The Fifteen-Fold Nobody Priced: Nvidia, SPIL, and the Physical Bottleneck Underneath Crypto's AI Trade

Fifteen times. Not fifteen percent. Fifteen times.

That is the number sliding out of Taipei this week — Nvidia and SPIL, the Taichung-based packaging arm of ASE Technology, reportedly planning to scale co-packaged optics capacity fifteenfold by mid-2027. No earnings call. No 8-K. No press release with a legal disclaimer stapled to the bottom. Just a supply-chain whisper that traveled from a Taiwanese trade desk into a crypto media outlet and landed in my feed at 2 a.m. Auckland time.

I read it twice. Then I opened a blank document and started typing, because this is the kind of number that gets repeated for eighteen months before anyone checks what it actually means.

Here is what it does not mean. It does not mean Nvidia is building fifteen times more GPUs. It does not mean the AI compute shortage is solved. It does not mean your favorite AI token is about to re-rate. What it means is far narrower and far more interesting: the most under-priced bottleneck in the entire AI infrastructure stack is not silicon, not memory, not wafer capacity — it is the sub-micron alignment of a glass fiber to a photonic chip, done at volume, with acceptable yield, by a company most crypto traders have never heard of.

And that bottleneck sits inside a supply chain with exactly one meaningful geographic concentration, one dominant customer, and one raw material that China controls roughly eighty percent of.

I didn't wait for the confirmation email on this one. In a bear market, confirmation arrives after the move.

So let's take it apart. Slowly. Because the crypto AI trade has been running on vibes for two years, and vibes do not survive contact with a fiber coupling machine that costs more than a Manhattan apartment and takes twelve months to deliver.

The Number Is a Packaging Number, Not a Wafer Number

First, the mapping problem. Every semiconductor story gets read through the wrong lens by default, and this one is worse than most.

When you hear "capacity expansion" in semiconductors, your brain reaches for wafer starts per month. Fabs. Cleanrooms. The Taiwan Semiconductor manufacturing cadence that the entire crypto mining industry learned to read during the 2021 GPU drought. That instinct is wrong here.

SPIL is not a foundry. SPIL is an OSAT — outsourced semiconductor assembly and test. Its job in the co-packaged optics value chain is the back end: taking a photonic integrated circuit that somebody else fabricated, taking a switch ASIC that somebody else fabricated, and physically marrying them into a single package with a fiber array attached to the edge, tested, burned in, and shipped.

The wafer-level work — the silicon photonics PIC, the TSMC COUPE platform — stays with TSMC. The assembly, the fiber attach, the optical coupling, the final test — that is SPIL.

So when the report says fifteenfold, it is saying something very specific: the fiber attach and optical packaging line capacity is going up fifteen times, and it is landing in the second half of 2027.

That timing is not random. Nvidia's photonics roadmap has Quantum-X shipping in the back half of 2025 and Spectrum-X Photonics in the back half of 2026. Rubin lands in 2026. The generation after Rubin — the one that will demand photonics at genuine scale rather than as a flagship curiosity — lands right around 2027. The fifteenfold expansion is not a response to current demand. It is a forward bet placed eighteen to twenty-four months ahead of the platform that needs it.

That is how this industry actually works, and it is why the number matters. Nobody builds fifteenfold optical packaging capacity for a sample program.

What Co-Packaged Optics Actually Replaces, and Why It Is Hard

Let me do the part that most crypto-native coverage skips, because it is the part that determines whether this trade is real.

Right now, the optical connection inside an AI cluster looks like this. A switch ASIC sits on a PCB. Electrical traces run from the ASIC to the front panel. At the panel, a pluggable optical module — QSFP-DD or OSFP, currently 800 gigabit moving to 1.6 terabit — converts the electrical signal into light and fires it down a fiber.

That architecture is beautiful because it is modular. A module dies, you pull it out and put in a new one. A vendor ships bad units, you switch vendors. Field service is a technician with a spare part and a torque screwdriver.

It is also power-hungry. The electrical path from ASIC to panel burns somewhere in the range of fifteen to twenty picojoules per bit. At 800G per port and tens of thousands of ports per cluster, that number stops being a datasheet footnote and becomes a thermal budget problem. In a modern AI cluster, the network layer — NVLink plus InfiniBand or Ethernet — has climbed to somewhere between ten and twenty percent of total cluster power draw.

