Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseInsight · On RadarX: @ZeeshhDecodes→ original

Rubin delivers 50 PFLOPS per chip, TPU 8t 12.6 — but Google wins on scale

Nvidia and Google both build AI chips. By the spec sheets, a single Nvidia Rubin delivers up to 50 PFLOPS (50 quadrillion operations per second) and a Google TPU 8t delivers 12.6, roughly 4 times less. But Google competes on numbers: one superpod (a huge cluster) holds 9 600 TPU 8t chips, while an Nvidia rack holds 72 GPUs. Labs don't buy chips, they buy cheap tokens, so compare the price per million tokens on your own workload, not the peak figure from a datasheet.

AI-processed from X: @ZeeshhDecodes; edited by Hamidun News
Rubin delivers 50 PFLOPS per chip, TPU 8t 12.6 — but Google wins on scale
Source: X: @ZeeshhDecodes. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @ZeeshhDecodes ↗
Editor-in-Chief
Rubin50 PFLOPS per chip, TPU 8t: 12.6, a gap of about 4x
Google superpod9 600 chips, Nvidia rack: 72 GPUs
Key metric #3Labs count the cost of a token, not a chip's peak

When a new AI accelerator comes out, the internet compares tables: who has more PFLOPS (quadrillions of operations per second). The author of the AI Decoded channel (@ZeeshhDecodes) compared Nvidia Rubin and Google TPU 8t. By the claimed peak, one Rubin delivers 50 PFLOPS and one TPU 8t delivers 12.

6, so Nvidia is about 4 times faster. But Google answers with scale: one TPU 8t superpod (a linked cluster) contains 9 600 chips, while an Nvidia rack holds 72 GPUs. For you, this means that picking a cloud by the figure for a single chip makes no sense.

What you need to calculate is the price of a token, that is, the price of each piece of text the model returns to the user.

What the per-chip numbers show

The author cites the spec pages. Nvidia Rubin is listed with a peak of 50 PFLOPS per chip. Google TPU 8t is listed at 12.6 PFLOPS. If a lab bought a single accelerator and put it on a desk, the choice would be obvious: in a chip-for-chip fight Nvidia still wins, and the author openly admits it.

But these chips don't sit on a desk. Large models are trained and served on hundreds or thousands of interconnected accelerators, and that is where the picture changes.

Where Google claws it back

According to the post, one TPU 8t superpod combines 9 600 chips, while an Nvidia rack holds 72 GPUs. Multiplying the peak numbers roughly gives about 121 000 PFLOPS (around 121 EFLOPS) per superpod versus 3 600 PFLOPS (3.6 EFLOPS) per rack. The comparison isn't quite fair: a superpod is a whole infrastructure unit, while a rack is only part of one. Still, it shows Google's logic: don't beat Nvidia on a single chip, but build a very large, tightly connected system and compete on price.

The author puts it this way: Google plays on scale and price. The chip is weaker, but there are a lot of them, and they are Google's own, without a third-party vendor's markup. Nvidia's strength is compute density: 72 powerful GPUs in a single rack.

Why the cost of a token matters

The post's main thesis: labs don't buy chips, they buy the cost of a token. This is the unit in which people count money if they sell access to a model or run one in production. Roughly, it works out like this: the price of the hardware and its rental per hour, plus electricity and cooling, plus networking, divided by the number of tokens the system actually produced in that hour.

Peak PFLOPS say little about this:

  • Real utilization. The peak on paper is almost never reached: memory, the network between chips and the quality of the software stack decide a lot.
  • Compute format. Peak numbers depend on precision (for example, FP8 or FP4). The post doesn't disclose this, so 50 and 12.6 can't be considered fully comparable without checking the documentation.
  • Purchase price and energy. A chip that is four times weaker but noticeably cheaper and more power-efficient can produce a token for less.
  • Software. Familiar tooling often weighs more than extra PFLOPS.

What to do in practice

If you are choosing a cloud or an inference provider, don't compare the theoretical peak. Run your own model with your context length and real traffic, then calculate the price per million tokens and the response latency. Nvidia may be faster and TPU may be cheaper. Only a measurement on your workload will settle it.

If the economics of your product depend on a cheap token, being tied to one camp is a risk, and it's worth testing both platforms.

Limitations

This is a short personal opinion on social media, not an independent benchmark. The figures come from spec pages, so they are a claimed peak, not measured training or inference speed. The post gives no real cost per token for Rubin or TPU 8t, and that is exactly what decides the argument.

Story

Key Questions Answered

What the per-chip numbers show?

The author cites the spec pages. Nvidia Rubin is listed with a peak of 50 PFLOPS per chip. Google TPU 8t is listed at 12.6 PFLOPS. If a lab bought a single accelerator and put it on a desk, the choice would be obvious: in a chip-for-chip fight Nvidia still wins, and the author openly admits it. But these chips don't sit on a desk. Large models are trained and served on hundreds or thousands of interconnected accelerators, and that is where the picture changes.

Where Google claws it back?

According to the post, one TPU 8t superpod combines 9 600 chips, while an Nvidia rack holds 72 GPUs. Multiplying the peak numbers roughly gives about 121 000 PFLOPS (around 121 EFLOPS) per superpod versus 3 600 PFLOPS (3.6 EFLOPS) per rack. The comparison isn't quite fair: a superpod is a whole infrastructure unit, while a rack is only part of one. Still, it shows Google's logic: don't beat Nvidia on a single chip, but build a very large, tightly connected system and compete on price.

Why the cost of a token matters?

The post's main thesis: labs don't buy chips, they buy the cost of a token. This is the unit in which people count money if they sell access to a model or run one in production. Roughly, it works out like this: the price of the hardware and its rental per hour, plus electricity and cooling, plus networking, divided by the number of tokens the system actually produced in that hour. Peak PFLOPS say little about this: Real utilization.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…