Gerald Yang

AI repriced everything. I'm documenting how. | Tokenomic author | Builder & Maker in Canada

[Gerald Yang @YangGerald]

@criscxuan woc

[Gerald Yang @YangGerald]

Claude Code 有桌面。Grok Build 一直蹲在终端里。

新版Grok 4.6到底有多强，我用一下午我给它做了个窗口，开源了。叫 Hull。不是官方的，也不重写模型——就是给 grok agent 套了个壳。

一个窗口能并跑多个 agent，登录不用先开 Terminal。中英都能用。

https://t.co/cwclKqpKtv

[Gerald Yang @YangGerald]

不是，这个opus 4.8 max是发生了什么事儿吗？用了几天fable，但是我记得原先4.8没有这么傻啊？疯狂出错，各种幻觉

[Gerald Yang @YangGerald]

https://t.co/kMdPNqjGu8

[Gerald Yang @YangGerald]

In the AI economy, there are five classes of citizen.

Tier 5 (enterprise): processes 100 books per minute. Has dedicated GPU clusters. No usage limits. AI works at the speed of thought.

Tier 1 (free): 10 requests to read one book. Constant rate limits. "Please wait 4 hours." AI works at the speed of patience.

Same model. Same technology. Completely different experience.

This is the new class system. Not first class vs economy on the same plane. More like private jet vs standby — you're not even guaranteed a seat.

The gap between what AI CAN do and what YOU can access is the new inequality. And it's widening with every model generation.

[Gerald Yang @YangGerald]

https://t.co/63gyfZsqUT

[Gerald Yang @YangGerald]

AMD's GPUs are competitive on raw performance. Sometimes they even benchmark higher than NVIDIA.

So why does NVIDIA still dominate 80%+ of the AI market?

Same reason iPhone dominates despite Android phones having better specs on paper: the ecosystem.

NVIDIA doesn't just sell chips. It sells CUDA — a software platform that every AI researcher, every framework (PyTorch, TensorFlow), every library, every optimization tool has been built on for the past decade.

Switching from NVIDIA to AMD isn't like changing a part. It's like changing your entire operating system — and rewriting every app you've ever used.

That's the moat. Not the chip. The ecosystem.

[Gerald Yang @YangGerald]

https://t.co/SnPjsmC62m

[Gerald Yang @YangGerald]

Training GPT-5 requires tens of thousands of GPUs working together. Each GPU computes at blistering speed.

But here's the bottleneck nobody talks about: the data transfer between GPUs.

Imagine giving 10,000 people supercars with 800 horsepower engines. Then making them all drive on a two-lane road. That's what GPU interconnect looks like today.

Each chip is an 8-lane engine stuck on a 2-lane data highway. NVLink and InfiniBand are trying to widen the road, but they're still not enough.

The limiting factor in AI training isn't how fast each GPU thinks. It's how fast they can talk to each other.

[Gerald Yang @YangGerald]

https://t.co/ljjrFS2oU0

[Gerald Yang @YangGerald]

A 12-inch silicon wafer produces about 100 H100 GPU dies.

How many are perfect? Maybe 20.

The other 80 have microscopic defects — a misaligned circuit, a contaminated layer, a particle smaller than a virus that landed in the wrong place. They get downgraded or scrapped.

This is why GPUs are always in short supply. It's not that TSMC can't make them. It's that 80% of what they make doesn't meet spec.

Improving yield from 20% to 80% takes years of process refinement. Not months. Years. There's no shortcut. No amount of money accelerates it.

The chip shortage isn't a production problem. It's a quality problem.

[Gerald Yang @YangGerald]

https://t.co/oOqHSjAVHi

[Gerald Yang @YangGerald]

Write TSMC a check for $50 billion today. The earliest you'll see chips from the new fab: 2028.

Not because TSMC is slow. Because building a cutting-edge semiconductor factory requires:

- 3+ years of construction
- Clean rooms 10,000x cleaner than a hospital OR
- 1,000+ sequential manufacturing steps per wafer
- Equipment that takes months to install and calibrate
- Engineers who take a decade to train

This isn't like opening a restaurant. You can't throw money at physics and make it go faster.

The AI compute shortage has an expiration date. That date is measured in years, not quarters.

[Gerald Yang @YangGerald]

https://t.co/remxFjtXri

[Gerald Yang @YangGerald]

There's exactly one company on Earth that makes the machines needed to manufacture cutting-edge AI chips: ASML, in the Netherlands.

Each machine costs $150 million. Weighs 180 tons. Takes months to install.

If one of TSMC's EUV production lines goes down for a week, NVIDIA's chip shipments get delayed. If NVIDIA ships late, OpenAI's expansion plan gets pushed back. If OpenAI can't expand, your usage limit gets tighter.

One machine in Taiwan sneezes. The entire AI industry catches a cold.

This isn't a hypothetical fragility. It's the actual architecture of AI compute today.

[Gerald Yang @YangGerald]

https://t.co/r5LzNrqNIO

[Gerald Yang @YangGerald]

In 2020, the world ran out of face masks. Not because we forgot how to make them. Because suddenly everyone needed one at the same time.

In 2025, the world is running out of GPU compute. Not because engineers don't know how to make faster chips. Because suddenly every company on Earth wants to train its own AI model.

The shortage isn't a technical failure. It's a demand explosion hitting a supply chain that was never built for this volume.

TSMC's fabs are booked two years out. ASML's lithography machines have a multi-year waitlist. NVIDIA can't ship fast enough.

We know how to make the chips. We just can't make them fast enough.

[Gerald Yang @YangGerald]

https://t.co/rK6zA5ig2e

[Gerald Yang @YangGerald]

"You've reached your usage limit. Please wait 4 hours."

You think this is a system error. It's not.

It's the AI equivalent of your bank saying "your credit line is maxed out." Not because you're a bad customer. Because the total supply of compute — GPUs, electricity, cooling — is physically finite.

TSMC's fabs are booked years out. NVIDIA ships a fixed number of H100s per quarter. Every person hitting "send" on ChatGPT is competing for the same pool of silicon.

When Claude tells you to wait, what it's really saying is: global demand for machine intelligence just exceeded supply, and you got rationed.

This is credit tightening. Not a bug.