In July I opened the Pokédex on Nebius and found a strange creature. A company trading at roughly 31 times its 1.92 billion dollar ARR, promising that ARR would reach 7 to 9 billion by year end. My conclusion was simple. Expensive today, potentially cheap tomorrow, assuming the buildout actually lands.

On August 12 Nebius dropped Q2. The stock ripped 34 percent in a single session and closed at 259.20, gave a little back the next day, and sits at 274 as I write this on August 17. The options market had been pricing a 13 percent move into the print. It got two and a half times that.

So no, the market did not sleep on this print. Time to check the stats again anyway, because a 34 percent day tells you what the crowd felt. It does not tell you what changed in the model.

_{NBIS USED Q2 EARNINGS}

Only the numbers that touch the valuation.

Stat upgrade

• Group revenue: 582.3 million, up 454 percent year over year, up 46 percent from Q1

• AI cloud revenue: 574.9 million, roughly 98 percent of the group

• ARR: 3.0 billion, up 56 percent from 1.92 billion at Q1

• AI cloud adjusted EBITDA margin: 49.7 percent. Group adjusted EBITDA margin: 41 percent

• Four landmark AI cloud deals, average TCV above 1 billion each

• Contracted power target raised again to 5 GW by year end 2026

• Total customer commitments: more than 40 billion

• Full year 2026 guidance reaffirmed across all metrics

Note the two margin numbers. The 50 percent everyone is quoting is the AI cloud segment. The group did 41 percent. Both are real, they are just not the same animal, and the 2030 model runs on the group.

Nebius didn’t just gain XP. It learned a few new moves.

_{THE NEW MOVES}

In July I named the things that had to go right. Three of them moved.

Move 1: demand stopped being theoretical

My July bear case was blunt. Seven to nine billion of ARR is guidance, and guidance is not cash.

Q2 put signatures behind it. Four deals above 1 billion each in total contract value, which is everything a customer commits to pay across the life of a contract. Reflection and Cohere are named, plus an unnamed US frontier lab and a large quant trading firm. Contract value signed in the quarter grew nearly 4x, and from new customers more than 9x. Management says it could sell all of 2027 today and is holding capacity back on purpose.

July question: can they sell it? August question: are they selling it too slowly?

That is a real upgrade, and it partly dissolves my worry about leaning on two customers who are building their own capacity. Partly. More below.

Move 2: the unit economics repriced

The most important slide in the deck, and almost nobody led with it.

Annual revenue per megawatt of capacity went from a 12 million dollar baseline for 2026, to above 20 million on Q2 deals, to a 40 to 50 million opportunity on the first short-term deals in Q3. Even older GPUs repriced more than 30 percent above Q1. Payback on Q2 deals fell to 1 year and 10 months, from two to three years.

Apparently Nebius found the Rare Candy of GPU economics.

That is what should move a valuation. Revenue growing was already in my model. Revenue growing at double the price per megawatt was not.

Two asterisks. The payback number is management’s own estimate, built on forecast costs and capacity that is not built yet. And the price of one new deal during a shortage is not the average price across a 5,400 megawatt fleet in 2030.

Move 3: the Capex Pokémon learned Self-Finance

Roughly 70 percent of Q2 deals came with prepayments, covering 50 to 60 percent of the capex they require. Nebius expects more than 9 billion of prepayments in 2026. In July it closed its first secured loan, 775 million at SOFR plus 2.50 percent, backed by deployed GPUs and the contracted cash flows of an investment-grade customer.

My July worry was that the bull case quietly assumes the money always shows up. It has not gone away. It has gotten cheaper.

_{BUT EVERY POKÉMON HAS A WEAKNESS}

Here I break from the victory lap, because the cash flow statement tells a different story than the infographic.

The prepayments did not fund the quarter. The ATM did.

Nebius brought in 2.25 billion of operating cash in Q2 and spent 5.66 billion on GPUs and data centers. That is a 3.4 billion hole, filled almost exactly by the 2.85 billion of stock it sold into the market. Cash still fell by 529 million.

The 8 billion cash pile deserves the same treatment. About 1.06 billion of it is restricted. Against it sits 5.97 billion of deferred revenue, which is customer money already collected for compute not yet delivered, and 8.5 billion of long-term debt, roughly double where the year started.

Prepayments are a better funding source than equity. They are still not profit.

Capex is still super effective

5.66 billion in one quarter, 8.13 billion in the half. At that pace 2026 capex lands near 20 billion, and the 9 billion of expected prepayments covers under half. The rest is debt, equity, or both.

Dilution

12.7 million shares sold at an average of 223.60, for 2.8 billion. Another 12.3 million remain available. Shares outstanding were 271.9 million at June 30, but the count used for earnings was 280.4 million because of pre-funded warrants, and the diluted six-month count was 302.6 million. Add 4.34 billion of convertible notes issued this year.

The business can beat every operating assumption I made while per-share value moves a lot less.

The depreciation clock got slower

Not in any highlight reel. Nebius now spreads the cost of its servers over five years instead of four. Same hardware, smaller annual charge, roughly 60 million less depreciation this quarter alone, and it grows with the fleet.

It does not touch adjusted EBITDA, which excludes depreciation by definition. It does flatter the GAAP loss of 190.4 million. And it matters for anyone valuing this on EBITDA in 2030, because when your biggest real cost is GPUs wearing out, depreciation is not an accounting artifact you get to look past. It is the business.

Same spirit: adjusted EBITDA also adds back 102.5 million of stock compensation. Net of that, group adjusted EBITDA is 133.7 million, not 236.2 million.

The counterparty mix changed, and not only for the better

In July my worry was leaning on Meta and Microsoft, two customers building their own compute in parallel. Q2 diversified that book, toward Reflection and Cohere, venture-funded model labs rather than investment-grade credits. Only the 775 million facility was explicitly backed by an investment-grade customer.

Longer contracts with weaker counterparties is a different risk, not a smaller one. It is the exact seam that punished other neoclouds.

Execution

Most of the Q2 monsters were signed against capacity arriving in late 2026 and land in 2027 revenue. We are still underwriting a construction schedule.

The demand question is slowly being answered. The new boss battle is execution.

_{TIME TO UPDATE THE POKÉDEX}

Same method on both dates, so the comparison is honest: shares outstanding at June 30, 271.9 million, and ARR as Nebius defines it, the last month of the quarter times twelve.

Price up 19 percent. ARR up 56 percent. The multiple down about a quarter.

The stock went up. It got cheaper anyway.

The path was not a straight line. Between my July piece and this print NBIS round-tripped hard in both directions, then ripped 34 percent on the numbers. Trading the narrative got you shaken out twice. Underwriting the business gave you nothing to do.

Analysts are still nowhere near each other. DA Davidson sits at 175 with a Neutral, BWS Financial at 200, and one five-star analyst just went to a Street high of 410 on pricing power. That is not a forecast. It is the same argument I laid out in July, still priced two completely different ways.

_{SO DOES MY BASE CASE CHANGE?}

Not because revenue beat. Here is the bridge, then the numbers.

What pushes it higher

• Revenue per megawatt above 20 million against the 12 million in my 2030 model. The biggest input change, and the one I was most conservative on.

• Payback at 1 year 10 months instead of two to three, which recycles capital faster.

• AI cloud margin near 50 percent, closing on the 52 my model assumed.

• 5 GW of contracted power instead of 4, and more than 1 GW a year of deployment from 2027.

• Prepayments covering half the capex on new deals, behind 40 billion of customer commitments.

What offsets it

• 5.66 billion of quarterly capex, and most of Q2’s contract value is a 2027 story.

• Long-term debt doubled to 8.5 billion in six months, with 5.97 billion of deferred revenue still to deliver. My model assumed debt and cash cancel out by 2030.

• My 52 percent is a group margin. The group did 41 percent, before stock compensation.

• And the share count, which deserves its own section.

First, the denominator

Every price target is a market value divided by a share count, and the share count is the part nobody models. Start from 280.4 million rather than the 271.9 million outstanding, because the pre-funded warrants already count, then add what the next year issues.

Here is the part that surprised me. The bull case carries the most shares. In the bear, Nebius sells stock cheap into weakness and the 4.34 billion of converts stay out of the money. In the bull it barely touches the ATM, but every one of those converts becomes equity. You get diluted in all three worlds. The question is whether you get paid for it.

Call it 15 to 20 percent more shares within a year. Against a 29 percent base case, that is not a rounding error.

The three cases, one year out

By August 2027 nobody is pricing this year’s guide. They are pricing the 2027 run-rate against 5 GW of contracted power. Pick the ARR visible by then, pick the multiple paid for it, divide by the shares that will exist.

Bear, the ramp slips. Capacity arrives late, ARR stalls near 7 billion against a 40 billion commitment book, and the funding question reopens. Equity gets sold into weakness and the multiple halves. Q2 made this take a demand shock, not just a bad financing week.

Base, guidance holds. ARR compounds to about 11 billion, 2027 capacity lands close to schedule, and the market keeps paying a growth multiple while the business grows into it.

Bull, the re-rating. ARR reaches 14 billion, prepayments and asset-backed debt carry the buildout, and inference starts earning software margin on top of rent. Nebius stops being priced as a compute reseller.

Notice what does the work: multiple compression, not growth. ARR more than triples in the base case and the stock returns 29 percent, because an 11 billion ARR company does not hold 25 times revenue. Most bull threads leave that out.

Skew is about three to one in my favor. Worse than July, when it ran closer to five to one. The business improved. The entry price improved faster.

A note on the 2030 model

I am not rebuilding it here. Same five inputs as July, same 674 billion of enterprise value, same 1,925 a share. The July piece has the full walkthrough.

Two numbers in it are stale, and only because the stock moved. Against today’s 74.5 billion, the base case is about 9 times the current company, not 11, and 7 times the share price, not 8.

One input I am deliberately not touching: 12 million of revenue per megawatt, despite Q2 printing north of 20. Twenty million is shortage pricing. Five thousand four hundred megawatts is what the world looks like once the shortage ends. You do not get to assume both at once. Raise the price and the capacity has to come down, and every honest version of that trade lands in roughly the same place.

This is the same discipline holding my one-year numbers where they are.

My provisional conclusion

Q2 does not force a higher target. It does something less exciting and more useful: it shortens the odds on the base and bull paths, and it tells me the input I was most conservative on may have been conservative by half.

What it does not improve is the price I pay for those odds. The business got better. The entry got worse. Those two cancel out more than anyone will admit this week. But if we can get a better entry, then we have a completely different story in our hands.

Here is the base case taken apart, because the arithmetic says something the headline does not.

The revenue argument everyone had this week is the one line that goes my way, and it is not the line that decides the outcome. Multiple compression costs roughly four times what dilution costs, and it is close to certain, because almost nothing scaling to 11 billion of ARR keeps a 25 times multiple. Dilution takes another 13 points quietly, while almost nobody models it.

Plainly: Nebius is a better company than it was in July and not a better buy, unless we get a better entry. If you own it, nothing here says sell, and the funding risk you were worried about got cheaper. If you do not, this quarter handed you a better business; just be mindful of your entry price.

And if I am wrong from here, it will not be because I misjudged the revenue. It will be because I paid 25 times for it.

_{FINAL VERDICT: NBIS EVOLUTION CONFIRMED?}

In July, owning Nebius meant betting that a pipeline of promised capacity would turn into revenue.

Q2 did not finish that story. It made the bet far less theoretical, and it repriced the economics of every megawatt that comes after it.

The thing that will cost you most is not a flaw and not a surprise. It is multiple compression, and it is close to unavoidable: almost nothing scaling to 11 billion of ARR keeps a 25 times multiple. Dilution was already on the board in July, when Nebius raised more than 6 billion in a single quarter. What changed is its shape. The at-the-market program went live this quarter, so the company now sells stock into the market as a matter of routine rather than in occasional financings. Same risk, steadier drip, and now sizeable enough to model.

What still has to happen has not changed. Roughly 20 billion of capex this year. A construction schedule across three continents. A funding mix that still includes selling stock into strength. And four more years of an AI cycle that has never once been priced for disappointment.

So, plainly. If you own it, nothing here says sell, and the funding risk you were worried about in July got cheaper. If you do not, this quarter handed you a better business at a worse price, and the patient version of this trade is waiting for the market to give back some of that 19 percent.

I am long. I’ll keep accumulating at every good entry.

Nebius has evolved. My valuation is evolving with it, just slower than the headlines.

Long NBIS.

_DISCLAIMER

This analysis is for educational and informational purposes only. It does not constitute financial advice or a recommendation to buy or sell any security.

August 18th, 2026 | Price: ~$274 | Sources: Q2 FY2026 Press Release · Q2 FY2026 Shareholder Letter · 6-K |NFA. DYOR.

The AI buildout has 12 layers. Interconnect is layer seven and it sits in the middle of the physical stack for a reason. Compute does the work. Memory holds the data. Interconnect moves it.

A GPU on its own is a very expensive space heater. It only earns its price when it can talk: to the memory stacked beside it, to the 71 other GPUs sharing its rack, to the tens of thousands more across the building, and increasingly to a cluster three states away.

Layer 5 is the muscle. Layer 7 is the nervous system that turns thousands of muscles into one organism instead of a pile of parts. And right now the nerves cannot keep up with the muscle.

