HBM built A Moat

SKHY is the market share leader in High Bandwidth Memory, which has been the story of 2025 and 2026, driving up profits at SK hynix and Micron. The customer base for Sk hynix has widened with HBM and has created a moat. With the HBM4 generation the base die moved to a TSMC logic process, where SKHY co-designed the base die with their end customers. Once a customer has engineered their chip around your base die, your memory orders are essentially locked in, as switching would mean redesigning everything. Sk hynix has co-design silicon with AWS, Google, Nvidia, Microsoft, Broadcom, and Meta. The valuable customers helped Sk hynix to surpass a $1 Billion valuation on the Korea exchange earlier this year. It however remained discounted to its memory peers due to its Korea primary listing. In an effort to expand its investor base and raise capital, the company issued a Nasdaq traded ADR. SHHY said the listing would expand its investor base, ultimately allowing its true corporate value to be properly evaluated. The listing comes at an interesting time as current memory products are insufficient for current long agentic workflows.

Two Memory Walls

According to Gholami et al., “AI and Memory Wall,” IEEE Micro, over the past 20 years, peak server hardware floating-point operations per second have been scaling at 3.0× per two years, outpacing the growth of DRAM and interconnect bandwidth, which have only scaled at 1.6 and 1.4 times every two years, respectively. The data supplied by the memory, the bandwidth, cannot keep up with GPU speed, and this is what is called the “memory wall #1.”

HBM was the fix for the first memory wall. It moved the memory on package, and then expanded it by stacking. However, HBM has reached its scale capacity, as it would face yield drops if stacks were to be increased. Current HBM running temperatures are near their limits. Nvidia’s latest Vera Rubin model which carries up to 288GB of HBM, is insufficient for some 1 terabyte size current models.

Now long context windows, and agentic multi-step reasoning all suffer from increasing Key-Value cache size. When you read KV cache think old chats (both your questions, and the AI response). Advanced models and their KV cache keep growing. HBM alone is increasingly insufficient to keep up with growing inference demand. More capacity is needed to hold new larger models, the KV cache, and allow agentic AI to run. All three of these utilize the HBM and together they have outgrown the capacity of HBM, which is “memory wall #2.” The solution is moving NAND up the stack and on package inside the AI chip. Enter two new flash tiers.

Shipping Now: SSD POD

In memory, it is the classic tradeoff triangle. Fast memory is expensive, and has lower capacity. High capacity memory is affordable, but it is slow. HBM read latency is about 15 to 30 nanoseconds while NAND flash takes between 20,000 to 100,000 nanoseconds. The AI inference market is a high transaction market with over 413 trillion tokens consumed daily. (I had to update the above number higher while writing, I wonder where it will be as you read this.) As token creation scales up, costs per token must decline along with them so the companies can remain profitable. Both speed and cost per token as well as function help to decide which memory type is used where. HBF and SSD POD have emerged as new layers in what is being called full stack memory. HBF is a next generation flash technology still under development. SSD POD is enterprise flash introduced by Nvidia earlier this year. It is basically a rack of SSDs dedicated to solve the KV cache bottleneck. SKHY is partners with Nvidia to supply the SSD. This new memory layer is already in production and can be seen in the Vera Rubin architecture, in the NVIDIA BlueField-4 STX The SSD POD Rack.

Trendforce reports that the enterprise SSD market for the top five companies, doubled, sequentially quarter over quarter. The SSD market saw both volume and higher contract prices. SK hynix is second by revenue in the SSD market, and in Q2 2026, they reported NAND ASP up mid-50% QoQ.

Coming soon - HBF

In 2025, SK hynix along with Sandisk formed a consortium to work on an open standard, high bandwidth flash, or HBF. Google and Tenstorrent joined the group as it releases its first technical specifications last month. HBF is exactly what you think it is, NAND stacked together, using the same packaging technology, on package, to create 8 to 16x the memory capacity of HBM.

