Welcome back.

Today, I want to talk about the sharp rebound in the Nasdaq—and why, even after a strong day, there may be no reason to feel FOMO.

Yesterday, I pointed out an interesting similarity in the QQQ chart.

After breaking out in June 2025, QQQ advanced about 17% before peaking in November. It then experienced a drawdown of roughly 13%.

The current sequence looks remarkably similar, although it has played out much faster.

Following its April 2026 breakout, QQQ also gained about 17%, before declining nearly 12%.

Markets never repeat themselves perfectly. But when the structure and percentages line up this closely, the comparison becomes useful.

Now, after today’s sharp rebound, investors who were not positioned may feel that they have already missed the move.

I do not think that is necessarily the right conclusion.

If the recent low holds and this develops into another sustained advance, the market may still be near the beginning of the move—not the end.

Using the previous advance as a rough analogue, a 30% to 35% move from the recent low would place QQQ somewhere in the 875 to 900 range.

But let me be very clear:

That is not a prediction. It is a scenario.

The market still needs to confirm it.

I would want to see four things.

First, the recent low must continue to hold.

Second, QQQ needs to reclaim and sustain the important resistance levels above it.

Third, market breadth needs to improve. A healthy advance should involve more than just a handful of mega-cap technology stocks.

And fourth, leading stocks need to break out—and then hold those breakouts.

The important point is that there is no need to chase a single strong session.

If a durable uptrend is beginning, there should be time to build exposure gradually as the market confirms itself.

The goal is not to catch the exact bottom.

The goal is to participate in the larger move while keeping risk clearly defined.

Right now, the market is offering early evidence that the correction may be ending.

If that evidence strengthens, the larger opportunity may still lie ahead.

So: no FOMO, no blind prediction, and no need to chase.

Watch the evidence. Define the risk. Build exposure deliberately.

Good luck, and keep learning.

Disclaimer: This podcast is for informational and educational purposes only and does not constitute financial, investment, tax, or legal advice. All opinions, market scenarios, and price targets reflect personal views and may change without notice. Investing and trading involve substantial risk, including the possible loss of principal. You are solely responsible for your own research, risk management, and investment decisions.

Imagine running a factory where every worker suddenly becomes ten times faster.

That sounds wonderful.

But now imagine that many of those workers are autonomous machines. They can build things, change things, and send those changes into production without waiting for a human.

Productivity rises. So does the risk of something going wrong.

Someone still needs to control the factory floor, inspect the work, enforce the rules, and keep a record of every decision.

That is the simplest version of the bull case for GitLab.

AI may change who writes the code. GitLab wants to remain the place where that code is planned, tested, secured, approved, and deployed.

What GitLab actually does

Building software involves much more than typing code.

Teams must store that code, test it, scan it for security problems, track changes, coordinate projects, and eventually release the finished product.

Historically, companies often bought a different tool for each job.

Think of it as a workshop assembled from several manufacturers. One company supplies the power tools. Another provides the security system. A third keeps the project schedule. Then someone has to make everything work together.

GitLab’s pitch is simpler: put the entire workshop under one roof.

One platform. One audit trail. One place to see what happened to the code from the moment someone planned a change until that change reached a customer.

For a startup, that can make development easier.

For a bank, pharmaceutical company, or other regulated enterprise, it can be essential. Those organizations need to know who changed the code, whether it passed security checks, and who approved its release.

GitLab can run in its own cloud or on a customer’s infrastructure. That flexibility helps it serve both modern software companies and large organizations with strict security requirements.

The direct AI opportunity

GitLab’s AI product is called the Duo Agent Platform.

These agents can help write, review, and test code inside the same platform customers already use.

Duo became generally available only two weeks before the quarter ended. Yet it immediately generated more new recurring revenue than GitLab’s two older AI products had produced together in any previous quarter.

Paid usage was running at an annualized rate of nearly $20 million by quarter-end.

That sounds exciting, but it comes with an important warning.

GitLab’s chief financial officer explicitly told analysts not to build that figure into their models yet. It represents only one quarter of data, and some of the demand could reflect excitement surrounding the launch.

That caution is healthy. The number is evidence of early demand, not proof of a durable revenue stream.

There are encouraging customer examples.

A top-10 American bank tested Duo and reported saving roughly 1.5 hours per coding task. The bank expects the number of active users to grow approximately twentyfold if the broader deployment proceeds.

GitLab is also selling Duo through the Amazon Web Services, Google Cloud, and Anthropic marketplaces.

That may sound like administrative detail, but enterprise purchasing can resemble airport security. Every additional checkpoint slows things down.

Selling through marketplaces customers already use removes checkpoints and makes the product easier to approve and purchase.

Most importantly, management assumes Duo will make no material contribution to its full-year revenue guidance.

The existing platform must therefore carry the forecast. If Duo adoption continues, that revenue could arrive on top of what management currently expects.

That is the AI optionality in the story.

The larger AI platform bet

GitLab is making a larger bet than simply selling its own coding agent.

It also wants to benefit when customers choose someone else’s.

Its proposed context service is called GitLab Orbit.

Think of an AI coding agent as a very talented new employee.

On its first day, that employee may understand programming, but it does not understand your company. It does not know why past decisions were made, how your systems connect, which policies must be followed, or what previous developers already tried.

That organizational memory is the context.

GitLab wants Orbit to supply it.

Outside tools such as Claude Code, Cursor, and Codex could plug into that context and pay GitLab based on usage.

If this works, GitLab would not need to win every competition for the best coding agent. It could become the tollbooth that different agents pass through to understand a customer’s software environment.

GitLab is also working with an unnamed AI lab on a major rebuild of Git, the underlying technology developers use to track changes to code. The goal is to support 100 times the current scale.

The identity of that AI lab has not been disclosed, but the partnership itself is meaningful.

An AI company has chosen to build foundational infrastructure with GitLab instead of simply routing around it.

The larger vision is one platform supporting three ways of developing software:

* Humans writing code manually.

* Humans working alongside agents.

* Autonomous agents performing more of the work themselves.

Across all three, companies still need identity, security, permissions, and an audit trail.

In fact, those controls may become more important as machines gain more freedom to alter production software.

More autonomous workers can mean more productivity. They can also mean more doors that need locks.

The financial foundation

The financial results suggest GitLab does not need to wait for this AI vision to support the business.

Quarterly revenue reached roughly $264 million, growing 23% from the previous year. That was four percentage points ahead of guidance.

Management, however, is guiding to only 16% to 17% growth for the full year.

That gap matters.

The optimistic interpretation is that management has set a cautious bar while the underlying business is performing better.

The portion of signed business expected to become revenue within the next 12 months grew 24%.

New customer signings increased 30% and reached their highest absolute count in 10 quarters.

Existing customers are spending about 17% more than they were a year ago, while gross retention remains above 90%.

The large-enterprise business is particularly strong.

GitLab now has more than 1,500 customers spending at least $100,000 annually. That group grew 18% and represents more than three-quarters of total recurring revenue.

The largest customers are not merely experimenting with GitLab. They are building more of their software operations around it.

GitLab is also improving profitability.

Its adjusted operating margin reached 14%, approximately two percentage points better than a year ago.

Free cash flow was unusually strong at nearly $147 million. Faster customer collections helped that figure, so it should not be treated as a normal quarterly run rate.

The company also holds roughly $1.36 billion in cash and short-term investments. It repurchased approximately 2.4 million shares during the quarter and still has $350 million available under its buyback authorization.

That balance sheet gives GitLab room to invest while the market changes around it.

The technical setup

There is also a technical setup behind the fundamental story.

GitLab’s stock has spent roughly six months building a base. Within that larger pattern, buyers have repeatedly stepped in at progressively higher levels.