_{THE NEW BOTTLENECK IS MOVING DATA}

Scaling AI means connecting more accelerators. Add enough of them and the talking grows faster than the computing, until another rack costs more than the work it contributes and the machine spends its day waiting instead of calculating.

Five things decide what is physically possible here, and no design gets to have all five:

• Bandwidth. Bits per second.

• Latency. How long a single hop takes.

• Distance. How far the signal has to survive.

• Power. Watts burned per bit moved.

• Cost. Dollars per port, per lane, per meter.

Everything else in this layer is an argument about which of those five to give up.

Copper is cheap and simple. No laser, no photodiode, no digital signal processor, just metal. What copper cannot do is go far at speed. Every doubling of the bits per lane shrinks the usable reach, and at 200 gigabits per lane copper has become a within-the-rack technology, and barely that. Light does not have that problem. It carries the same bits across kilometers. The catch is that light needs a laser, and lasers are the hardest thing in this layer to build.

So the trade gets made in the same order every time: copper while you can, optics when you must. The story of 2026 through 2028 is that “when you must” keeps arriving earlier, and each time it does, optics moves one step closer to the GPU.

_{THREE INTERCONNECT MARKETS}

Interconnect covers three distances: a meter, a building, and a continent. They share a name and not much else. Different physics, different buyers, very different competition.

Scale-up. Inside the AI machine.

Accelerators pretending to be one giant computer. NVLink-style fabrics, copper flyover cables, rack-scale architectures, and eventually optical engines sitting on the package. The question this market answers is how thousands of chips behave like a single machine.

At GTC 2026, Huang answered the copper-versus-optical question with both, and said the first layer of the network stays copper inside the rack. Then the schedule moved. SemiAnalysis reported the Kyber NVL144 rack slipping to 2028 because the PCB midplane at the heart of it remains hard to manufacture, and the larger NVL576 configuration likely delayed or shipping in small volumes. A copper stopgap called NVL72x2 was reportedly cancelled after customers rejected it, and Nvidia says its roadmap is intact.

That cuts both ways, and neither camp gets to claim it. Optical scale-up is coming. It is also harder to build than Nvidia’s March roadmap made it look.

Scale-out. Connecting the AI factory.

Rack to rack inside the building. Pluggable transceivers, the 400G to 800G to 1.6T ladder, Ethernet against InfiniBand, linear pluggable optics. This is the biggest pool of investable dollars today and the one with the clearest 2026 revenue attached to it. BofA sizes the CY25 global datacom Ethernet transceiver market at roughly $22bn.

Scale-across. Connecting AI factories to each other.

Coherent optics, dense wavelength division multiplexing, metro and long-haul, data center interconnect, submarine fiber. This was a telecom story for twenty years. It is becoming an AI story because frontier training runs are outgrowing what a single campus can power. Once a training job spans two sites, the fiber between them is part of the computer.

Chip, rack, data center, campus, region, world. The same physics problem six times over, and each one costs more than the last.

_{WHAT IS ACTUALLY INSIDE THE NETWORK}

Most people picture a cable. What a hyperscaler is actually buying is a small machine, and the bill for it is not split the way you would guess.

On Nomura’s teardown of an 800G transceiver, the transmit sub-assembly holding the laser die, photodiode die, isolator and receptacle is roughly 27% of the module, the electronic chips around 20%, the receive sub-assembly about 15%, and the circuit board about 7%. The remaining third is housing, optics, assembly and test.

So close to half the value sits in the laser and the electronics, which happen to be the two pieces that are hard to make and hard to qualify. The rest is putting parts together.

Keep that in mind when you read market share tables. Share of modules shipped and share of value captured are not the same number, and the gap between them is the whole reason components trade at a premium to modules.

_{WHY INTERCONNECT KEEPS GETTING MORE VALUABLE}

Three of the drivers are arithmetic. The fourth is the one that changes the size of the prize.

More accelerators. Traffic grows faster than headcount. Double the GPUs and you more than double the conversation between them.

Faster accelerators. A faster chip is a hungrier chip. Keeping it fed takes more bandwidth per socket, not the same bandwidth spread thinner.

Bigger domains. Clusters measured in thousands are becoming domains measured in hundreds of thousands, and every expansion adds ports.

Optics moving closer to the compute. Optics used to sit at the edge of the data center. Then it moved into the rack. Now it is heading for the package. Every step along that path attaches more optical content to each individual GPU, and that is the difference between interconnect being a line in the networking budget and interconnect being part of what a GPU costs.

The expert BofA hosted in April put 2026 demand at 50 to 60 million 800G units and 30 million 1.6T units, with the industry able to supply only 70 to 80% and 50 to 60% of that. That is one specialist’s estimate, not a number any hyperscaler has confirmed, and unit forecasts in this industry have been wrong before. But a shortfall of that shape is the single claim I most want confirmed or killed over the next two earnings seasons, because it decides whether pricing holds or whether 2027 turns into a discount war.

_{THE TECHNOLOGY WARS}

Six fights are running at the same time. None of them is settled, and each one moves money between the names in section 10.

The form factor fight is the one people misprice. Pluggables are field-serviceable and boring, which in a data center is a compliment, because when one dies a technician swaps it in a minute. Linear pluggable optics rips out the DSP to save power and cost, and pays for it in reach. Co-packaged optics moves the optics onto the switch package for the best power per bit and the worst serviceability, since one failure now means touching the switch itself. The expert view is that pluggables stay mainstream into 2030, and Barclays argued in July that broad CPO adoption may take until 2029 or 2030. If you are paying up for the CPO story alone, you are early by years.

The light source fight is a packaging fight in disguise. External continuous-wave lasers are easier to cool and easier to replace, and they add parts. Integrated is denser and harder to yield. Neither wins on optics. Thermal reliability decides it, which means the winner gets picked by packaging engineers rather than by laser designers.

The materials fight has a challenger worth naming. Thin-film lithium niobate supports bandwidths above 100 GHz and single-channel modulation above 240 Gbaud, against roughly 60 to 90 Gbaud for silicon and about 130 Gbaud for indium phosphide, which makes it a candidate for 1.6T and 3.2T. Silicon photonics still owns the volume, at roughly 50% penetration for 800G and 60% for 1.6T in 2026. Watch TFLN. Do not underwrite it yet.

_{WHERE THE BOTTLENECKS ACTUALLY ARE}

Read more

What memory is, in plain language

An AI chip cannot compute on data it does not have in front of it. Memory is what holds that data and feeds it to the processor. Two dimensions matter: capacity, meaning how much fits, and bandwidth, meaning how fast it moves.

There are four types worth knowing, and they are not interchangeable.

Why memory became the bottleneck

The industry calls it the memory wall. Past a point you cannot get more out of your compute unless you also expand memory bandwidth. Adding GPUs without adding memory just gives you idle silicon.

Three shifts made this acute in 2026:

• Inference overtook training. Training is a burst. Inference runs forever, with growing context windows.

• Agents replaced chatbots. An agent observes, plans, calls tools, and remembers. Longer sessions, parallel work, and a KV cache that has to be retained rather than recomputed. Hot KV cache sits in DRAM or HBM. Idle KV cache gets offloaded to NVMe and QLC enterprise SSDs. Every layer of the stack gets pulled at once.

• Token volume is compounding. Goldman Sachs forecasts a 24x increase in token consumption by 2030, and the striking part is the mix: almost all of that growth is agent tokens, not chat tokens.

The supply side cannot answer

Memory fabs take two to three years to build. Most announced projects will not add meaningful HBM capacity before 2027.

Worse, HBM cannibalizes everything else. One additional bit of HBM takes roughly three times more silicon than standard DRAM, moving toward four with HBM4. Every wafer sent to HBM is a wafer that does not become DRAM for a car or a laptop. Clean room space is finite, so manufacturers are choosing what not to build.

Here is the number that makes the case better than any narrative. Bit supply growth, per Gartner and UBS estimates as of June 2026:

DRAM supply is running at the bottom of its own historical range. NAND supply is running below its historical range in every single year. That is not a cycle, that is a structural shortfall, and it is the strongest quantitative argument that NAND is the next leg rather than a story.

The rest of the evidence on the ground:

• Micron has said it can serve only 50% to two thirds of what its key customers ask for, and calls the demand-supply gap the largest it has ever seen.

• As of August, roughly 60% of big tech’s memory and AI substrate needs are being met. Unmet 2026 demand rolls into 2027, and 2027 into 2028.

• Micron, Samsung and SK hynix have effectively sold out DRAM and HBM through 2027. NAND capacity for 2026 is sold out too.

• Sell-side models put undersupply running into at least late 2027 or 2028, with real relief only after 2029.

Price is doing what price does. Spot DRAM is up roughly 793% and NAND roughly 442% year over year, per Bloomberg data as of June 2026 (measured on inSpectrum DDR5 1Gx16 and TLC 1TB). Standard DRAM contract prices were forecast to rise 58 to 63% quarter over quarter in Q2 2026, with NAND contracts up 70 to 75%.

The spillover is macro now. US PPI for electronic components ran +26.9% year over year in May, up from +5.9% in January. Lenovo, Dell and ASUS have flagged 15 to 20% price increases. Deutsche Bank models DRAM inflation as adding $150 to $300 to the price of a luxury vehicle and $400 to $600 for higher-autonomy models. Memory went from commodity to macro variable.

Why invest, and why now

The demand is contracted, not hoped for. Hyperscaler capex commitments for 2026 total about $680B: Amazon $200B, Google $186B, Microsoft $157B, Meta $138B. NAND flash market revenue alone is forecast to go from $70B in 2025 to $147B in 2026 and $176B in 2027. Total memory revenue is forecast to roughly triple, from $216B in 2025 to about $750B in 2027.

Industry structure changed too. Enhanced long-term agreements now fix volume two to three years out of five, covering 60 to 70% of hyperscaler DRAM volumes, with take-or-pay clauses and prepayments. Lower profits at the peak, higher profits at the trough. That is the difference between a commodity cycle and an infrastructure annuity.

And the market still does not believe it. Forward P/E as of June 24, per Bloomberg: Samsung 4.9, SK hynix 5.7, Micron 7.2, Kioxia 7.5, SanDisk 9.5. The Nasdaq-100 sits at 22.9.

Two things I would not swallow whole. First, the widely circulated $12 trillion of cumulative AI funding capacity through 2031 is a Coatue estimate built on 10% annual EBITDA growth and 3x debt to EBITDA, with $2tn of it labelled sovereign funds, private credit and neoclouds. That is a capacity estimate, not committed money. Second, the same Tema deck cites Gartner’s 2027 memory revenue as both $748B and $758B on different pages. Small thing, but if a number moves $10B between two slides, use it as a direction, not a decimal.

The bear case is one sentence and it is the one that matters: if AI demand moderates even modestly, all that HBM and silicon capacity has to go somewhere, and the market can swing from severe shortage to oversupply fast. July 2026 was the preview. The Philadelphia Semiconductor Index fell 19% in the month, its worst since 2008.

Which brings us to the two funds.

Read more

_{BEFORE WE DIVE IN}

Three distances, one company at all of them

Lumentum sits on Layer 7 of the AI buildout. Interconnect. The floor that decides whether ten thousand GPUs behave like one machine or like ten thousand very expensive space heaters.

Layer 7 is not one market, though, and that is where people get lost. It is the same problem solved three times, at three different distances. Here is where Lumentum’s products land on each one.

Inside the rack, chip to chip. Copper still owns this. When co-packaged optics changes that, the light has to be fed in from outside the package, and that is the piece Lumentum sells: CW and ultra-high-power lasers, not the optical engine itself. Management expects the crossover in late 2027. Revenue today is roughly zero, which makes it the biggest unmodeled line in this piece.

Rack to rack, inside the building. Where the money is, and also where the crowd is. The 800G and 1.6T pluggable transceivers, the modulated lasers inside them (EMLs and DMLs, which also go into plenty of competitors’ modules), and the optical circuit switches that reroute the fabric.

Building to building, across a city or an ocean floor. The oldest business and the steadiest one. Pump lasers, narrow linewidth lasers for data center interconnect, coherent components for long haul. In the March quarter pump lasers grew 80 percent year over year and narrow linewidth more than 120 percent. Nobody writes about this line. It compounds anyway.

Most companies on this floor pick one distance and defend it. Lumentum shows up at all three with the same object every time: a semiconductor laser grown on indium phosphide, sold as a bare chip, packaged into a module, or wired into a system. The distance changes, the customer changes, the margin changes. The thing underneath does not. That is why I file this as a light source company and not a module maker, and it is the reason the margins look the way they do.

On March 20, 2026, CEO Michael Hurlston said Lumentum is “under shipping the market” by 25 to 30 percent. Everything below follows from that one sentence.

_{SECTION 1 · COMPANY SNAPSHOT}

What LITE is, and how a dot-com casualty became critical AI infrastructure

Start with what it is today. Lumentum makes the lasers and the light-handling hardware that move data through modern networks. There is also a smaller industrial laser business, the kind that cuts and welds things in factories. San Jose, Nasdaq, ticker LITE.

The positioning matters because AI has run into a wall that has nothing to do with chips. Clusters got big enough that moving data between processors became harder than processing it, and at these speeds, moving data means light.