HBF works alongside HBM, and won’t replace it. Flash has limited write cycles and is anywhere from 100 to 1,000x slower than HBM. Therefore HBM keeps the current working KV cache, while HBF could hold the static model weights, so its function is more read oriented. The model could be read from HBF without harming the write limited flash. The technology will offer a multiple of data storage for the same chip size. Bandwidth, the speed of data, is vastly improved as the stacked flash allows HBF to reach bandwidth levels that regular flash could not achieve. The specifications aims for a bandwidth of 1.6 TB/s for generation 1, slower than an HBM4 stack, but generation 2 & 3 are expected to equal and eventually surpass current HBM speed. Additionally, both capacity and bandwidth could grow, and a single HBF stack could offer terabyte scale capacity. Initial product samples are targeting 2027, with mass production coming in 2028.

Valuation

Before this summer, SKHY traded only on the Korean market and with less investor exposure the company traded at a lower valuation. The ADR has been here for two months, and it is trading at a premium to the Korean shares. As investors become aware of the stock, I believe its discount to Micron will disappear.

Memory valuation is inverted. When the p/e is low, investors believe the top is in. So with forward P/E at 5.4 and Micron at 6.26, investors believe the top is in, and earnings growth doesn’t continue. However, Q2 2026 saw DRAM ASP up 30% QoQ and NAND ASP up mid 50% QoQ. Those numbers come from a quarter with long term agreements. These agreements span multiple years. With supply sold out, and the supply tied to individual hyperscalers, it appears the financial strength continues.

Traditionally P/B is how memory companies were valued. It has ranged from 1x at bottoms and 5x at peaks, but now we have MU at 10.96 and SKHY at 7.68. Well above historical values. However, next year’s earnings are predicted at $34.03 a share, so $34.03 × 7.11Billion = $242 Billion in next year’s earnings. The current equity is $169 Billion , so next year’s earnings eclipse the entire equity on the balance sheet. Micron is similar. Current equity is $100 Billion, and has a forecast of $177 Billion in earnings.

After next year’s earnings then, SKHY P/B would be trading around 3x, MU trading around 4x Book value. Both within range and reasonable for the growth we are seeing.

Current values don’t really reflect the change HBM had on memory companies, creating very high margins, consistent demand that outpaces supply and the actual profitability. With the market share in HBM and NAND, SKHY should not trade at a discount to MU. When you consider their HBM market position and the customer relationships, I think SKHY should trade at a premium to MU. SKHY trades at $184 today, and if it were to trade at Micron’s forward P/E of 6.26 it should be priced at $213. If it traded at a 10% premium that price would be $234. Wall Street’s average target is $253.91, a 38% discount to today’s price.

Risks to the thesis

SK hynix has been slowly losing share to Samsung and MU. (Insert photo) It is likely the trend will continue and will likely end with SK hynix holding less than 50% share, but holding some form of a duopoly with Samsung. UBS believes by 2027, Samsung will lead with an estimated 41% of HBM while SKHY falls to 39%. The share is decreasing in an ever increasing market.

HBM market share by company Q2 2026 (Couterpoint)

Additionally, HBF has not received official adoption as the standard in future AI chips. Nvidia and AMD, the largest AI chipmakers, have not stated their future chips will use the HBF architecture. To address the growing capacity needs of modern AI many companies are actively building different solutions. The shape of the AI inference chip is still being developed and the final design to the memory capacity problem will have many solutions. However, we understand more memory is needed in all of its forms. With HBM, it was originally SK hynix that successfully developed the technology with AMD. SK hynix is behind the HBF and has teamed up with Google, Sandisk and Tenstorrent. The execution risk is tempered by its previous success.

Last, the stock is Korean based, and so the ADR carries Korean currency risk along with additional foreign government risk. There could also be foreign fees assessed by your brokerage.

Conclusion

OpenAI recently claimed to solve one of the Millennium Prize math problems, a group of seven of the hardest math problems that exist today. While the proof is still being verified the news release shows the promise of AI inference. In this case OpenAI is said to have used roughly 10,000 agents running together for about 88 hours. AI leaders additionally promote healthcare breakthroughs in the future. These advanced AI inference chips need large and fast memory to run continuously. The read, think, act loops all run through memory. As more memory layers are added to the next generation of chips, companies like SK hynix that sell both the DRAM and NAND will not reach peak anytime soon. With a conservative $234 target, anything below $200 is a buy.