Co-packaged optics attacks that directly. Instead of driving electrical signals across a PCB to a panel module, the photonic engine sits in the same package as the switch ASIC. The electrical trace shrinks from inches to millimeters. Power per bit drops toward five picojoules, potentially lower. Bandwidth density goes up. Latency goes down. Signal integrity problems that plague high-speed copper largely evaporate.

On paper, it is an obvious win. In practice, four things make it brutally difficult, and all four live inside SPIL's scope.

One: photonic IC yield. Silicon photonics is not logic. You are building waveguides, modulators, photodetectors, and grating couplers on a silicon wafer, and every one of those structures has process sensitivity that logic designers find alien. Yield on advanced photonic PICs is improving but remains well below what a mature logic process delivers.

Two: fiber attach. This is the killer. You are aligning an array of optical fibers — a fiber array unit, FAU — to grating couplers on the PIC with sub-micron precision, and you are doing it at production volume, with an adhesive cure step, with thermal expansion mismatch between glass and silicon, and with a tolerance window measured in fractions of a micron. The equipment that does this — high-precision die bonders and fiber alignment systems from ASMPT, BESI, ficonTEC, MRSI — has lead times commonly running six to twelve months and costs that make a crypto miner's GPU budget look like pocket change.

Three: thermal management of the laser. The light source itself, typically an external laser source using EML or DFB lasers, is thermally sensitive and generally kept outside the hot package. Routing light from an external source into a co-packaged engine without losing too much of it is its own engineering problem.

Four: maintainability. This is the one that decides the adoption curve, and almost nobody in crypto talks about it. In a pluggable world, a broken fiber means replacing a module. In a co-packaged world, a broken fiber can mean scrapping an entire switch package — an ASIC that costs five figures, a photonic engine, and a precision-aligned fiber assembly, all written off because someone yanked a cable too hard during a rack move.

Hyperscalers are obsessive about field serviceability. They have built their entire operational model around it. Telling them that a single fiber break now means a multi-thousand-dollar package replacement is not a small ask.

That is the real adoption gate on co-packaged optics — not whether it works in a lab, but whether a data center technician can service it at 3 a.m. without destroying five figures of silicon.

The Capacity Math Nobody Published

The original report gives us exactly two anchors: fifteen times, and mid-2027. It gives us no absolute numbers. No ports per month. No wafer starts. No capital expenditure figure. No customer commitment.

That absence is itself information.

Fifteen is not a number that comes out of a market model. Market models produce round numbers — two times, three times, ten times. Fifteen comes out of a supply chain conversation where somebody said "we are going from X to fifteen X" and the X was small enough that the multiple sounded dramatic. Which is exactly what I would expect if the current baseline is a sample-scale or pilot-scale line producing thousands to low tens of thousands of ports per month.

Read that way, fifteenfold expansion is not a story about co-packaged optics going mainstream. It is a story about a technology moving from zero to one. From engineering samples to a real production line. The absolute number in 2027 may still be small relative to the total optical interconnect market, which is dominated by pluggables shipping in the tens of millions of units annually.

This is the single most common error I see in crypto infrastructure coverage: treating a dramatic multiple as a dramatic absolute. Fifteen times a small number is still a small number. If the baseline is fifteen thousand ports a month, the target is two hundred twenty-five thousand. Meaningful. Not transformative. Not yet.

The capex picture reinforces this. ASE Group's overall capital expenditure runs in the low billions of US dollars annually. A co-packaged optics line is a fraction of that — the equipment is expensive per unit but the line footprint is nothing like a fab. What matters is not the dollar figure. What matters is the depreciation schedule. Assembly line equipment typically depreciates over five to seven years. If the line is installed through 2026 and 2027 and does not reach high utilization until 2028, the depreciation drag lands squarely on ASE's 2026 and 2027 gross margin, which currently sits in the sixteen to eighteen percent range at the group level.

That is a margin compression story dressed up as a growth story. Investors who buy ASE on the headline will be buying revenue growth with deteriorating profit quality. Those are not the same thing, and the market routinely confuses them.

The Demand Side Is Real, and That Is the Problem

Here is where I have to be careful, because the bear case writes itself and the bear case is not quite right.

