The reason for NVIDIA’s fall was a report by The Information saying that Google, the new darling of AI stocks, was heavily pushing to sell its TPUs to customers, specifically mentioning Meta and Anthropic, crucial buyers for both NVIDIA and Google, as early adopters.

Rather, The Information simply made accessible ‘common knowledge’ what we analysts knew way before (as the Spanish saying goes, it was “un secreto a voces”, a “secret” that wasn’t really a secret and was being shouted for anyone who actually knows how to follow this stuff to hear, like seeing Anthropic opening a TPU Kernel Engineer, as we discussed a few weeks ago).

Furthermore, TeraWulf announced a deal with Fluidstack, in which Google acted as a backstop in case Fluidstack couldn’t pay rent to Wulf… just as Fluidstack was announcing its agreement with Anthropic.

All this goes to prove that markets aren’t remotely aware of what is going on until mainstream media catches on. But let’s get into the weeds to describe why this matters and why it is happening.

The narrative for NVIDIA bearishness is dead simple: NVIDIA was “untouchable”, and it isn’t anymore, thanks to Google’s chips, the TPUs that not only gave “birth” to the widely considered best model around, Gemini 3 Pro (this is a speculative claim at best, but it’s nonetheless “common knowledge”), but they are also (alongside Amazon’s own ASIC, Trainium), behind Claude 4.5 Opus’s training.

Now, Google wants to sell these chips to third parties rather than just renting them, making it an official competitor to NVIDIA and AMD.

NVIDIA seems nervous enough that they published a statement, pumping their chests, saying that while they were congratulating Google, their GPUs were better. But the truth is, Google is serious about this.

But how serious? Google’s TPU business is real, and it looks pretty good.

As we’ll see today, I believe this is the last thing Google needed to become the world’s most valuable company, something I predicted would happen before the end of 2026… but might be coming much sooner.

But let’s get to business. Why would anyone choose Google’s chips over NVIDIA’s? How good are TPUs when compared to NVIDIA’s GPUs?

From a technical standpoint, the fundamental differences between them are notable. The first question we must answer is what makes them different. Both are accelerators, meaning hardware specifically designed to accelerate computations. How, you ask? By parallelizing them.

Instead of executing a set of computations in sequence, we do many of them in parallel. Of course, this only works for workloads that are actually parallelizable.

In our case, both GPUs and TPUs focus on a particular type of maths, linear algebra, and in particular, matrix multiplications (‘matmuls’), which can be broken into parts and, thus, executed faster.

And here’s where the big difference between each product emerges. I’ll cut to the chase:

Crucially, the other big characteristic difference is that TPUs require lower power, simply because they are meant for one thing and one thing only, while GPUs retain some of the original flexibility, making them less optimal for any given workload.

But theoretical efficiency and reduced power requirements are basically rounding errors in a game where capital costs govern the total cost of ownership (TCO).

In layman’s terms, in a market where buying the actual hardware accounts for 80% of the total cost of buying and using it over its entire lifetime, being a few hundred watts more efficient doesn’t cut it.

So, time to find out the answer. What option is, economically speaking, better?

Before we proceed, the nitty-gritty. AI accelerators (i.e., AI hardware) are measured in “number crunching” capacity. That is, how many operations per second they can deliver.

This number is measured in FLOPs (Floating Operations Per Second), where a floating-point operation is a term for a number in scientific notation.

Therefore, the best way to compare both is to use metrics such as performance per dollar, particularly relative to TCO. In other words, how many operations on average are we going to get for every dollar we have paid and will pay?

Another critical thing to consider is that, in AI, we always, always look at system-level metrics. I won’t go as far as to say chip-level metrics are irrelevant, but let me put it this way:

The reason is that AI workloads, both training and inference (running the AI models), are distributed, so we almost always require two or more accelerators per workload. This means the value customers are actually looking at is the metrics of your AI servers, not the performance of a single chip.

All things considered, to perform the economic comparison, we will compare both companies by TCO/TeraFLOP (one trillion operations per second) of compute using FP8 (the most common precision used in AI, where each number has 1 byte (or 8 bits).

For the analysis, we are going to compare Google and NVIDIA’s top servers, the Ironwood TPUv7 pod and the Blackwell Ultra, respectively.

Finally, we'll start with the analysis, then move to the ‘so what’ part, where I’ll share what I believe is clearly the outcome (winners and losers) for 2026.

To make this as accurate as possible, since we’ll be relying on estimates in some cases, I’m going to use two popular analyst sources instead of one and compare them. These are The Next Platform and SemiAnalysis, as well as my own estimates using Jensen Math (more on that later). One is “more right” than the other, but I still want to include it so that you’re aware of the variability.

According to the semiconductor expert blog The Next Platform, they estimate a 9,216-TPU Ironwood pod, the largest scale-up size Google offers for its latest TPU generation (i.e., how many TPUs can they scale a single system to), costs $445 million to build (including all costs, not just the chips themselves). Assuming a 20% markup (my estimate), that gives us a retail price of $534 million.

This system has a theoretical FLOP throughput of 42.52 ExaFLOPs (official, published by Google). That is, the pod can deliver 42 quintillion operations per second.

On the other hand, an NVIDIA GB300 NVL72 (72 GPUs), the so-called Blackwell Ultra, has an unknown retail price and offers a peak throughput of 0.72 ExaFLOPs, or 59 times less than Google’s system.

Luckily for us, Jensen Huang, NVIDIA’s CEO, was happy to share the numbers we can use as guidance. On one occasion, Jensen mentioned that the capital cost of 1 GW was around $50 billion to $60 billion, of which approximately $35 billion was going directly to NVIDIA in direct GB300 value.

Assuming $55 billion for 1 GW in NVIDIA server equivalents and using Schneider Electric’s 142 kW per-rack form factor (meaning it’s a GB300 NVL72 built by Schneider requiring that much power), this would mean around 7,142 GB300 servers, or $4.9 million per rack of GPU direct investment, and around $7.7 million per rack once counting all data center costs.

Thus, on a very rough comparison estimate, the resulting values are:

All in all, even though we are talking just about capital costs, they represent the overwhelming majority of the TCO, so it seems costs look clearly better from the NVIDIA perspective.

But wait, weren’t TPUs cheaper? Why would anyone even consider TPUs then?

Under the paywall, we’ll answer the following: the “tricks” NVIDIA plays on us; the role Broadcom plays in all of this and what to make of it as an investor; the outlook for the big three, NVIDIA, Google, & AMD; the picture once considering NVIDIA’s newer GPU, Vera Rubin, Google’s upcoming TPU, TPUv8, and AMD’s MI450; and the takeaways to navigate AI hardware in 2026, the most dynamic market in the space, with key events to watch out for.