Now the odd part. Lumentum was spun out of JDS Uniphase in 2015, which means it walked out of the building carrying the optics business of the company that became shorthand for the dot-com collapse. For the next decade it was a deeply cyclical telecom components supplier, plus the VCSELs behind the face scanner in your phone, plus industrial lasers. It bought Oclaro, NeoPhotonics and Cloud Light along the way. As recently as fiscal 2024 it did $1.36 billion of revenue and lost $546 million.

So this is not a startup catching a wave. It is a decades-old manufacturing base that kept building indium phosphide capacity through a market that kept disappointing it, and then found itself holding the scarcest asset in AI infrastructure.

What happened next is not subtle. Quarterly revenue went $533.8 million, then $665.5 million, then $808.4 million. Up 58, then 66, then 90 percent year over year. On March 2, 2026, NVIDIA bought $2 billion of convertible preferred at $695.31 a share and committed to multi-year purchases of advanced laser components. Two weeks later Lumentum closed a multi-year, multi-billion-dollar optical circuit switch agreement with an existing hyperscaler. S&P 500 in March. Nasdaq-100 in May.

The old story is a rounding error now. 3D sensing, the iPhone business that used to drive this stock, is below 5 percent of revenue. The market has repriced accordingly: roughly $70 billion fully diluted today, against about $6 billion a year ago.

Worth sitting with for a second: NVIDIA paid $695.31 a share in March, and the stock is at $693 today. Whatever you conclude about the multiple, you are buying it at the price the most informed customer in the industry agreed to pay._{SECTION 2 · FUNDAMENTALS}

How it makes money

Two reporting buckets, and the difference between them is the difference between selling ingredients and selling meals.

• Components: Laser chips, pump lasers, wavelength management, sold to companies that assemble them into something bigger.

• Systems: Finished products sold to end customers: cloud transceivers, optical circuit switches, industrial lasers.

  

Underneath both, this is a manufacturing business. Lumentum owns indium phosphide fabs, the scarce asset in this industry, and it is increasingly dropping its own laser chips into its own transceivers instead of buying them from somebody else. That vertical integration is the margin story, and it is why Components matters more to my thesis than the faster-growing Systems line does.

Management’s framing has been consistent, and the numbers back it: this is a supply problem, not a demand problem. Capacity is effectively spoken for into 2027, and what caps any given quarter is how many wafers come out of the fabs, not how many customers call.

The management team

Three seats, each mapped to something this company specifically needs.

Michael Hurlston, CEO since February 2025, is the pivot itself. Seventeen years at Broadcom, then CEO of Finisar, which he sold to the company now called Coherent. He has sold optics into hyperscalers and he has run a semiconductor P&L, and that is the exact combination it takes to turn a telecom-era supplier into a merchant chip business. He also sold a company at the bottom of this cycle, so he knows what the turn looks like from the inside.

Wajid Ali, CFO, has the hardest job in the building. Roughly half the NVIDIA proceeds go to strategic capex, the rest to working capital and possible vertical-integration deals. Financing a multi-fab buildout without shredding the share count is the whole game, and as the financials show, the count is already moving.

Wupen Yuen, President of Global Business Units, owns the roadmap. EML output is up eight-fold since fiscal 2023 and the company has demonstrated 400G-per-lane optics. When supply is the binding constraint, whoever converts fab capacity into shippable product is the person actually setting revenue.

Who actually buys from Lumentum

Hyperscalers and AI infrastructure builders. Some buy direct, some through the module makers and equipment vendors sitting in between. Lumentum does not name them, but the concentration is not in doubt: the same handful of companies funding the entire buildout are absorbing this output.

The commitments are unusually concrete for a component supplier. NVIDIA’s $2 billion investment came attached to a multi-year purchase commitment for advanced lasers. An existing hyperscaler signed a multi-year, multi-billion-dollar optical circuit switch agreement in March, against an OCS backlog already past $400 million. A separate multi-hundred-million-dollar co-packaged optics order is booked for delivery in the first half of 2027.

NVIDIA is the one to sit with, because it is now investor, customer, roadmap partner and anchor tenant of the new Greensboro fab, all at once. That is enormous validation and a real dependency in the same sentence. Both are true. NVIDIA made the point itself by putting an identical $2 billion into Coherent the same morning. It wants two qualified suppliers, not a favorite.

_{SECTION 3 · FINANCIAL PROFILE}

The quarter operating leverage stopped being a promise

Two fiscal years ago this company lost $546 million. Last quarter it ran a 32 percent operating margin. That is the section in two sentences, but how it got there is the part worth your time.

Start with scale rather than growth rates. Nine months of fiscal 2026 produced $2.01 billion of revenue against $1.65 billion for all of fiscal 2025, and the June quarter alone is guided to $960 million to $1.01 billion, more than the company did in the entire first half of fiscal 2025.

Here is the arithmetic behind the margin story. Across those three quarters revenue rose 51 percent while non-GAAP operating expenses rose 14 percent, taking opex from 20.7 percent of revenue down to 15.6. The factories and the engineers were already paid for, so most of each new dollar drops through. Non-GAAP operating margin ran 10.8, 15.0, 18.7, 25.2 and 32.2 percent across five quarters, with 35 to 36 percent guided for Q4. Revenue did not quite double over that stretch. Operating profit rose almost six-fold.

That leverage has a price, and I would rather point at it than let you find it on your own. R&D fell to 9.7 percent of revenue from 12.9 percent two quarters earlier. Spending rose, it just rose far slower than sales. This is a company whose entire position depends on staying ahead on the roadmap, so some of today’s margin is tomorrow’s R&D going unspent.

Read the profit line carefully, because the headline figures are non-GAAP. Q3 GAAP net income was $144.2 million against $225.7 million non-GAAP, and the largest single bridge item is $46.8 million of stock compensation, a real cost paid in shares rather than cash. GAAP is improving fast, $4.2 million to $78.2 million to $144.2 million this year, though the middle figure was flattered by a one-time $27.5 million escrow settlement. The gap should keep closing on its own, since amortization of acquired intangibles is fixed near $34 million a quarter while revenue climbs.

Which brings me to the number almost nobody discusses. Diluted shares went from 72.2 million a year ago to 95.2 million last quarter, guided to 102 million for Q4. The natural assumption is that NVIDIA caused it. It did not. The preferred is under 3 million shares. Almost all the rest is convertible notes issued when the stock traded at a fraction of today’s price, now deep in the money and sitting on the balance sheet as $3.24 billion of current debt.

So the $53.9 billion market cap on your screener is the common count. Fully diluted is nearer $70 billion. That is a 30 percent difference, and the bigger number is the one that belongs in every multiple you run.

Cash shows where the profit goes. There is $3.17 billion on hand after NVIDIA’s check, property, plant and equipment went from $726 million to $964 million in nine months, and consensus puts fiscal 2026 free cash flow near $209 million against $619 million of nine-month adjusted EBITDA. That is the tell. Earnings are converting into fabs, not into cash. It is the right call at this point in the cycle, and it is also the reason this will not be a cash-return story for years.

The number that matters most is wafers. Everything above is downstream of indium phosphide output: EML volume up eight-fold since fiscal 2023, capacity planned up another 50 percent through the end of calendar 2026, a fifth fab in Greensboro for 2028. When a company is under-shipping demand, revenue stops being a sales forecast and becomes a manufacturing schedule.

Bottom line: margins are expanding faster than revenue, and the balance sheet has been recapitalized by the most important customer in the industry. The question was never whether Lumentum grows. It is whether the fabs come online fast enough to turn demand into revenue before the cycle turns.

_{SECTION 4 · PRODUCT AND TECHNOLOGY}

Four businesses, three clocks, one bottleneck

Start with the light. EML and CW laser chips are the source inside every transceiver, and this is where Lumentum is structurally advantaged. EML shipments doubled year over year in Q3, and 200G EML revenue more than doubled sequentially.

Cloud transceivers are the volume business and the crowded one. 800G and 1.6T modules sold to data center operators at structurally lower margins. Which is exactly why the 1.6T ramp matters more than the revenue attached to it. Lumentum is putting its own lasers inside its own modules at scale for the first time, and if that works, a commodity product starts carrying chip-level margin.

Optical circuit switches are the newest engine. Arrays of tiny tilting mirrors that steer light from one port to another without ever converting it to electricity, saving latency and a serious amount of power in buildings that are already fighting the grid. Backlog passed $400 million, a multi-year multi-billion-dollar agreement landed in March, roughly $400 million ships in the back half of this year, and management points toward a $1 billion run rate in 2027. It is also, in their own words, the tightest rope they are walking.

Ultra-high-power lasers for co-packaged optics are furthest out and possibly largest. CPO puts the optics right beside the switch chip, and that design needs light fed in from outside the package. Lumentum sells exactly that. Meaningful revenue is expected exiting 2026, with a multi-hundred-million-dollar order already booked for the first half of 2027.

Then the one with no revenue line at all. Inside a rack today, GPUs talk over copper. Management expects a hybrid setup from the second half of 2027, where anything longer than about three meters goes optical. Optics have never shipped inside a cluster at scale, so this is not a bigger slice of an existing market. It is a market that does not exist yet.

The other half of the moat is a crystal

None of it happens without indium phosphide, the compound semiconductor these lasers are grown on. You cannot conjure a wafer fab, and the entire industry is competing for the same substrate supply. Lumentum’s answer has been to lock it down: a capacity reservation with AXT running through 2031, and a former Qorvo plant in Greensboro, North Carolina being converted into its fifth InP fab, targeted for 2028, with NVIDIA named as a customer.

Wafers reserved, capacity committed, customers signed, all years before the revenue arrives. Whoever secured substrate early owns the chokepoint.

_{SECTION 5 · GROWTH CATALYSTS}

What upcoming milestones could send this stock significantly higher?

Lumentum’s catalysts share an unusual trait. Almost every one of them gets settled in a factory rather than in a sales meeting. The demand is already contracted. What is open is whether the company can build.

Each milestone below answers a different version of that question. Margin first: do its own lasers, inside its own modules, actually lift gross margin. Then volume: can it ship the switches it has already sold. Then the open-ended one, whether optics move inside the rack at all, which nobody has done at scale.

The order is the useful part, because the cheapest doubts retire first. Guidance for fiscal 2027 arrives in eleven days. The margin and volume questions both land over the two quarters after that. The one worth the most, and the one least visible in today’s price, does not resolve until deep into 2027.

_{SECTION 6 · MACRO AND INDUSTRY BACKDROP}

The capex weather system, and where Lumentum sits inside it

Hyperscaler capital spending is the climate this stock lives in. Alphabet’s $195 to $205 billion capex guide for 2026 knocked the entire optics group down on free cash flow worries. A week later Microsoft held its own spending outlook and the same names ripped. Lumentum swung from down 9 percent to up 13 percent inside three sessions this week, without a word of company news. Own this and you own a leveraged bet on other companies’ budgets. Make peace with that or do not own it.

The competitive field

It sorts into three groups:

• Coherent is the one to beat. Larger, more diversified, scaling indium phosphide in Texas, and the recipient of an identical $2 billion NVIDIA investment on the same day. NVIDIA kept both relationships explicitly non-exclusive, which is the tell: these are the two qualified suppliers, being played against each other on purpose.

• The module tier is where the price pressure lives. Applied Optoelectronics, Fabrinet, InnoLight and Eoptolink compete on assembled transceivers, the layer with the most entrants and the strongest Chinese presence.

• The switch vendors are integrating optics themselves. Broadcom and its peers moving to co-packaged optics threatens pluggable modules directly, while creating demand for the external light sources Lumentum sells.

That last one is the nuance worth holding onto. Co-packaged optics is bearish for transceivers and bullish for laser chips, and Lumentum sells both. As optics move closer to the chip, the module shrinks and the light source grows. A pure module maker meets that transition as a threat. Lumentum meets it as a mix shift toward its best margin.

The sector bear case is one word: cyclicality. This industry burned investors in the 2001 telecom collapse and again in the 3D sensing hangover, and Lumentum’s own half-billion-dollar loss came only two fiscal years ago, with the same fabs and much of the same portfolio. Optics turns faster than its factories get built.

_{SECTION 7 · RISKS}

What could break this

When a business runs this hot, the risks do not disappear. They just get easier to skip. So let us not.

1. Concentration, and a loop that funds itself. A handful of hyperscalers plus NVIDIA drive the demand, and NVIDIA is at once the largest strategic investor, a major customer and a co-development partner. Nothing improper in that, but part of what looks like independent demand is the supply chain financing itself, and that structure unwinds fast if the sponsor slows.

2. Supply is the governor, not demand. The constraints are electrical components, laser diodes and indium phosphide wafers. Under-shipping protects pricing today, and it also means any execution slip costs revenue that never comes back. The structural fix in Greensboro does not produce until 2028.

3. Dilution is not a footnote. Diluted shares are up about 40 percent year over year and guided to 102 million. Per-share growth is fighting a headwind that never shows up in the revenue line.

4. Valuation prices the ramp, not the results. At roughly 84 times consensus fiscal 2026 earnings, you are paying for 2027 and part of 2028 before either arrives.