The demand pull is genuine. Hyperscaler capital expenditure — Microsoft, Google, Meta, Amazon combined — is tracking toward the three hundred billion dollar range annually across 2025 to 2027. AI training clusters have crossed the ten thousand GPU threshold and are heading higher. At that scale, the network stops being plumbing and becomes the binding constraint. You cannot build a hundred-thousand-GPU cluster on pluggable optics without drowning in power and thermal problems at the network layer.

So co-packaged optics is not a solution looking for a problem. The problem is real, it is measured in megawatts, and it is getting worse every generation.

But real demand is not the same as realized demand, and there is a specific mechanism by which this trade goes wrong.

Co-packaged optics has a competitor that is not standing still. Linear pluggable optics and linear receive optics — LPO and LRO — remove the DSP from the pluggable module, cutting power substantially while keeping the modular, serviceable, multi-vendor architecture that hyperscalers love. If LPO gets good enough, fast enough, a meaningful chunk of the co-packaged optics use case evaporates before it ever scales. Not because co-packaged optics failed technically, but because the boring incremental solution arrived first and was good enough.

That is how most technology transitions actually die. Not with a dramatic failure. With a good-enough alternative that ships two years earlier.

And there is a second mechanism: ecosystem lock-in resistance. Co-packaged optics ties the optical engine to a specific switch ASIC vendor. A hyperscaler that builds around Nvidia's co-packaged photonics is building around Nvidia's photonics, full stop. No second sourcing the optics. No swapping module vendors when one underperforms. No price competition at the module layer.

Hyperscalers have spent a decade deliberately structuring their supply chains to avoid exactly this. They will pay a premium for optionality. They will pay a larger premium to avoid lock-in.

So the fifteenfold number is a capacity bet, not a demand guarantee. Those are two different things, and the gap between them is where the money gets made and lost.

The Geography Problem, and the Material Nobody Prices

Now the part that should worry anyone holding AI infrastructure exposure through a token, an equity, or a fund.

SPIL's co-packaged optics capacity is concentrated in Taichung, Taiwan. Not distributed. Not diversified. Concentrated.

Nvidia's public position on supply chain diversification is well documented and entirely reasonable. The company has been pushing packaging capacity toward Amkor's Arizona facility and ASE's Kumamoto operations in Japan. Those are real moves. They are also slow moves, and they are not where the initial fifteenfold expansion lands.

So we have a situation where the single most important new interconnect technology for AI infrastructure is being stood up, at scale, in one city on an island with a well-known geopolitical risk profile. That is not a prediction of catastrophe. It is a statement about tail risk concentration, and tail risk is exactly what markets are worst at pricing until it is too late to price it.

Then there is the material layer, which almost nobody in crypto is tracking.

Gallium. China controls roughly eighty percent of global gallium production. Gallium is a critical input for compound semiconductors — the EML and DFB lasers that sit inside every photonic engine, the VCSELs used in shorter-reach optical links. China has already placed export controls on gallium and germanium. Those controls have been calibrated rather than absolute, but the lever exists and everyone in the industry knows it exists.

Follow the chain. Chinese gallium export restriction tightens. Laser source availability tightens. Photonic engine production tightens. Co-packaged optics ramp slows. Nvidia's photonics roadmap slips. AI cluster network deployment slips. And every crypto asset whose thesis is "AI compute demand goes up forever" takes a hit from a direction nobody modeled.

That is a second-order dependency chain that runs from a mine in Guangxi to a token chart in your portfolio. Most people holding the AI narrative have never drawn that line.

There is one more layer. Advanced packaging equipment — the high-precision bonders and alignment systems — comes predominantly from Europe and Singapore. Not currently restricted for Taiwan delivery. But the export control framework around advanced packaging has been tightening steadily across 2022, 2023, and 2024, and co-packaged optics assembly is not explicitly carved out as permanently exempt. It is simply not yet named.

The supply chain risk here is not a single point of failure. It is four points of failure stacked vertically, and each one is controlled by a different government.

Broadcom Got There First, and That Matters More Than It Sounds

Competitive positioning, because this shapes which standard wins and standards determine who captures the margin.

Broadcom shipped co-packaged optics switching silicon first. Bailly, built on Tomahawk 5, went into production in 2024. Davisson followed. Broadcom is roughly a year ahead of Nvidia in co-packaged optics switching, and it has been selling into the same hyperscaler base for longer.