The area around $35 has become an important technical boundary.

A convincing move above that area would suggest the stock is leaving its base. Failure to hold the recent higher lows would weaken the setup.

The broader software sector has also been improving. Several software stocks have held up well even during weaker trading in the Nasdaq, suggesting that investors may be rotating back toward the sector.

None of this guarantees a breakout.

The chart simply provides a way to judge whether the market is beginning to agree with the fundamental thesis.

The counter-case

First, growth beneath the headline is uneven.

Bookings grew only 12%. Revenue from customers running GitLab on their own servers was flat, while smaller customers grew just 7%. The cloud and enterprise businesses are carrying more of the load.

Second, software seats remain under pressure.

Layoffs at GitLab’s customers are reducing seat counts, and price-sensitive customers represent about one-fifth of recurring revenue. GitLab’s shift toward usage-based pricing may help, but it has not yet been proven.

Third, the AI evidence is very early.

The Duo figures represent one launch quarter, and the company’s own chief financial officer has warned investors not to extrapolate them.

Fourth, execution risk is rising.

GitLab is cutting 14% of its workforce and exiting 22 countries while attempting an ambitious technical rebuild. A leaner organization could become faster, but it also has less room for mistakes.

And this is not a cheap stock in the traditional sense.

GitLab remains unprofitable under standard accounting rules and trades at more than 40 times forward adjusted earnings. The existing business must continue performing for the AI optionality to matter.

What to watch

There are six things worth monitoring from here.

First, revenue growth compared with management’s 16% to 17% guidance. If growth remains closer to the latest 23% result, the cautious-guidance argument becomes more credible.

Second, watch bookings, near-term contracted revenue, and new customer signings. Together, they tell us whether today’s demand can become tomorrow’s reported growth.

Third, watch the split between cloud and self-managed subscriptions, along with growth among smaller customers. The enterprise business is working. GitLab still needs a healthy pipeline beneath it.

Fourth, watch Duo—but do not extrapolate one launch quarter. The important evidence will be sustained usage, broader deployments, and recurring customer expansion.

Fifth, watch the restructuring. The bull case requires GitLab to improve efficiency without damaging sales or delaying its platform rebuild.

Finally, watch the $35 area and the relative strength of the broader software sector.

The chart should confirm the thesis, not replace it.

Bottom line

The GitLab story comes down to one question:

As AI agents write more software, does the platform surrounding the code become less important—or more important?

The bull case says more important.

When one human writes code, governance is useful.

When thousands of autonomous agents can write, test, and deploy code around the clock, governance becomes essential.

GitLab does not need to create every worker in the AI software factory. It needs to control the factory floor.

The existing business provides the foundation.

Duo provides direct AI revenue.

Orbit could allow GitLab to collect a toll from outside agents.

And its governance layer may become more valuable as software development becomes increasingly autonomous.

But this remains a growth-plus-optionality case, not a deep-value case.

The durable platform must keep compounding for the AI upside to matter.

This article is for educational and informational purposes only. It is not financial advice. Do your own research.

A GPU that can’t talk to the other ninety-nine thousand, nine hundred and ninety-nine GPUs is a space heater.

That’s the entire thesis in one sentence.

For the past three years, investors have argued about who makes the best AI chips.

NVIDIA.

AMD.

Google’s TPUs.

Amazon’s Trainium.

But powerful chips sitting alone are like brilliant musicians who can’t hear the rest of the orchestra.

The performance comes from coordination.

And that coordination depends on the network.

As AI clusters get larger, the connections between the chips are becoming more valuable, more complicated, and more essential.

This is the other AI trade.

The companies that build the roads, intersections, bridges, and express lanes carrying data through the AI factory.

They can get paid regardless of which model wins, and sometimes regardless of which accelerator wins.

Before we begin, this discussion is educational. It is not financial advice, and I’m not predicting stock prices. The goal is to understand the technology, the competitive landscape, and why it matters.

There is one framework I want you to remember.

Copper inside the rack.

Optics between racks.

Coherent optics between buildings.

Three boundaries.

Think of an AI data center as a city.

Copper handles the short streets within a neighborhood.

Optics runs the highways connecting neighborhoods.

And coherent optics operates the high-speed rail linking separate cities.

Almost every company in AI networking sits somewhere along those three routes.

And much of the industry’s competitive struggle comes down to where each boundary falls, how quickly it moves, and who collects the toll.

Copper inside the rack.

Optics between racks.

Coherent optics between buildings.

Keep that framework in mind, and the rest becomes much easier to understand.

The first thing most people get wrong is imagining an AI data center as one enormous network.

It is actually three separate networks, each designed for a different job.

The first is called scale-up.

Scale-up is the network inside a single rack, where GPUs communicate directly with other GPUs.

Imagine seventy-two chefs trying to prepare one enormous meal.

It is not enough for every chef to be individually talented.

They need to exchange ingredients, coordinate timing, and avoid getting in one another’s way.

If communication is slow, the whole kitchen slows down.

Scale-up networking is the communication system inside that kitchen.

Its goal is to connect a group of accelerators so closely that software can treat them as one enormous computing engine.

This is the highest-bandwidth and most tightly controlled layer in the entire data center.

NVIDIA’s technology here is called NVLink.

The current generation moves roughly one point eight terabytes of data per second, per GPU.

A flagship seventy-two-GPU rack can move around one hundred and thirty terabytes per second across the full system.

The next generation is expected to increase that substantially.

That one-point-eight-terabyte number matters because the connection inside the rack is roughly ten times faster than the network connecting one rack to another.

It is the difference between handing a document to the person sitting beside you and shipping it to another office across town.

And surprisingly, the connection inside the rack runs primarily on copper.

Actual copper wire.

We’ll come back to why.

The second network is called scale-out.

This is also known as the back-end network.

It connects one rack to another, turning individual systems into clusters containing ten thousand, one hundred thousand, or eventually even more accelerators.

If scale-up turns one rack into a single machine, scale-out turns an entire warehouse into a single computer.

This is the central battleground in AI networking.

It is where NVIDIA’s proprietary technology competes against the open merchant ecosystem.

The third network is the front end.

That handles storage, data ingest, system management, and ordinary enterprise traffic.

Think of it as the loading dock and administrative office.

It matters operationally, but it is not the most differentiated or strategically contested part of the AI network.

So our focus is scale-up inside the rack and scale-out between racks.

The central scale-out battle is InfiniBand versus Ethernet.

In one corner is InfiniBand.

InfiniBand is NVIDIA’s proprietary networking fabric.

It is purpose-built, lossless, extremely low-latency, and supplied by a single vendor.

NVIDIA became the only major commercial supplier after acquiring Mellanox in 2019.

Think of InfiniBand as a private railway.

One company owns the tracks, the trains, the signaling system, and the stations.

Because everything is designed together, the system can run with extraordinary precision.

InfiniBand won the first phase of the AI build-out for a straightforward reason.

For tightly coupled training workloads, it worked reliably and delivered exceptional performance.

In the other corner is Ethernet.

Ethernet is the public highway system.

Many companies can build the vehicles.

Many vendors can supply the roads and traffic-control equipment.

Customers are not locked into one operator.

But ordinary office Ethernet was not originally designed for tens of thousands of GPUs trying to communicate simultaneously.

That would be like putting Formula One cars onto suburban streets and wondering why traffic backs up.

So the industry began rebuilding Ethernet for AI.

The Ultra Ethernet Consortium brings together much of the non-NVIDIA ecosystem, including AMD, Broadcom, Arista, Cisco, Meta, Microsoft, and Oracle.

Its purpose is to make Ethernet behave more like a purpose-built AI fabric while preserving the benefits of an open, multi-vendor standard.

