For twenty years, Google had the perfect business.
The beauty of its business model was that it did not have to spend.
One more search cost almost nothing. No extra factory and no new machine to replace every few years.
Then last quarter, Sundar Pichai, the man behind the machine, decided selling answers is no longer fun.
For the very first time since its founding, after 87 consecutive quarters of cash generation, Alphabet has a negative cash flow.
Pichai placed a $200 billion bet on building the shovels for AI: chips, servers, and data centers. And says the spending will rise to an even more stunning level again next year.
Immediately, this triggered a sell-off.
But the deeper problem is not the stock price. It’s that Google may be changing the economics of its entire business. I’ll answer these questions in this one:
How did Google redeem itself from being an AI underdog?
What is Google becoming?
How real is Gemini’s demand?
Is the overall AI demand even real?
What to watch next?
Let’s start with a little bit of drama.
After three years of being an AI underdog
Three years ago, Sundar Pichai put a “Code Red” on ChatGPT.
Because ChatGPT is one of the firms that had reached 100 million users in the shortest time, and the first time in 20 years, it’s this challenger that might actually replace Search.
While Google had the most established AI lab, the data, and the users, etc. The moment ChatGPT was released, it suddenly looked slow, confused, and late to the party.
Even though Pichai explained that he didn’t say "code red", the fact that he had to clarify this says everything you need to know.
Pichai then spent the years that followed carefully steering this ship on the brutal AI roadmap.
From 2022 onwards, they’ve released some stunning AI products: Nano Banana, Search Assistant, Gemini Business, new-gen TPUs, and more data centers.
With all his hard work, finally, in late 2025, the tables flipped.
The rebirth?
In late 2025, Sam Altman sent his own Code Red memo.
Because Gemini had become a serious threat to OpenAI. He even delayed some of its other plans to focus on fighting Gemini.
For Google, this was supposed to be the rebirth.
The underdog had finally earned recognition from its biggest rival.
Here are Google’s wins that it celebrated in its latest earnings:
Search revenue grew, same as YouTube advertising rev. And then there was Cloud, up 82%, a serious acceleration.
This 82% matters because Google Cloud was already enormous.
Assuming Google Cloud sustains its current quarterly run rate, it has added the equivalent of a $45 billion-a-year business in just one year—more revenue than Starbucks generated in all of 2025.
But it’s not all unicorns and rainbows. The investors smelled something wrong.
In particular, the kind of company that they see Alphabet is becoming.
What kind of company is Alphabet becoming?
Conclusions first, two things:
From search-heavy to AI and cloud,
From capital light to really intense.
This is its segment revenue trend between 2023 and 2025. Search was the main revenue. And it was still free cash flow positive.
However, the trend is clear: starting in 2025, its CapEx has increased dramatically, to the point that people wonder whether it is now becoming an infra company with negative free cash flow?
In 2023, Google spent about 10 cents on data centers and chips for every dollar it made.
In 2025, that became 23 cents.
This year (2026), it could exceed 40 cents.
All this means that Google still makes most of its money from ads. But it is now spending like a company that needs giant factories to grow.
That is why investors are nervous. Because building infrastructure costs money long before you know whether the products coming out of them will pay for themselves.
Besides, the lifetime of that capex is also an issue.
Spending $200B on GPUs that are realistically useless in 3 years is very different than spending $200B on CPUs that can be written off over a much longer life.
Once the data centers are built, how often do they need to refresh the GPUs? Is it now forever burning $100B in capex just to renew GPUs?
This brings us to the depreciation accounting game.
The depreciation accounting game
The whole AI boom may come down to one extremely boring question:
How long can a GPU really make money?
Many believe that after two or three years (Google initially said four), a GPU may no longer be good enough for the most valuable AI work. If Google has to replace it to keep up, but still only depreciates it over six years, this creates a serious accounting loophole.
The issue has two folds:
Issue #1: Think of two identical laptops.
One sits in an office, handling email, slides, and web browsing for four years, all fairly lightweight tasks.
