the user's starting point. Xerox PARC helped create foundational personal-computing ideas without capturing the PC platform. Kodak built important digital imaging technology without preserving the economics of film. Google faces the same separation between invention and capture. The largest value may go to the company that owns the habit through which people ask, decide, and act.
Search advertising creates organizational gravity
The phrase "Google's AdWords cash cow" captures the basic incentive but uses an older brand name. AdWords became Google Ads in 2018. The economically important system is broader: Search and other advertising, YouTube advertising, and network advertising together remain Alphabet's dominant revenue source.
Alphabet's second-quarter 2026 filing shows why the incumbent business exerts so much gravity:
| Q2 2026 revenue | Amount | Share of Alphabet revenue | Year-over-year growth |
|---|---|---|---|
| Google Search and other | $63.3 billion | 52.8% | 16.8% |
| Total Google advertising | $81.6 billion | 68.1% | 14.4% |
| Google Cloud | $24.8 billion | 20.7% | 81.8% |
| Alphabet total | $119.8 billion | 100% | 24.2% |
Search is still growing, which makes the transition harder. Replacing a business already in visible decline would be easier. Every quarter of strong Search growth raises the internal opportunity cost of a product decision that could reduce ad inventory, lower click-through, or make monetization less predictable.
Generative answers change several parts of the unit economics.
Inference adds cost. In 2023, Alphabet chair John Hennessy said an exchange with a large language model could cost roughly ten times as much as a conventional keyword search. That estimate is too old to use as a current cost ratio. Hardware, model architecture, caching, and serving efficiency have improved rapidly. In July 2026, Sundar Pichai said the cost of serving AI Mode had fallen to its lowest level since launch. Google does not disclose a comparable current cost per query.
A synthesized answer can reduce the need to click a list of links. Traditional search creates many measurable surfaces for ads and auctions. A single answer, and especially an agent that completes a task, concentrates the interaction. Google must invent advertising and transaction formats that remain useful without corrupting trust in the answer.
AI also changes the success metric. Search historically monetizes queries and clicks. An agent is judged by completed outcomes. If a system books the trip, chooses the software, or purchases the product, the most valuable commercial event may occur after several hidden reasoning steps rather than beside ten blue links.
The revenue mix creates an obvious incentive for caution, but public evidence cannot show that executives deliberately suppressed large models to protect advertising. Reliability, legal exposure, brand risk, and safety standards also mattered. A company that derived 68% of quarterly revenue from advertising had more to disrupt than a startup with no installed profit pool. The available record does not let us assign a precise weight to each cause.
Did Google really hold generative AI back?
The record contradicts the claim that Google stopped Transformer research or refused to deploy AI. It funded successive model families, built TPUs, applied BERT to Search, published research, and integrated machine learning across its products.
There is evidence of productization delay. Reuters reported that LaMDA researchers Noam Shazeer and Daniel De Freitas pushed Google to release a conversational product before leaving in 2021, and that management declined because the system did not meet the company's standards for safety and fairness. Google publicly demonstrated LaMDA in 2021 but did not open Bard broadly until March 2023.
That episode shows that Google possessed conversational technology and chose a cautious public-release path. It does not establish advertising protection as the decisive reason.
A startup could gain attention from a surprising model and survive visible errors. Google Search had a global trust position to defend, and an incorrect generated answer could look like a failure of the information system itself.
The economics also differed. A startup gained users by increasing model usage. Google risked replacing a cheap, proven, auction-supported query with a more expensive answer whose monetization was initially unclear.
Google's approval chain was longer. A frontier model had to cross research, product, safety, legal, policy, advertising, and brand boundaries before reaching users. Those checks protect a global platform, but they slow public learning.
Each source of caution was rational on its own. Together they delayed public learning. OpenAI accumulated consumer habit, attracted developers, and forced Google to respond on a timetable it did not choose.
The pattern fits the innovator's dilemma without requiring a story about managerial stupidity. Google optimized against real liabilities. OpenAI optimized for learning speed. Once the new interface became credible, speed became more valuable than Google's risk model had anticipated.
The Kodak analogy fits, but not in the popular way
The simplified Kodak story says the company invented the digital camera, hid it to protect film, and was destroyed by its own cowardice. The historical record is more instructive.
Kodak engineer Steven Sasson built an early digital camera prototype in 1975. Kodak later invested in digital research, introduced a commercial digital SLR in 1991, manufactured Apple's QuickTake, and held a leading share of the US digital-camera market in 2004 and 2005. Its digital business struggled to reproduce the economics of film, and Kodak filed for bankruptcy protection in 2012.