5. Backlog language is looser than it sounds. A “multi-billion-dollar purchase commitment” is not contracted revenue with fixed volumes and fixed prices. Read it as serious intent backed by deposits, not as a receivable.

6. The volatility is structural. The stock sits about 37 percent below its May high after falling roughly a third peak to trough this summer, and leveraged long and inverse ETFs now trade on it. Much of this float is being rented, not owned.

Bear case in one sentence: you are paying more than 80 times this year’s earnings for a historically cyclical component supplier whose demand comes from a few customers, whose growth is gated by wafers and parts it does not fully control, and whose share count grew 40 percent in a year.

The counterweight is not small either. The constraint here is physics, not ambition. Copper stops working at these speeds and light does not, and the number of companies that can make an indium phosphide laser at volume is very close to two.

_{SECTION 8 · VALUATION AND PRICE TARGETS}

Cheap is not the word. The question is whether the ramp earns the multiple.

Lumentum’s market cap is $53.93 billion. On the 102 million diluted shares the company guides to, the equity is closer to $70 billion, roughly 23 times fiscal 2026 sales. Per share, at $693 the stock trades near 84 times consensus fiscal 2026 EPS of $8.23. Annualize the Q4 guide instead and it falls to 59 times. Use the fiscal 2027 consensus of $18.35 and it falls to 38.

All three are defensible, which tells you the argument is not really about the multiple you pick.

So here is my frame, with the assumptions stated out loud. EV/EBITDA is the better measure for a company pouring its profit into fabs, since it carries heavy depreciation and a balance sheet full of converts. Run it against fiscal 2028, not 2027, since a buyer twelve months from now is paying for the year in front of them. Build from the guided Q4 midpoint of $985 million: 14 percent quarterly growth puts fiscal 2027 at $5.5 billion, in line with consensus and exiting near $1.66 billion a quarter. Slow to 10 percent and fiscal 2028 is roughly $8.5 billion. Assume 105 million shares and net debt near zero.

The differences across those rows are almost entirely operational. The bear is growth slowing to 18 percent with a few points of margin handed back, and you make nothing for a year. The base is the ramp continuing, with margins landing where the Q4 guide already points. The bull adds the scale-up transition arriving on time.

One input is doing more work than any revenue line, though. Every row holds the multiple above 30 times forward EBITDA, near where it trades today. Compress it to 18 times, which is what an optics downcycle has always done, and the bear is not flat, it is about $390, while the base falls to $635.

Read that one twice, because it is the actual risk. The business can be exactly right and the multiple can still take half your money. That is the part the table cannot show, and it is the part you size for.

The argument on this name was never about whether Lumentum grows. It is about what you pay for growth that has a cycle attached to it.

_{SECTION 9 · TRADING THE NEWS}

The substrate deal that did not make the headlines

On July 26 Lumentum signed a capacity reservation agreement with AXT for indium phosphide wafer substrates, running through December 31, 2031, with two deposits of $43.5 million applied as credits against future purchases.

Why it matters. Guidance can be sandbagged. Backlog can be front-loaded. Cash sent to a supplier six years forward is not a talking point. A company that thought this was a 2027 phenomenon does not prepay for wafers through 2031. It also confirms where the real chokepoint sits, and it is not the module, or the switch, or even the laser. It is the crystal the laser is grown on.

The bear note. AXT manufactures in China, and indium sits on the list of materials that has already been used as trade leverage. Lumentum secured its supply and added a geopolitical seam with the same signature. Greensboro is the hedge against exactly that, and Greensboro is two years away.

The week the stock traded on somebody else’s earnings

On July 28, Lumentum fell about 9 percent, Coherent 11 percent, and Applied Optoelectronics 10 percent. No news from any of the three. The trigger was Alphabet’s capex number and the free cash flow question underneath it. On July 29 Microsoft reported 43 percent Azure growth and held its capital spending outlook. On July 30, Lumentum opened up about 13 percent.

My take. Nothing about this business changed in 48 hours. What changed was the market’s confidence in the people writing the checks. That is the trade you are actually in, and it will stay that way until Lumentum’s own numbers are large enough to stand on their own. Right now the earnings confirm the price. They do not set it.

_{SECTION 10 · THESIS SUMMARY AND VERDICT}

My verdict on LITE: forever long, and never for a second forgetting where it came from

Strip away every chart and the story is one sentence. At 1.6 terabits copper stops working past two meters, and somebody has to make the light.

Anyone with capital can rack GPUs. An indium phosphide laser at volume takes fabs, epitaxy expertise, wafer supply and years of qualification. NVIDIA did not write $2 billion checks to Lumentum and to Coherent on the same morning because these two are replaceable. It did it because they are not, and because two is an uncomfortably small number.

Two things separate Lumentum inside that pair.

The first is where it builds. A fifth indium phosphide fab going up in North Carolina with NVIDIA named as a customer, at a moment when the module tier’s cost advantage runs through China and export controls are a live policy question. American fabs are not a marketing line here. They are a supply chain that survives a trade war.

The second is the size of the lasers. Co-packaged optics needs light fed in from outside the package, at powers ordinary transceiver lasers cannot reach, and ultra-high-power CW is precisely what Lumentum sells. As optics move onto the switch package, the module shrinks and the light source grows. Lumentum makes the part that grows.

A word on how I got here, because it explains the size more than the thesis does.

I opened LITE at the beginning of this year at $344, and I didn’t find it by being clever about lasers. I found it by walking down the building. In January the word that kept surfacing in every supply chain note I read was photonics, so I went floor by floor: Lumentum, Marvell, Tower, Applied Optoelectronics, then one deeper into advanced packaging, then equipment and testing, then the substrates underneath all of it. LITE was the first name on that list and the one I understood least at the time. It’s now the one I intend to hold longest.

And since I don’t get to only tell you the parts I got right: I was wrong about co-packaged optics this year, and it cost me real money. In the spring I put close to 30 percent of the book into the Taiwan CPO chain, optical engines, fiber array units, external light sources, on the assumption the transition arrived roughly on schedule. Then the SemiAnalysis work pushed the timeline out by more than twelve months and that whole basket came apart in June. What I got wrong wasn’t the direction. It was the calendar. On this floor the calendar is the trade, and I’d been treating it as a detail.

That mistake is why Section 5 is ordered the way it is, cheapest doubts first. It’s also why the scale-up story is not what I’m paying for here. And yes, I see the irony of owning a name whose upside leans on the same 2027 timeline that just cost me money. The difference is what the timeline was holding up. In the Taiwan basket the schedule was the position, and when it moved there was nothing underneath it. Here it’s a free option sitting on top of a business already shipping $808 million a quarter. If CPO slips another year, LITE still has the lasers, the switches and the fabs. That basket had a delivery date and a hope.

On August 11, I’ll be at the dining-room table at 4:00 p.m., earnings deck open and baby monitor beside the laptop, waiting for Lumentum to tell us whether the fabs kept up.

Now the part the run obscures, and the part I keep repeating to myself. This is still a component business. Customers who hold real buying power, cycles that turn faster than fabs get built, margins that compress on the way down as beautifully as they expand on the way up. The same company, with the same fabs, lost half a billion dollars two fiscal years ago. People want this to be a software-style compounder because the growth rate looks like one. It is not. It is a manufacturer with a chemistry advantage, and you own a chemistry advantage through the cycle, not through a quarter.

Supply, not demand, is the near-term ceiling, which makes every quarter a manufacturing report card. The two lines I grade are operational: whether vertical integration shows up in the 1.6T gross margin, and whether OCS ships what it promised for the back half. Both get answered within two quarters, an unusually short wait for a thesis this size. Behind them sits the thing no model carries, scale-up optics inside the cluster, and if that lands in late 2027 this is a bigger company than anyone has written down.

So here is the thesis as a single choice. If you want exposure to how AI clusters will talk to each other for the next decade, you can buy the switch, the module, or the light.

I want the light. Lumentum is the light.

Forever long, held alongside GLW as structural optical infrastructure rather than a swing. I do not chase it on a green day, and I do not sell it because somebody got spooked on a mechanical unwind; on the contrary, I will try to add on good macro dip. Forever long is not the same as never worried, and position size is what makes both of those true at once.

DISCLAIMER

This analysis is for educational and informational purposes only. It does not constitute financial advice or a recommendation to buy or sell any security.

July 31st, 2026 | Price: ~$693 | Sources: Q1–Q3 FY2026 Earnings Releases · Q1–Q3 FY2026 Earnings Presentations · NVIDIA Strategic Investment (Mar 2026) · AXT Capacity Reservation (Jul 2026) · March 2026 Investor Day | NFA. DYOR.

Somebody sells the wire. In May you could not buy a basket of them. Now you can buy four. EUV, FOTO and LAZR listed between May and June. On August 6, Roundhill’s LYTE finally went live, the one I told you to watch and could not buy. Four funds, one summer, one very crowded idea.

I published this piece on July 27 with three funds and a footnote about a fourth. The footnote is now the most interesting fund on the board, so here is the updated version. If you read the original, skip to LYTE and to the new ranking at the end. Everything else stands.

Photonics is just light doing a job electricity used to do. Lasers, fibre, transceivers, silicon photonics, all of it moves data as photons instead of shoving electrons down a copper wire, and that swap is the whole investment case. AI compute has scaled about 60,000 times in twenty years while the links between the chips scaled maybe 30 times. The processors outran the wires, copper cannot close a gap that wide, so the connections are going optical. That is the trade.

And the demand is already funded, not forecast. TSMC is guiding to 52 to 56 billion dollars of capex in 2026, chip equipment sales just set a record, and the silicon photonics market alone is modelled to roughly quadruple by 2032. When the buyers have already committed the cash, demand is the easy part. Price is the hard part.

Which is the catch. Every name in these funds sits downstream of somebody else’s budget, selling to hyperscalers, fabs and telcos, and those buyers all cut capex together when the cycle turns. The bottleneck is structural, but the revenue that fixes it is cyclical, so expect drawdowns that have nothing to do with whether you were right. FOTO is eleven weeks old and has already swung from 18.87 to 27.86, a 48% range. Size accordingly.

One more warning built into a ticker. LAZR used to be Luminar, the lidar company that went bankrupt and had its stock wiped to zero this April. Tema recycled the symbol of a photonics blowup for a photonics fund. Not a red flag, just a reminder that this theme has already eaten people.

Four funds trade today, they all claim the same word, and under the hood they share almost nothing. Except Lumentum. Three of the four have Lumentum as their single largest position, at 15.67%, 14.83% and 14.52%. The fourth has it seventh. Whatever else these funds disagree about, they agree on that.

_{WHY INVEST, AND WHY NOW}

Three reasons it is a now trade, and a basket trade.

• The tech left the lab this year. Co-packaged optics stopped being a whitepaper. At March’s OFC conference, Marvell and Lumentum demonstrated optical switching built for AI racks, not slideware. Demos are not volume, but in this industry the jump from a working demo to a paid design win is measured in quarters, and that clock started.

• The ETF wrapper is itself the catalyst, and LYTE just proved it. Photonics stayed a stockpicker’s corner because the best names were too small or too foreign for big money to touch. Funds change that plumbing. In July I wrote that a fourth issuer filing into this theme was confirmation the category is real, and that you should assume LYTE arrives with size. It traded 72 million dollars on day one, more than Roundhill’s own DRAM fund did on its debut, and Roundhill’s release says DRAM went on to become the fastest ETF ever to 25 billion in assets. Six trading days later LYTE held about 239 million dollars, up from a 1 million dollar seed on August 6. EUV took eleven weeks to gather 485. LYTE is pulling money in more than four times faster than the biggest fund in the category did. Getting swept into a fund is how a small photonics name re-rates, and there are now four vacuums doing the sweeping.

• Nobody knows which layer wins yet, and now you also have to guess which country. Even if you could buy every name, you would still have to pick where the profit lands: the module assemblers, the laser chip makers, the wafer houses upstream of both, or the Chinese incumbents shipping the most volume today. This early, that is a coin flip with four sides. A basket lets you own the whole question until the market answers it, which is the honest case for a fund over a single stock.

_{THE THREE FUNDS WORTH YOUR TIME}

_{COMPOSITION: WHAT YOU ARE ACTUALLY BUYING}

One tell explained the first three funds: each calls itself pure play and each defines it differently. FOTO sets a 50% revenue bar, then reserves the right to ignore it. EUV’s screen has a fourth test that ignores photonics revenue entirely, a company qualifies just by ranking top ten in the activity, which is how TSMC gets in. LAZR simply bought a private AI lab.

LYTE adds a fourth definition and on paper it is the strictest of the group: a hard 50% revenue test with no listed escape hatch, plus a 1 billion dollar minimum market cap and a 10 million dollar daily volume floor. The catch is that LYTE’s screen is rigorous about the business and completely silent about the passport. Same phrase, four definitions, all legal. So trust the holdings, not the label. Here is what the holdings say.