Where each one stands and where they finish the year

CoreWeave has posted $1,212M, $1,365M, $1,572M, $2,078M and $2,575M across five quarters with no sequential decline. It’s full year guide is between $12.4 to $13.2 billion for 2026. If you back out the first half and the Q3 guide, the implied fourth quarter is $4.1 to $5.1 billion. This means QoQ growth is running at 24%, 37%, and 31%. Its exit run rate is $18.5 to $19.5 billion. The company added about 500 megawatts of active power in Q2, with 300 of those arriving in June.

Nebius grew 454% year over year. It guides to $3.0 to $3.4 billion for 2026 with an exit run rate of $7 to $9 billion. It booked $981.3 million in the first half of the year, so the second needs to deliver $2.0 to $2.4 billion. The company is now at the point where signed contracts are becoming invoices. So far it is working, and sales grew roughly fourfold in the quarter, four separate deals over $1 billion each, and the AI cloud business has a 50% adjusted EBITDA margin. Now the company has to continue to execute at this pace.

IREN is the only company with declining revenue, which is from this long bitcoin to AI compute transition. Revenue fell $7.6 million sequentially as bitcoin miners were swapped out for AI GPUs. AI cloud revenue doubled during the quarter to $70.5 million. Meanwhile, mining revenue fell to $66.7 million and $638.8 million of hardware was written off. Adjusted EBITDA has fallen for five straight quarters, from $121.9 million down to $19.2 million. The next big event is Horizons 2 - 4 landing in Q4 2026, with revenue arriving mostly in the March 2027 quarter.

Applied Digital reported revenue up 407% to $258.7 million in its May quarter, but $152.4 million of the $203.0 million HPC line was one-time tenant fit-out billed at construction margins. Recurring base rent was $44.1 million and full-year NOI was $90.4 million.

Compared to where they are today, these compute sales targets are all ambitious. CoreWeave is at roughly $10.3 billion on Q2 revenue annualized, so it needs to essentially double. Nebius is at $3.0 billion and needs to nearly triple. IREN is at $1.0 billion and needs to quadruple. Applied Digital publishes no run rate, only NOI, which is why it is not on the chart. How they can do it is by the rising price of AI compute.

Nebius currently bills between $9 million and $14 million dollars, with $12 million average. They are signing new deals at over $20 million dollars per megawatt. Q2 deals closed at $20 to $25 million, and the first short-term deal, signed in early Q3, priced above $40 million. Compute prices are rising rapidly, while also delivering new capacity quickly. Nebius raised prices on older-generation GPUs by more than 30% over Q1. Most of the step from $12 million to $20 million is longer term

IREN recent three-year contracts are priced above $20 million per megawatt with active discussions near $25 million, but the contracted book sits well below that. Microsoft is $9.7 billion over five years against 200 MW of critical IT load, or $9.7 million per megawatt per year, and the $1 billion of ARR operating today runs near $10 million.

CoreWeave does not disclose its new contract values to investors. However, they did disclose that they raised prices 25% across teh board in July. And they are compleylty sold out. So we can understand that gimilir to NBIS, their older GPUs are sold out and prices have gone up. Overall, we can calculate a backlog of $104 billion against 3.7 GW of contracted power works out to $28 million per megawatt of total contract value. Roughly in line with what NBIS signed in the quarter.

Applied Digital is a landlord, renting out space for a fixed rate of about 1.7 million dollars per megawatt. Because they use long-term leases, they collect exactly what they sign for. Their lower price reflects the fact that they provide the building, not the computers.

Layers = Millions = Value

The difference between earning $40 million per megawatt and $1.7 million is a function of how many layers a company owns. Applied Digital sells two layers while Nebius sells six. The value can be seen clearly to exist in these upper Neocloud layers. For that reason, I’m going to drop Applied Digital from our conversation. Too much debt, too little to gain from an all consuming AI compute future. So let’s look to value the three remaining companies.

