Alphabet just reported the best quarter in its history. Revenue beat. Earnings beat. Cloud grew 82% year over year, the fastest growth the segment has ever printed, and management said they still cannot build capacity fast enough for the customers lining up.
The stock dropped 7%.
Somewhere between those two facts sits the most important disagreement in markets right now. And the best version of that disagreement this week played out in a Reddit comment section.
Something new
Yesterday I replied to a thread asking whether the Google dip was a buying opportunity, and the reply got more attention than I expected. By morning it had been read over 100,000 times, and my inbox was full of very solid pushbacks. A few hundred of you joined from there. Welcome.
The discussion was too good to leave buried in a comment section, so I decided to cover it properly. These debates play out on Reddit and X hours after the news and days before the narratives settle, and staying on top of them is a full time job. So this is the first of a recurring series: the most heated market debate of the week, distilled. The thesis, every serious counterargument, where each one lands.
The point of the series is simple. Most of us are stuck in our own bubbles, following the accounts that agree with us, reading the takes that confirm what we already hold. Every trade has two sides, and the person on the other side of yours is not stupid. Seeing how both sides actually think about the same thesis is the closest thing to a full picture markets offer.
Here is everything worth knowing from the Google war.
The thesis that started it
Alphabet raised full year capex guidance by $15 billion, to $195-205 billion. That single line did the damage. Everything else in the report beat.
So the drop is a referendum on one question: is $200 billion of AI capex investment or destruction?
Alphabet's own answer is hard to argue with. Cloud grew 82% and they are turning customers away for lack of capacity. The classic accusation when capex explodes is "empire building," management chasing scale for its own sake. Empire building requires spending without demand. This spend has a waitlist.
That was my comment, more or less. Then the bears showed up, and they were better than the usual "AI is a bubble" crowd. Five objections stood out. Some of them I could answer. One of them caught me flat.
Objection 1: The depreciation wall
This one the bears get right, and it deserves a straight answer instead of a counter.
All of this capex hits the income statement over the next several years as depreciation, and margins compress before the revenue fully catches up. Roughly 60% of AI infrastructure spend goes to chips (McKinsey), which depreciate on 3-5 year schedules. Reported margins will look worse for several quarters even if the underlying business is fine.
Two things soften it without erasing it. Accounting life and economic life keep diverging: frontier training moves to new silicon while five year old chips run inference at full utilization, so assets carried at zero on the balance sheet are still generating invoices. And the other 40% of the spend, buildings, power, cooling, land, depreciates over 10-25 years and survives every chip generation.
But the muddy stretch is coming either way. It is the reason the multiple is reasonable. You are being paid to sit through it.
A cousin of this argument came up repeatedly: each dollar of capex is one-time, but the program is permanent, so capex functionally becomes opex and free cash flow never normalizes. The test for that is maintenance versus growth spend. If growth stopped tomorrow, opex stays but growth capex drops toward zero, which is exactly what Amazon proved in 2022 when it throttled the AWS buildout and cash gushed. The spend is a dial tied to demand, and the quarter demand stops justifying it, the dial turns. Whether management actually turns it is a fair thing to doubt, which is why capex growth versus cloud revenue growth is one of my triggers below.
Objection 2: "The cloud growth is subsidized token demand"
The commenter who made this argument had held Google since last year and sold his entire position in June, so he had done the work. His case: cloud growth comes from AI labs whose users pay less than the tokens cost. When tokens reprice to true cost, demand evaporates, and the capex was destruction after all.
The premise has a crack in it. Inference is generally sold above marginal cost. Serving tokens runs gross margin positive across the industry, and the giant lab losses come from training runs and payroll. Cost per token keeps collapsing, and prices fall alongside it.
What actually runs on subsidy is the R&D, funded by venture and corporate money that could dry up. A lab funding winter would dent cloud growth, genuinely. It would leave three things standing: enterprises running open source models, inference workloads that are already unit-positive, and the ordinary cloud migration underneath everything. Frontier models are open source now. A dead lab's workloads migrate to whoever serves the same models, and the compute gets rented either way.
There is also a live test running. If the subsidy thesis were right, usage should drop every time labs tighten free tiers and raise prices. Every price increase so far has been met with usage growing anyway.
The same commenter had a second point that deserves its own line: Anthropic reportedly accounts for more than 40% of Google's cloud backlog, and neither Google nor Microsoft splits AI cloud revenue from the rest. Even with healthy token economics, one customer being nearly half your backlog is concentration risk, full stop. Anthropic pays in cash rather than investor credits, which separates it from the circular Microsoft-OpenAI setup, but the honest tell to watch is whether backlog keeps growing in quarters without a headline lab deal. If it does, the demand is broadening. If it stalls, this objection graduates.
Objection 3: The $80 billion equity raise
This is the one that caught me. Someone asked whether negative free cash flow meant more equity offerings were coming, and I answered that Google would issue debt before it ever touched equity. Wrong. Alphabet announced an $80 billion equity raise in June, including a $40 billion at-the-market program that starts selling in Q3. A reader corrected me within the hour, and he was right to.
So I went through the actual filings. Roughly 2% dilution against a $4 trillion market cap. The ATM's stated primary use is covering tax obligations on employee stock vesting. And Berkshire Hathaway anchored the raise with $10 billion, which is not something Berkshire does for capex stories it doubts.
There is even a case the raise is smart: selling your own stock near all time highs to fund the scarcest asset in the economy is buying low and selling high with your own paper. But the correction stands, and it sharpened the piece. Dilution is on the table now in a way it never was in Google's history. Worth watching, not worth panicking over.
One more thing the bears miss
Google, and every other hyperscaler, would buy even more compute if they could. They are buying everything the supply chain can physically produce, and the waitlists say it still is not enough. The $200 billion is the constrained number.
Strange as it sounds, that constraint helps the bull case. Bubbles die from supply overshooting demand, and right now overshooting is being limited. TSMC wafer starts, memory allocation, and power hookups ration the buildout whether the buyers like it or not. In 1999 telecoms could order unlimited fiber, and did, and that ability to overshoot is what built the crash. Here the discipline is forced. Nobody chose it, but it is helping with the risk management for the whole trade.
What would change my mind
Every thesis needs falsifiable triggers, otherwise early and wrong are indistinguishable. Mine, in order:
Empty racks. The whole buildout rests on demand exceeding supply, and every player currently turns customers away. The first meaningful unused capacity anywhere in the system means the cycle is turning.
Capex outgrowing cloud revenue. Cloud grows 82% today against roughly 50% capex growth. The quarter that relationship inverts meaningfully, the red queen argument wins and the multiple deserves to compress.
A lab funding failure with visible utilization impact. If a major AI lab fails to raise and compute utilization drops with it, the subsidized demand thesis gets its proof.
None of the three is flashing. Until one does, the setup is what it was before the drop: one of the best businesses ever assembled, at a reasonable multiple, discounted because the next four to six quarters look muddy on the models that quarterly-judged fund managers live and die by.
They cannot wait. You can. That gap is the trade.
Read the Kimi K3 piece on why the whole AI infrastructure trade is accelerating
This is not financial advice. neym is an independent research newsletter. The author holds a position in Alphabet. Do your own research.