# A Contract Signed Through 2031 — Seven Days of Upstream Posts, Laid Side by Side

> An independent researcher who watches only the upstream of AI infrastructure left forty-odd public posts this week. Put them in order and what surfaces isn't a ticker — it's four lines of reasoning about capital picking winners, growth rates truncated by supply, time turning into an asset, and which country the real chokepoints sit in.

Published: 2026-09-07
Locale: en
Tags: AI infrastructure, supply chain bottlenecks, optics, memory, reading signals


![A long workbench receding into deep perspective, small optical components laid out in a single row under one hanging lamp, a working plant visible at the far end of the corridor](/covers/serenity-x-week-2026-09-07-cover.png)

> "The sage sees the faint and knows what is sprouting; sees the tip and knows the end."
> — Han Fei, *Han Feizi*, "Forest of Persuasions I" (3rd century BCE)

Han Fei is not describing prophecy. He is describing where judgement comes from: you do not have to wait for a thing to finish growing, you recognise it while it is still a beginning. The hard part is that a beginning looks exactly like noise. The difference only shows up when you arrange them — one news item is noise, five items pointing the same way are a beginning.

This week's public information contained several of those beginnings.

## What this piece is

There is an independent researcher on X, writing as Serenity, who covers only the upstream of AI infrastructure: optics, memory, packaging, power. Over the past seven days he left more than forty public posts — earnings-call summaries, industry news, and his own inferences.

What I did is simple. I put the posts in order, did not retell them one by one, and asked a single question: treating these seven days as raw material, how many reusable ways of judging can I pull out of it?

A word on the boundary of the material. Every piece used here comes from public posts that anyone can open; I took nothing from behind a paywall. The figures cited come from public financial reports, public earnings-call remarks, public news, and public regulatory filings. Also: no tickers appear here, no recommendations, no price targets. This is about industry structure, not about what to buy.

I pulled out four lines.

## Four lines from the week

### 1. Capital has started doing the job of industrial policy

The first turn of the week has nothing to do with chip specifications and everything to do with who owns a piece of whom.

The largest accelerated-computing supplier in the world did three things of the same nature within days: invested 3.5 billion dollars in a large Asian chip design house (public news); acquired the world's largest open-model and dataset community platform for 12.9 billion dollars (public news); and its balance sheet already carries a long list of holdings up and down the chain — a processor maker, custom-silicon vendors, laser and optical component suppliers, compute landlords.

Taken separately, each has a commercial rationale. Laid side by side, they reveal a playbook that was run once before: cultivate a new set of winners, align them financially and strategically with yourself, then use that new set to dilute the incumbent whose position had become too comfortable.

The cloud compute round went the same way. Every hyperscaler was building its own silicon, which is bad news for a merchant supplier; so capital flowed toward a class of independent compute landlords, and the market gained a group of customers who buy accelerators instead of designing them. The custom-silicon round has the same shape: the pie was being eaten by a handful of firms, and now the pie is bigger — but so is the number of people at the table.

![Two horizontal bars compared: the pre-investment bar is shorter and the incumbent winner takes up most of it; after the investment the total bar grows longer, and although the incumbent's segment keeps the same length, three new entrants dilute it into a minority share.](/figures/capital-dilutes-the-incumbent-en.svg)

There is a reusable judgement here. When you assess a company's moat, look past what it sells and read what its investment positions are setting up. A holdings list is a public statement of intent. Unlike guidance it does not get revised, and unlike a prepared remark it needs no rhetoric. Money sits where it sits.

The second-order effect is worth noting. The party being diluted will not sit still, and firms squeezed by the same force tend to move toward each other. So the thing to record from this week is not "who invested in whom" but "what alliance this pressure eventually forces into existence."

The community-platform acquisition is not a straightforward purchase of scale either. Talks reportedly began after other buyers approached the target (public news), which makes it closer to a defensive move. Eighteen million developers, three million models, half a million datasets — none of those numbers generate revenue. What they generate is a default: the path a new project takes the first time it downloads a model. Another developer-platform acquisition years ago worked on the same logic. You are not buying the code, you are buying the habit.

### 2. When a growth rate is truncated by supply, the number changes meaning

One detail this week carries, in my reading, the highest information density of the seven days.

At a public meeting with an investment bank, the accelerated-computing supplier explained that unconstrained, next fiscal year's revenue could grow more than one hundred per cent — and that the roughly seventy per cent it guided to publicly is what it believes it can actually supply.