The other runs a Bitcoin miner at full power, day and night.
Both laptops may still turn on for another four years. But the second one has had a much harder life.
That is the exact same question for AI chips.
An advanced GPU used constantly for frontier-model training may not have the same economic life as one running lighter inference or batch jobs.
Of course there’s also this optimistic case that the old GPUs move down the ladder: from frontier training to inference, then to cheaper models and internal workloads.
But a chip being usable is not the same as a chip earning enough money.
Issue #2: Profits depend on how long Google says a GPU lasts
Not how long it actually lasts.
This is where its CapEx issue comes in … and one of the reasons why investors are worried.
Even if Google can keep old GPUs running, it still needs to buy new ones every two or three years to maintain Gemini’s competitiveness.
That means the question now becomes: how much must it keep spending? How often should it replace the machines at the frontier? And how should its accounting practice change accordingly?
Imagine a cab company buys a $60,000 car.
If it expects to use the cab for four years, it records a $15,000 cost each year in its accounts. If it decides the same cab will last six years, it would record only $10,000 each year.
But that results in two different accounting profits for the same $12,000 in revenue from driving passengers. If the cost is $15,000 per year, then the profit is $ -3,000. If depreciated over 6 years, however, they would make $2,000 in profit.
So Alphabet may be playing the same number game since 2023 with a few more zeros:
we completed an assessment of the useful lives of our servers and network equipment and adjusted the estimated useful life of our servers from four years to six years
The change reduced that year’s depreciation expense by $3.9 billion and lifted net income by $3 billion. And makes the business more ‘profitable’ than it actually is.
The question about its AI cloud ops is whether Google can keep it busy, at a profitable price, for six years?
So that was cloud infra; what about its AI model, Gemini?
How real is the Gemini demand?
For three consecutive quarters, Sundar Pichai disclosed a growth rate for gen-AI product revenue:
In Q3 2025, he said: “revenue from products built on our generative AI models grew more than 200% year-over-year,” then “400% year-over-year” in Q4 2025, and “800% year over year” In Q1 2026.
Wow, that’s quite impressive! But…
He never talked about the baseline number.
Think someone running a lemonade stand. In his first year, only his mom buys a glass, so he makes $1. The next year, his dad (cause the mom asks him to) and a neighbor each get a glass too, bringing his total to $3.
He proudly tells the whole neighborhood, “Our sales skyrocketed by 200% year-over-year!” Since going from $1 to $3 really is a 200% jump.
But he conveniently leaves out the fact that he started with a single dollar.
That is why not showing the baseline can mean genuine success, though more often than not (especially when you never talk about the baseline) it just means you started with almost nothing.
Then Pichai changed tactics in its Q2 2026 earnings call.
This time, Pichai emphasized 22 billion tokens per minute and the 950 million monthly active users.
While the numbers sound big and outstanding again, they are still not proof that Gemini is generating revenue.
Gemini is placed in front of almost every search engine and Android user; these two metrics may perfectly reflect interactions with a free trial rather than with users paying for its AI product.
So until now, we only know that some people are using Gemini (and even some of that might be forced on them), but we don’t know how much of that usage is paid for and profitable.
So perhaps Cloud demand gives us a cleaner answer?
How real is the cloud demand?
The thing is… Nobody can yet explain the full economics of the AI buildout.
For example, the arrangement with Anthropic (and similar cases) makes that especially difficult.
Alphabet has committed up to $40 billion to Anthropic. Anthropic has, in turn, agreed to use Google’s TPUs and Google Cloud capacity.
Imagine there’s a propane company called Gooogle (with three Os). Claudius owns a crematorium, but it does not yet have enough money to buy fuel to burn much of anything.
So out of the kindness of their heart, Gooogle gives Claudius $30 billion on credit for Claudius to buy propane from Gooogle to burn some bodies.
Gooogle then announces:
Demand for propane has exploded!! 🥂
Which, is technically true…
The propane is real, same as the $30 billion in revenue. There may even be plenty of dead people. However, the question is whether there are enough paying funerals to justify $30 billion of propane, or whether Gooogle has financed its own best customer.
