Zuck Cheats on Jensen: The Silicon Valley Herd Just Decided Nvidia is Optional

If you listened closely yesterday, you could actually hear the sound of Jensen Huang panic-scrolling Vinted to see how much his leather jacket is worth second-hand. Rumour has it he was last seen trying to trade it for a Google fleece and a "Noogler" hat.

The news that Facebook is in talks to rent and eventually buy Google’s TPUs is the final nail in the "infinite growth" narrative. Facebook is currently Nvidia’s absolute biggest whale, the company actively bulldozing half the planet to build data centres the size of small countries. Now they are shopping at Google.

We are deep in late 2025 and the reality is setting in. The whales are migrating, and they aren't swimming towards Nvidia.

The "Training" vs "Inference" Trap:

Before we laugh at the bagholders, let’s get the technical boring bit out of the way because it explains why Nvidia is about to hit a brick wall.

There are two stages to this AI grift. First is Training, which is building the brain. This requires massive, raw power and is Nvidia’s H100/Blackwell territory. It costs billions. Second is Inference, which is using the brain. This is you asking ChatGPT to write a passive-aggressive email to your landlord. This requires efficiency and low cost.

For the last three years, everyone was panic-buying Nvidia cards for training. But now the models are built. We are entering the inference era. Using a £30,000 Nvidia GPU to generate a picture of a cat is like using a Ferrari to deliver a Domino's pizza. It works, but the fuel costs will bankrupt you.

The Great Plateau:

Let's be real. When GPT-5 dropped back in August... It was better, sure. But was it "change civilisation" better? No.

We have hit the law of diminishing returns. Throwing another 100,000 GPUs at a model isn't making it 100% smarter anymore. It is making it 5% smarter for 500% of the cost. The training gold rush is slowing down because the shovels have become too expensive for the amount of gold we are actually finding.

The Traitors are Leaving the Ship:

Look at the guest list for Nvidia’s pity party. It is literally everyone who matters.

Facebook: They are pivoting to Google TPUs. For them to consider buying Google hardware is a massive shift. Google famously never sells its chips. This shows just how desperate Zuckerberg is to stop paying the Jensen Tax.

OpenAI: The poster child for the boom is now two-timing Microsoft. They are designing their own daily driver chips for inference and renting Google TPUs. They know the future is cheap inference rather than brand loyalty.

Microsoft: They are pushing their own Maia chips for 2026. Are they as fast as Nvidia’s? No. Do they need to be? Also no. They just need to run Outlook summaries without costing £4 per email. Efficiency is the new benchmark rather than raw speed.

Google: The smug winner. The release of Gemini 3 Pro and Nano Banana Pro proved a fatal point. You do not need Nvidia to be State of the Art. Google’s entire stack is home-cooked, it is cheaper, and it works better. The "Nvidia Moat" myth is officially dead.

The Bagholders: Oracle and The Ego

This leaves us with the "Greater Fools" left holding the silicon bags when the music stops in 2026.

First is xAI. Elon Musk is currently building the world’s largest cluster to run Grok. Why? Because his ego demands it. Let’s look at the numbers. Grok 4.1 launched just last week, claiming to beat Gemini in benchmarks. But the numbers that matters are the real shock.

Reports from earlier this year showed xAI burning through nearly $1 billion a month in expenditures to build out the "Colossus" cluster, all to chase a market share that is still sitting at less than 1%. Spending $13 billion a year to generate $500 million in revenue isn't a business plan, how's this guy "world's richest man" again? Unless Elon pivots to Tesla’s Dojo chips, he is just setting money on fire to measure contests with Zuck and Altman.

Then there is Oracle, the biggest sucker at the table. Larry Ellison is buying up every Nvidia chip he can find, building massive "superclusters", and renting them out on a "first come, first served" basis. Larry’s entire business plan is based on owning the Ferraris and hoping people pay £10/hour to rent them.

But what happens in 2026 when Google and Microsoft offer "good enough" vans for £2/hour? Who is going to rent Larry’s Ferraris? Oracle has no chips of their own. They have no GenAI models of their own. They are a glorified landlord in a market where all the tenants are buying their own houses. When the backlog clears and the hype dies, Oracle is going to be left with warehouses full of depreciating Blackwell chips that nobody can afford to turn on.