For Want of a Nail: What a Week of Public Posts on AI's Upstream Looks Like Side by Side

For want of a nail the shoe was lost; for want of a shoe the horse was lost; for want of a horse the rider was lost; for want of a rider the battle was lost; for want of a battle the kingdom was lost.

— English proverb, as printed in Benjamin Franklin’s Poor Richard’s Almanack (1758)

What this piece is about

Serenity is an independent researcher on X who follows the upstream of AI infrastructure: optical networking, memory, packaging, and power. Over the past seven days he published 28 public posts. They include upgraded numbers from brokerage reports, on-site checks that followers brought back from an optics trade show in Shenzhen, some component-price arithmetic on an old memory chip, and a complaint about paid engagement networks on X.

I laid all 28 side by side and pulled out four threads that interlock. What I want to keep is how the reasoning moves: how far down a “demand upgrade” has to be broken before it means something; whether downstream buyers can absorb a component price hike; and, with demand this strong, who takes the risk first.

A word on sources first. Every piece of material here comes from posts on X that anyone can open for free; I did not look at any of his paid content. The numbers are ones he relayed from public filings, public talks, and public news, plus publicly quoted passages from brokerage reports and public component price listings. I did not go back and verify each original report, so I cite where each number came from and readers can trace them. This piece deliberately lists no tickers, says nothing about what to buy, and gives no price targets. It stays with industry structure.

This week’s threads

1. Optics: from “more modules” all the way down to “laser prices”

The most complete chain of reasoning this week was in optical networking. It breaks into four layers.

Layer one is the total. He cited a recent Goldman Sachs note that raised its global optical module market forecast: $67.7 billion for 2026 (+33%), $131.4 billion for 2027 (+81%), and $148.5 billion for 2028 (+115%).

A total on its own carries limited weight. An upgrade can come from price assumptions or from volume assumptions, and those two differ a lot in reliability.

Layer two shows where the upgrade comes from. Goldman raised the number of optical modules modeled per GPU on a leading GPU maker’s next platform: 1.6T modules went from 0 to 2, and 3.2T modules from 3 to 5. It also raised its 2027 forecast for 1.6T-and-above by 61%, with an estimated 80% silicon photonics penetration rate.

This step turns the question from “how big is the market” into “how much light does each chip consume.” The second question is sturdier because it is tied to hardware architecture: once the architecture is set, content follows chip shipments, and there is no need to guess at prices.

Layer three is components. With more optical modules per chip and a higher share of silicon photonics, Serenity’s inference is that demand for continuous-wave lasers and silicon-on-insulator substrates scales up with it. Silicon photonics puts the light path on a silicon chip, and the light has to come in from an external source; that external source is a continuous-wave laser. So the silicon photonics penetration number connects straight to laser volume.

Layer four is price, and this is where the week got interesting. A Citi TMT note described lasers and optical fiber as being positioned as “the next HBM,” and mentioned tight supply. The same week, a follower did an on-site check for him at CIOE 2026 in Shenzhen and came back with three findings: a laser supplier is actively engaged with Chinese pluggable module makers; that supplier has raised laser prices because supply is tight; and 70mW lasers are the tightest, since they are associated with 800G pluggables. Two days later he summarized another check, this one on a large Chinese optical module maker, which explicitly confirmed that laser average selling prices are rising across the 70mW–200mW range, and the laser supplier itself disclosed that it had raised prices.

Laid out in order:

Total forecast raised → more content per chip → more laser demand → prices already raised → gross margin expansion in coming quarters

Serenity’s wording for the last step is that the higher pricing “should support gross margin expansion for laser suppliers in future quarters as higher pricing flows through.” The price increase happens this quarter; the margin shows up in later quarters. My understanding is that old contracts and inventory have to run off before the new prices reach the financial statements. He did not spell that part out.

What I value most in this chain is knowing where it would break:

• The upgrade stops at the total. If content per chip doesn’t move, price assumptions are holding the number up, and it deserves a discount.
• Price increases show up in only one spec. If only 70mW were rising, that would be a single-spec shortage that fades once capacity catches up. The second check says the whole 70–200mW range is rising, which is a wider footprint than one spec. That is what the second check adds over the first.
• Small sample. The checks come from one trade show and a handful of followers. He said himself that this was the first one he saw and he may have missed others.

These three conditions decide whether this is a structural shortage or one quarter of price movement. This week’s evidence leans toward the former, but the sample can’t carry a conclusion yet.

Why Citi’s comparison carries weight becomes clearer in the second thread.

2. Memory: whether a price hike passes through depends on its share of the finished product

Memory had two examples this week, one expensive and one cheap. Only side by side do they turn out to be making the same point.