This hands money from your left hand to your right has a name: circular deal.
Now, back to Google.
Since the public disclosures do not tell us how much of Google’s demand is funded by… Google.
So you can’t really blame investors for worrying how much of this real demand comes from the outside world, and how much is Google’s own money making the factory look busy.
But here’s the more fundamental issue with the entire AI ecosystem.
How real is the AI demand?
Both the AI circular deals and the Gemini user numbers lead to the same but a much more profound question:
How real is the AI demand to justify this hundred-billion-dollar market?
I’m not doubting if there’s anybody using AI; of course there are! Just think how many times you’re secretly asking AI to summarize my work w/o engaging it ;-)
Joking aside, five clues that no one has figured out if AI is even a net positive investment. Here’s a Morgan Stanley’s leaked internal report and other materials that help explain what I meant.
Clue #1: We’re still far from seeing profit in AI
It says enterprises first spend money modernizing data and getting ready for AI (where we’re at). And the productivity boost, like development automation (human workers replacement), only arrives later.
And even then, monetization depends on companies finding a new way to charge for it.
So Google’s 82% Cloud growth may be real. But this is still the stage where the companies are buying the plumbing before they have proved the house is even worth building.
Clue #2: Usage is not the same as durable revenue.
Companies are beginning to scrutinize token spending, especially in prototypes, agent workflows, and code that never reaches production.
Also, they’re stepping away from using the expensive model for every task. Instead, companies route simple work to cheaper models and reserve the frontier model for high-value ones (which arguably, only a handful that really need a frontier model).
That is why 22 billion tokens per minute and 950 million Gemini users are not enough. They show activity, not revenue. Gartner reaches a similarly awkward conclusion: customers have stalled on basic AI use cases.
Clue #3: The backlog proves contracts, not returns.
Pichai also emphasized the $514 billion Cloud backlog proves that customers are ordering capacity.
It doesn’t, however, tell us how quickly that backlog turns into revenue, what margin Google earns, how much is hardware rather than software, or how much of the demand is ultimately funded by Google itself.
Clue #4: The hidden risk is commoditization
Morgan Stanley says the cost of reaching a fixed level of intelligence is falling by roughly 5-10x a year. It also points to the rapid growth of open-weight models.
What’s worse is the birth of Kimi K3. The choice is no longer dumb but open or smart while closed. Kimi K3 is smart AND open.
On top of this, if anything, 2026 is the AI routing war. Even the traditionally payment companies (that you don’t view as an AI company) want a piece of the AI routing pie. Indicating no one, again, no one, cares which model, but the outcome.
AI’s ‘intelligence’ is worthless unless the fundamental flaws of LLMs are one day corrected; otherwise, this continues.
So where does this put Gemini?
Clue #5: Google may capture the demand without good economics.
The question is:
Can revenue grow faster than power, depreciation and data-center costs?
Basically, for Google, it’d be all about whether it can make money selling what they bet on?
The Disclosure Dilemma
Now, major AI and semiconductor stocks have experienced massive gains in the first half of the year (January through June), followed by a notable decline in July.
And it’s clear that the market has concerns about the current value of these companies. Because keep in mind, the issues I pointed out aren’t Google’s alone; the rest are, different to some degree, but all run on a very similar business model.
Here’s the credit spread, which is a measure of how risky investors think debt is. Higher spreads mean more risk.
So it’s very clear that Google has lost investors’ trust, especially if you compare it with Microsoft’s.
From what I see, these AI companies (in this case, Google) have two choices: either
Keep AI folded within Cloud and ask investors to trust that Cloud’s growth reflects healthy AI economics.
Or it can disclose AI revenue and margins separately, and bring it to light.
From my view, disclosure is the better approach, of course, but it’s also the riskier move, a move that no AI company dares to try yet.
Next week I’ll be comparing Google’s strategy with Microsoft’s to see how the market treats these two somewhat similar businesses differently.