Notice what that sentence does to the metric. Normally a growth rate measures demand: a high number means the market is buying. But when a company says its number is truncated by its own capacity, the metric now measures something else. It no longer tells you how strong demand is; it tells you how tight the bottleneck is. Real demand is hidden inside the clause "if we were not constrained," and the gap between the two numbers is the size of the hole still unfilled upstream.

![Two bars: the left one, labelled unconstrained by supply, is tall, while the right one, the guidance given publicly, is short, and a double-headed arrow on the right marks the height difference as the upstream gap.](/figures/growth-capped-by-supply-en.svg)

The same week, another custom-silicon and networking supplier guided to the same shape on its earnings call: AI revenue roughly doubling in each of the next two fiscal years, well above what the street had modelled, with management adding that actual demand exceeds even that outlook and the company will work to improve supply (public earnings-call remarks).

Both companies guided far above expectations. Both were sold after reporting. The contradiction dissolves once you apply the reading above: the market has already priced the supply ceiling as a known constraint, so it has stopped trading "how fast will you grow" and started trading "did this quarter's increment beat the last thing you said." When growth is locked by capacity, long-range guidance carries less marginal information, because everyone already knows the answer will be "as much as we can build."

That yields a practical move. When you see supply-constrained guidance, do not stop at the growth rate — trace backwards to the thing doing the constraining. Who is the choke? When does that layer loosen? And until it does, who is collecting the rent?

One more line from the same session gets skipped often: inference is now larger than training, where the two were roughly even eighteen months ago (public meeting notes). Training is one-off, capital-project work. Inference tracks usage, like an operating expense. That flip means compute demand has changed character from "project" to "electricity bill" — not something that ends when the build ends, but something paid for every day. For the whole upstream, that means lower volatility of demand and higher persistence.

![Two stacked bars of equal length: the upper one splits training and inference evenly, while the lower one gives inference clearly more than half, showing the ratio has flipped.](/figures/training-to-inference-flip-en.svg)

### 3. Visibility has been signed through 2029 to 2031

The third line is the easiest to overlook this week and the longest in duration.

Gather the contract years scattered across the posts: a major memory maker with roughly seventy per cent of capacity committed under long-term agreements through 2031; advanced packaging capacity booked into 2030; even legacy memory suppliers starting to allocate 2029 and 2030; optics-side long-term agreements landing in 2029 to 2030 (all from public news and public earnings materials).

Together those years say one thing. Buyers are willing to lock price and volume four to six years ahead in exchange for being able to get the goods at all.

![A timeline running from 2026 to 2032 with four horizontal bars extending from today out to 2029, 2030 and 2031, showing how long capacity is locked up.](/figures/locked-in-until-2031-en.svg)

Let me translate that into plainer language. Visibility itself has become an asset. "Good visibility" used to be an adjective in an industry note; now it is a contract with an expiry date, something you can value and something you can default on. When a company says its capacity is booked through 2030, it is handing over more than a revenue expectation — it is handing over collateral it can take to a lender. That is a large part of why this round of data-centre and compute operators can support capital spending numbers that look frightening on their own.

The same fact has a second face that gets discussed less: a long-term agreement is both protection and a lock. Capacity signed through 2031 also means pricing headroom signed away through 2031. So "won a long-term agreement" and "will make money" are two separate judgements. The contract locks volume, not margin.

Turned around, the most profitable seat often sits where there are no long-term agreements at all. This week produced a clean example. A Taiwanese legacy-memory maker reported monthly revenue of about 153 million dollars in June, 214 million in July, and 249 million in August, against roughly 35 million in the same month a year earlier — while, according to industry media, its wafer allocation this year is similar to last year's (public monthly revenue disclosures and public media reports). Allocation flat, revenue up several times over: the only variable left in the middle is price. July net income of about 109 million dollars says that increase drops through to profit largely intact.

![Four rising monthly revenue bars with a horizontal dashed line above them marking a wafer quota close to last year's, contrasting a flat quota against a revenue staircase.](/figures/quota-flat-revenue-up-en.svg)

Why would obsolete-generation product price like that? Because the new generation is lucrative enough that capacity migrated upmarket, and because Chinese memory suppliers are moving upmarket too, leaving the old product to whoever still wants it. Demand did not disappear — industrial, automotive and consumer electronics still consume it. The remaining suppliers hold pricing power not because their technology is good, but because nobody else wants the job.