Nvidia's counter is not technical superiority at the package level. Nvidia's counter is end-to-end control. Nvidia owns the GPU, owns NVLink, owns the network fabric, and can bundle. If you are buying a rack of GB200 NVL72 systems, the network decision is not fully yours. That bundling power is a real moat, and it is the reason Nvidia can enter a market a year late and still take meaningful share.

But bundling power has a boundary. It works where Nvidia controls the full stack. It works less well in open Ethernet deployments where a hyperscaler is deliberately mixing vendors to maintain leverage. And the Ethernet side of the AI network market is exactly where Broadcom is strongest.

So the realistic picture is dual leadership. Broadcom leads in open, merchant co-packaged optics. Nvidia leads where it can bundle. Both expand. Neither dominates. And the OSAT layer — SPIL, Amkor, and eventually others — captures a modest, technically scarce, but structurally weak slice of the value.

Weak because of customer concentration. SPIL's co-packaged optics capacity is being built for one customer. That customer represents, by any reasonable estimate, more than half of the line's addressable output. The moment Nvidia qualifies a second source — Amkor Arizona, or a Foxconn entity, or an in-house line — SPIL's utilization risk becomes acute.

Being the sole qualified supplier to a monopolist buyer is a wonderful position right up until it is the worst position in the industry. There is no pricing power in that relationship. There is no volume guarantee. There is only the privilege of having been chosen, revocable at will.

Why Any of This Matters to a Crypto Portfolio

I know what you are thinking. This is a semiconductor supply chain story. I hold tokens. What does a fiber alignment machine in Taichung have to do with my bags?

Everything, if your bags are in the AI-adjacent crypto sector. And in a bear market, that is most of the sector that still has a bid.

Let me be concrete about the transmission channels, because vague macro hand-waving is how people lose money while feeling informed.

Channel one: the compute-DePIN tokens. Every decentralized physical infrastructure network selling GPU compute is, whether it admits it or not, competing on the margin with centralized AI infrastructure. If co-packaged optics reduces network power per bit by two to three times, the cost structure of a large centralized AI cluster improves. That makes centralized compute cheaper, which compresses the price umbrella that decentralized compute networks have been hiding under. The DePIN compute thesis depends partly on centralized infrastructure being expensive and bottlenecked. Every efficiency gain at the network layer tightens that spread.

Channel two: the AI agent tokens. The entire narrative of autonomous agents transacting on-chain assumes inference gets cheap and ubiquitous. Cheap inference requires cheap, high-bandwidth, low-power networking at scale. Co-packaged optics is a genuine enabler there. But it is an enabler on a 2027 timeline, which means the tokens pricing that future today are pricing something that has not yet been de-risked.

Channel three: the mining sector pivot. Bitcoin miners who have been repositioning as AI hosting providers are, functionally, real estate and power companies now. Their economics depend on how much compute they can host per megawatt. Network power efficiency improvements change the calculus of what a megawatt is worth. A miner who signed a ten-year hosting deal at a fixed power envelope is exposed to the possibility that the same megawatt hosts substantially more compute in 2028 than it does today, meaning their contract is worth less than they modeled.

Channel four: the reflexive narrative layer. This is the one that actually moves prices in the short term, and it is the least rigorous. When Nvidia photonics news hits, AI-adjacent tokens pump. Not because anyone did the transmission analysis above. Because the word Nvidia appeared. That reflex is a trading opportunity and a trap simultaneously.

I have watched this movie. In 2021, I watched DeFi tokens pump on the announcement of a partnership that, when I actually read the terms, amounted to a logo on a website. Community buzz wasn't the signal. It was the noise that preceded the signal by about six weeks.

That is what I expect here. A pump on the headline. A slow deflation as people realize the fifteenfold number is a small base and a 2027 date. Then a real move, much later, when the actual capacity comes online and the actual revenue shows up in an actual quarterly report.

The Contrarian Read: The Bottleneck Is Not Where the Market Is Looking

Here is the angle I have not seen anyone publish.

Every crypto and tech analyst covering AI infrastructure has spent two years fixated on CoWoS — TSMC's chip-on-wafer-on-substrate advanced packaging — and HBM supply. Those are the famous bottlenecks. They are the ones that show up in every "AI supply chain" thread, every research note, every token whitepaper that wants to sound technical.