FOTO is the concentrated pure-play, the tightest optical bet on the board. Fifteen names, led by Lumentum at 15.7%, then IPG 13.4%, nLIGHT 11.3%, Fabrinet 11.0%, Coherent 10.6% and Ciena 8.7%. No TSMC, no ASML, nowhere to hide, and the top ten is about 90% of the fund. If the interconnect trade is what you came for, this is the undiluted version in a US-only wrapper. One wrinkle: FOTO holds most of these through total-return swaps against cash, not the shares outright, so a raw holdings screen shows a fat collateral line and double-counted legs, which is why vendors quote its weights so differently. The critique that survives all that: the top six are liquid US names you could buy yourself in ninety seconds, so you are paying 0.75% for a basket you could replicate. The defence: most people won’t, and won’t rebalance it either.

The financials. 136M to 175M in assets depending on the vendor, 0.75% fee, no distributions yet. Options listed on Cboe on June 16, a real liquidity signal this early. It has swung from 18.87 to 27.86 and trades at 21.02 today. Read eleven weeks of tape as noise.

LAZR is the picks-and-shovels basket, and the August file makes it the most differentiated book on the board. It owns the suppliers, not the brand on the module. Lumentum still tops it at 14.8%, then AXT at 9.9%, the wafers laser chips are grown on, and Aixtron at 9.9%, the German machine that grows them. Below that sits equipment and materials nobody else packages in one ticker: Aehr 5.1%, LandMark 4.8%, Furukawa 4.5%, Tower 4.5%, then Soitec, IQE, IntelliEPI, WIN Semiconductors, Sivers, Chroma and MPI. Twenty-seven names, top ten 66.2%, which makes this the least concentrated fund here after EUV. If you think the real chokepoint is wafer and laser-chip supply rather than assembly, this is still the only fund that expresses it.

Then the number that changed my mind. In July, Anthropic was 10.64% of this fund and the main reason I would not rank it first. It is now 2.72%. The stake did not shrink. It is still 2.0 million dollars, the same position, unchanged. The fund quadrupled around it, because creations buy the public basket and cannot buy more of a closed private stake. Mark it down from an implied 1.8 trillion to July secondaries near 1.2 trillion and the fund loses 0.9% of NAV, not 3.5%. To the 965 billion primary, 1.3%, not 4.9%. The tax on the best portfolio here just got refunded, and Tema did not have to do anything to refund it.

Be clear about what that is, though. Dilution is mechanical, not a decision. If Tema adds, or if the fund shrinks, the weight climbs straight back. There are now twelve points of headroom under the 15% cap, so the live question flips from whether they will be forced to trim to whether they choose to buy more.

The financials. Still the smallest of the four at 73.6M as of August 13, though it has roughly quadrupled since the July file. 0.75% fee. The spread was 64 basis points in July, though that was measured when the fund was a quarter of its current size, so treat it as an upper bound rather than a current quote. 45.9% US-domiciled, 12.01% in five Chinese names, unhedged, and roughly a quarter spread across Germany, Taiwan, Japan, Israel, France, the UK and Sweden.

EUV is the big one, and barely a photonics fund. Cheapest at 0.35%, largest at 485M, tightest-trading of the four. But its top five, TSMC, ASML, Applied Materials, Lam Research and KLA, are 39% of the fund and the exact names already topping the semiconductor ETF you probably own. Add Corning at six and the first optical name, Lumentum, is only seventh at 3.89%. The defence is stronger than the critique: EUV’s tail is the most complete in the category. Below the chip-equipment block sit Novanta, Viavi, Ouster, nLIGHT, Camtek, Aehr, IPG, AXT, Hesai, Aeva, Tower, ADTRAN, POET and Lightwave Logic. Nobody else puts POET and Lightwave Logic in one ticker for 35 basis points. So it is a barbell: fab equipment up top, real photonics microcaps in the tail, not much in between.

The financials. 485.5M across 41 holdings, 0.35% fee, annual distributions, trading within about 4 basis points of NAV. Top ten 57%, the most diversified of the four. One flag the marketing skips: Corgi is a newly registered adviser running its first fund, and the prospectus says so under New Adviser Risk.

LYTE is the China trade, and the most literal version of this thesis on the board. Thirteen companies, split almost in half. The US side is 47.7%: Lumentum 14.5%, Ciena 13.9%, Coherent 13.7%, Applied Optoelectronics 2.2% and AXT 1.3%. The other side is the one no other US-listed fund weights this way: Eoptolink 14.0%, Zhongji Innolight 13.1%, Suzhou TFC 8.9%, Yuanjie 5.9%, Accelink 4.8% and T&S 1.1%, which is 47.6% in mainland China, plus LandMark in Taiwan at 1.8%. Fabrinet at 4.1% is the Thai balance.

That split is the whole reason to care. This article argues the connections are going optical. The two companies shipping the most optical transceivers into AI racks are Innolight and Eoptolink. FOTO does not own them. EUV does not. LAZR does, but at about 6% between the pair. LYTE holds roughly 27% in those two names alone, four times LAZR’s weight, which is the difference between a nod to the module leaders and a bet on them.

Open the holdings file and it looks like a third of the fund sits in Treasury bills. It does not. Same swap structure as FOTO, different reason. The 27.1% T-bill and 3.6% money fund are collateral, offset by a minus 29.8% cash line, so net cash is under 1%, and each Chinese name shows up twice, once as a swap leg and once as equity. Add the two for the real weight, or use Roundhill’s factsheet, which already does. Why bother at all? Tax rules. A fund like this has to get most of its book into sub-5% positions at quarter end, and thirteen names with four sitting near 14% cannot.

Which means the swaps are plumbing, not leverage, and I owe you a correction. In July I called this the most aggressive prospectus of the group. The permissions are aggressive. The practice is not. Gross exposure is 99% invested with no leverage at all. Wide language, plain long book. Watch the first rebalance.

Concentration is the real risk. Top ten is 95%, the tightest of the four. Top five is 69%. Four names sit near 14% each, so one bad guide from Lumentum, Coherent, Ciena, Eoptolink or Innolight moves the whole fund. And there is no tail: the 1 billion dollar market cap floor screens out POET, Lightwave Logic and every other lottery ticket EUV holds. Thirteen incumbents, nothing else.

One warning before you buy two of these. Roughly 46% of LYTE is Lumentum, Ciena, Coherent and Fabrinet, all names FOTO already carries near the top. Owning both is not diversification, it is doubling four stocks and bolting China onto the side. If you want both exposures, pair LYTE with LAZR.

The financials. 0.65% fee, cheaper than FOTO and LAZR, nearly twice EUV. Options from day one. About 239 million in assets on August 13, up from a 1 million dollar seed six trading days earlier. Traded at 26.16 against an implied NAV near 26.20, essentially flat to NAV in week one. Active in name, quarterly rebalanced in practice.

The flag the marketing skips: half this fund trades in Shenzhen and Shanghai, which close thirteen hours before New York does. The 4pm NAV is struck off stale marks, A-shares halt often and in size, and every swap faces one dealer. None of it is hidden. It is all in the prospectus. It is just not in the pitch deck.

_{WHO EACH ONE IS FOR}

FOTO is for the investor who wants the optical interconnect trade in dollars and nothing else. No fab equipment diluting the thesis, no private marks, no China headline risk, no overnight gap between what you own and what is priced. The tradeoffs are real: the swap structure takes a minute to understand, the top six are replicable in ninety seconds, and you are buying manager discretion rather than a rule.

LAZR is for the investor who wants the upstream supply chain and will hold names he has to look up. Highest torque, most differentiated portfolio, the widest spread of upstream materials and equipment names in a single wrapper, and the private-mark problem that used to define it is down to 2.72%. It is still the smallest fund here, likely still the widest to trade, and half its book is small foreign industrials. Size it like a sector bet, not like a core holding.

EUV is for the investor who wants cost, liquidity and a real tail, and treats photonics as a tilt rather than the thesis. If you are more “the whole chipmaking chain” than “optics specifically,” this is your fund and it is the cheapest on the board. Do not mistake it for a pure photonics play. Roughly 39% of it is fab equipment you may already hold twice over in a chip fund.

LYTE is for the investor who believes the module thesis literally and will accept a jurisdiction as the price of it. It is the only fund that owns the companies actually shipping the most optical transceivers into AI racks, at a fee below FOTO and LAZR, with day-one options and near-NAV trading. You are buying thirteen incumbents, no tail, no discovery, and a book that reprices overnight in a market that halts. If a China headline would make you sell at the open, this is not your fund, and there is no version of it with the China removed.

_CONCLUSION

Same bottleneck, four different bets. FOTO is the concentrated swing at US optical interconnect. EUV is chip equipment that happens to touch light. LAZR reaches upstream to the companies that supply the module makers. LYTE buys the module makers that actually ship, wherever they are incorporated.

For most readers, still buy FOTO. One fund whose fate rides on the copper-to-light thesis and nothing else, liquid, dollar-denominated, no overnight gap, no mark you cannot audit, no headline that halts it. It is the one I would still hand a first-time buyer. Set and forget.

My personal ranking changes more than I expected. LAZR first, LYTE second, FOTO third, EUV fourth, and the gap between the top three is small.

LYTE moves straight into second and is the fund I find hardest to argue against on the stocks alone. If the thesis is that the world is replacing copper with light, the companies doing the most of that replacing are in this fund and in no other. Eight days ago you could not buy them. That is a real gap in the market being closed, not a repackaging exercise, and it is the only one of the four where I can point at the portfolio and say those are the volume leaders rather than those are the tickers we could get.

LAZR goes to first, and I want to be precise about why, because nothing in the portfolio got better. In July I wrote that on the stocks alone I would rank it first, and that the only thing stopping me was one dollar in nine parked in a mark I could not audit. That dollar is now one in thirty-seven. The book I already liked is unchanged and the wrapper stopped charging me five points of NAV to hold it. It is still the smallest fund on the board and was the widest to trade when I last measured it, so size it accordingly, and watch the private weight after the next rebalance.

Which is the pattern worth naming. Every fund on this board charges a tax for its best idea. LYTE’s tax is a jurisdiction you cannot hedge. FOTO’s tax is 75 basis points for six stocks you could buy yourself. EUV’s tax is 39% semicap you already own. LAZR’s tax was the steepest of the four and it got refunded by nothing more than the fund getting bigger. That is the whole reason the order moved, and it is worth remembering that a tax cut you did not negotiate can be taken back the same way.

EUV lands fourth: cheapest, most liquid, best tail, but a photonics tilt rather than a photonics fund.

_{THE BEAR CASE, BECAUSE I ALWAYS GIVE YOU ONE}

Every fund here is under four months old, none has a track record, and the largest is run by a first-time adviser. FOTO is six tickers you could buy for free, wrapped in a swap. LAZR is the smallest fund here, half of it is small foreign industrials, and the private mark I just stopped worrying about shrank as a share of the fund only because new money arrived, not because anyone sold it. EUV is the semis you already hold. And LYTE, the one I just promoted, is thirteen stocks where the top five are 69% of the book, half of it settles in a market that halts, and every swap leg faces one dealer. Seven days of trading is not a track record, it is a launch.

The whole theme is AI with more leverage and less liquidity. In a capex pause the upstream names fall hardest, LAZR most of all, and FOTO already printed a 48% range in eleven weeks to show you what that feels like. The newest fund inherited its ticker from a photonics company that went to zero in April. The bottleneck is real. The path through it is not.

_DISCLAIMER

This analysis is for educational and informational purposes only. It does not constitute financial advice or a recommendation to buy or sell any security.

August 14th, 2026. Updated from the July 27 edition | FOTO $21.02 · LAZR $47.11 · EUV $26.76 · LYTE $26.16 | Sources: FOTO, LAZR, EUV and LYTE fund factsheets, holdings files, decks and prospectuses · SEMI / GS / Research and Markets TAM research | NFA. DYOR.

_{BEFORE WE DIVE IN}

Most companies in the AI buildout occupy one layer. Penguin occupies three and sells into a fourth.

Layer 6: Memory. The Integrated Memory segment, up 111% year over year. But Penguin sits at Layer 6 differently than the names usually parked there. Micron, SK hynix, and SanDisk own fabs and make the raw dies. Penguin runs a “Fab-Light” model: it buys their commodity memory and engineers it into specialty modules built for a specific customer’s machine.

The fabs make the bricks. Penguin makes the wall.

That has three consequences. No fab, so no multi-billion-dollar capex cycle. It earns a spread rather than the full price move, which is why Integrated Memory carries below-average margins even while it doubles. And a memory upcycle cuts both ways, lifting revenue while raising what Penguin pays its suppliers.

What keeps this from being a middleman story is CXL, an interconnect protocol that lets memory hang off the fabric instead of sitting captive beside the processor. Penguin’s CXL expansion cards and MemoryAI KV Cache products sit exactly on that seam, where memory stops being a component and starts being part of the fabric. Memory becomes a system rather than a part, and systems carry pricing power.

Layers 5 and 3: Compute Hardware and Software. The Advanced Computing segment. Penguin makes neither GPUs nor CPUs. It assembles them with memory and networking into working racks (Layer 5), then runs them on its own ClusterWareAI cluster software (Layer 3).

Layer 4 is the customer, not the business. Penguin does not rent capacity. It builds the AI factory that enterprises, sovereign programs, and neoclouds operate. Its name for that whole vertical slice is “the AI Factory Platform.”

The third segment, Optimized LED, is the Cree LED business, outside the AI stack entirely. It grew 7% in Q3 and management guides it down about 5% for the year. Cash-generative, not why anyone owns the stock.