A newer approach called M R C was also introduced by a group including OpenAI, Microsoft, Broadcom, AMD, and, notably, NVIDIA itself.

Its goal is to create much larger and more efficient switch configurations, scale beyond one hundred and thirty thousand computing engines, and reduce the number of switches required by roughly sixty percent.

Imagine replacing a maze of connecting flights with one enormous airport hub.

Fewer stops.

Fewer handoffs.

Less equipment.

Lower cost.

Now here is the data point that captures the direction of the market.

In the first quarter of 2026, data-center Ethernet switch revenue grew sixty-one percent year over year, surpassing ten billion dollars.

And the number-one vendor in data-center Ethernet was NVIDIA.

Its Ethernet revenue reached roughly two point one billion dollars, nearly three times the prior-year level, placing it ahead of Arista and Cisco.

Think about what that means.

NVIDIA has the strongest economic interest in preserving its proprietary InfiniBand ecosystem.

Yet one of its fastest-growing networking businesses is Ethernet.

It is like the owner of the private railway becoming the biggest supplier of trucks for the public highway.

That does not mean the railway is disappearing.

But it tells you NVIDIA has no intention of watching the open market grow without participating.

The current AI back-end market is approximately two-thirds Ethernet and one-third InfiniBand.

But this is not a clean victory.

InfiniBand revenue also rebounded sharply during the same period.

NVIDIA is not abandoning its proprietary fabric.

It is playing both sides of the board.

This is a long competitive grind, not an overnight displacement.

Ethernet is gaining ground for three main reasons.

First, merchant silicon reduces dependence on a single vendor.

Hyperscalers do not want one company controlling the engine, the transmission, the roads, and the tollbooths.

Second, Ethernet network designs can be more efficient.

Some can reach full cluster scale in three switching tiers, while comparable InfiniBand architectures may require four.

Think of each tier as another connection at an airport.

Every additional connection requires more gates, more baggage transfers, more time, and more opportunities for delay.

Removing one tier can reduce the number of optical transceivers by roughly one-third.

And those transceivers cost real money.

Third, every hyperscaler wants negotiating leverage against NVIDIA.

Even customers that depend heavily on NVIDIA GPUs do not necessarily want NVIDIA controlling every surrounding layer.

But inside the rack, NVIDIA remains in a much stronger position.

This is the scale-up layer.

And so far, the open ecosystem has not cracked it.

NVLink is deployed, mature, and approximately twice as fast as the emerging alternatives.

NVIDIA has also made a strategically clever move called NVLink Fusion.

Instead of reserving NVLink only for NVIDIA-designed systems, the company will license portions of the interconnect so that third-party processors and custom chips can connect to NVIDIA’s fabric.

Imagine a country realizing it cannot stop neighboring countries from building their own cars.

So instead, it invites all those cars onto its roads and charges them to use the highway.

Rather than simply losing customers who develop custom silicon, NVIDIA is trying to pull those chips into its own networking ecosystem.

The open alternative is called U A Link.

The second version of the standard was published in April 2026.

It has broad industry support, including AMD, Broadcom, Google, Intel, Meta, Microsoft, Apple, and Amazon.

The architecture is designed to connect as many as one thousand and twenty-four accelerators within a pod.

But the current competitive position remains clear.

U A Link is a blueprint and early construction.

NVLink is a finished bridge already carrying traffic.

And NVIDIA keeps extending that bridge while competitors are still completing theirs.

The merchant ecosystem is making progress in the open scale-out layer while remaining behind in the proprietary scale-up layer.

Now we move between racks.

This is where copper runs out of road and optics takes over.

Inside a rack, the distances are short enough for copper.

But once data needs to travel from one rack to another, copper begins losing signal quality and consuming too much power.

Fiber solves that problem by carrying the data as light.

Think of copper as a delivery van.

It works extremely well for short trips around the neighborhood.

Optical fiber is the freight train.

It costs more to load, but once the distance and traffic increase, it can carry vastly more data much more efficiently.

This is why optics has become one of the highest-growth parts of AI infrastructure.

Every AI accelerator needs to communicate with accelerators in other racks.

That requires optical connections.

And the number of those connections rises as AI clusters get larger.

At the same time, the industry is upgrading from eight-hundred-gigabit connections to one-point-six-terabit connections.

In simple terms, the road is becoming twice as wide while the number of vehicles using it is also increasing.

That gives optical suppliers two growth drivers at once.

More connections.

And more expensive connections.

To understand the companies, it helps to think of an optical module as having three basic jobs.

First, someone has to create the light.

That is the laser.

Second, someone has to translate the electrical signal from the chip into a clean optical signal, and then translate it back at the other end.

Third, someone has to assemble all those pieces into the finished transceiver that plugs into the network switch.

The laser is the hardest part to manufacture.

Many companies can assemble a module.

Far fewer can reliably produce the advanced lasers required for the newest one-point-six-terabit connections.

That makes the laser one of the key bottlenecks in the optical supply chain.

Coherent and Lumentum are two of the important Western suppliers here.

Coherent has an advantage because it manufactures both lasers and finished optical products.

It is like a restaurant that owns the farm supplying its most important ingredient.

When the ingredient is scarce, controlling your own supply matters.

Lumentum offers more concentrated exposure to lasers and optical components.

That can give it greater upside when optical demand accelerates, but it also means the business is more exposed when demand slows.

Then comes the signal-processing chip inside many optical modules.

This chip is called an optical D S P.

Think of it as a translator combined with noise cancellation.

At very high speeds, the signal becomes distorted as it travels.

The D S P cleans it up so the receiving equipment can understand it.

Marvell is the leading merchant supplier of these chips.

Broadcom is the other major player.

Together, they control most of this part of the market.

NVIDIA’s strategic investment in Marvell shows how important this technology has become.

Even NVIDIA, with its enormous internal engineering resources, wants a close relationship with one of the leading optical-signal specialists.

Finally, someone has to assemble the finished module.

Fabrinet is one of the major contract manufacturers doing that work.

But assembly usually has lower margins and fewer technological barriers than producing the laser or the signal-processing chip.

Think of the optical module as a premium smartphone.

The assembler matters.

But the companies supplying the advanced processor and camera sensor often capture more of the economics.

There are also major Chinese transceiver suppliers, including Innolight and Eoptolink, that lead the market by shipment volume.

For an investor, the simple map is this.

Coherent and Lumentum help make the light.

Marvell and Broadcom help translate and clean the signal.

Fabrinet and other manufacturers assemble the finished product.

The most durable bottleneck appears to be the laser.

The D S P is another attractive layer because two companies dominate it.

Assembly benefits from booming demand, but it is generally the most competitive part of the chain.

There is one potential disruption worth understanding.

It is called linear-drive optics, or L P O.

L P O tries to remove the D S P from the optical module.

The attraction is lower power consumption and lower cost.

But removing the D S P is like removing noise-canceling technology from a phone call.

It works when the connection is short and the environment is controlled.

It becomes much harder as speed, distance, and signal noise increase.

That is why L P O may win some shorter-distance connections without replacing D S Ps everywhere.

The larger point is straightforward.

AI clusters need more optical connections.

Those connections are moving to faster and more expensive technology.

And within that supply chain, the greatest value is likely to sit with the companies controlling the hardest components to manufacture, not necessarily the companies assembling the final box.

Now let’s return to copper.

Why does the fastest network inside the data center still rely on copper rather than optics?

The answer is physics.

Every time signaling speed doubles, copper’s practical reach is cut roughly in half.

It is similar to trying to shout a message through a crowded room.

The faster you speak, the harder it becomes for someone far away to understand you.

At current speeds, passive copper can carry a signal for about one meter.

Active copper, which includes a small signal-conditioning chip inside the connector, can extend that range to roughly two and a half or three meters.