They are also, increasingly, the bottlenecks that are being solved. CoWoS capacity has been expanding aggressively for three years. HBM supply has attracted enormous capital. The famous bottlenecks attract capital, and capital dissolves bottlenecks.

Co-packaged optics fiber attach is the opposite. It is unfamous. It is unglamorous. It is a room full of precision alignment machines and process engineers arguing about adhesive cure profiles. Nobody is writing threads about adhesive cure profiles.

And that is precisely why it is the binding constraint.

The bottleneck in AI infrastructure has moved from the front end — wafers, memory, logic packaging — to the back end: optical assembly, fiber coupling, and test. The market has not repriced this because the market does not have a mental model for it.

This is structurally identical to a mistake I watched the crypto industry make repeatedly. Everyone obsessed over layer-one throughput. Transactions per second. The famous bottleneck. Meanwhile the actual constraint was data availability and settlement cost, and the industry poured billions into the wrong layer before figuring that out.

I have said for a while that most rollups do not generate enough data to justify dedicated data availability layers. That is not a popular view. It is, however, the view that the on-chain data supports. The famous bottleneck gets the capital. The actual bottleneck gets ignored until it breaks something.

Same pattern here. The famous bottleneck is CoWoS. The actual bottleneck is a fiber array unit aligned to a grating coupler at sub-micron tolerance, produced at volume, by a company with one customer and one factory.

And there is a deeper structural point. Co-packaged optics does not just move the bottleneck. It changes the failure mode. With pluggables, a bad batch is a warranty claim. With co-packaged optics, a bad batch is a scrapped switch package. The industry is trading modularity for efficiency, and it is doing so at exactly the moment when AI clusters are getting large enough that field failures scale linearly with deployment size.

A hundred-thousand-GPU cluster has, conservatively, tens of thousands of optical links. At any meaningful per-link failure rate, the service burden is enormous. Co-packaged optics makes each failure more expensive to remediate. That math works in a lab and it works in a pilot. It gets uncomfortable at scale.

Nobody in the AI token complex is pricing the serviceability problem. They are pricing the picojoules.

What I Actually Did, and What It Taught Me

Based on the week I spent running autonomous trading agents on a testnet earlier this year, I have a bias I should disclose.

I did not audit the code. I am not that person. I ran the agents, watched them make irrational trades, lost money to a neural network that had clearly never encountered a liquidity crunch, and wrote about how absurd the whole thing felt. What I learned in that week was not technical. It was structural.

The agents were fast. They were not wise. They optimized for the metric they were given and they were catastrophically wrong about everything adjacent to it. One of them kept buying into thin order books because the momentum signal said buy, and the momentum signal did not know what a thin order book was.

AI infrastructure investing has the same failure mode right now. The market is optimizing for the metric it can see — capacity multiples, capex announcements, partnership headlines — and it is blind to everything adjacent. Serviceability. Customer concentration. Depreciation drag. Gallium. Fiber breakage in a rack move at 3 a.m.

Speed isn't a substitute for seeing the whole board. It's about feeling the market before the market feels itself, and the market is not feeling the fiber coupling constraint yet.

I have been doing this for twelve years. I was in an Austin hacker house in 2017 during the Ethereum Classic hard fork, and I broke the story in fifteen minutes not because I understood the protocol better than anyone else in that room — I did not — but because I was listening to the Telegram voice chats instead of reading the technical documentation. I heard the panic in someone's voice before the block explorer confirmed anything.

That instinct has served me. It has also burned me, because sometimes the panic in someone's voice is just a person having a bad day. The discipline is knowing which is which. Here, the discipline says: the number is real, the timeline is real, and the market's interpretation of both is wrong in a specific and identifiable direction.

The Three Things I Am Watching, and the Timeline for Each

The bear market has taught me one thing above all: distraction is a luxury we can't afford. So let me be surgical about what would change my mind in either direction.

Signal one, near term. Whether ASE discloses co-packaged optics capital expenditure in its quarterly reporting, and at what magnitude. If the line is genuinely fifteenfold, the capex has to show up somewhere. A vague mention in an investor deck is not confirmation. A specific number with a specific timeline is. Watch the quarterly calls, not the trade press.

Signal two, medium term. Whether Nvidia names a second co-packaged optics assembly source by the end of 2026. If it does, SPIL's scarcity premium collapses and the whole "fiber coupling is the moat" thesis weakens. If it does not, the concentration risk becomes the dominant story — one factory, one customer, one island.