An aside worth keeping: Taiwan requires monthly revenue disclosure, which hands you a high-frequency thermometer twelve times a year. Financial statements are quarterly and late. Monthly revenue is monthly, and it takes no inside knowledge. The "allocation flat, revenue multiplied" combination above is derivable from two months of public numbers.

### 4. The chokepoints are not located between the US and China

The fourth line is the most counterintuitive of the week.

The surface narrative is US–China technological conflict, and this week's events fit it: China reportedly halted some rare-earth shipments to the US (public news); the US defence establishment announced 174 million dollars to secure gallium supply, benefiting an American aluminium producer and a Japanese trading house as offtake partner (public news).

But list the links in semiconductor manufacturing where only one or a few suppliers exist, and most sit in neither country. The only company able to build extreme-ultraviolet lithography systems is Dutch, with its optics and laser source coming from two German suppliers. Four Japanese firms hold over ninety per cent of extreme-ultraviolet photoresist. One Japanese firm holds over ninety per cent of coat-and-develop equipment; patterned mask inspection is at one hundred per cent; two Japanese firms hold about ninety-three per cent of mask blanks; packaging films, glass cloth, and wafer thinning and dicing equipment cluster in Japan as well. This list requires no secrets — it comes from laying public annual reports and industry statistics side by side.

In other words, the genuine monopoly positions sit in Europe and Japan while the visible fight runs between Washington and Beijing. Using "US versus China" as your frame for supply-chain risk will therefore miss an entire row of the most fragile nodes, and miss them systematically.

![A narrow band at the top shows the United States and China facing off, while two large cards below list the segments the Netherlands, Germany and Japan hold alone, each with a bar that is almost completely full.](/figures/chokepoints-are-not-us-china-en.svg)

The week also demonstrated a better reading method: rather than tracking what China has banned, watch Japan's import figures. Japan's industrial structure determines which rare earths it consumes most, and its dysprosium and terbium come almost entirely from a single source. In June, Japanese imports of gallium, dysprosium, terbium and yttrium from China all stood at zero; one Japanese chemical company suspended acceptance of new orders for magnets containing dysprosium; another magnet producer had received no Chinese dysprosium export licences as of last month (all public news).

Policy announcements leave room for interpretation, wording preserves options, and suspensions can be reversed. Customs numbers spare nobody's feelings. When you want to know whether an export control is a gesture or an action, find the third country most dependent on the material and watch how far its imports fall. That ruler requires no inside information at all.

![Two horizontal bars of equal length: the upper one, representing how much is needed, is almost full, while the lower one, representing actual June imports, is completely empty with a zero marked at the left end.](/figures/full-need-zero-imports-en.svg)

## Going further

### Your numbers come in three different hardnesses

My guess is that after reading the above, the question in your head is not "are these conclusions right" but "with this many numbers, how do I know which one can carry an inference." I was stuck on that one for a long time.

I ended up sorting information into three layers, softest first.

The softest layer is prediction: price targets, company guidance, anything containing the word "expected." It can be revised, and revision requires an explanation to nobody. Its use is telling you what the market currently believes, not what will happen.

The middle layer is commitment: signed agreements, placed orders, announced acquisition amounts, appropriated government funding. It can be broken, but breaking it has a cost, which makes it harder than prediction. Every one of this week's 2029-to-2031 contract years lives here.

The hardest layer is physical quantities that already occurred: tonnes exported in a month, revenue for a month, units of annual capacity, customs data, container counts. It will not be revised, because it is already past tense. Its weakness is lag; its strength is that it does not lie.

![A two-axis chart where three points run diagonally from upper left to lower right: the softer the information, the more immediate it is, and the harder it is, the more it lags.](/figures/three-hardness-layers-en.svg)

There was a control case for that lag this week. Someone compiled a well-known AI-focused fund's holdings as disclosed for 30 June, concentrated in memory and compute landlords. A filing is hard information — but it was published in mid-August, describing end-June, and liquidation had reportedly occurred in July. You are looking at a photograph more than two months old. Useful for understanding what that investor was thinking in the first half. Useless for deciding what to do now.

My own practice now: build the foundation of an argument on layer three, take direction from layer two, and treat layer one as a thermometer only. If a chain of reasoning stands entirely on layer one, I go back and rewrite it.