The valuation is hard to do as their multiples change rapidly. Over the past twelve months, CoreWeave’s market capitalization traded between $39 billion and $87 billion, while IREN ranged between $12 billion and $24 billion. Nebius moved 30% in a day after reporting Q2 earnings. The price to earnings are nonexistent to give us value, and we could use a price to sales, but that is also moving quickly. So while researching enterprise value for the valuation, one thing stood out: total debt.

IREN & NBIS show balanced balanced sheets. Both of their short-term cash and receivables is very close to their total debt. They’re both around 7- 8 billion.

CRWV has $51.4 billion across debt and leases, against $7 billion of cash with total liabilities over $72 billion. If compute prices stall or change direction or new GPUs come out and lower the price per token, basically if anything happens this company just so over leveraged that I have to eliminate it here.

Down to IREN and NBIS

Shares

We talked about the massive share issuance of IREN during its lasts 12 months, growing over 70%. In addition to that behind the current 380 million shares sit another 146 million potentially dilutive shares. up from 26.3 million a year earlier. This compares to NBIS’s 17.5% selective dilution, after a large run up in the stock.

Capital expenditure commitments

Comparing each company’s 2026 capital program to its enterprise value highlights the scale of funding required:

While both have plans to spend around 25B, one company has $582 million in AI revenue last quarter and the other has $70 million. NBIS has revenue growing each of the last four quarters and IREN’s has fell for that many quarters, wait actually five quarters.

Overall the capex guide for NBIS is 10x its current revenue rate, the real rate for actual earnings, not a planned future. IREN is planning to spend 98x its current income and is pointing you to look at the run rate they will achieve by year end. Or will they, in looking at their 10k (end of August 2026) I found this: Horizon 2-4 is targeted for delivery in phases in calendar Q4 2026, with grace periods under the Microsoft Agreement for delivery extending from mid-Q4 2026 to the beginning of Q2 2027.

IREN carries too much risk as it’s capex is 98x it’s current revenue rate and asks you to trust they will be earning $1 billion by the end of the year. It did say currently they are at a $1 billion run rate, so that would still be 25x to 30x their revenue, on spending. With possible construction delays, along with non stop share dilution makes me eliminate IREN.

Why I would buy Nebius

NBIS has capacity growing to end 2026 with about 800 and 1,000 active megawatts. From there, they plan to double again to roughly 2,000 active megawatts by the end of 2027. Contract values have roughly doubled from a $12 million MW to over $20 million. On top of that, they are also holding back some space and signing some short term contracts where buyers pay between $40 million and $50 million a megawatt. The margins are growing. The AI cloud adjusted EBITDA margin rose from 24% in December to 45% in March and 50% in June. Much of its funding comes from customer prepayments at zero percent interest. In a buildout where everyone is losing money on capital, the winner or survivor is whoever loses the least. In an industry where most players are burning massive cash or diluting shareholders to buy chips, Nebius is executing financially, operationally and is doing right by shareholders. I would recommend Nebius as the best of the four.

If you read this I'll make you the following promises: you’ll have a complete understanding of this whole new thing called A Neocloud. You will understand each of these companies business models, Applied Digital (APLD), Nebius (NBIS), IREN (IREN) and Coreweave (CRWV). You will understand which ones you might want to invest in. The reason for this article is that a lot of people on Reddit and across financial media believe these companies have the potential to be re-rated much higher. That fits our thesis here. If you’ve been even loosely paying attention to news headlines, you’ve seen Nebius, NBIS and IREN signing multi-billion-dollar contracts with Microsoft and Google, so they certainly warrant a much deeper look.

Applied Digital signs $5.2 billion AI data center lease

So a Neocloud is a specialized computing company that rent out high performance GPU’s. They utilized specialized software to help run massive machine learning workloads and AI-specific training and inference loads.

Why do Neoclouds exist? Why doesn’t Microsoft just build these themselves?

Speed and Power. Getting a new facility running can take years with permitting, building, grid connections, and so on. So since OpenAI and Anthropic and the rest need compute right now, they are leasing it from these Neoclouds.