Signal three, long term. Whether linear pluggable optics closes the power gap faster than co-packaged optics scales. This is the quiet killer. It will not announce itself. It will show up as a series of conference papers at OFC, then as a product line, then as a design win, and by the time it is obvious the trade is over. Watch the pluggable roadmap with more attention than the co-packaged roadmap, because the incumbent is the one with the installed base.

And one meta-signal that matters more than all three. Where the fifteenfold number came from. The original report originated in a crypto-native outlet, which almost certainly means it was adapted from Taiwanese supply chain reporting — DigiTimes, the economic dailies, the trade press that actually talks to the equipment vendors. That is a second-hand chain. Second-hand chains distort. A "capacity planning target" becomes a "confirmed expansion" somewhere between the third and fourth retelling. Treat the number as a directional signal from an unnamed source, not as a disclosed fact.

I would rather be slow and right about this than fast and wrong. Which is not something I say often.

The Part Where I Tell You What It Means for Your Bags

The bear market question is always the same question. Not "what is the upside" — "is my position safe."

So here is the honest answer, structured for that question.

If your exposure is to AI-adjacent tokens whose thesis rests on centralized AI infrastructure being permanently bottlenecked and expensive: this news is mildly negative for you, on a two-year horizon. Every efficiency gain at the network layer compresses the spread that decentralized compute has been living in. Not fatal. Not immediate. Structural.

If your exposure is to tokens whose thesis rests on AI inference becoming cheap and ubiquitous: this news is mildly positive, but the timeline is longer than the price implies. Co-packaged optics at scale is a 2027 to 2028 story. If your token is pricing 2028 in 2025, you are early in a way that looks identical to being wrong.

If your exposure is to the mining-to-AI-hosting pivot: watch your power contracts. Efficiency gains at the network layer do not change your megawatt envelope, but they change what a sophisticated counterparty will pay for that envelope in a renegotiation. Do not sign long-dated fixed-compute contracts against a technology curve that is bending down.

If your exposure is to nothing in this sector and you are just reading to understand the machine: the takeaway is that the AI trade's physical constraints are migrating backward through the stack, away from the famous chokepoints and toward the unglamorous ones. The people who understand fiber coupling will outperform the people who understand tokenomics over the next three years. That is uncomfortable and I believe it.

The Thing Nobody Says Out Loud

Co-packaged optics is going to happen. I am not a skeptic on the technology. The power math is too compelling, the bandwidth density is too compelling, and the AI cluster scaling curve is too steep for pluggables to carry alone forever.

The question was never whether. It is when, and who captures the margin when it does.

And on that question, the fifteenfold number tells a story that is much less flattering than the headline. It tells you that the technology is at the zero-to-one stage in mid-2025, that the scale line lands in 2027, that the assembly is concentrated in one city, that the customer is concentrated in one company, that the critical materials run through one country, and that the whole thing depreciates over five to seven years regardless of whether demand shows up.

That is not a slam dunk. That is a leveraged bet placed by a company with weak bargaining power, on behalf of a customer with total bargaining power, on a technology whose main competitor is improving every quarter.

It might work beautifully. ASE is a serious operator with real engineering depth, and SPIL's fiber coupling know-how is genuinely scarce. Scarcity is worth something.

But scarcity in a supplier with one customer is not a moat. It is a hostage situation with good optics.

Takeaway

The signal is not the fifteenfold. The signal is that the AI infrastructure buildout has moved past the era where the constraints were famous. CoWoS was famous. HBM was famous. Both got capital and both are dissolving.

What is left is a room in Taichung full of alignment machines, a fiber array unit with sub-micron tolerances, a laser source that depends on a metal China mostly controls, and a serviceability problem that no hyperscaler has signed off on at scale.

That is the actual frontier. Not the token. Not the chart. The room.

When the next AI narrative cycle arrives — and it will, they always do — the people who will be right are the ones who spent this bear market learning where the room is, not the ones who spent it refreshing the chart.

I don't wait for the signal. At some point, if you are paying attention, it becomes the signal.

So watch the quarterly calls. Watch the equipment order books. Watch whether a second source gets named. Watch the gallium.

And when someone tells you the AI trade is a straight line up, ask them one question.

Ask them what happens when a fiber breaks.