What You Can Calculate, and What You Do Anyway — Notes on Gooaye EP695

What you learn from books stays shallow; to truly know a thing, you must do it yourself.

—— Lu You, “A Winter Night, Written for My Son Ziyu” (Southern Song, 1199; translation mine)

What this episode is about

Gooaye EP695 (2026-09-09) opens with the host laughing at himself. Wearing his gaming-VC hat, he had arranged to meet an overseas studio at TGS, assuming all along it meant the Taipei Game Show. One day he decided to sketch out his schedule early, and discovered TGS is the Tokyo Game Show — happening the following week. He admits he normally does everything at the last minute. This time the impulse to plan a week ahead saved him a flight.

Then the market: Taiwan stocks have chopped sideways for over a month, which a friend of his called “hell for day-flippers.” The heaviest segment covers a Google Cloud executive’s remarks at a Goldman Sachs conference — customer velocity, overspend, and payback periods on AI servers. Listener questions close it out, from the CPO value chain to a kid who refuses to eat meat.

The main points

1. Start first, fill in the map later. He describes his investing, his small ventures, and the gaming fund as things he jumped into out of enthusiasm without seeing the whole picture. The TGS mix-up is the other side of the same coin: learning as you go means you miss things everyone else considers obvious. What saved him was one early impulse to check.

2. In a chop, speed gets taxed. A stock locks limit-up, opens down the next morning, gets slaughtered to limit-down, then recovers days later. He says this tape isn’t hard — it’s hard if you keep using the March-to-June playbook where anything you bought went up. Each round trip costs a little; stacked over months it can eat 5–10% of your annual return. Two responses: keep your method and cut size, or adjust the method. He chose the latter — wait even when he wants in, and take it when it comes down.

3. Three numbers from Google Cloud. New customer acquisition running at twice last year’s pace. Deals above $100M more than doubling. And customers who commit to spend $100 typically spending north of $150. That third one carries the most information: overspend above 50% means the demand wasn’t just polite contract language.

4. The payback figure is what detonated the segment. AI servers pay back in under two years overall; their in-house silicon in roughly one. That overturns a popular story — that the hyperscalers write huge checks, post impressive backlogs, never really earn on it, and blow up on the turn. Turns out the money comes back fast.

5. Which makes negative free cash flow legible. Plenty of people had been staring at Google’s cash flow turning negative with no buybacks. But if you can find a one-year payback, spending more is the rational move. His inverted version is the valuable part: the day they start buying back stock, that’s the signal — it means they can’t find anything outside better than their own shares.

6. The price is a narrower imagination premium. Going from frontier-model contender to infrastructure supplier moves you from “nobody knows what the future looks like, here’s a rich multiple” to “utility provider, lower multiple, but the cash is real.” He uses that to explain why Buffett would personally sign off on the position: this has fallen back into territory you can actually calculate.

7. Don’t crown a model winner. Not long ago the market decided only one lab could still ship. Weeks later another one launched and the whole feed flipped. He puts the half-life at two or three months, not six. Meanwhile API prices keep converging — when what each lab collects starts to look alike, hardware choice collapses back onto cost-efficiency.

8. Who is actually paying, in three layers. Chip and component suppliers ship and get paid; cash lands immediately. Cloud providers front the money and recover it over five-year contracts. Model companies burn shareholder capital — and a good share of those shareholders are the clouds upstream. The only money genuinely arriving from outside is from end customers. He thinks the setup has room to run, and he thinks the day supply meets demand it fires backward. He declines to play doomsayer: the point is whether you participated on the way up.

Going further

”If the news is this good, why is the index still grinding?”

This was my first question. A one-year payback should move something, yet the index sits near highs refusing to declare.

Here’s how I break it down. Industry and price are two lines running at different speeds — he uses the old image of a dog and its owner. The dog runs ahead and behind, but it follows the direction the owner is walking. A choppy tape means the dog is circling, not that the owner turned around.