That one-to-three-meter range corresponds almost perfectly to the dimensions inside a rack.

Beyond those distances, optics becomes necessary.

The reach limitation is not a minor technical detail.

It defines the addressable market for an entire category of companies.

NVIDIA’s flagship seventy-two-GPU rack illustrates the point.

It contains approximately five thousand copper cables, totaling around two miles of copper, with no optical connections inside the rack.

That was a deliberate engineering decision.

Using one-point-six-terabit optics inside the rack would have increased power consumption, reduced reliability, and added tens of kilowatts to the system.

Copper remains inside the rack for the same reason people take an elevator inside a building but use a train to travel across a city.

The right transportation technology depends on the distance.

Credo is one of the cleanest public-market exposures to active electrical cables.

Its revenue increased more than two hundred percent during its most recent fiscal year.

Astera Labs sells retimer chips and fabric switches.

A retimer is like a relay runner receiving a tired signal halfway through the race, refreshing it, and sending it onward at full speed.

Astera has one of the richest valuations in the networking ecosystem, reflecting extremely high expectations.

Amphenol is the diversified incumbent.

It sells connectors, cables, and backplane systems across many markets, while its data-center business has been growing exceptionally quickly.

The trade-off is customer concentration.

Credo’s largest customer represented roughly two-thirds of annual revenue, while its top ten customers represented around ninety percent.

That pattern appears throughout the highest-growth parts of AI networking.

The more direct the AI exposure, the more fragile the customer base can become.

It is like owning the only food truck outside one enormous factory.

Business is fantastic while the factory is running three shifts.

But if that customer changes its schedule, your revenue changes overnight.

A single delayed hyperscaler program can materially alter a supplier’s growth trajectory.

Now let’s look at two technologies on the horizon.

The first is co-packaged optics, usually shortened to C P O.

Today, optical transceivers generally plug into the front of a network switch.

Co-packaged optics moves the optical engine much closer to the switch chip itself.

Think of today’s design as placing an airport several miles outside the city.

Passengers must first travel across town before they can board the plane.

Co-packaged optics moves the airport next door.

The electrical signal travels a much shorter distance before becoming light.

That can reduce power consumption and increase bandwidth density.

Both NVIDIA and Broadcom have announced co-packaged optical platforms.

This has led some investors to assume that pluggable transceivers will soon be displaced.

The timing matters.

The most defensible current view is that co-packaged optics becomes a meaningful volume technology in 2027 and beyond.

The first deployments may occur earlier, but pluggable modules are expected to remain above ninety percent of switch ports in the near term.

Co-packaged optics is likely to complement the pluggable market before it begins materially displacing it.

It is similar to electric vehicles.

The long-term direction may be clear, but the existing installed base does not disappear the moment the new technology arrives.

That makes C P O an important future risk for transceiver companies, but not necessarily an immediate threat to the current upgrade cycle.

A broader technology land grab is already underway.

Marvell is acquiring Celestial AI for approximately three and a quarter billion dollars, aiming to bring optical connectivity into the scale-up domain.

Private companies including Ayar Labs and Lightmatter are also developing optical input-output technologies.

That may become the next major architectural contest.

The second forward-looking opportunity is data-center interconnect.

Individual data-center sites are beginning to reach physical power limits.

In many locations, operators cannot obtain enough additional electricity to continue expanding one building or one campus.

Imagine an airport that has run out of land for new runways.

Instead of continuing to expand one location, the operator links several nearby airports and manages them as one system.

AI operators are beginning to do something similar.

They connect multiple data centers across a metropolitan area and operate them as one logical training fabric.

The industry often calls this scale-across.

This is the third boundary from our opening framework.

Coherent optics between buildings.

High-end router demand is becoming increasingly influenced by data-center connectivity rather than traditional telecommunications.

That represents a major structural shift for the optical-transport industry.

Ciena is one of the clearest public-market exposures to this trend.

Its revenue recently grew around forty percent year over year, while demand for coherent optical products has reportedly exceeded available supply.

So how should investors organize the competitive map?

The cleanest combination of AI exposure and durability may be in switch silicon and network systems.

Broadcom represents the silicon side.

Arista represents the branded-systems side.

Think of them as selling the traffic lights and highway interchanges.

They benefit whether the vehicles are powered by one kind of engine or another.

They can participate across multiple optical architectures.

They do not need to predict the exact timing of co-packaged optics, the winning transceiver design, or which laser supplier gains share.

They sell the switching layer at the center of the network.

The optical laser layer may offer the largest absolute-dollar growth combined with a legitimate manufacturing bottleneck.

Coherent and Lumentum are two of the clearest examples.

The module itself may become more standardized.

The laser remains harder to produce.

The highest-beta portion of the market is copper cables and retimers.

Credo and Astera Labs have some of the fastest growth rates, but they also carry extreme customer concentration and demanding valuations.

They are speedboats.

They move quickly when the water is calm and demand is strong.

But they are less stable when conditions change.

The value-oriented corner includes companies such as Cisco, where AI networking is an additional source of growth rather than the entire investment thesis.

Cisco is more like a cargo ship.

It may not accelerate as quickly, but one AI program is less likely to determine the entire voyage.

Data-center interconnect suppliers may also offer a later-cycle and more diversified way to participate.

The area requiring the most caution is thin-margin contract manufacturing that has re-rated primarily because of an AI narrative without a comparable improvement in its underlying margin structure.

A company assembling the shovel does not necessarily earn the same economics as the company that owns the scarce metal used to make it.

Highly valued pure plays can also become vulnerable when market prices move beyond even the optimistic assumptions of the analysts covering them.

Now let’s state the bear cases clearly.

The first is that NVIDIA’s full-stack strategy may continue winning.

NVIDIA offers the GPU, NVLink, InfiniBand, Ethernet switches, network-interface cards, and software.

It owns the engine, transmission, road network, and navigation system.

That is the most integrated and highest-attach networking stack in the market.

The merchant ecosystem offers openness and multi-vendor flexibility.

Which approach ultimately controls the AI back end remains unsettled.

The second risk is customer concentration.

Credo, Arista, Fabrinet, Astera Labs, Celestica, and several other suppliers depend heavily on a small number of hyperscalers.

The greatest advantage of a pure play is also its greatest weakness.

When one customer is the wind filling your sails, a change in direction can become a problem very quickly.

The third risk is the white-box model.

Hyperscalers can combine Broadcom switch silicon with open-source networking software and generic hardware.

That puts pressure on branded network-system vendors, especially among the largest and most sophisticated customers.

It is the networking equivalent of buying high-quality ingredients and cooking the meal yourself instead of paying a restaurant.

Arista’s primary defense is the quality and consistency of its software platform.

The fourth and most important risk is capex digestion.

Valuations across this group assume that hyperscaler capital spending continues rising.

It does not require a collapse to create problems.

A pause can be enough.

Think of an escalator that stops moving.

Nobody has fallen through the floor.

The building is still standing.

But everyone who assumed they would keep moving upward now has to walk.

When growth expectations are extreme, merely flattening capital expenditures can produce falling earnings estimates and multiple compression at the same time.

So let’s bring the map back to its simplest form.

Copper inside the rack.

Optics between racks.

Coherent optics between buildings.

Networking is the other AI trade.

It is the infrastructure layer that gets paid regardless of which model wins, and in many cases, regardless of which accelerator wins.

The cleanest and most durable exposure may be in switch silicon and network systems because those companies can benefit across multiple architectural paths.

The highest operating leverage sits in optics, particularly around the laser bottleneck, where the transition from eight hundred gigabits to one point six terabits combines higher prices with rising unit demand.

The highest-beta companies also carry the richest valuations and the greatest customer concentration.