The expensive end. He relayed a Reuters report: high-bandwidth memory has pushed up AI accelerator card prices in China by 20% to 50%. One major Chinese chipmaker raised indicated pricing on a new accelerator 20–50% above quotes from just two months earlier; another’s upcoming card is reportedly up 20–30%. High-bandwidth memory obtained through grey-market channels in China can cost several times the regular price.

The cheap end. One old DDR2 memory chip lists at about $0.96 on a public component pricing platform, with some specs as high as $2.50. Serenity did the arithmetic: even if that chip’s price tripled, the finished product’s bill of materials would rise by about $2, roughly the price of a warehouse-club hot dog. Chips like this go into products such as network security cameras that sell for around $33. Add $2, and buyers won’t notice.

Both examples are about pass-through. A component’s share of the finished product’s price decides whether downstream buyers pull back when the supplier raises prices.

• The old chip is about 3% of a $33 camera. After tripling, it’s about 8%. The camera still gets built and still gets sold.
• High-bandwidth memory is a far larger share of an accelerator card, so when it rises, the card’s quoted price moves up 20–50%. Grey-market prices several times normal show that someone is still willing to pay after the markup.

Together they give one test: when a component’s share is small, downstream absorbs the increase quietly; when its share is large, the increase shows up in the finished product’s price, and the question becomes whether the product’s buyers have alternatives. This week, buyers of AI accelerator cards in China accepted the markup.

Back to optics. Read through this test, Citi’s “next HBM” comparison says lasers are moving from “a small part nobody thinks about” toward “tight supply, rising quotes, starting to affect finished-product prices.” Where the comparison finally lands depends on how much of an optical module’s cost the laser represents, and this week’s posts didn’t give that number. It’s the gap I most want to fill after reading.

3. Demand visibility stretches out; the risk lands on whoever carries debt

Three posts this week were about how strong demand is.

• Usage side: he said demand for Astra is unprecedented, to the point that OpenAI might pause new Pro subscriptions. His translation: “AI buildout + compute go brrr.”
• Supply side: he said many semiconductor companies now have visibility into 2029–2031, and capex shows no sign of slowing.
• Timeline: in an economic scenario publication from Anthropic, the most extreme scenario has GDP growth reaching 15.4% in 2030, around 7.3 times the current rate of about 2.1% a year. What Serenity noticed was the shape of the curves: the scenarios barely separate in 2026–2027, then bend sharply apart in 2028–2030. His read is that 2026–2027 are buildout years and 2028–2030 are years of economic acceleration.

With demand this strong, an easy-to-skip question comes up: where is the risk? His answer was direct. The risk sits in the segments with leverage or financing risk, such as neoclouds or frontier labs with too many obligations. Semiconductor firms at 3 to 15 times forward earnings, bringing in large operating income every quarter, are not on his list of worries.

The reasoning behind that runs like this: in a buildout, money goes out first and comes back later. If the spender is using borrowed money or has signed long-term payment obligations, a revenue delay of a year or two is enough to break its cash flow. The side selling equipment and components gets paid every quarter, so a delay hurts it much less.

The next step is my own reading, connecting two of his posts; he didn’t link them: if Anthropic’s curves put “the economy starts accelerating” after 2028, then “who makes it through 2026–2027” is the question that matters most in this stretch.

Two deals from the same week fit into this frame:

• A cloud giant received warrants to buy roughly a $4 billion stake in a mobile chipmaker, tied to up to $60 billion in custom silicon milestone revenue. He joked that with equity or warrants already held in several chip and connectivity suppliers, the cloud giant is turning into a semiconductor ETF. His verdict: good for both companies, more so for the chipmaker.
• A neocloud partnered with a data software company that has deep ties to the US government, and was named that company’s preferred sovereign AI infrastructure partner. His verdict: given the partner’s existing government and enterprise relationships, the implications could be significant.

My reading (again, not something he wrote): a large buyer that trades purchase commitments for warrants gets back a slice of the supplier’s future gains in exchange for the money it pays out, which hedges the buyer against the supplier’s pricing power. A neocloud that teams up with a partner holding government customers is looking for steadier revenue to stand behind its debt. Both deals answer the same question: in a buildout, who pays first, and who carries the risk of delay.

The timing gap shows up on the supplier side too. He mentioned that one optical company has a $600 million at-the-market offering, which creates near-term selling pressure (usually a month or so), while he puts that company’s fundamental inflection point around the end of Q2 2027. I read the two timelines separately: pressure from the supply of shares is measured in weeks, and the turn in product supply and demand is measured in quarters.

4. Information has a supply chain too

He spent several posts this week on something that has nothing to do with chips.

He described how paid engagement networks on X work: companies pay influencers to comment on or quote their posts (now sometimes labeled); they spin up affiliated accounts named things like ”___ Finance” or ”___ [country]”; once each one grows, they quote each other to boost engagement. In the end every affiliate account has tens or hundreds of thousands of followers, which then feed visibility back to the main account. He also separated two kinds of inauthentic engagement: farming engagement to sell courses or build a network, and short sellers using dozens of old crypto marketing accounts to spread disinformation about a company.