Management’s own framing:

“We are seeing very strong AI-driven customer demand for memory and AI infrastructure solutions. As inference and agentic AI workloads become more persistent and context-rich, memory is increasingly becoming one of the primary performance and scalability bottlenecks.”

Kash Shaikh, CEO, Q3 FY26 earnings call, July 7, 2026

Two housekeeping notes: the fiscal year ends the last Friday in August, so “Q3 FY26” ended May 29, 2026. And the parent redomiciled from the Cayman Islands to Delaware on June 30, 2025. Same business, new passport.

Read more

Meanwhile in Seoul, where the same company has traded since 1996, it’s down 15.7% since the ADR priced. Two exchanges, one company, and a 23% gap between them that was 42% two days ago.

Prices as of July 16, 2026, and moving violently.

BEFORE WE DIVE IN

SK hynix sits at Layer 6: Memory. If you’ve read the Micron and SanDisk pieces, you know why this layer matters: Layer 5 gets the headlines, but every accelerator in it is dead weight without the chips that hold the data and feed it fast enough.

SK hynix is the only company that sits in four of those five. Neck and neck with Samsung in DRAM, trading the top spot quarter to quarter. Number two in NAND at 20.9%. LPDDR going into everything from phones to NVIDIA’s next platform. And in HBM, the bucket that is actually scarce, roughly 56% of a market only three companies on Earth can serve and the first to reach HBM4 mass production.

That breadth is the point. SanDisk is a NAND story. Micron is the American DRAM and HBM story. SK hynix is the whole layer, and it leads the one bucket everybody is short of.

Which brings us to why this piece exists now. As of July 10, 2026, you can finally buy it. SK hynix listed ADRs on Nasdaq, raising $26.5 billion, the largest US listing a foreign company has ever done. Until last Friday, the best-positioned company in Layer 6 was effectively unbuyable for most American investors.

“SK hynix seeks to be wherever AI is, continually demonstrating our technology leadership.”

Kwak Noh-Jung, CEO, at the Nasdaq opening bell, July 10, 2026

The short version of the thesis: the buildout hit a wall money cannot fix, one company sits on the far side of it, and the stock trades at eight times what it will probably earn this year.

Read more

_{BEFORE WE DIVE IN}

Nebius sits at Layer 4: Cloud Infrastructure. This is the layer that takes all the silicon underneath, the GPUs, the memory, the interconnect, and assembles it into something a customer can actually use: powered, cooled, networked, and rented out by the hour.

Layer 4 has four buckets:

• Hyperscaler clouds (AWS, Azure, GCP) rent general-purpose compute at massive scale and run their own full stack.

• Data center colocation providers own the buildings and the power, then lease that space to whoever fills it.

• Edge and inference clouds push compute close to the end user to serve models with low latency.

• NeoClouds, the bucket Nebius lives in, are GPU-specialized from the ground up: built for one job, training and running AI, and nothing else.

Nobody trains a frontier model on a chip. They train it on a cluster: thousands of GPUs wired together, fed by memory and interconnect, sitting inside a building that draws the power of a small city. Building that building, filling it, and keeping it running at high utilization is its own bottleneck, and it is the one the NeoClouds solve.

The shortest version of the thesis: the hyperscalers committed to spend hundreds of billions on AI through 2027, and they physically cannot pour concrete and rack GPUs fast enough to keep up with their own demand. So they are outsourcing the overflow. Microsoft and Meta, two companies that build data centers for a living, looked at their own buildout timelines and decided it was faster to rent capacity from Nebius than to wait on their own. That is the name of the game. The compute crunch is so severe that the buyers who build are now renting.

Nebius is one of a small handful of independent “neoclouds” positioned to catch that overflow. Nebius is the one that owns its full stack, from the data center design down to the servers and racks, and it is the one that walked into this market with a Nasdaq listing already in hand.

Before we go further, a note on what makes this name different from the rest of the AI trade. Most of the supply chain is a parts business: you sell a chip, a substrate, a laser, a stack of memory. Nebius sells time on a machine it owns. That changes the economics, the risks, and the valuation math completely, and it is the whole reason the stock looks the way it does.

_{SECTION 1 · COMPANY SNAPSHOT}

What is NBIS and how did a Yandex spinoff end up in the Nasdaq-100?

First, what it is today. Nebius calls itself "the AI cloud company": a single platform that covers the whole AI journey, from data and model training to tuning, inference, and deployment, built on deep in-house engineering and already serving enterprises across healthcare, robotics and physical AI, financial services, media, and retail. It is headquartered in Amsterdam and trades on Nasdaq as $NBIS. That positioning matters because AI is shifting from research into large-scale production, and production is where cloud infrastructure gets reshaped.

Now the strange part: where it came from. Nebius is what was left of Yandex, the “Google of Russia,” after the 2022 war forced a split. Yandex sold its entire Russian business and kept the assets that sat outside Russia: a data center in Finland, a self-driving unit, a data-labeling business, an edtech platform, and roughly $2.5 billion in cash. Founder Arkady Volozh renamed those pieces Nebius Group and resumed trading on Nasdaq in October 2024.

So this is a “startup” that is already public, with real operating history and a founder who has built a continental-scale tech company once before. Volozh took Yandex to a $30 billion peak. He is now doing it a second time, in a new country, in the most capital-intensive corner of the hottest market in tech.

For most of 2025 the stock flew under the radar. Then the contracts landed: $17.4 billion from Microsoft, a Meta deal that grew from $3 billion to $27 billion, and a $2 billion equity check from NVIDIA. The market repriced the company in months. Shares ran from the low $40s to a high near $300, Nebius joined the Nasdaq-100 in June 2026, and a hedge fund run by former OpenAI researcher Leopold Aschenbrenner disclosed a 5.6% stake worth about $2.6 billion, its single largest position. The leftover asset became one of the most fought-over names in AI infrastructure.

_{SECTION 2 · FUNDAMENTALS}

How does it make money and who’s running the show?

Nebius organizes around one core business and a cluster of optionality. The core is Nebius AI Cloud: large-scale GPU clusters, a full software platform on top, and developer tools for training and running models. In Q1 2026 it did $389.7 million of revenue, about 98% of the group. Everything else is upside that has not been priced yet.

The model is straightforward to describe and brutal to execute. Nebius secures power and land, builds or leases data centers, fills them with NVIDIA GPUs, wraps its own software stack around the hardware, and sells access. The differentiator it leans on is vertical integration: it designs its own servers, racks, and connectivity rather than just renting bare metal, which it argues lets it ship capacity faster and run it at higher margins. Volozh frames the constraint the same way every quarter: this is a supply problem, not a demand problem. On the Q1 call, management said demand was vastly exceeding capacity and that pipeline generation hit a record, up roughly 3.5x quarter over quarter.

The management team

Three seats matter most, and each maps to something Nebius specifically needs.

Founder and CEO Arkady Volozh built Yandex into a $30 billion company over two decades. Having built a continental-scale tech company once already, he is the rare operator with the conviction and the credibility to bet tens of billions on a GPU buildout and make hyperscalers believe he can deliver it.

CFO Dado Alonso joined in 2025 from Amazon, Booking.com, and Naspers/OLX. At Nebius the funding model is the business model: someone has to structure asset-backed financing against multi-year hyperscaler contracts while managing dilution on a company raising billions a quarter. That makes the finance seat the most important operational seat in the building, and it is held by someone from high-growth tech finance, not a sleepy treasury desk.

CRO Marc Boroditsky joined in 2025 after growing Twilio’s revenue more than 10x to $4 billion. You do not hire a closer of that caliber unless you plan to keep signing contracts the size of the Microsoft and Meta deals already on the books.

Who actually buys from Nebius?

Two names dominate, and that is both the bull case and the bear case in one breath.

Microsoft was first. In September 2025, Nebius signed a five-year deal worth $17.4 billion, expandable to $19.4 billion, supplying dedicated GPU capacity out of a new data center in Vineland, New Jersey. It was Nebius’s first contract with a major tech player, and Volozh said at the time he expected “more to come.” He was right.

Meta came next, and the structure is worth reading closely because it is more than a headline number. The latest agreement is a $27 billion, five-year deal, the largest in Nebius’s history, and it comes in two parts: a $12 billion, five-year purchase of dedicated compute capacity scheduled to begin in early 2027, plus up to $15 billion more that Nebius can sell to Meta on pre-agreed terms, or to its own AI cloud customers at market rates. That second leg is the clever bit. It is committed capital and long-term visibility, but Nebius keeps the option to reallocate the capacity if it finds a better home for it. The same month, NVIDIA invested $2 billion directly into Nebius, the same playbook it ran with CoreWeave, securing a strategic foothold in a buyer of its next-generation platforms.

Stack those together and Nebius entered the back half of 2026 with a backlog around $46 billion, the overwhelming majority of it Microsoft and Meta.

Sit with that. A roughly $3 billion revenue company is sitting on a $46 billion order book, and two customers are most of it. That is demand visibility most infrastructure businesses would kill for. It is also a dependency most would lose sleep over. Both are true at once. That is what owning Nebius means right now.

_{SECTION 3 · FINANCIAL PROFILE}

Nebius has crossed from obscure spinoff into one of the fastest-growing names in public markets, and Q1 2026 is where the growth became impossible to argue with.

Revenue hit $399 million, up 684% in a year. The core AI cloud business did better still, up 841% to $390 million, now 98% of the company. The number the Street actually tracks is ARR (annual run-rate, basically the last month’s revenue times twelve), and it jumped 54% in a single quarter to $1.92 billion. For scale: Nebius did $530 million in all of 2025. It just exited a quarter running at nearly four times that.

The profitability turn is real, but read it carefully. On an adjusted-EBITDA basis (rough operating profit), the AI cloud unit cleared a 45% margin, almost double the prior quarter, and management is steering toward 20-30% operating margins. The headline GAAP profit of $621 million, though, is mostly a $781 million paper gain on its ClickHouse stake, not money earned. Strip that out and the business still runs a loss, and the cleaner figure actually got worse: an adjusted net loss of $100 million versus $84 million a year ago. Scaling toward profit, not there yet.

Cash flow is the interesting part. Operating cash flow hit a record $2.3 billion, swinging from an outflow a year earlier, almost entirely because customers pay upfront. Those prepayments dropped $3.2 billion into the quarter and sit at $4.8 billion on the balance sheet. Cash on hand is $9.3 billion. But the appetite is enormous: Nebius raised $6.3 billion in Q1 alone ($4.3 billion in convertible debt plus NVIDIA’s $2 billion equity check), and debt doubled to $8.4 billion in three months. That is the tell. This business burns capital as fast as it grows, and the whole thesis rides on it raising more on friendly terms.

For 2026, management guides to $3.0-3.4 billion in revenue, $7-9 billion ARR by year-end, and roughly 40% margins, funded by a staggering $20-25 billion capex budget aimed at 2027 capacity.

The number that matters most is power. Contracted capacity is now 3.5 GW and climbing past a raised 4 GW target, with Nebius owning over 75% of it. Live, revenue-generating capacity should hit 800-1,000 MW by year-end, up from just 170 MW. That is a five-to-six-fold expansion in twelve months, and the footprint went from one large site to seven.

Bottom line: few companies in public markets are compounding this fast, the core engine is turning profitable, and the balance sheet is being levered hard to build capacity that is mostly pre-sold. The question was never whether Nebius is growing. It is whether the money keeps showing up without crushing the stock.

_{SECTION 4 · PRODUCT AND TECHNOLOGY}

Full-stack, vertically integrated, and racing up the value chain

Most neoclouds just rent you a GPU. Nebius rents you the whole machine around it. It designs its own servers, racks, and networking, then layers its own software on top, so customers get a managed AI platform instead of bare hardware they have to wire up themselves. The flagship layer, Nebius AI Cloud, got more enterprise-ready with its March release, Aether 3.5: run a model without planning capacity in advance, move data across clouds at massive scale, and tighter security and audit controls. That range is the point. It can serve a two-person startup and a hyperscaler renting a gigawatt off the same platform, and management bets that software is what keeps customers sticky and margins high.

This is the difference between a landlord and a commodity wholesaler. Anyone with cash and a NVIDIA allocation can stack GPUs. The moat, if there is one, is speed, utilization, and software that makes the compute easy to use. It helps that Nebius is a NVIDIA reference partner with early access to the newest chips, including being among the first in line for the next-generation Vera Rubin platform in late 2026.

The more important shift is happening up the stack, in inference, the part where a trained model actually answers requests. Nebius spent 2026 climbing there through a mix of acquisitions and team-and-IP deals: Tavily (agentic search), Eigen AI (inference and model optimization, with technology NVIDIA ranked #1 at GTC 2026), and Clarifai, where it brought in the core team and licensed the inference IP for system-level orchestration rather than buying the company outright. Each added tools to run models faster and cheaper. The logic mirrors what Micron learned: training is the loud, headline-grabbing workload, but inference is the one that runs forever and pays a recurring, high-margin bill. Whoever drives down the cost per answer and makes deployment painless embeds themselves in the part of AI that never stops running. Nebius is bolting all of it onto Token Factory, its inference platform, which already counts Revolut and monday.com among its customers.

