Gooaye EP702: Engineers Want the Best Tool, the Boss Reads the Invoice

Contents

1. Why would a company ration how much AI its engineers use?
2. Who does that optical-module rumour report actually restrict?
3. October turned broadly strong. Should I chase whatever hasn’t moved?
4. Smart people go into quant. Do they make more than we do?
5. So how do I use any of this?
6. Worth a look
7. The one thing to take with you

A good bow is hard to draw, yet it can reach high and strike deep; a good horse is hard to ride, yet it can carry weight over distance; a good talent is hard to command, yet he can bring his lord honour.

—— Mozi, “Qin Shi” (Warring States period; my own translation)

In Gooaye EP702, released 3 October 2026, host Hsieh Meng-kung talks about MediaTek tightening the AI budget its engineers can draw on. The figure he heard is spend running at the order of one to two hundred million a month; the company has been building its own data centre to cut that bill, and engineers have been complaining online about the result. He compresses the whole thing into one line: most supply-chain arguments end up back at cost. The other thread is a Morgan Stanley report dated 1 October, which says an optical transceiver module stays clear of restriction if 65% of its bill-of-materials cost is American. His read is that the threshold writes down the status quo, and what it blocks is China’s future path to cutting cost with its own parts. The numbers and the reasoning come from him and his industry contacts, and he says himself the report reads more like a rumour than a published rule, so there is no official text to check it against.

Why would a company ration how much AI its engineers use?

Because the invoice grows faster than the productivity shows up.

The situation in the episode: usage kept climbing, spend reached the one-to-two-hundred-million-a-month range, and the company built its own data centre to bring that down. Everyone then piled into the in-house local models, the servers couldn’t take the load, response times dragged, and the engineers pushed back — why won’t you let me spend the money?

He doesn’t take a side, and starts by admitting he hits the brake himself. Open the newest model in unlimited mode, ask one question, and it thinks for five minutes; halfway through you want to cut in and say just give me whatever you’ve got so far, because the meter is running at thousands a question. If someone with no real budget constraint wants to throttle it, a company answering to shareholders and an income statement has a case.

He also states the other half of the dilemma: no company has a reason to not want its engineers armed with better weapons and delivering better results. The stuck part is measurement — proving that the spend came back as output. His analogy is a city-building game. In sandbox mode with infinite money everyone lays perfect utilities, builds several power plants and sets the streets out on a grid; play it for real and the budget bites, and the cities that survive are the ones hacked and re-routed one layer at a time.

His forecast is that this is a growing-pains phase with a way out: on-premise servers paired with frontier models sitting at the customer’s own site — a large bill up front, no recurring subscription stacking up forever. He also draws the boundary: open models are enough for people writing copy, not enough for chip layout and design work, which is why the expensive models exist and will keep existing until specialised ones arrive.

The line that stayed with me was the one he added at the end: enterprise buyers are already the group most willing to pay, so when even they start braking, the supply-chain imagination shouldn’t be running too far ahead.

Who does that optical-module rumour report actually restrict?

First, plain language. An optical transceiver module is the small box in a data centre that turns the electrical signal from a switch or GPU server into light fired down a fibre, and turns returning light back into electrical signal. The episode pulls one 800G module apart, citing a teardown of Innolight’s EML DR8 unit with a selling price of 315 dollars.

Outbound, the path is: electrical signal into the DSP — the digital signal processor that cleans the signal up and re-times it — then to a driver that amplifies it enough to push a laser, out through eight EML lasers, and through lenses and a fibre array into the fibre. Inbound runs the other way: light becomes current, gets amplified, and returns to the DSP. He stops to point out that aligning the light sounds trivial and isn’t, which is why the market keeps arguing about whose alignment technology is better.

The report says a module is fine to ship as long as 65% of its BOM cost is American. His read: an 800G module already runs in the low-to-mid sixties on US content — Broadcom or Marvell for the DSP, Lumentum and Coherent for lasers — so the threshold writes down the status quo, and no obvious winner falls out of it. What it locks is something else. If China later pushes its own DSPs, drivers or photonic chips to cut cost, raising the domestic share drops the US share below the line and the product can’t ship.

That inversion is the most useful thing in the episode. The rule reads like a punishment aimed at the present, and it lands on future pricing power. The market pulled DSP and laser names up on 1 October betting the screws tighten later. He notes the report targets the 3.2T generation, while today’s volume sits at 800G and 1.6T — in a hot market people will tell you volume is imminent, and in practice it takes years.

For Taiwan, he files it under indirect speculation. US suppliers are safe on the BOM maths for now but would rather not wait for trouble, so mechanical parts, optical components and even assembly moving to Taiwan is plausible, and EML laser packaging could route around China. Taiwan’s only DSP house, a MediaTek subsidiary, ships at 400G, so it isn’t on any first-wave list. He also explains why none of this moves fast: the furthest-upstream materials sit in China, the US can’t stand that up quickly, and pushing to the kill shot invites retaliation that takes both sides down.