The Layers

Before we profile the companies, we need to briefly discuss layers. This below slide comes from NBIS. It will be helpful for us to better understand each company. They show six layers of a neocloud company.

Power. The power layer is control of the interconnect and substation. Quite confusingly, none of these companies actually generate their own power. They own substations, and connections, so they own the power access and some of the infrastructure.

Land and buildings. Specialized, purpose built AI focused buildings.

Hardware. The GPUs and racks, Nvidia and Broadcom and AMD chips, memory and the storage.

Cloud platform. The software that allows the GPU’s to be rented and to become a product for sale. In general, this is the basic software for general computer services and rental.

Service and software. This might seem like an additional software layer, but this is the actual neocloud layer. The cloud platform is simply access to the GPU, however the GPU service and orchestration software is running the massive amounts of data to train a new AI model, and make sure large jobs with massive data can’t crash or can be saved and will operate to completion without fail utilizing scheduling, optimization, failure detection and many other advanced tools. This is what allows them to charge a premium on top of the actual compute and electricity.

GTM or Customer. A positioning layer. Layer one through five are things you own, but layer six is the position you hold. Basically, are you the one that owns the customer relationship. Since so many of the above layers can be rented or outsourced, who holds and sells the compute from all the above?

The 4 Company Profiles

IREN (Iris Energy) —All the layers

Layers Owned: Layers 1 through 6, they call it three layers, but below you can see from their deck, they operate in all layers across the neocloud framework.

Business Model: Formerly a pure-play Bitcoin miner, IREN is executing a complete transition into a fully vertically integrated AI Cloud platform. They own their power interconnects and land, build their own data centers, buy and host GPUs, and operate their own software stack. Fully owns and finances GPU clusters, directly capturing the margin of high-density AI compute training and inference.

Applied Digital (APLD) - The Landlord

Layers Owned: Layers 1 and 2

Business Model:Applied Digital operates like a specialized AI data center REIT, although they are not a REIT. They secure land, build power interconnects, secure the power contracts, and build AI focused facilities. They rent the facilities to tenants on 15-year contracts.

They own the power & land & buildings but do not own any GPUs. They do not carry hardware depreciation risk, nor face the steep costs of GPUs but also don’t benefit if compute prices swing way up. Their job is simply to collect rent checks from major hyperscalers, including CoreWeave, and remaining legacy hosting tenants, including a Bitcoin operator.

Nebius Group (NBIS) — The Full-Stack Software & Infrastructure Cloud

Layers Owned: Full Stack — Layers 1 through 6, mostly

Business Model: Nebius builds and operates its own data center infrastructure while it operates AI software and developer tools to run AI training and inference. It owns the GPUs and physical data center footprint, catering directly to hyperscalers, frontier labs, enterprises and tech companies requiring AI cloud compute. Additionally as seen below they also do some colocation , layers 3-6, and have introduced the asset light business model. This gives partners who may have the hardware, facility and power to now add Nebius’s systems architecture, software stack and customers to begin to offer AI compute and gain more dollar per GPU rented.

CoreWeave — The Specialized Compute Provider

Layers Owned: Layers 3 through 6

Business Model: CoreWeave is the largest pure-play neocloud operator. Unlike competitors, CoreWeave generally avoids real estate development or direct power infrastructure ownership. Instead, they lease facility space and power from developers ,like APLD, and direct their capital into buying GPUs and building AI orchestration software. They have the largest amount of GPUs and GPU compute, that operating as an AI specialized cloud and compute provider for AI developers and hyperscalers.