The other half of the moat: power and land

You cannot conjure a gigawatt overnight, so Nebius plays two timelines at once: bring capacity online now, and lock land and power for years out. Its owned sites span the original Finnish data center, a New Jersey site serving Microsoft, two 1.2 GW “AI factories” in Missouri and Pennsylvania, and a 310 MW build in Finland set to be one of Europe’s largest. All told, five owned sites should deliver 3 GW. To power them without waiting in the grid-connection queue, now one of the industry’s worst bottlenecks, Nebius struck a deal with Bloom Energy for 328 MW generated on-site.

Power is the new oil here. Land with a live grid connection and a signed power contract is getting scarcer than the GPUs themselves, and whoever locked it early, before the rush and at the right price, owns the chokepoint. That is the whole Nebius pattern: capacity contracted, owned, and largely pre-sold before the first rack ships.

_{SECTION 5 · GROWTH CATALYSTS}

What upcoming milestones could send this stock significantly higher?

Here is the thing about Nebius’s catalysts: they stack. That is what separates a hot quarter from a lasting re-rating, and it is why the timeline below reads as one chain rather than six scattered events.

The logic runs in order. Power comes online, fills up, and turns into run-rate revenue. That revenue compounds toward the multi-billion-dollar guide. A swelling contract book then converts the company from a volatile compute reseller, exposed to spot pricing, into something closer to a contracted utility, where every signed name lifts the revenue floor. And the part almost no one has put in a model yet is the move up the stack: as Nebius starts earning software and inference margin on top of pure rental margin, a capacity business quietly becomes a platform.

Each step de-risks the next. The capacity makes the revenue real, the contracts make it durable, and the software makes it more valuable. Miss one and the chain still mostly holds; hit all of them and a 684% quarter stops looking like a spike and starts looking like the first chapter. The timeline below is that chain, from the next earnings print to the destination it all points toward.

_{SECTION 6 · MACRO & INDUSTRY BACKDROP}

The compute supercycle, and where Nebius sits inside it

The hyperscalers are spending hundreds of billions on AI through 2027, and chips are no longer the bottleneck. The real constraint is physical: building, powering, and cooling data centers fast enough to actually rack those chips. A GPU sitting in a warehouse, not plugged into a live, powered cluster, earns nothing. Neoclouds exist to close that gap, and demand is so fierce that even Microsoft and Meta are renting the overflow.

Geography cuts both ways for Nebius. Its Amsterdam base and European engineering core position it perfectly if Europe gets serious about building its own sovereign AI capacity instead of renting from American hyperscalers, an underpriced tailwind. The flip side is the Yandex heritage and the geopolitical baggage that trails it, which we will get to in the risks.

The sector’s bear case is simple and has played out before: compute pricing is not guaranteed. GPUs depreciate fast, everyone is building at once, and the moment supply catches demand, rental rates can roll over while the debt and depreciation keep grinding. Nebius’s answer mirrors Micron’s: contracted, supply-constrained, and largely spoken for today. Whether that holds as the whole industry builds is the entire question.

Where Nebius sits in the neocloud field

The trick to this sector is seeing that the companies lumped together as “neoclouds” run three different businesses. Picture a stack. At the top, you operate the AI cloud and sell software. In the middle, you own the hardware and rent out raw GPUs. At the bottom, you own the building and the power and lease it to someone else. Where you sit decides how much you earn per megawatt and how much risk you carry. Nebius lives at the top.

CoreWeave (CRWV) is the one to beat. It is the largest independent neocloud and the most entrenched with NVIDIA, with Q1 2026 revenue of $2.1 billion and a $99.4 billion backlog, and it earns the most per megawatt in the group. But the model is leveraged and asset-light: it leases its buildings from the likes of Equinix and Digital Realty and carries roughly $30 billion of GPU-backed debt, which is why its net loss widened to $740 million in Q1 even as revenue jumped 112%. Nebius is smaller but growing faster, owns over 75% of its own power, and runs a cleaner balance sheet, trading scale for ownership and durability.

IREN is the closest match, minus the software. Formerly bitcoin miner Iris Energy, IREN owns its power and sites like Nebius and has landed a $9.7 billion Microsoft contract plus an NVIDIA partnership targeting 5 gigawatts. The gap is the top of the stack: IREN rents bare-metal GPUs rather than selling managed software, and it shows in the economics. Nebius’s edge is the software layer IREN has not built.

The former miners (APLD, WULF, CIFR, HUT) are landlords, not rivals. They don’t operate an AI cloud or sell compute. They build campuses, deliver power, and lease it on long, hyperscaler-backed contracts: Hut 8’s 15-year, $7 billion Fluidstack lease backstopped by Google; Cipher’s nearly $5.5 billion AWS lease; TeraWulf’s Fluidstack deals with a $3.2 billion Google backstop; Applied Digital past 1 gigawatt contracted, much of it to CoreWeave. Sitting low means cleaner, longer-duration economics that dodge the GPU refresh cycle, but far less value per megawatt, lighter software, and heavy reliance on outside capital. These are Nebius’s suppliers and build-out peers, not competitors for the same customer.

So what sets Nebius apart? It is the only name that owns all three layers at once: power like the landlords and IREN, the cloud like CoreWeave, and a managed software and inference platform almost none of them have. That is why it earns more per megawatt than the pure-infrastructure plays and runs a cleaner balance sheet than the scale leader.

The honest caveat: none of this is a monopoly. Anyone with capital and a NVIDIA allocation can rack GPUs. Nebius’s moat is execution speed, owned power, utilization, and software, not physics the way it is for memory. Real edge, but only while it keeps out-executing a field racing just as hard.

_{SECTION 7 · RISKS}

What could break this

Nebius just printed 684% revenue growth, a backlog near $46 billion, and a stock up more than 450% in a year. When a business is this hot, the risks do not vanish. They just get easier to ignore. So let us not.

1. The funding model is the business model. This one overrides the rest. Nebius is guiding to $20-25 billion of capex against barely $3 billion of revenue, raised over $6 billion in Q1 alone, and doubled its debt to $8.4 billion in a single quarter, with more borrowing already flagged. The entire machine depends on raising capital, over and over, on terms that do not gut the equity. If AI funding tightens or the cost of capital spikes, the buildout stalls and the contracts get harder to honor. The bull case quietly assumes the money always shows up. History says be careful with that.

2. Two customers are most of the backlog. Microsoft and Meta dominate that $46 billion book. Wonderful while they are hungry, dangerous the moment one slows, renegotiates, or decides its own data centers have caught up, and both are building in parallel. There is not much cushion underneath. Note too that up to $15 billion of the Meta deal is capacity Nebius can reallocate, flexible, but not a hard, dedicated purchase.

3. The financing loop looks circular. NVIDIA puts $2 billion into Nebius, then Nebius buys NVIDIA GPUs. It is the same loop running through most of the AI buildout, and it makes demand look more independent than it is. Nothing improper, but part of the “demand” is the supply chain funding itself, and that can unwind fast.

4. Compute is closer to a commodity than memory. Only three companies on Earth can make HBM. Far more can rack GPUs. Nebius’s edge is speed, utilization, and software, not physics. CoreWeave and the hyperscalers are in the same race, so if rental rates compress, there is no structural monopoly protecting margins.

5. Depreciation, and a small accounting tell. GPUs lose value fast, on roughly a four-year clock. Starting Q1 2026, Nebius stretched the assumed life of its servers from four years to five. That lowers depreciation each quarter and flatters margins, defensible if the gear really lasts, but exactly the kind of estimate a skeptic watches. And if AI demand pauses, Nebius is left with billions in hardware depreciating against contracts that may not all get honored. The asset base does not forgive an air pocket.

6. Valuation prices perfection. At roughly 65 times sales, the stock cannot survive a single soft quarter unscathed. It ran from a 52-week low near $44 to $300, and a stock that climbs like that can give a lot back fast on any wobble.

7. Heritage, governance, and execution. The Yandex origin carries geopolitical baggage outside management’s control, governance scores poorly, and Volozh keeps outsized voting control through special shares. On top of that, scaling live power roughly six-fold in a year while digesting several acquisitions is a lot going right at once. Any one slip is survivable. Several together compress the story.

Bear case in one sentence: you are paying 65 times sales for a capital-hungry compute provider whose backlog leans on two customers, whose margins depend on a commodity staying scarce while everyone builds, and whose growth plan needs billions more raised on friendly terms.

The counterweight is not small: demand for compute is outrunning supply so badly that the companies who build data centers for a living are renting from Nebius instead. If that holds, a stumble just slows the growth rate. If it cracks, it is a long way down from 65 times sales.

_{SECTION 8 · VALUATION & PRICE TARGETS}

Cheap is not the word. The question is whether the ramp earns the multiple.

Near term, value it on ARR. At about $230 (early July 2026), Nebius is worth roughly $60 billion, about 31 times its current $1.92 billion run-rate. Expensive on its face. But here is the bull’s move: management guides that run-rate to roughly quadruple, to $7-9 billion exiting 2026. Hit that, and the same $60 billion price tag is only about 8 to 10 times run-rate for a company still compounding fast. Suddenly it is not expensive at all. It is cheap for the growth. The argument in one line: do not value Nebius on the run-rate it has, value it on the one it is about to reach.

The bear’s answer is just as simple. A guide is a promise, not money in the bank. That $7-9 billion rides on capacity not built yet, funded by capital not raised yet, sold to two customers building their own in parallel. Pay 31 times today’s run-rate and you are not buying the ramp, you are underwriting it. The market has watched compute get overbuilt before. This multiple is not the market calling Nebius cheap. It is the market betting the buildout lands.

You can see that bet in how far apart the price targets sit: a low of $180, a high of $475, a median of $329.40. That is not a forecast, it is the argument priced two ways. The high is the ramp landing clean; the low is it slipping while debt and depreciation grind on. The median sits about 43% above today’s price, a real but modest base case, with all the real action still in the tails. A year ago the average target was near $55. The easy mispricing is gone. From here, Nebius has to earn it on delivery.

Long term, the yardstick changes. Sales multiples are a near-term lens. As Nebius matures and its depreciation bill balloons, all those GPUs and buildings wearing down on schedule, the better measure becomes EV/adjusted EBITDA, the standard for capital-heavy infrastructure. So the 2030 case has to be built on profit, not revenue, with one caveat stamped on every line: these are assumptions four years out, and even the base case bakes in near-flawless execution.

Three inputs drive it. Capacity: roughly 5,400 MW of active, revenue-generating power by 2030, up from a few hundred today, which already assumes no serious delays at the big US sites (Pennsylvania, Independence, Vineland), the single biggest risk to the number. Revenue per megawatt: about $12 million, a modest lift on today’s rough $10 million sector standard. That math is roughly $65 billion of 2030 revenue. Margin and mix: the subtle one. The Meta and Microsoft deals are lower-margin bare-metal work, fixed pricing with the customer doing everything except plugging in the GPUs. The case needs Nebius to keep tilting revenue toward its own higher-margin cloud (today the split is around 50/50). Do that, and the margin rises as high-margin cloud and inference grow, landing near 52%, which on $65 billion is about $34 billion of adjusted EBITDA.

Put a 20x multiple on that, the “AI infrastructure becomes a utility” case, rich but defensible for a grower this central, and you land near a $674 billion enterprise value. Since debt and cash should roughly cancel by 2030, that is also the market cap, before counting strategic stakes like ClickHouse (Nebius owns about 25%), Avride (about 83%), and Toloka, which the model pencils at another ~$30 billion.

Against today’s ~$60 billion, that base case is about eleven times the current company, and even after dilution the implied 2030 share price is roughly eight times today’s price, near $1,925 a share on ~350M shares. That is the number that makes people lose their minds, so here is the discipline: it still assumes the buildout lands, the mix shifts, the funding stays cheap, and the AI cycle compounds for four more years. Move any one input and the output swings hard.

So is it cheap? Not on anything you can hold today. It is cheap only if you trust the ramp, the buildout behind it, the funding behind that, and the customers anchoring it all. The question was never whether the business is growing. It is whether that growth is fundable and durable enough to grow into a price this rich, this year and for the four after it.

_{SECTION 9 · TRADING THE NEWS}

Meta wants to be a cloud provider, too

The two-way read. Bloomberg reports Meta is building a cloud business to sell access to its AI computing power and models, aimed straight at AWS, Azure, and Google Cloud. It cuts both ways for Nebius, and both are true. Bullish: even Meta, sitting on one of the largest GPU fleets on earth, now sees selling compute as a business worth entering. That is the whole Nebius thesis, compute is the product, said out loud by a hyperscaler. Bearish, and sharper: Meta is Nebius’s largest single contract at roughly $27 billion, and a customer that starts reselling compute is a customer drifting toward competitor. It hardens a risk already on the board, your two biggest names are building in parallel.

Bottom line. Near term, nothing breaks. Meta signed Nebius because its own buildout is behind, and standing up a commercial cloud business takes years, so the $12 billion of dedicated capacity is still needed on day one. This is the risk turning concrete, not detonating. It also makes the flexible leg of the deal, the up to $15 billion Nebius can resell, look smart in a world where Meta wants optionality on its own compute. The line to watch: whether building in parallel quietly becomes buying less from Nebius.