A few other posts from the same week read well alongside:

• His “bottleneck thesis” reportedly made it into a quant fund’s research pipeline; he said he hoped the institutions enjoy his takes on DDR2, continuous-wave lasers, and indium phosphide substrates.
• He thanked followers for bringing back on-site checks from the Shenzhen show, saying the small community reaches all the way around the world.
• He said retail deserves the same research and early thematic ideas that institutions get first, and that even at a million followers, that remains his priority.
• He joked that short-dated tanker freight futures ETFs outperformed all of his ideas, and that he hadn’t known they existed until the week before. His takeaway: there are always more obscure things to learn.

Side by side, I see one researcher’s output amplified through three channels in a single week: an institutional research pipeline, followers’ on-site checks, and paid engagement networks. The first two bring back things that can be checked (report passages, trade-show prices). The third brings back volume. For readers, the bottleneck in information has moved from “can’t get it” to “can’t tell it apart.”

I want to keep the tanker post too. How deeply you think about a theme and what happened to rise most in that window are two separate things. He closed with a joke at his own expense and didn’t defend his ideas.

Further thoughts

When an institution raises its numbers, how do I tell real demand from mood?

A report says the market forecast is up 115%. The first reaction is excitement, the second is doubt, and then you’re stuck, because there’s no tool at hand to judge it. I get stuck there often.

The laser chain this week gave me a way to break it down, as three questions:

1. How much did the total change? This layer is the cheapest. Change one assumption and the forecast goes up, so it gets the lowest weight.
2. Did content change, or price? Content is tied to hardware architecture and moves only when the architecture does. Price is an assumption and can be rewritten overnight.
3. At the bottom of the stack, have component prices moved? A price quote means someone is paying. This layer costs the most to produce and is the hardest to fake.

Lasers had something in all three layers this week. If I can only fill in the first layer, I treat the upgrade as mood.

The third layer has limits of its own. Trade-show checks are small samples and may carry the checker’s hopes. They are evidence and can’t yet carry a conclusion. So even with all three layers filled, I don’t drop my doubt. I shift it from “is there demand” to “how long can this shortage last.”

What I’ve tried: every time I see an upgrade, I open three boxes in my notes and leave blank whatever I can’t fill. The more blanks, the more reserved I am about the number. The method doesn’t tell me what to do. It keeps one big number from pulling me along.

If a component gets more expensive, will it scare off buyers downstream?

When you see “some part is up 3x,” it’s easy to jump straight to worry: downstream can’t afford it, so demand dies.

The two memory examples helped me split that worry into two questions:

• What share of the finished product’s price is the component? A 3% part that triples adds about 5 percentage points of cost to the product, and most buyers won’t notice. A part that makes up 30% of the price and rises 20% forces the finished product’s price to move.
• Do the finished product’s buyers have alternatives? When a large-share component rises, it comes down to whether buyers will swallow it. This week, buyers of AI accelerator cards in China swallowed it, and grey-market prices are the evidence.

There’s also a timing gap. In the quarter prices go up, the supplier’s financial statements may not show it yet, because old contracts and inventory delay the effect. So “checks say prices are rising” and “the financials haven’t moved” can both be true at once. There’s no need to rush into deciding which side is wrong.

My own uncertainty, stated plainly: I don’t know what share of an optical module’s cost the laser represents, and this week’s posts didn’t give it. That number decides whether the “next HBM” comparison holds up, and it’s the first thing I’ll go looking for.

When a crowd on my timeline is pushing the same theme, whom do I believe?

When ten accounts talk about the same theme on the same day, it looks like consensus. The mechanism Serenity described this week shows that those ten accounts might be one network.

The filter I’ve tried is one question: does this post contain anything someone else could check? A number with a source, a trade-show price, a report passage whose original can be found. If yes, it’s information. If no, I treat it as volume first, however many times it’s been shared.

The material for this piece went through the same filter. Among Serenity’s 28 posts this week were research, jokes, condolences to strangers, and his views on one brokerage’s strategy. I built the threads only from the ones carrying checkable numbers.

One thing to take with you

Next time you see news that some component is short or getting more expensive, find three numbers and write them down: the component’s unit price, the price of the finished product it goes into, and how much the latest quote changed.

That’s how the old memory chip arithmetic worked: $0.96, $33, up threefold. Put those three numbers next to each other and you can see whether the increase will pass through. If I can’t find even one of the three, I set the news aside until the numbers show up.

Serenity publishes his original reasoning and follow-ups publicly on X. If you want to see them earlier than my summary, go follow his account directly.

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.