Back in June, I wrote a MoneyMorning article titled “This could soon become the hottest AI chip on the market!”. It was all about Google’s Tensor Processing Unit (TPU) – a custom-built chip designed for one thing and one thing only: running AI models at lightning speed. Here’s what I said then…
Unlike Nvidia’s GPUs, which were originally built for video game graphics and only later repurposed for AI, TPUs were created from scratch for the AI era. They’re purpose-built for maximum efficiency, low latency and blistering performance. And while GPUs remain the gold standard for training AI models, TPUs are engineered to dominate inference – the day-to-day work where AI actually thinks, reasons and responds to users in real time.
Training happens once.
Inference happens billions of times.
That’s why I argued back in June that TPUs might quietly become the most important AI chip of all. Not the flashiest, not the most hyped, but the one that ultimately powers the real-world usefulness of AI.
So why bring this up again today?
Because the prediction I made is no longer theoretical.
It’s happening.
Meta just put TPUs on the global stage
In the past week, a story broke that sent a shiver through the AI market: Meta is reportedly negotiating a multi-billion-dollar deal with Alphabet (Google) to buy TPUs directly and deploy them inside its own data centres as soon as 2027.
Until now, Google has used TPUs almost exclusively inside its own AI empire – powering Search, YouTube, Ads, Maps, and now Gemini. The only way outsiders could touch a TPU was by renting them through Google Cloud.
So, selling them directly to another tech giant is a complete shift in strategy.
It signals that Google is no longer treating TPUs as an internal secret weapon, but as a major commercial product line.
So, why is Meta doing this?
Simple…
Diversification, cost control and performance optimisation.
Meta’s AI workload is exploding. Llama models are scaling. Reels, advertising, recommendation engines and generative features all require real-time inference across billions of users.
Nvidia GPUs are phenomenal – but they’re expensive, power-hungry and in short supply.
TPUs offer something different:
• Lower cost per inference
• Better energy efficiency
• Deeper optimisation for certain workloads
• Supply that isn’t bottlenecked
• Direct hardware + software integration with Google’s AI ecosystem
Meta doesn’t want to be dependent on a single supplier for its next decade of AI growth – especially when Nvidia’s dominance gives it enormous pricing power.
Meta is quietly building one of the largest AI compute networks on the planet. So, if you’re going to deploy hundreds of billions of dollars in infrastructure, you don’t want all your eggs in one basket. You want options. You want flexibility. And you want cost leverage.
Now, let’s get this clear…
Nvidia isn’t “finished”…
Nvidia still sells the world’s best AI training chips. Full stop.
Everyone knows it. Every frontier model – OpenAI, Anthropic, Meta, xAI, DeepMind – is still trained on GPUs.
Nothing about that has changed.
What is changing is the shape of the AI stack
For the past decade, AI was dominated by training.
But the next decade? It will be dominated by inference, deployment, and real-time, everywhere AI.
That’s where custom chips will thrive. Think…
• Alphabet TPUs
• Microsoft Maia
• Amazon Trainium & Inferentia
• Apple’s next-gen on-device chips
These aren’t Nvidia killers. They’re simply slices of a massive market Nvidia was never going to own entirely.
It also means smaller specialised custom chip makers such as Marvell could emerge as a niche winner.
Simply put – we’re transitioning from a GPU-only world to a multi-silicon world.
Inference is becoming the main battlefield. Cost efficiency is becoming the competitive weapon. And custom chips are stepping into the spotlight.
The Meta-Google deal is just the beginning.
Not a subscriber to Money Morning?
You can get free daily recommendations like these with Money Morning eletter. Just sign up here.