Maximum torque and maximum fragility often arrive together.

From here, there are three developments worth watching.

First, the share shift between Ethernet and InfiniBand in the scale-out back end.

Second, the timing of co-packaged optics, which currently appears to be a 2027-and-beyond volume event, leaving more runway for the pluggable upgrade cycle.

And third, hyperscaler capital-spending guidance.

Because every valuation across this ecosystem depends, directly or indirectly, on that line continuing to rise.

The memorable takeaway is simple.

The AI factory is not defined only by how many chips it contains.

It is defined by how effectively those chips can communicate.

A room full of geniuses who cannot exchange ideas is not a team.

It is just a crowded room.

And a GPU that cannot talk to the rest of the cluster is not an AI accelerator.

It is a space heater.

Thanks for listening to Inder’s Desk.

If you found this useful, subscribe on Substack for more deep dives into the businesses, technologies, and market forces shaping tomorrow’s winners.

Until next time, keep learning, keep questioning, and keep investing with conviction.

This episode is for educational and informational purposes only. It is not financial advice. Do your own research.

Almost nobody understands memory.

But here’s the thing.

A GPU without memory is like a Formula One car without fuel.

It doesn’t matter how fast the engine is...

If it can’t access the data it needs.

And that’s why, quietly, memory has become one of the biggest investment stories in artificial intelligence.

Today I want to give you a mental model.

Not a specification sheet.

Not a semiconductor textbook.

A framework.

Because once you understand the memory hierarchy...

You’ll never look at AI hardware the same way again.

Let’s start with a simple question.

What exactly is memory?

Most people think memory is one thing.

It isn’t.

There are really four different families.

DRAM.

HBM.

NAND.

And NOR.

Each solves a completely different problem.

Think of your own desk.

The papers spread out directly in front of you...

That’s DRAM.

It’s your working memory.

Fast.

Easy to access.

But the moment you leave the desk...

Everything disappears.

That’s exactly how DRAM behaves.

It’s incredibly fast.

But it’s volatile.

Turn the power off...

Everything is gone.

Now imagine you replace that desk with an entire wall of filing cabinets.

The information stays there forever.

It’s slower to reach...

But nothing gets lost.

That’s NAND.

It’s storage.

Your SSD.

Your laptop.

Your cloud storage.

Not particularly fast...

But it remembers everything.

Now here’s where AI changes the game.

Traditional DRAM was good enough for CPUs.

It isn’t good enough for modern AI.

Today’s GPUs consume data so quickly...

That ordinary memory simply can’t feed them fast enough.

And that’s where HBM enters the picture.

High Bandwidth Memory.

If normal DRAM is like driving across a city using roads...

HBM is like living in a high-rise apartment with an express elevator.

Instead of traveling long distances...

The memory is stacked vertically.

Right next to the processor.

The data barely travels at all.

It simply moves up and down.

The result is extraordinary bandwidth.

That’s why every modern AI accelerator...

Whether it’s an NVIDIA Blackwell...

An AMD MI350...

Or Google’s TPU...

Uses HBM.

Without it...

Those chips simply couldn’t operate at today’s scale.

Now here’s the part investors often miss.

HBM isn’t replacing DRAM.

It’s consuming it.

HBM is actually built from DRAM dies...

Stacked on top of each other.

Which means every wafer diverted into HBM production...

Is one less wafer producing traditional DRAM.

That matters.

Because supply tightens.

Commodity DRAM prices rise.

And suddenly...

Even businesses that aren’t selling HBM begin benefiting from the AI boom.

That’s one of the most important second-order effects in the entire memory industry.

Now let’s talk about the companies.

At first glance...

The market looks competitive.

In reality...

It’s remarkably concentrated.

Three companies dominate.

Samsung.

SK Hynix.

And Micron.

Together...

They control roughly ninety percent of global DRAM production.

But they’re no longer equal.

Samsung remains the largest overall memory company.

Micron is the only major American pure play.

And SK Hynix has quietly become the leader where it matters most.

HBM.

Today...

If you’re buying an NVIDIA AI accelerator...

There’s a very good chance the HBM inside came from SK Hynix.

That’s one of the biggest shifts in the semiconductor industry over the past few years.

Leadership changed...

Without many investors noticing.

Storage tells a different story.

NAND remains a much more competitive market.

Samsung.

Kioxia.

SanDisk.

Micron.

SK Group.

And increasingly...

Chinese manufacturers.

Competition keeps margins lower.

Which is why DRAM has historically produced much better profitability than NAND.

One is an oligopoly.

The other...

Is much closer to a knife fight.

Now let’s talk about the investment opportunity.

Memory has always been one of the most cyclical industries in technology.

Prices rise.

Everyone expands production.

Supply catches demand.

Prices collapse.

Then the cycle begins again.

It’s happened for decades.

AI has changed some of those dynamics.

But it hasn’t eliminated the cycle.

In fact...

It may simply have created a more profitable version of the same cycle.

Here’s where many investors get trapped.

Memory stocks often look cheapest...

Exactly when they’re most dangerous.

Why?

Because earnings are temporarily inflated.

Peak pricing creates peak profits.

Peak profits create low P/E ratios.

The stock looks cheap...

Right before the cycle turns.

That’s why experienced memory investors spend as much time looking at book value...

Capacity additions...

And supply growth...

As they do earnings.

This is one of the few industries where the lowest multiple can actually represent the highest risk.

So where does that leave us today?

Demand remains exceptionally strong.

AI infrastructure continues expanding.

HBM remains sold out.

Server DRAM markets remain tight.

And NAND pricing has improved dramatically.

Those are all positives.

But investors should always remember...

Memory rewards patience.

Not excitement.

Because eventually...

Every cycle turns.

The challenge isn’t predicting whether that day comes.

It’s recognizing where we are before everyone else does.

If you remember one thing from today’s discussion...

Make it this.

AI isn’t just creating demand for faster chips.

It’s creating demand for faster access to data.

And in computing...

Moving data has become just as important as processing it.

That’s why memory has moved from being a commodity...

To becoming one of the most strategic pieces of the AI stack.

And understanding that...

May give you an edge long before the next earnings report does.

This brings up a tough situation for investors: What happens to the broader AI stock ecosystem if this capex engine simply stops accelerating, flattens, or begins to pull back?

Thanks for reading Inder's Desk! Subscribe for free to receive new posts and support my work.

To understand the risks, it helps to look at how a slowdown trickles down, who gets hit hardest, and where the counter-arguments lie.

The Ripple Effect: How a Capex Plateau Hurts

Stay with me, if hyperscaler spending stalls, the damage isn’t just a simple line-item reduction. It hits the market through three distinct mechanisms:

* Direct Hit to Supplier Revenue: Companies like NVIDIA and Broadcom aren’t selling routine replacement gear—they sell infrastructure for new builds. If capex flattens, top-line growth for direct AI suppliers doesn’t just slow down; it flattens, or as I like to say, runs into a brick wall.

As seen in the chip stock index performance below, valuations skyrocketed alongside the capex expansion, making high-multiple stocks particularly vulnerable if growth slows:

* The Double Whammy (Slower Growth + Multiple Compression): Highly valued names like Astera Labs trade at eye-watering multiples (around 97x forward earnings) because the market expects endless acceleration. When growth slows, you get hit twice: analyst earnings estimates fall, and the multiple investors are willing to pay shrinks at the same time.

* Debt Doesn’t Shrink When Growth Does: Tech companies are taking on fixed debt to build out capacity today. If revenue growth fails to materialize at the expected pace, those fixed interest obligations remain, turning a simple growth slowdown into a real balance-sheet predicament.

Breaking Down the Ecosystem: Who Is Most Exposed?