Kodak's failure was business-model migration. Its highly profitable film and chemical system resembled a razor-and-blades model. Digital cameras removed recurring film consumption, lowered barriers to entry, compressed hardware margins, and shifted value toward electronics, software, and eventually smartphones. Kodak participated in the new technology without building a durable new profit engine around it.
That is where the comparison with Google becomes useful:
| Kodak | |
|---|---|
| Invented an important version of the disruptive technology | Invented the Transformer architecture with a Google research team |
| Continued investing in the new field | Continued building BERT, LaMDA, PaLM, Gemini, TPUs, and AI products |
| Depended on exceptional economics from film consumption | Depends heavily on advertising attached to search and attention |
| Entered digital products but struggled to replace the old profit pool | Has entered generative AI but still must prove agent-scale monetization |
| Lost value as the user workflow moved away from film | Could lose value if user intent begins in an external agent rather than Search |
Google has options that Kodak lacked. It can place AI directly inside Search; Kodak could not turn chemical film into a digital consumable. Google also owns global distribution through Search, Android, Chrome, YouTube, Workspace, and Maps, plus model infrastructure through TPUs and Google Cloud. Its core business is growing, its balance sheet can finance enormous compute investment, and AI itself is accelerating Cloud revenue.
Kodak does not predict Google's fate. It shows why technical participation is insufficient when a new technology changes where profit is captured.
The Nokia analogy is weaker but still useful at the interface layer. Nokia lost control as value shifted from handset engineering toward smartphone operating systems and app ecosystems. Google already controls Android, so it is not starting from Nokia's position. But an AI agent could become an operating layer above the browser and mobile OS. If users ask another company's agent to choose services, the underlying Google distribution may become less strategically visible.
Google is behind in some races, but it is not out of the race
Claims that Google has simply "fallen behind" compress several different contests into one phrase.
On frontier model quality, leadership is unstable. Stanford's 2026 AI Index showed the leading public chatbot scores tightly clustered in March 2026: Anthropic at 1503, xAI at 1495, Google at 1494, and OpenAI at 1481 on the cited Arena measure. Such leaderboards change quickly, depend on evaluation design, and do not measure product reliability or economics. They show that Google was competitive, not dominant.
On release cadence and talent, the evidence is less comfortable. Axios reported delays around Gemini 3.5 Pro and concerns from current and former employees. The departures of Dean, Shazeer, and Jumper intensify the issue. Google disputes a broad morale explanation and argues that model leadership naturally leapfrogs among labs. Both can be true: a company can remain near the frontier while developing organizational problems that surface first as slower releases and selective exits.
On distribution and monetization, Google remains unusually strong. In its July 2026 earnings remarks, management reported 950 million monthly Gemini app users, more than one billion monthly users for AI Mode, and roughly 22 billion tokens processed each minute through its model APIs. It also reported that Google Cloud revenue grew 82% year over year to $24.8 billion and that Cloud backlog reached $514 billion. The usage and backlog figures are company disclosures rather than independently audited adoption measures, but the Cloud revenue appears in Alphabet's regulatory filing.
Google also controls the full technical stack: chips, data centers, models, APIs, consumer surfaces, enterprise software, and advertising demand. Few competitors can subsidize consumer inference, sell the same capability through a cloud platform, and distribute it to billions of existing users.
That advantage has a corresponding weakness. The full stack is also a coordination problem. A model release can affect Search quality, publisher traffic, ad auctions, cloud customers, Android partners, regulators, and the corporate brand at once. Google has more assets than a frontier startup, but also more internal constituencies capable of slowing a decision.
Google remains technologically competitive, but invention and distribution no longer guarantee leadership. It has considerable power and less room for organizational delay.
Losing the gateway to intent matters more than a benchmark loss
Search is valuable because it sits at the moment when a user declares intent. Some intentions are informational. Others lead to purchases, travel, local services, software selection, entertainment, or professional work. Advertising converts that stream of intent into auctions and measurable commercial outcomes.
Generative AI can alter the path in three stages:
- AI as a Search feature. Google synthesizes answers within Search and preserves the starting point. This may change page economics, but the intent still enters Google's system.
- AI as a destination. Users begin with ChatGPT, Claude, or another assistant, which decides when to retrieve information. Google may become an unseen source, an ad supplier, or not participate at all.
- AI as an agent. The system compares options and completes actions. The owner of the agent can control recommendation, routing, payment, and the customer relationship. Traditional search becomes one tool inside a larger workflow.