Power Plan

So much is happening, I’ll cover everything I think is important. So let’s start with power. Whenever they mention power, you need to pay attention to the details. There is “secured power” which means they have a right or are in a line, but it could be simply dirt right now. “Contracted” power could mean many things but is often a substitute for is in process. Many of the largest numbers you will read are contracted power, because it simply means it’s underway. “Connected” means its connected but doesn’t mean there is a customer attached or that it is earning anything. So the ones we care about are “Active or Live” power. That is the actively running amount, that a company is billing for with cash is coming in. The active can also show IT load vs gross, meaning that not all power is sold, some power is used by the facility for cooling and so IT load is what is again the real number here. Much of the below power numbers are changing and coming online in mere months, so here is a close snapshot of how ACTIVE power looks across the companies currently and in the very near future:

IREN

They have about 75-100 MW active power with a goal to have 300 MW active by the end of the calendar year. This company along with others has fiscal years that are not aligned with calendar years. So I am always talking about calendar year. They are building and activating their data center in blocks, so that revenue can commence asap. The upcoming quarter shows them delivering more MW throughout the quarter, reaching up to 300 MW by the end of the year, from their current estimated 75-100MW. The 300 is also IT power, not gross power. They have a stated goal of 800 MW Active, by the end of 2027, and announced plans that total over 5 GW.

IREN has global data centers, and in fact all of the companies are global except for APLD.

APLD

Applied Digital is also delivering in phases, and currently has shown progress. From last quarter ending in May this year to the last June quarter it increased from 100 MW to 175 MW. It still has 286 MW that is rented to a bitcoin operator and they have stated no current intentions to change that. However, it is important to note that the current contract runs out in 2027, so they have optionality with this space. They have data centers exclusively in the United States with many additional sites coming online during the remaining half of this year, 2026, and throughout 2027 and 2028.

Also important is APLD is a landlord, and their income has no connection with the price of bitcoin. Their future plans are shown here in their recent company deck. It summarizes as reaching 1.41 GW all delivered by 2028.

NBIS

Nebius currently has between 175-250 MW active with 800–1,000 MW expected to be reached by the end of the year. The CEO states that their capacity build-out is accelerating, and he is expecting to deploy over 1 GW per year in 2027, and each year after. They have talked about 5GW as “contracted power” by year end, but it will be years before that is active MW. They are the most geographically diverse of the group, and has many sites throughout Europe and the US.

Coreweave

Currently it has 1,500 MW active power and forecasts to reach 1,850 MW by year end. With their focus on the GPU and software layers, it’s spending less on facilities, and they have used that extra capital on GPUs allowing it to scale up to the highest active MW in the group. The forecast is also the highest in the group with plans to reach 3 GW active by the end of 2027, and more than 8 GW active by the end of 2030. Behind this sits an active 4.2 GW in “contracted power” which is up from 3.7 GW is contracted power, so it’s great to see the future active power under the initial power stage.

Financing the buildout

Buying, converting, cooling, building all of this is very expensive, and all of these companies need to finance their plans. GPUs are crazy expensive, and since they are losing money, where is all the money to build these data centers coming from? So we need to look into how much each is paying to build their neocloud future.

Applied Digital has the most expensive cost of capital. They took on $6.9 billion in project bonds paying high interest between 6.75% and 9.25%. Over 4 Billion of this will come due in 2030 to 2031, so will likely need refinancing. On top of that, APLD set up a preferred program that brought in 1.8 billion and compounds at 12.75%, and go up from there, and comes with guarantees of 1.8x to 2.0x return. They are also issuing stock, and diluting shareholders by adding 53 million new shares, and there was another 10 million for stock based compensation so, 63 million in newly minted common shares. This brought in roughly $1 billion over the past thirteen months. I also need to include that 21.5 million warrants, 7% of the company shares, were given to CoreWeave just to get them to sign the initial leases. This is the worst financing of the group.

IREN somehow tops APLD in stock dilution. Iren basically used its own stock as a limitless credit card. Over the past year, their share count grew by more than 70%, jumping from around 224 million shares to over 380 million. They brought in over $2.5 billion from these issues, and they are also borrowing additional capital in the form of convertible bonds paying virtually 0% to 1% interest. These bonds are covertible to more IREN stock down the road depending on value. In some good debt news, Wall Street lenders finance 90% to 96% of their actual GPU hardware costs at single-digit rates. This is due to the already contracted AI compute agreements signed with Microsoft and other high investment grade clients.

