Disclaimer: This post is machine-translated from the original Chinese article: https://ai-coding.wiselychen.com/mtk-ai-budget-cut-cost-efficiency-trap/
The original work is written in Chinese; the English version is translated by AI.

Translation Note

This article is translated from the Chinese original. The author writes primarily in Traditional Chinese, and the original version is the canonical source — including the latest updates, comments, and follow-up discussions.

Read the Chinese original: [聯發科砍 AI 預算，來看降本增效的誤區 影片逐字稿](https://ai-coding.wiselychen.com/mtk-ai-budget-cut-cost-efficiency-trap/)

If you spot any translation errors or have feedback, please refer to the Chinese version as the source of truth.

Author: Wisely Chen Date: October 2026 Series: AI Coding Field Notes — YouTube Transcript Keywords: MediaTek, AI budget, token maximizing, enterprise AI adoption, cost reduction, efficiency, AI transformation

Transcript

Hi everyone. The biggest AI news in Taiwan lately isn’t whatever OpenAI or Anthropic did this week. It’s that MediaTek cut every employee’s AI budget — cut it in half, then in half again. It’s now down to roughly US$2 to 3 per person per month.

That set off a big debate online. Here’s a company this large, a semiconductor design house that frankly isn’t short on budget. It went from pushing everyone to use AI — to maximize AI token usage, “token maximizing” — to suddenly saving on AI spend.

What employees said on Threads: AI-written test programs, a jammed local GPU farm

A lot of people, mostly on Threads from what I saw, were angry. The company had been pushing them to use AI for slides and for coding, and they’d built the habit. Anyone could easily use Claude to write fairly good test programs. Quite a few people mentioned using AI for exactly that. I’m not in that industry, but I believe those programs matter a lot to their work.

Now that’s gone, so everyone is flooding the company’s internal local GPU farm. The result is a massive traffic jam, with everyone stuck waiting.

Two interesting points

There are two really interesting things here.

First, whatever you think of it, AI has entered our working lives. People now know how to use it, and it’s changed how efficiently we work — and even how we work.

Second, why did the boss suddenly swing from strongly encouraging AI investment, even encouraging employees to max out AI and tokens, to cost control and saving?

That reflects the single most important thing inside any organization.

The boss’s priority order: stock price, orders, cost, efficiency

Think about it. When a boss looks at an organizational change or a transformation, what do they look at?

The first thing is probably the stock price. Set that aside, and the first thing they care about is orders: where do orders come from? Will this investment help me grow orders, or move straight into a new market? Second, they care about cost. Third, they care about what we call efficiency.

So stock price is number zero, orders number one, cost number two, efficiency number three. That’s the order they weigh things in, and it’s a very important order: one always beats two, and two always beats three.

Orders: did AI help MediaTek win Google TPU business?

With that in mind, let’s go back to this case.

First: did AI help MediaTek (MTK) win new orders? Look at their investor conferences — they keep saying the big push right now is to win Google TPU orders. Put simply, that’s a B2B business. So optimizing employee efficiency doesn’t directly affect orders. That’s harsh, but it needs to be said clearly.

In MediaTek’s case, AI can’t decide whether they win that order or not. To some extent, whether the boss has a strong relationship with Google probably matters more than whether the company uses AI.

Cost: no layoffs, so AI only added spending

Second: what did AI do to cost?

We can see clearly that during this period — at least in the public news I’ve seen — there were no big layoffs, and no layoffs driven by AI. Instead, the company put money into cloud budgets and a self-built local GPU farm, which actually raised costs.

Efficiency: the complaints prove it’s useful, but verification takes longer too

Third: given how loud the complaints are, we can say this directly — if people are complaining, it means AI is genuinely useful. At least somewhat useful. So there must be some enhancement in efficiency.

How much enhancement, though, isn’t clear. A lot of companies now say that while AI clearly increases production speed, the time spent verifying the output goes up too. Add a bit here, subtract a bit there, and you don’t necessarily get a multiple-fold gain. That’s the reality right now.

Two beats three: the boss’s decision is logical

So here’s what we see:

One, AI doesn’t help MediaTek win more orders.

Two, AI clearly adds cost for MediaTek.

Three, efficiency does improve — let’s just assume it does.

But two beats three. So in this situation, there’s nothing wrong with what the boss did. It’s entirely consistent with their logic.

A signal for everyone: beyond cutting costs, AI has to open new territory

So, bringing it back to today: I think this is an important signal. It shows the main theme of AI adoption across every company. Any company adopting AI has to ask: beyond cutting costs and boosting efficiency, how does AI give you new territory and new orders? Or raise your stock price, or expand the territory you already hold, or let you build a new product — so that things you couldn’t do before become things you can do.

Otherwise you end up where MediaTek is: costs went up, efficiency went up, but unless MediaTek eventually does layoffs, nobody can see AI’s results. Not the boss, not the capital markets, not the shareholders. They won’t see it. That’s the hard reality.

What I tell clients: things you couldn’t do before, but can now

That’s why, when I talk to clients or friends about AI adoption, I always stress this: beyond cutting costs and boosting efficiency, what can you do now that you couldn’t do before? Or what new market can it open up for you? At the very least, it should help you win customers.

If you don’t push for that kind of aggressive innovation, or for real business outcomes, and you only follow the traditional “cut costs, boost efficiency” story, AI will most likely end in one of two things:

The first is layoffs.

The second is what happened at MediaTek: token maximizing at the start, then back to token minimizing, or to rationing every token.

Pull AI out — can people fill the gap?

At that point, it’s everyone on the ground who suffers. As the last commenter put it: today’s deadlines, schedules, and KPIs are all set around AI. Pull AI out, and it’s very hard for people to make up the difference.

So what do you do? You’ve already scaled up. You can’t go back to the old days of hand-writing every line of code. The question is whether people can fill the gap. Pretty clearly, that’s very hard.

So what now? That’s the problem we’re facing.

That’s all for today. Thank you.

Editor’s note: The figures “US$2-3 per person per month” and “cut in half, then in half again” come from MediaTek employees’ posts on Threads. MediaTek has not publicly explained why it adjusted the quotas. “Competing for Google TPU orders” reflects market reports.

Further Reading