The agent stage carries the greatest risk. Google does not need to lose most navigational or simple factual searches to suffer relative decline. It could retain enormous traffic while losing a smaller set of high-value commercial and professional tasks to agents. Revenue quality may deteriorate before query volume does.
This also explains why benchmark leadership is not the decisive metric. A model can rank second or fifth and still win if it becomes the default place where users begin tasks. Conversely, the best model can fail commercially if it lacks distribution, trust, or a sustainable route to outcomes.
Google's defense should be measured at the intent layer as well as the answer layer. Useful indicators would include:
- the share of high-value tasks that begin and finish in Gemini or AI Mode;
- revenue and gross profit per successfully completed task instead of per query;
- inference cost per satisfied user outcome;
- advertiser and merchant conversion rates inside agentic journeys;
- retention of frontier technical leaders and the time from research milestone to public product;
- the share of external AI workloads won by Google Cloud, including workloads whose consumer interface belongs to a competitor.
If Google wins those measures, fewer traditional link clicks need not mean strategic decline. If it protects query volume while losing task initiation and completion, reported Search strength could become a lagging indicator.
Three futures for Google
Google's next phase is unlikely to resemble a single dramatic collapse. Several outcomes can coexist for years before one becomes dominant.
1. Search becomes the agent
In the strongest outcome, Google willingly changes Search from a page of ranked results into a trusted system that researches, recommends, and acts. It develops monetization around qualified outcomes, transactions, subscriptions, and useful sponsored options while preserving a visible boundary between paid influence and model judgment.
The cash cow becomes transition capital. Search advertising funds compute, Gemini gains product autonomy, Google Cloud captures external demand, and the company accepts temporary pressure on revenue per query to retain control of intent. Google's distribution then becomes an overwhelming advantage rather than an asset to defend.
2. Google remains profitable but loses the primary interface
In the middle outcome, Search, YouTube, and Cloud continue to grow, but another assistant becomes the preferred starting point for complex work. Google supplies infrastructure, models, maps, video, or retrieval behind the scenes while a competitor owns the user relationship.
This would not look like Kodak's bankruptcy. It would look like relative strategic demotion: an exceptionally profitable company supplying components to a layer it no longer controls. Cloud growth could partially offset weaker interface power, but bargaining power and data advantages would migrate toward the agent owner.
3. The cash cow becomes a cage
In the weakest outcome, Google repeatedly calibrates AI products to preserve existing search economics. Decision cycles remain slow, elite researchers continue to leave, and users shift their most valuable tasks elsewhere. Advertising stays strong long enough to disguise the loss of future relevance, then weakens after habits and developer ecosystems have already moved.
That is the genuinely Kodak-like path. It requires more than missing one model release. It requires a repeated inability to convert invention into a new profit system.
The first two outcomes appear more plausible than the third. Alphabet's revenue growth, AI adoption, infrastructure, and distribution make extinction rhetoric analytically weak. Scale does not remove the choice facing management. It makes an early decision more expensive.
Google should measure whether AI wins user intent even when it reduces conventional result-page activity. That is the test of deliberate cannibalization.
It also needs commercial systems for agent-completed tasks that do not turn recommendations into disguised ads. Frontier teams need clear product ownership, release-speed targets, and enough autonomy to learn in public without discarding safety controls. Gemini, Search, Android, Chrome, Workspace, Maps, YouTube, TPUs, and Cloud should reinforce one another rather than defend separate internal mandates.
Jeff Dean's exit belongs in the same operating review. Management needs to understand why exceptional builders prefer a new institution even when Alphabet remains their investor and compute partner.
Google is unlikely to disappear on Kodak's timetable. Companies of this scale usually lose the most valuable layer before they lose revenue. Google's fate turns on whether it can make traditional Search less central while Search is still strong.
The Transformer changed the interface and economics of Search. Google's organizational response will decide whether it leads that change or finances someone else's.
Sources
- Google Research, Attention Is All You Need, 2017.
- Alphabet, Quarterly report for the period ended June 30, 2026, filed July 23, 2026, and Q2 2026 earnings remarks, July 2026.
- Axios, Google's AI leadership shuffle, August 6, 2026, and Discovery Loop financing and cloud partnership, August 2026.
- Reuters reporting on Google's cautious path from LaMDA to a public chatbot, March 2023.
- Stanford Institute for Human-Centered AI, AI Index Report 2026, April 2026.
- Willy Shih, Kodak and the Digital Revolution, Business History Review, 2025.
Financial shares and growth rates are calculated from Alphabet's filing. User counts and API usage are company disclosures. Leaderboard results are snapshots. The discussion of organizational incentives is an inference from public evidence.