As they say, not all tech companies are created equal. The risk varies wildly depending on where a company sits in the food chain:

Tier 1: The Hyperscalers (GOOGL, MSFT, AMZN, META)

* Risk Level: Low.

* They have massive, highly profitable core cash cows (Search, Office, AWS, Ads) to cushion the blow. The real risk here is not company returns, it can be chalked up to capital dilution and drag on overall returns, rather than risking the company’s survival.

Tier 2: Chips & Networking Suppliers (NVDA, AVGO, MRVL, Memory)

The VanEck Semiconductor ETF (SMH) serves as a proxy for hardware and chip suppliers. As shown below, valuations have traded near 52-week highs, leaving suppliers heavily exposed if hyperscaler capex flattens:

* Risk Level increased to moderate.

* While carrying fairly strong balance sheets, their stock prices also reflect huge growth expectations. Keep in mind they face significant valuation multiple compression, despite their underlying business remaining stable.

Tier 3: The Neo-Clouds (CoreWeave, Nebius)

* Risk Level: High, fragile.

* These pure-play GPU clouds are essentially giant levered bets on endless capex. Without non-AI fallback businesses, a drop in incremental demand makes their heavy debt loads dangerously fast.

Tier 4: Private Credit Lenders (Blue Owl, PIMCO, BlackRock)

* Risk Level: Systemic / Contagion.

* Private credit has underwritten roughly $800 billion in data center debt—much of it off-balance-sheet. If projects stall, credit contagion becomes a real threat. This ripple effect is perhaps the most unfavorable scenario of the bunch.

Where Balance Sheets are Getting Stretched

When we take a glance at recent company guidance, we can observe just how aggressive the spending race has become.

* Alphabet: Raised its 2026 capex target to $195-$205B, tapping both equity ($49.6B) and debt ($20.3B) in Q2 alone. `

* Microsoft: Guiding to ~19-B in FY26 capex (+61% YoY).

* Amazon: Leading the pack by setting a ceiling of a flat 200 billion dollar single-year capex guide for 2026.

* Meta: Pushing capex to $125B-$145B. Analysts currently project Free Cash Flow could actually dip all the way into the negative territory within this year. Note that Meta is also utilizing off-balance-sheet Special Purpose Vehicles (like Hyperion) carrying high debt-to-equity-ratios.

* CoreWeave & Nebius: CoreWeave’s debt leaped 3.5x in a single year to just over 17 billion dollars. Reminder that it carries ~$1.2B in annual interest), while Nebious doubled its non-current debt in one quarter to $8.4B while relying heavily on anchor clients such as the likes of Meta and Microsoft.

The Counter-Case: Why the Bulls Aren’t Panicking Yet

While the bear case is structurally sound and stable, several real-world factors imply the AI trade isn’t about to just collapse overnight:

* Improving Monetization: The industry is currently generating about $1.19 in AI revenue per each dollar of infrastructure depreciated. This is an increase from the sub one dollar standing from last least. Monetization seems to be pulling ahead of the cost curve.

* Deceleration is Already Price In: Most comprehensive models aren’t really predicting to have infinite growth over 70% forever. Wall Street itself expects capex growth to cool down to ~13% in 2027 and ~5% in 2028.

* Power Constraints (Not to be Confused With a Lack of Demand): Backlogs remain massive. For example, Microsoft’s $80B Azure backlog is largely constrained by physical power availability for data centers, it isn’t just the enterprise losing its appetite.

* Tripwire Haven’t Fired, Knock On Wood!: Key indicators of a legitimate crash–an unexpected 20% or higher cut in capex, or enterprise AI adoption stalling to under 15%. Simply put, neither of these have happened (yet).

The Bottom Line

The market has shown us just how sensitive and reactive it is to capex jitters–whether it’s Alphabet dropping short of 7% after raising its spending guidance or Nevius taking a 13% hit on competitive fears.

Hyperscalers might have the cash flow to survive through a miscalculation, but high-multiple chipmakers and heavily indebted nep-clouds don’t have that luxury. Moving forward, they key metric to watch won’t just be how much these giants spend, but whether their revenue per dollar of depreciation continues to rise alongside it.

Thanks for reading Inder's Desk! Subscribe for free to receive new posts and support my work.

Live Snapshot (as of 2026-07-22/23)

* Micron (MU): $959, market cap ~$1.08T, trailing P/E ~22, P/B 10.75

* SK Hynix (000660.KS / SKHY Nasdaq ADR): ₩1.78M / ~$165 (ADR), market cap ~$854 to 903B, trailing P/E ~17, P/B ~8.0

* Samsung (005930.KS / SSNLF): ₩268.5K, market cap ~$1.12T, trailing P/E ~21, P/B n/a

* Kioxia (285A.T): ¥64.7K, market cap ~$205 to 216B (down 52% from its late-June peak, extremely volatile), trailing P/E ~64, P/B n/a

* SanDisk (SNDK): $1,599, market cap ~$237B, trailing P/E ~54, P/B n/a

* Context, not memory names: Nvidia (NVDA) ~$5.1T market cap, Broadcom (AVGO) ~$1.9T market cap

1. The Map: Memory Types at a Glance

* DRAM: Volatile working memory (DDR5 servers/PCs, LPDDR5X mobile, GDDR7 graphics). AI relevance is indirect but big: HBM eats DRAM wafers, which tightens commodity DRAM and pushes prices up. This is the core of the Big-3 thesis.

* HBM: Stacked DRAM dies wired by through-silicon vias (TSVs), sitting next to the GPU. This is the AI story, full stop, every Nvidia H100/B200/GB200, AMD MI300X/MI350X, and Google TPU uses it. Highest-margin, fastest-growing memory line.

* NAND: Non-volatile 3D storage (SSDs). Real second-order AI demand exists here too: HDD lead times now run over 52 weeks, pushing cloud providers toward QLC SSDs (TrendForce, Sep 2025). Structurally lower-margin than DRAM (see section 3).

* NOR: Byte-addressable code storage. No meaningful AI relevance. A durable niche in auto/industrial: Macronix, Winbond, Infineon.

* Emerging (MRAM, ReRAM, FeRAM, CXL): Mostly optionality. Small and speculative (see section 6).

Kill list: PCM / 3D XPoint / Optane is dead (Intel and Micron shut it down 2021 to 2022). And note CXL is not a memory cell, it’s an interconnect protocol over PCIe for attaching extra DRAM. Standards are mature; hyperscaler deployment is still early.

HBM in two paragraphs (the part worth actually understanding)

A good analogy I like to use to better understand the difference between DRAM and HBM, is comparing things we see in our everyday lives. DRAM transfers information very similar to a city-road, multiple lanes traveling the distances to send information back and forth from the CPU. HBM, on the other hand, is like stacking a block of apartments on top of each other, and using high-speed elevators rather than city roads. This means the data will not travel far, but in the small space it can travel, it travels instantly, like an elevator would.

HBM stacks 8, 12, or 16 DRAM dies vertically (TSVs through the silicon), on a base logic die, on a silicon interposer beside the GPU (TSMC CoWoS). The trick is width, not clock speed: a 4-stack HBM setup gives a 4096-bit bus versus GDDR’s 512-bit. The JEDEC generations run HBM3 (Jan 2022, 819 GB/s), then HBM3E (up to 48GB 16-Hi, ~1.2 TB/s), then HBM4 (spec finalized Apr 2025: 2048-bit interface, doubled, 2.0 TB/s floor; vendors claim top bins of 2.8 to 3.3 TB/s, treat as a range). GDDR7 (JEDEC Mar 2024, 32 to 48 Gbps/pin) is the cheaper-per-GB alternative, and Nvidia’s inference-focused CPX GPU uses GDDR7, not HBM. HBM is not for everything.