October turned broadly strong. Should I chase whatever hasn’t moved?

He didn’t rotate. His book is weighted to optical this cycle, the positions are behaving, so he left them alone.

The tape he describes: overnight futures at a new high, Nasdaq futures strong, and the first-half feeling of everything running at once coming back. The rotation order is clear in his telling — before July the themes took turns, and the last leg belonged to passive and power components; this cycle they were the last to move again, and they finally went last week, by which point the fast crowd already had their hands full of passives.

His reservation sits on the squeeze-out trade. The shortages and price increases are concentrated in AI-related parts, so picking something off that line means betting that displaced capacity eventually reaches it. He puts that lower in his ranking and says it makes him nervous, while the market enjoys it and sends those names limit-up quickly.

There’s one test in this part worth keeping: on calling the top early, his bar is that the index moving averages have to roll over before the discussion is even worth having, and shorting before they do gets you squeezed beyond recognition. He then names the retail loop — treating bearish commentators as prophets on the way down and trashing them on the way back up. And one line for your own accounting: when the index is working and your positions aren’t, that’s a stock-picking problem.

Attribution deserves the same care. He cites SSD-related names falling the previous day, blamed on hard-drive makers expanding capacity, and points out that the capacity base there is low enough that even a big expansion barely moves the industry. Markets react to small things, and in hindsight nothing happened.

Smart people go into quant. Do they make more than we do?

A listener asked whether the people doing this are all maths and computer science prodigies, and whether a good enough brain earns more there. His answer: the proportion is high, and whether it pays depends on the format of the contest.

His split: our side runs on points. You don’t have to be outstanding in any one area — a decent all-round profile, a long time in the market and steady compounding can produce a good record. Quant and high-frequency run as a ranking ladder, where the fastest system with the lowest latency can take the profit, and being equally smart with worse hardware or a worse strategy means your share gets taken. He says he couldn’t do it; listening to them discuss capturing one or two percent, he wonders where the point is, until you remember they turn the same capital over repeatedly and lever it up, and a certain small percentage levered hard is a lot of money.

What I take from it has little to do with quant. Identical ability compounds in a points format and gets absorbed by the leader in a ranking format.

So how do I use any of this?

Readers are usually in one of three spots: wanting to upgrade a tool without knowing how to justify it, looking at good news and not understanding the price action, or watching a group that just started moving and wondering whether to switch.

On the first, the budget story says to price a year. Engineers wanting the strongest tool isn’t wrong, and the company applying the brake isn’t being difficult; the difference is who sees the invoice. I’ve been in that hole — testing something, pushing the settings to maximum out of habit, five minutes gone before I look up. My fix was replacing “what does this run cost” with “what does a year of this cost”, and the answer tends to write itself.

On the second, the module segment gives a two-layer test: whose cost does this change, and how many years out. The 65% threshold changes nobody’s cost under today’s BOM, changes China’s future cost-down path, and lands in the 3.2T generation. When price action looks wrong, I’m usually reading those two layers as one.

On the third, his reason for not rotating is worth copying: his positions still work. A rotation should come from something of yours breaking, not from something else starting to run. For what counts as broken, borrow the moving-average test — write down one fact somebody else could verify, and use it as the switch.

The stickiest answer in the episode has nothing to do with investing. A listener wrote in about a vague emptiness, maybe wanting a sports car and feeling it was out of reach. He said to treasure the stretch where you still want the car, because the happiest part is while you’re saving for something; once it’s in the garage it turns from a thing you want into a thing you have, and the things you have don’t make you happy. I laughed, because it connects to the cost argument running through the rest of the episode — while we want something we’re happy to assume an infinite budget, and only start reading the invoice after it arrives.

Worth a look

• Gooaye EP702 (3 October 2026), the source for these notes
• Morgan Stanley’s 1 October 2026 note on optical transceivers, relayed in the episode; the original has to come from your broker
• US Commerce Department and FCC releases, to see whether this restriction ever becomes a published rule
• Earnings decks and filings from Broadcom, Marvell, Lumentum and Coherent, to check the claims about BOM share
• MediaTek’s earnings materials, to check the descriptions of AI spend and the in-house data centre

The one thing to take with you

Every argument that something “should use the best available” ends at one question: whose invoice is it?

Engineers want top-spec models and the boss wants top-spec results; the collision is over which income statement the bill lands on. Neither side in that company is wrong — what’s missing is a way to convert spend into measured output, and that’s the same gap in most of the upgrade decisions we make at home.

Here’s a version I’ve tried that works away from money too. This week, pick one thing you treat as an obvious upgrade — an appliance, a work subscription, a class for your kid — write its full annual cost on paper, and ask yourself one question: if that money came straight out of my own pocket today, would I still upgrade?

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.