The price hike nobody had to model

Most of this thesis runs on guidance and backlog, numbers management chooses to show you. This one is different. In May, Nebius customers received an email informing them that on-demand GPU prices were going up, and it laid the demand story bare: effective June 1, standard on-demand pricing rose roughly 30% across the board. H100 went from $2.95 to $3.85 an hour, B200 from $5.50 to $7.15, with similar jumps on H200 and B300. The spare-capacity tier (preemptible, the cheaper option you take when you can tolerate interruptions) rose even more, some rates up 70%.

Why it matters. In a commodity market, you do not raise prices 30%. You cut them to win volume. Nebius did the opposite: it raised them, told customers it still offers some of the most competitive pricing around, and nudged them toward locking in reserved capacity instead. That is not a company fighting for demand. It is a company with more demand than supply, rationing it by price. And in this model the increase is almost pure profit, because the data center is already built, the GPUs already racked, the power already paid for. Charge 30% more for the same hour of compute and most of that extra dollar falls straight to the bottom line. It is the cleanest confirmation yet of the “supply problem, not a demand problem” line management repeats every quarter, except this time it came straight to customers, not from a podium.

The bear note. It is one snapshot, on a handful of chips, and it does not translate cleanly to the blended price across the whole book. The Microsoft and Meta deals, most of the contracted revenue, are locked at fixed terms and do not move when the on-demand sheet does. A list-price bump on spot capacity is real, but it is the smaller, more volatile slice of the business. Read it as a signal, not a margin model.

My take. Guidance can be sandbagged and backlog can be front-loaded, but a price increase sent to paying customers is the one number management cannot dress up. It is the market clearing in real time, and it cleared higher. Until the on-demand sheet starts moving the other way, the crunch is real and the pricing power is Nebius’s. The customers are paying up. That is the tell.

_{SECTION 10 · THESIS SUMMARY & VERDICT}

My verdict on $NBIS: own it, but size it like the high-beta name it is.

Strip away every chart and quarter and the whole story is one sentence: the companies that build data centers for a living have decided it is faster to rent from Nebius than to wait on themselves. That is the tell that matters. You do not outsource your own core competency to a startup unless the shortage is brutal and the startup is genuinely better at the one hard thing. Nebius is better at it because it sits at the top of the neocloud stack, owning its power, running the cloud, and building into software and inference, where almost nobody else holds all three. You cannot fake a powered gigawatt, and you cannot bolt on a software layer in a quarter. That is the moat.

Now the part the run obscures. This is not a business that wins by being right once. It wins by raising money, building, filling, and then raising again, dozens of times, without ever paying too much for the capital. Capex is guided to $20-25 billion against barely $3 billion of revenue. The contracts are gorgeous, but two customers carry most of the book, and both are building their own capacity beside it. So the thing you are actually buying at roughly $230 is not the business as it exists. It is your confidence in the next twelve raises and the next two years of execution. Get that wrong and the most beautiful backlog in the sector becomes a liability schedule.

That is why I think about this position as a dial, not a switch. The stock has already traveled $44 to $300 and back to $230 in a year, and it will keep moving like that, because every print either confirms or threatens the ramp with nothing in between. So I hold a core and I feed it on the ugly red days, when the funding fear is loudest and the price is cheapest, never on the green ones. A name like this can shed 30 to 40% on a single soft quarter, and the whole point is to still be holding when it does.

One myth to kill, because it sets you up to be disappointed. People want this to be the “next Nvidia,” a company protected by a product only a handful of firms on Earth can physically make. It is not. Anyone with capital and a NVIDIA allocation can stack GPUs. Nebius has no chemistry monopoly, and its edge is execution, not physics: it ships faster, owns its power, and holds the cleanest contracted book among the independents. That edge is real but fragile, because it lasts only as long as it keeps out-running a field sprinting just as hard. Do not buy this expecting a moat made of physics. Buy it expecting a moat made of speed and software.

So where does that leave the price? The market is paying up for the contracts it can already see and almost nothing for the part I believe in most: Nebius is becoming the AWS of inference, and inference is forever. Training is the loud one-time headline. Inference is the workload that never switches off, every query and every agent, billed by the second for as long as the model runs. And here I stop being an analyst and become a customer. The shortlist for serious AI compute used to be Microsoft, Google, and AWS, full stop. Not anymore. Nebius has entered that conversation, and I am seeing it happen from the buyer’s side of the table, not off a pitch deck. That leg is in almost nobody’s model. Neither are ClickHouse, Avride, or Toloka. If the buildout lands and money stays cheap, today’s price is a fraction of what this becomes. If either breaks, it unwinds fast, and debt and depreciation do not wait for the story to heal. Both paths are live. That is the trade.

I am long, with conviction and a leash. Sized to survive a bad print, never chased on a good day, never left alone.

Long NBIS. High conviction, high beta, sized accordingly.

DISCLAIMER

This analysis is for educational and informational purposes only. It does not constitute financial advice or a recommendation to buy or sell any security.

July 1st, 2026 | Price: ~$230 | Sources: 20-F · 6-K · Q1 FY2026 Earnings Call · Q1 FY2026 Shareholder Letter | NFA. DYOR.

_{BEFORE WE DIVE IN}

Micron sits at Layer 6: Memory. The model weights, the training data, the context window your prompt just filled up, the cache that lets a chatbot answer you in milliseconds: none of it lives inside the GPU itself. It lives next to it, on chips built for one job, holding data and moving it fast enough that the GPU never has to sit around waiting.

Micron doesn’t just sell a component in the AI stack. It controls one of the scarcest inputs in the entire buildout, the memory every accelerator, every hyperscaler, and every frontier AI lab on Earth cannot get enough of.

There is no AI training run without DRAM feeding the GPU in real time. There is no AI accelerator at scale without HBM stacked directly on the package. Micron makes both.

Sanjay Mehrotra, Chairman, President and CEO, put it plainly on September 23, 2025: “As the only U.S.-based memory manufacturer, Micron is uniquely positioned to capitalize on the AI opportunity ahead.”

Before we go further, a quick note on the alphabet soup. Headlines throw around DRAM, NAND, and HBM like they’re interchangeable. They’re not, and the difference is the whole reason Micron’s stock looks the way it does right now.

The fastest GPU in the world is useless if it’s sitting idle waiting for data, and that’s exactly the problem with training and running today’s massive AI models: moving information back and forth eats up more time than the actual computing. HBM fixes this by stacking memory chips directly on top of the GPU, like moving your desk into the room where all the files are kept instead of down the hall. That proximity is why it’s become the real bottleneck in the AI race: you can have all the computing power in the world, but without enough HBM, it just sits there waiting. And unlike regular memory chips, HBM can’t be swapped in from any supplier on short notice. It takes months of testing to qualify a new one for a specific chip design. That’s why the companies that make it have buyers locked into contracts years out. HBM didn’t just get faster, it became the thing standing between “we have an idea for an AI model” and “we can actually build it.”

_{SECTION 1 · COMPANY SNAPSHOT}

What is MU and why did it just cross $1 trillion?

Founded in 1978 in Boise, Idaho, by brothers Ward and Joe Parkinson and two colleagues, it spent the next 47 years as one of the most cyclical names in tech: prices spike, the industry overbuilds, prices crash, and the stock round-trips with it, profitable one year, bleeding cash the next. Wall Street priced it like a commodity, because for most of its history, that’s what it was.

AI broke that pattern. Once memory became the bottleneck standing between every AI lab and the chips it needed, Micron’s customers stopped haggling on the spot market and started locking in multi-year supply contracts instead. Micron’s entire 2026 production sold out before the year even started.

Headquartered in Boise, Idaho, with fabrication and assembly operations across the US, Taiwan, Japan, Singapore, Malaysia, China, and India, Micron trades as $MU on Nasdaq. Fiscal 2025 revenue hit a then-record $37.4 billion, up 49% year over year. Fiscal Q2 2026 came in at $23.86 billion, up 196% year over year. Then fiscal Q3 2026 didn’t just hit the $33.5 billion guide. It blew through it. Revenue landed at $41.5 billion, up 74% sequentially and 346% year over year. One quarter of revenue now exceeds Micron’s entire annual revenue in any fiscal year through 2024.

Wall Street noticed long ago. The stock ripped over 1,000% in a year, and in 2026, Micron crossed a $1 trillion market cap, becoming a trillion-dollar company almost nobody would have predicted for a memory chipmaker a few years earlier.

The 47-year-old commodity company is now one of the most fought-over names in the AI supply chain.

_{SECTION 2 · FUNDAMENTALS}

How does it make money and who’s running the show?

Micron organizes its business into four units: Cloud Memory (hyperscale cloud and all data-center HBM), Core Data Center (mid-tier cloud, enterprise, OEM, and storage), Mobile and Client, and Automotive and Embedded. Right now, the data center is the entire story. Data center revenue exceeded $25 billion in fiscal Q3 alone, an annualized run rate of over $100 billion. Data center SSD revenue exceeded $5 billion, more than doubling sequentially.

Read more

The three ways an AI cluster talks to itself and where AOI actually lives

AOI sits inside Layer 7 of the AI buildout, arguably the most argued-about layer in the entire stack. Every GPU cluster on Earth depends on it. Wall Street has built competing trillion-dollar TAM models around it. It’s where the sharpest disagreements in AI infrastructure investing are playing out in real time. AOI is in the middle of that argument, making two components nobody else makes enough of.

Layer 7 isn’t one problem. It’s three different distance problems, each with its own physics and its own winners.

Scale-Up — chip-to-chip, inside the rack. NVLink territory, co-packaged optics. AOI isn’t here.

Scale-Out — rack-to-rack, inside the data center. This is AOI’s turf. The pluggable transceivers that turn electricity into light so GPUs can talk to each other, plus the laser chips that power them. AOI grows those in-house while most of the competition is stuck waiting in line for supply.

Scale-Across — data center to data center, long-haul and submarine. Different game. Not AOI’s lane.

AOI doesn’t just play in interconnect. It owns the one slice of it that’s the actual bottleneck on how fast hyperscalers can get GPUs talking to each other..

Layer 7 is the highest risk/reward layer in the entire AI stack. Technically demanding. Supply-constrained. Three different problems wearing one name. AOI sits right in the middle of Scale-Out, making the two components nobody else makes enough of.

Read more

Here's what almost everyone gets wrong about AI infrastructure

When people talk about the AI buildout, they talk about chips. NVIDIA. GPUs. That’s Layer 5. It gets all the headlines.

But look one layer down. Layer 6 is Memory -- and every GPU in existence is useless without it. The datasets AI trains on, the model weights it runs on, the caches that let it answer your question in milliseconds -- all of it lives in NAND flash storage. Not RAM. Not the cloud. Physical chips, in physical servers, that hold data even when the power is off.

There is no AI without storage. And there is no storage at this scale without NAND. SanDisk is one of two companies in the world that can supply it at hyperscaler volumes. The other is their manufacturing partner, Kioxia.

“This quarter marks a fundamental inflection point for Sandisk — where our technology leadership is enabling a deliberate shift in our mix toward the highest-value end markets, led by Datacenter.” — David Goeckeler, CEO, April 30, 2026

Before we go further -- the AI memory stack has a lot of acronyms and they’re often used interchangeably in headlines even though they’re completely different technologies. Knowing the difference matters because it’s the difference between understanding why SanDisk is uniquely positioned and just taking someone’s word for it..

AI & COMPUTE INFRASTRUCTURE: A MEMORY & STORAGE PRIMER

Read more

SNDK for $150. Everyone was still staring at NVDA.

I kept buying. Kept reading. Kept going deeper into why memory was the real bottleneck in AI -- not the chips, not the models, the memory. The thing every GPU needs and nobody was building fast enough.

That position is up over 1,000% today.

That’s one of many. The full portfolio is up +500% YTD.

I'm not telling you that to flex. I'm telling you that because the framework that got me there is the whole reason this Substack exists.

Who I am

By day, I’m the head of AI at a major enterprise. I manage a multi-million dollar budget and deploy it across the AI stack -- infrastructure, tooling, compute, all of it.

That means I’m not analyzing this buildout from the outside.

I’m inside it. I am the customer.

The companies on my watchlist are often the same companies my employer is buying from. When I tell you memory was the bottleneck in 2025, it’s because I was the one waiting on it. When I tell you photonics is the next chokepoint, it’s because the people selling it to me have already told me where the next constraint is forming. It’s building to scale.

That is an unfair advantage, I must admit, but this gives me the confidence and the edge most retail investors are looking for.

The other half of my edge is older. 10 years as a PM at tech and fintech. Started as a software engineer. Running AI at 30. The career path wasn’t an accident -- it was always about being in the right place at the right time. That’s the only investing philosophy that has ever made me serious money.

Being first is overrated. Being at the right place at the right time pays ten times more.

The thing that’s missing from most AI coverage

The AI buildout is a stack. It goes all the way from the application you use on your phone down to the gallium and indium being mined out of the ground. Every single layer has companies in it. Most of them are unknown. Several of them are the actual chokepoint -- the thing the whole pyramid above them depends on and can’t replace.

SIVE. TRT. XFAB. SHT. PENG. BRUN. don't show up on mainstream forums. They show up in the supply chain notes of the companies that do.

You can’t trade what you can’t see.

Read more