2. Real-World Pricing Numbers

Standard DRAM, DDR5:

* A standard kit of two 16GB sticks (32GB Kit) of decently rated desktop RAM on the long run will generally cost anywhere from $180 all the way up to $250 USD. Prices of RAM often shift so expect RAM prices to be on the higher side due to the state of the memory market.

HBM (High Bandwidth Memory):

* Contract / Enterprise Pricing: HBM is sold exclusively in multi-layer stacks (usually 24GB or 36GB cubes) directly to GPU makers like NVIDIA and AMD.

* A 24GB HBM3e stack costs approximately $300 – $400 USD ($12.50 – $16.00+ per GB), making it more than an astonishing double to triple the cost per gigabyte of standard DRAM.

* When integrated into an AI chip, such as the NVIDIA Blackwell GPU which carries 193GB of HBM3e, the memory alone accounts for thousands of dollars of the chip’s total price tag.

NAND Flash (Storage):

* Consumer SSDs (PCIe 4.0): 1TB M.2 SSDs can run anywhere from $110 to $160 USD, just over 10 cents a GB.

* Enterprise SSDs, which tend to be higher in density, can run around 20 cents on the GB.

3. Manufacturers: Who Makes What

* Samsung (005930.KS / OTC: SSNLF): Makes DRAM, HBM, and NAND. #1 in DRAM and NAND, and the only fully integrated player (memory plus foundry plus logic).

* SK Hynix (000660.KS Seoul, plus SKHY Nasdaq ADR since Jul 10 2026): Makes DRAM, HBM, and NAND (via Solidigm). The HBM leader, and Nvidia’s #1 supplier.

* Micron (MU, US): Makes DRAM, HBM, and NAND. The only US pure-play, and a CHIPS Act beneficiary.

* Kioxia (285A.T, IPO’d Dec 2024): NAND only. Invented NAND (Toshiba, 1987). Bain holds 51%.

* SanDisk (SNDK, spun from WD, Feb 2025): NAND only. The cleanest US-listed NAND pure play, with a Flash Ventures JV with Kioxia.

* CXMT (private, China; pursuing a ~$4.2B Shanghai IPO): Makes DRAM (DDR4/DDR5 legacy). China’s commodity-DRAM entrant.

* YMTC (private, China, state-backed): NAND. Roughly 12% NAND share, sitting outside the official top-5 tables.

4. Market Share (date and source on every figure)

Flag: the percentages below are the last quarter TrendForce published to press. A newer 2Q26 table could not be sourced this pass. Treat the split as directionally stable, not this-second precise.

DRAM, 1Q26, TrendForce: - Samsung: 38.5% - SK Hynix: 28.8% - Micron: 22.4% - CXMT: roughly 6 to 10% of output (not revenue-equivalent) - Nanya: 1.9% - Winbond: 0.6%

Big 3 combined is roughly 90%. Notable: SK Hynix was briefly #1 by revenue in 1Q25 (first time in 33 years) and beat Samsung in FY2025 operating profit (first time ever) on its HBM-heavy mix.

HBM, mid-2026, mixed analyst sources (ranges; the volume split remains open): - SK Hynix: ~50 to 65% (durable leader) - Samsung: ~25 to 30% (recovered from a ~17% low in Q2’25 after its Nvidia-qualification lag) - Micron: ~15 to 22% (up from roughly 0 in 2024)

Nvidia qualified all three for Vera Rubin HBM4 (June 2026), so the Samsung-qualification saga is resolved. But qualification does not equal volume: who wins the HBM4 volume allocation is still an open question (could not be pinned to a hard split this pass).

NAND, Q4’25, TrendForce (sources noisy): - Samsung: ~28% - SK Group (Hynix plus Solidigm): ~22% - Kioxia: ~15.6% - Micron: ~14% (some sources say ~12%) - SanDisk: ~14% (some sources say ~13%) - YMTC: ~12% (outside the TrendForce top-5)

The structural point: NAND never consolidated like DRAM. Five-plus players, plus state-backed YMTC, means NAND margins run structurally below DRAM/HBM. DRAM is an oligopoly; NAND is a knife fight. (Chinese suppliers are also rising toward ~19% of global NAND bit output, the eventual rebalancing driver, see section 7.)

The two numbers that explain the whole HBM thesis

* HBM is ~30% and rising toward ~50% of DRAM revenue but only ~8 to 13% of bits (end-2025: ~18% of wafer input, ~8% of bits). Disproportionate dollars per wafer is the whole point.

* Sourced supply mechanism: wafer displacement, roughly a 3:1 HBM-to-DDR5 wafer-conversion ratio (Micron). The popular “TSV yield / CoWoS packaging is the real bottleneck” framing could not be confirmed here, treat it as an open question.

5. Strengths and Weaknesses

SK Hynix - Strength: HBM leader, first 12-layer HBM3E (Sep 2024), HBM4 development complete, and 1Q26 operating margin of 72%, above Micron’s 67.6% and TSMC’s 58% (verified) - Weakness: Smallest of the Big-3 by total DRAM revenue, heavy Nvidia/AI concentration, enormous capex commitments

Samsung - Strength: #1 in DRAM and NAND, deepest balance sheet, only vertically integrated player, HBM catch-up resolved (HBM4 mass production Feb 2026), and DS division operating profit ~$36B in 1Q26, up ~49x year over year (verified) - Weakness: Lost roughly 2 years of Nvidia HBM leadership, foundry division a persistent drag, and ~40% of NAND comes from Xi’an, China (geopolitical exposure)

Micron - Strength: Only US pure-play (CHIPS Act, the default non-Korea/non-China supplier), HBM sold out through 2026 with over $100B in signed contracts, hit a $1T market cap on May 26 2026 (verified), and Q3 FY26 (ended May 28) revenue of $41.5B / EPS $24.67, a record quarter, larger than all of FY25 ($37.4B), with revenue near-doubling three quarters running: $13.6B, then $23.9B, then $41.5B (verified) - Weakness: Smallest of the three, still building HBM credibility, highest capex intensity relative to its balance sheet

Kioxia - Strength: NAND heritage plus Flash Ventures JV scale (shared Japan fabs with SanDisk), BiCS10 at 332 layers - Weakness: 100% NAND-cycle exposure with no DRAM to smooth earnings, Bain’s 51% PE overhang, thin float, and whipsaw volatile (down 52% from its late-June peak) (verified)

SanDisk - Strength: Clean US NAND vehicle, and the JV survives the WD split - Weakness: Same all-NAND cyclicality as Kioxia

CXMT - Strength: Real competitiveness at commodity/legacy DDR4-DDR5, targeting an HBM back-end by end-2026 - Weakness: Not competitive at the leading edge/HBM yet, and on the US Entity List: interagency-approved in 2025 but still not published or delayed as of mid-2026, the key China swing variable (verified)

Full-year 2026 profit figures, labeled honestly: the big numbers circulating (Samsung ~$200 to 257B, SK Hynix ~$40 to 56B+) are analyst projections, not company guidance (verified). Anchored reality: SK Hynix’s reported Q1’26 operating profit was ₩37.6T (~$27B) at that 72% margin; Samsung’s DS division came in around $36B. SK Hynix reports Q2 on July 23, so these will move.

NAND layer counts, the most volatile spec in the space: Samsung is at 400+ in production with a 900-layer R&D demo in May 2026; SK Hynix is at 321; Kioxia and SanDisk are at 332. TLC means 3 bits/cell (mainstream); QLC means 4 bits/cell, roughly 20 to 30% cheaper per TB, and is the AI bulk-storage workhorse.