There’s a caveat you can’t drop, or this becomes a sedative for anyone who’s underwater. “Focus on the industry, ignore the price” only works if what you hold are the owner’s clues, not the dog’s. Order visibility, overspend ratios, where the big customers are pointing capex — those are the owner. Whether a limit-up holds overnight is the dog. I’ve caught myself assembling a pile of dog evidence to argue about the owner’s direction, and calling it long-term thinking.

One test I’ve tried: write down the reason you changed your mind this week, and check whether it’s price or something else. If it’s price, you’re trading short-term while telling yourself you’re long-term, and you’ll do neither well.

”A new theme appears with twenty related names. Which one do I buy?”

A listener asks where the value sits in CPO: fiber components, high-precision active alignment, the packaging platform, or the optical engine itself. The host’s answer is honest in a way that cuts against instinct — you can’t work this out in advance, so you buy the whole basket and watch where the value migrates.

Why counterintuitive? We’re trained to finish the homework before entering. He’s saying that while value distribution is still unresolved, “picking the right one” is false precision.

The basket isn’t a free-for-all, though. Two constraints came with it. Keep the theme at 30–50% of the book, spread across a few names. And “most exposed” is not “biggest mover” — he points at TSMC, a bottleneck almost everywhere, that refuses to squeeze the way memory makers do because it wants to grow with its customers. So being a bottleneck doesn’t mean collecting rent. Whether you collect depends on whether you’re willing, and whether your customers respond by cultivating a second source.

That corrected something for me. I used to hunt bottlenecks by asking who’s constrained. Now I add: did this constrained party actually raise prices over the past year? If not, the value flows into somebody else’s pocket.

”How do I tell a one-year payback from marketing?”

The same conference had plenty of marketing language — training efficiency multiples, cheapest inference in the industry. He says he’s gone numb to that; every new piece of hardware benchmarks itself against the leader.

So why did he believe the payback number? The difference, I think, is that this one is tied to an action the speaker can’t fake. Claiming great cost-efficiency costs nothing. Claiming a one-year payback commits you to keep spending and to not buy back stock — because the moment you buy back, you’ve conceded that nothing outside beats your own shares. Words bolted to behavior are harder to lie with.

Which makes his falsification condition the most useful thing in the episode: when they start returning cash via buybacks, returns have drifted toward mediocre. I wrote it down because it has a nice property — it comes to find you. No daily monitoring required.

The bear case survives, of course: a meaningful slice of this demand is circular, with model companies spending cloud money on cloud compute. The only outside money is end customers. So his claim is “it can run a while longer,” not “it won’t break.” The gap between those two sentences is the gap between this piece and a promotional one.

Where to look next

• The episode aired 2026-09-09 and is on the usual podcast platforms; the opening chat and the listener Q&A carry real weight, so it’s worth hearing whole
• The Google Cloud remarks at the Goldman conference are public, as are the company’s recent capex and free cash flow lines — worth reconciling yourself
• The dog-and-owner metaphor comes from André Kostolany; the original framing is worth reading in full

One thing to take with you

One idea: to judge whether a person or a company truly believes something, look at what they’ve paid for it, not at what they said.

Google saying it offers great value costs nothing. Choosing to pour cash in rather than buy back stock costs plenty — negative free cash flow, market skepticism, a multiple falling out of the imagination range and into the calculable one. Paying that price is what gives “one-year payback” its weight. And the day they stop paying it, the signal announces itself.

Something I’ve tried, offered rather than assigned: pick something you half believe — a colleague swears they’re slammed, a friend says this is the year they start a company, your employer says a direction matters, a family member says they’ll change. Don’t rule on it. Write one sentence instead: “If this is true, what concrete price will they pay this month?” The price has to be visible — money, time, or giving up something else they wanted. Pick one.

Then put it away and spend the month watching only that, not listening.

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.