Why is Google renting GPUs when it has TPUs?
Google recently made a deal with SpaceX to rent GPUs. But this might raise a question in your mind: why would Google want to have GPUs?
There can be broadly 3 reasons; the last one is more interesting.
Google is running out of compute, so it wants to get it wherever it can.
Google Cloud’s customers don’t want TPUs; they want GPUs.
The more interesting one can be Google is deliberately moving its team off the TPUs so that they can free up the space for their customers. Once they are on the TPUs, then it becomes a lock-in for them.
So it can happen that the shortage isn’t the problem, but it can be a good Sales tool for Google.
Why did Siri fail, and what is Apple doing to improve?
Remember the 2024 demo: “When does my mom land in San Francisco?”
Siri is supposed to figure out that your mom has sent you her flight details at any point in time, maybe in messages or email and answer based on that.
But Siri was not able to do so and became the biggest embarrassment in the history of Apple.
So, to answer this question, Siri needs to do these things
Find out the flight details - Search across your messages, mail, and calendar. Locate the one place your mom mentioned her flight.
Understand and answer. Read what it found, work out the landing time, and phrase a reply.
The second step is about the AI, but Apple failed at step 1.
Why is Apple Search bad? Because search was never important before. Apple has a lot of standalone app which was built over 15 years ago.
Each app stores the data in its own way. There was no reason to share them. This was enough for humans, but AI needs something completely different.
One index across all apps, not per-app search.
Semantic-based search, not an exact keyword match.
This retrieval should be fast enough.
So Apple tried to solve this.
This is why WWDC 2026 was strangely boring. Apple spent stage time on infrastructure, including rebuilding the search index.
Now Siri would be able to search for information without knowing which app it is in.
Still a long way to go for Apple; this needs to be more agentic rather than just retrieval.
A model that runs on the phone for privacy and speed. Bigger tasks go to Apple's private cloud, which now runs on Nvidia chips inside Google's data centres.
Now, again, why does Siri run on Nvidia Chips inside Google Data Centres?
So every AI model has two phases in its life
Training - this is to build the model
Inference - this is to use the model
So training is completed in a timeframe, but inference doesn’t have an end date.
So even if you are using TPUs for training, the training needs to end, but inference has to continue.
Chips aren't interchangeable. Software written for Nvidia doesn't just run on TPUs, and vice versa. You optimise for the specific hardware: how it handles memory, how it splits work, how it batches requests.
Since Nvidia chips are everywhere, it will not create that kind of lock-in.
I hope you have found these case studies useful; you can go super deep into our flagship
AI PM Course (PMs at Microsoft, Coinbase, Indeed & 800+ PMs rated 4.9/ 5).
See testimonials and course details OR the AI Builder Cohort
About Author
Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking. AI Product Manager/Builder Cohort