6. Biggest Customers

* HBM: Nvidia is dominant (~60%+ of demand), and SK Hynix’s relationship with Nvidia is the axis of the whole trade. AMD is second. Hyperscaler ASICs (Google TPU, AWS Trainium, Microsoft Maia) are real demand too, but note Broadcom and Marvell are ASIC design partners, not HBM buyers, they integrate the Big-3’s HBM on hyperscalers’ behalf.

* DRAM: Hyperscalers (servers), phone OEMs (Apple, Samsung mobile, Chinese OEMs), and PC makers.

* NAND/SSD: Hyperscalers (datacenter eSSD, the surging segment), enterprise, Apple, and client PCs/phones.

Concentration cuts both ways: multi-year contracts give revenue visibility, but the highest-margin product line hangs substantially on one customer, Nvidia.

7. How to Invest

* MU, Micron (pure-play memory): Cleanest US access to the whole thesis.

* SKHY / 000660.KS, SK Hynix (pure-play): Biggest HBM beneficiary, now directly buyable in the US via the SKHY Nasdaq ADR (since Jul 10 2026; 10 ADR = 1 share), no more Seoul-only friction.

* 005930.KS / SSNLF, Samsung (diversified): Memory exposure diluted by foundry, mobile, and display businesses.

* 285A.T, Kioxia (NAND pure play): NAND-cycle leverage, but thin float, a PE overhang, and high volatility.

* SNDK, SanDisk (NAND pure play): US-listed NAND-only vehicle.

* LRCX, Lam Research (equipment): TSV etch, direct HBM/3D-NAND leverage.

* AMAT, Applied Materials (equipment): Broad deposition/etch exposure.

* KLAC, KLA (equipment): Multi-die stack inspection/yield, more critical as stacks rise.

* 6857.T, Advantest (equipment): HBM test, dominant share, one of the most direct HBM picks-and-shovels plays.

* ASML (equipment): Weak HBM-specific link, HBM runs on mature DRAM nodes, not EUV. Don’t buy it for this thesis.

* ALAB, Astera Labs (adjacent): CXL/PCIe interconnect and memory pooling, an AI-infra beneficiary, not an “HBM controller.”

* MRVL / AVGO, Marvell / Broadcom (adjacent): ASIC integrators, so indirect HBM demand pull-through.

* MRAM, Everspin (emerging): The only discrete-MRAM pure play, small, an aerospace/auto niche.

* WBT.AX, Weebit Nano (emerging): ReRAM IP licensing (TI license, tape-outs), early and speculative.

* 688008.SS / 6809.HK, Montage (adjacent): Memory-interface chips, with China exposure.

* SOXX / SMH, ETFs (broad semis): Memory is a minority slice, and no pure-memory ETF exists.

Emerging-tech reality ranking: NOR/FeRAM (established niches) beats MRAM (real, small) beats ReRAM (early licensing) beats CXL (protocol, deployment early) beats PCM (dead).

8. The Cycle and Risk Framework (the judgment section)

Why memory is brutally cyclical: commodity product, plus oligopoly, plus lumpy multi-year fab additions, plus inelastic short-run demand, equals boom/bust. Recent history: the 2017 to 2018 boom (DRAM prices roughly tripled), then the 2018 to 2019 bust, then the 2022 to 2023 trough (Samsung’s chip profit down over 90%), then the 2024 to 2026 AI recovery.

Where we are (verified 2026-07): - Server DRAM contract prices up 13 to 18% quarter over quarter into Q3’26 (verified, TrendForce, 7/9/26), and TrendForce now guides DRAM tightness into 2027 too (RDIMM bit-supply growth only +15 to 20% year over year versus faster server-CPU growth). No 2027 DRAM glut signal. - NAND: roughly 4 to 5% undersupply in 2026, with relief only in 2H27 (verified, TrendForce, 7/21/26), not “2027” broadly. SLC NAND is up 120 to 170% in 2H26. - 2026 HBM/DRAM/NAND capacity is described as “sold out” by both Micron and SK Hynix. - The glut-risk vector is specifically Chinese suppliers (CXMT/YMTC) rising toward ~19% of global NAND bit output, and that’s what eventually rebalances NAND in 2H27.

The supercycle debate, live and unresolved (hold both sides):

Bull case: 3 to 5 year contracts dampen volatility, HBM demand is committed multi-year, and there’s talk of a “supercycle through 2028.” Long-term agreements (LTAs) make up roughly half of SK Hynix’s revenue.

Bear case: all three makers are expanding into the same demand signal, a classic synchronized-overinvestment glut setup, and “every memory cycle ends the same.” Also, LTAs dampen HBM’s cyclicality specifically, NOT the commodity DRAM/NAND majority of the business.

A margin reality check: HBM briefly lost its margin premium to DDR5 64GB RDIMMs in Q1 2026, because its annual contract pricing lagged a DDR5 spot spike (TrendForce). “HBM equals permanent fat margins” is a structural tailwind with real quarterly wobble, not a law.

The valuation trap (the most important paragraph in this document): memory stocks look cheapest on P/E exactly at the earnings peak, because cycle-high prices inflate the E. Concretely: Micron’s P/B has re-rated from its historical 1.3 to 2.5 range to 10.75, 4 to 8 times normal. Its trailing P/E of ~22 “looks cheap” only because annualizing the peak quarter ($24.67 EPS times 4, about $99) implies a ~9.7x run-rate, meaning it’s cheap only if you extrapolate record DRAM/HBM pricing forever. SK Hynix’s ~8x P/B on a 72% operating margin that exceeds TSMC’s is itself the peak-of-cycle tell. Normalize on P/B versus history and mid-cycle earnings, never on a single peak-year multiple.

Other risks, ranked: 1. Cycle timing, the valuation trap above 2. Customer concentration, Nvidia via HBM 3. China, CXMT/YMTC ramping at the commodity end (NAND toward ~19% of bits), with export-control policy (the CXMT Entity List still unresolved) as the swing factor 4. Capex intensity, record profits get recycled into record spending 5. FX and access, the KRW/JPY overlay on the Korea/Japan names (though SK Hynix’s SKHY ADR now removes that friction for the one name) 6. The share-confusion trap, Samsung leads total DRAM, SK Hynix leads HBM. Don’t conflate them.

Before You Act: Verification Checklist (status as of this pass)

* Live valuations pulled, done. See the Live Snapshot above; the “$41.5B Micron quarter” and “$205B Kioxia cap” both verified as real, not errors.

* Latest TrendForce quarterly share tables, NOT done. Could not re-source a newer quarter this pass; DRAM 1Q26 / NAND Q4’25 shares stay as the last-published, directionally stable figures.

* HBM4 volume allocation (Vera Rubin), NOT done, still open. Qualification is done (Jun 2026), but that does not equal volume share, no hard split available yet.

* CXMT Entity List, done. Still interagency-approved-but-unpublished/delayed as of mid-2026 (last dated source Jun 17 2026), remains the key China swing variable.

* SK Hynix US access, done, resolved. SKHY trades on Nasdaq since Jul 10 2026 (10 ADR = 1 share).

* Valuation on P/B vs history, done. MU P/B is 10.75 versus its 1.3 to 2.5 norm; peak-quarter-annualized run-rate P/E is ~9.7x, the cycle trap, quantified.

* 2027 supply, done. NAND relief pushed to 2H27 (2026 deficit ~4 to 5%); DRAM tightness now guided into 2027; watch Chinese NAND (~19% of bits) as the rebalancing vector.

Compiled 2026-07-22; dollar figures, listings, and status items verified against live sources 2026-07-22/23. Two items could not be re-verified this pass and stay explicitly flagged: the exact quarterly market-share percentages and the HBM4 volume split. Everything else is current as dated. Memory is violently cyclical, all prices/caps move daily, re-pull before acting.

Educational primer, not financial advice. For informational purposes only. Do your own research.