You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Amazing to see how much capex Google intends to spend in 2026. If we assume Google's revenue growth in 2026 will be equal to Q4-2025 (18%) is has surpassed TSMC capital spend ratio:
Here the overview since 2020. TSMC seems rather modest in capex!
However, Google depreciates some 40% of capex over very long times, as the CFO states:
Anat Ashkenazi
CFO
The question with regards to the CapEx and the kind of what makes up the total that we've announced for this year and last year, approximately 60% of our investment in 2025, and it's gonna be fairly similar in 2026, went towards machines, so the servers, and then 40% is what you refer to as long-duration assets, which is our data centers and networking equipment. And I think you're probably referring to the depreciation delta between them. Those long-term duration asset depreciate over, you know, the building could be 40 years or longer. Other components may be maybe less than that.
Will be interesting to see if this massive capex by Google these years will bring back their revenue. For TSMC their management predicted recently that from 2024 to 2029 TSMC will grow with a 25% CAGR, that would be quite an accomplishment. With their gross margins around 60-65% and their enormous efficiency of operation (0.75 net profit margin / gross margin), TSMC will be an enormous "oil producer" for the AI infrastructure buildout:
"relative capex" is not the big news IMO. Google is ADDING 100B in new spending over the next year.... we did not see this coming in August. Hence the current boom. 100B added to the semiconductor hardware ecosystem by one company.
Side note: We are not at equilibrium so we dont know what the impact of constraints are. Qualcomm is not growing because there is not enough memory. Small companies cannot get memory. PC OEMs are having memory issues.
Datacenters are the priority and the winners. Who are the losers?
"relative capex" is not the big news IMO. Google is ADDING 100B in new spending over the next year.... we did not see this coming in August. Hence the current boom. 100B added to the semiconductor hardware ecosystem by one company.
Side note: We are not at equilibrium so we dont know what the impact of constraints are. Qualcomm is not growing because there is not enough memory. Small companies cannot get memory. PC OEMs are having memory issues.
Datacenters are the priority and the winners. Who are the losers?
It seems Wall Street is somewhat concerned about these huge capex investments for the large hyperscalers and wonders about the ROIC for these companies.
I can imagine OpenAI & Microsoft are having a tough time? Gemini and the integration with Apple and Samsung phones seems to be a strong force towards consumers. I guess Google needs capacity to enable that huge growth, besides all the enterprise AI that Google is developing.
Google has the monopoly on search and that seems a very good start to get hooked on Gemini.......
This is what Gemini states about Search and AI:
Google’s deep integration of Gemini into Search (now moving into AI Mode) creates a massive feedback loop. Every time you ask a follow-up question, give a "thumbs up," or refine an AI Overview, you are effectively acting as a volunteer trainer for the model.
How Google Uses Your Search AI Data
Google’s current privacy policies for Gemini and Search Generative Experience (SGE) break down the usage into three main buckets:
Machine Learning Training: For personal accounts, Google uses your prompts, the AI’s responses, and your subsequent "engagements" (clicks or follow-ups) to fine-tune Gemini’s logic. The goal is to teach the model which answers actually satisfied the user.
Human Review: To ensure safety and accuracy, Google uses a pool of human reviewers. They look at anonymized samples of conversations to "annotate" them. Google explicitly warns: “Do not enter anything you would not want a human reviewer to see.”
The "Personal Intelligence" Layer: As of early 2026, Google has rolled out a feature called Personal Intelligence. If you enable this, Gemini can look at your Search history, Gmail, and Photos to answer you. While Google claims they don't train the model on the private content itself (like your actual photos), they do train on the prompts you use to find them and the responses the AI gives you.
"relative capex" is not the big news IMO. Google is ADDING 100B in new spending over the next year.... we did not see this coming in August. Hence the current boom. 100B added to the semiconductor hardware ecosystem by one company.
Side note: We are not at equilibrium so we dont know what the impact of constraints are. Qualcomm is not growing because there is not enough memory. Small companies cannot get memory. PC OEMs are having memory issues.
Datacenters are the priority and the winners. Who are the losers?
How Google Uses Your Search AI Data