![A timeline marking the end-of-June holdings position, the July liquidation reports, the filing published only in mid-August, and where you are reading it now, with the two-month-plus gap marked out.](/figures/filing-lag-photo-en.svg)

### Borrow the ruler, not the ticker

The second common pain point is harder to say out loud: you read a well-argued analysis, you buy along with it, and you buy the top.

One post this week spoke to exactly that. Some readers took the reasoning behind the legacy-memory work and carried it to a similar opportunity in a different market; others took the reasoning about laser supply-demand imbalance and applied it to industries they already understood. The author's reaction was pleasure — pleasure that people were learning a way of thinking rather than copying his positions.

I would sharpen that distinction. Any piece of analysis contains two things: a conclusion and a ruler. Conclusions expire, because prices move daily and you have no idea how much time separates the moment he formed the view from the moment you read it. Rulers do not expire.

Take this week. Four rulers fall out: read a company's investment positions, not just its product line; check whether guidance is truncated by supply; check what year the contracts run to; check the import data of the third country most dependent on a given material. You can carry all four into food, construction, medical devices — any industry where you know a little more than the next person.

![One piece of analysis branches downward into conclusions that expire and rightward into four rulers, which then fan out into four boxes: food, construction, medical devices and your own field.](/figures/take-the-ruler-not-the-pick-en.svg)

Here is what I've tried. After finishing an analysis that excites me, I skip the conclusion, read the piece backwards, and count how many rulers the author used, writing them on a notepad. A few days later I look at that list again. If I cannot use a single one of them, then what I took from that article was emotion.

### So much happens in a week — how do you separate the noise

The third pain point is volume. Forty-odd posts, and far more industry news than that. You cannot chase all of it.

I sort by duration: does this affect days, quarters, or years?

A stock gapping up is days. A guidance revision is quarters. Capacity contracted through 2031 is years. Within the same week, the first two routinely take ten times the attention of the third, while the third is what changes the shape of the industry. So now, reading a week's worth of material, I ask one question first: when does the effect of this end? If the answer is short, I let it go.

![Three rows compared: the bars on the left, showing how long the effect lasts, grow longer down the rows, while the bars on the right, showing the attention received, grow shorter, running in exactly opposite directions.](/figures/span-versus-attention-en.svg)

There is one category, though, that looks like noise and is worth keeping.

Two absurd items surfaced this week: an old-line consumer electronics company known for action cameras announced a merger with a photonics firm to build high-speed optical transceivers for AI data centres, and rose nearly eighty per cent in a day; and a footwear company reinvented itself as a compute landlord (both public news).

When a crowd of companies with unrelated core businesses all try to squeeze into the same layer, it tells you two opposite things and both hold. First, that layer is now collecting rent — enough of it that outsiders want a share. Second, capital has begun paying for a story rather than for capacity, which is a reading on the temperature of the cycle. The first confirms where the bottleneck sits. The second reminds you that everyone else can see it too.

My handling: write it down, act on nothing. An absurdity signal makes a good thermometer and a bad map.

## Sources, and where to go next

Every thread in this piece comes from Serenity's public timeline on X; all I did was arrange and infer, and the original judgements are his. If you want the primary posts and want to watch him work through them one at a time, following his account will serve you better than reading me — his posting density is high, he corrects himself in public, and he wrote a line I agree with: he puts this out because he finds it interesting, not as advice.

## One thing to take with you

Out of forty-odd posts, the thing I most want to leave you with is a year.

When a contract is signed through 2031, it tells you three things at once: the buyer believes the shortage lasts that long, the seller will trade that many years of pricing for certainty, and until then this position is not easily displaced. One year, three judgements.

Here is something I've tried that you can do in ten minutes. Pick an industry where you know a little more than most people — your own line of work, the street you live on, a category you have bought for years. Find one piece of public news or one public financial report from this year and look for a sentence of the form "booked through such-and-such a year." Write that year down, with today's date beside it.

If you find one, you now hold a free, public measurement of how tight that industry is. If nothing in the whole industry contains any such sentence, that is an answer too: nobody here needs to secure supply in advance.

The first time I did this, it was in a carmaker's annual report, where a component supply agreement ran three years out. I thought nothing of it at the time. A year later that component was in the news for being unavailable, and I realised the sentence had been sitting on a page I had already read. The difference was never about who got the information. It was about who treated the year as a number.