Nebius has the best funding news in the entire group because their customers are paying the bill upfront. About 70% of their customer deals require 50% to 60% upfront payments to reserve cluster capacity. The prepayment funds are on their balance sheet as about $6 billion in deferred revenue, giving them a total around $8 Billion in cash at the end of last quarter. They use some low-rate convertible notes as well but they don’t need to take on high interest debt to fund their future. They have also done some dilution but at an understandable rate. First, they did it after their shares had risen to new highs, and did a reasonable 14%, or about 17.5 million shares. Funding a 1 GW buildout / year will take more capital than their $8 billion so this is understandable.

CoreWeave is running with maximum financial leverage. They have over $16 billion in lease liabilities for the buildings it rents, and $35 billion in total debt. The Debt to equity ratio is 10.87, compared to APLD at 2.86, IREN at 1.87 and NBIS at 0.97. Coreweave utilizes convertible bonds, alongside billions in senior notes paying 8.5% to 9.75% interest. So the company relies entirely on outside funding to work, and their current debt leaves them a quarterly interest bill of more than $640 million. They can service that debt as long as their contract backlog continues to convert into revenue, but readers should know their rating status is currently at junk. Their shares have also been diluted as some of the convertibles have come due. They recently issued more debt, and the financiers demanded a higher rate.

Coreweave raises yields to entice investors

Comparing them all, customer prepayments king, NBIS, comes out on top as the cheapest financing, and they have 2 billion in cash on the books, left over from the Yandex corporate split. Next is IREN with its 0 to 1% coupon convertible debt, then the highly leveraged CoreWeave, and last the high interest construction bonds with warrants, and due dates, APLD.

That's part one, tomorrow we will look at what all four are actually making in revenue right now, and what each one makes per megawatt that's live and billing, and how that is changing rapidly. I’ll give you the insights I learned while doing this research and then I'll tell you which ones I'd put money into and why. Stay tuned, subscribe.

The Opening Pop

On Friday morning, at the end of the worst week for chip stocks in months, the biggest American listing ever completed by a foreign company listed at $149 and opened at $170. Demand ran seven times the shares available. Baillie Gifford, Coatue, and Leopold Aschenbrenner’s Situational Awareness fund anchored about $5 billion of it before the shares were even open for the rest of us.

I’ve been around new issues long enough to know better than to buy the opening pop. If you are old enough to remember the SpaceX IPO, you’ll remember the stock opened at $170, spiked to $225 within a day, and has already handed back the entire move. SpaceX loses about $5 billion a year and trades above 100 times sales.

SK hynix (SKHY) is the other trillion-dollar listing of the summer, and it is a completely different kind of stock. This is a company that makes lots of money right now, not a bet on an ideal future where electric cars and spaceships fill our roads and skies. Memory stocks, including Micron (MU), have seen net income explode along with their share prices over the past year, so a new ADR listing from the biggest AI-memory maker on Earth has my full attention.

The numbers that got my attention specifically: SK hynix earned about $64 billion in revenue last year at a 49% operating margin, and has since improved it to a massive 56%. Its profit margin was larger than Samsung Electronics, larger than MU, and makes it one of only three trillion-dollar companies with profit margins above 50% this year. Plus, it came to the Nasdaq priced between five and seven times this year’s expected earnings, depending on whose estimate you use.

(Korean won figures in this article are converted at the ₩1,523.5 per dollar rate the prospectus itself uses.)

A quick review before we go further. Last October I published the case for Micron on Seeking Alpha when it traded at $212 and the market still priced it like a commodity cyclical. The argument was that high-bandwidth memory (HBM), with its stacked memory design, had quietly transformed memory manufacturers into critical suppliers for the AI infrastructure. Micron closed Friday at $979—up about 360% since that article. I’m still long from $212, and I continue to hold. SK hynix is the same story but now is available to U.S. investors without setting up a Korean account or having to buy a Korean ETF with all of the other holdings.

What This Company Actually Is

SK hynix is the world’s #1 supplier of HBM, the high-bandwidth memory stacked next to every Nvidia GPU, with 56.4% of that market by revenue in the first quarter, per IDC figures cited in the prospectus.

The broader story looks like this. In DRAM, they are the #2 maker globally at a 29.1% share. In NAND Flash, they are also the #2 maker at an 18.5% share, with their strength heavily concentrated in enterprise SSDs for data centers. Essentially, they sell the full memory stack: HBM for accelerators, DDR for servers, LPDDR for power-sensitive machines, GDDR for graphics, and eSSD storage.

Two structural points from the prospectus are worth passing along, because they’re the kind of thing that separates this business from the average performer.

First, the barriers to entry are incredibly high. Gartner projects total global semiconductor capital spending of $237 billion in 2026, with the top 20 spenders accounting for 87.6% of it. You cannot start a memory company anymore; the capital requirements are tens of billions a year.

Second, the customer base has fundamentally changed. Enterprise buyers, specifically data centers, now account for approximately half of all memory demand, according to IDC—up from about a quarter in 2020. Memory used to be brutally cyclical, but we know that’s not the case anymore. Hyperscaler capital spending is planned and funded years in advance.

Global Memory Semiconductor Market Size

(In billions of U.S. dollars)

The prospectus indicates the memory market growing from $216 billion in 2025 to $633 billion in 2026 and $748 billion in 2027. If that seems like a lot, it’s actually all of the available memory on planet earth. All of it is all sold out, sorry if you wanted to buy a lot of memory.

SK hynix revenue went from $43 billion in 2024 to $64 billion in 2025, up nearly 50%, with HBM revenue more than doubling. Then the first quarter of 2026 showed $34.5 billion—up 198% from a year earlier. Micron’s numbers are similar; its May quarter came in at $41.5 billion, up 346%, and it guided the current quarter to roughly $50 billion. This is a three-player market where demand from a handful of hyperscalers has outrun the planet’s supply, and both listed pure plays are sold out into next year.

Having half of last year’s revenue in just one quarter this year makes me worry about the market’s sustainability. However, earlier this year Sanjay Mehrotra, MU CEO, said they signed guaranteed revenue through 2030 via long-term contracts, which was great news. Now with Kwak Noh-jung, SK hynix CEO, stating that “customer demand for AI memory chips will outstrip production capacity even beyond 2030”, the sustainability is confirmed to me. The most important thing to remember is the HBM architecture itself, in that every advancement uses more and more memory. The AI needs more and more memory to execute complex tasks, provide true utility with larger context windows, and build upon previous tasks.

How the Buildout is Being Paid For

With every AI name, you have to ask where the money for the plans comes from, because the answer lately is the bond market. The four big hyperscalers plan roughly $725 billion of combined capital spending this year, up 77% from last year’s record. Morgan Stanley expects AI-related debt issuance to approach $570 billion in 2026.

Alphabet has announced a significant $85 billion equity dilution to finance its AI data center expansion. Amazon has borrowed roughly $64 billion across offerings this year, set records in the Canadian and European bond markets doing it, and still opened a $17.5 billion credit line. Even Nvidia sold $25 billion of bonds in June.

SK hynix funded its expansion with this new ADR. It generated about $28 billion net, with demand seven times the offering, diluting existing holders by about 2.5%—and it barely needed to do that. The business generated $35 billion in operating cash flow last year against $18 billion of capital spending. The first quarter alone produced $17.3 billion in operating cash against $5 billion of capex.

The proceeds have a specific job: roughly $30 billion for new Korean fabs including the first Yongin cluster plant, $7.8 billion of EUV lithography tools through 2030 (ASML anyone?), and a $12.5 billion advanced packaging plant breaking ground next spring, along with packaging capacity in Indiana. Notice what’s not on that list: China. Every dollar of growth capital lands in Korea or the United States, which matters for the risk section below.

Profitability: Numbers Memory Has Never Printed

The first quarter’s gross margin was 79.3%. The operating margin was around 72%. Return on equity runs above 60%, on a balance sheet with almost no debt and about $14 billion in cash.

For reference, Micron recently reported gross margins exceeding 80%. This unprecedented economic performance by two competitors indicates a structural scarcity issue, not some great management team. It also brings to mind my second question that matters for the investment: Do margins like this continue? Are memory margins permanently changed?

Valuation: MU as a Comparison

On earnings, the stock is startlingly cheap. It sits between five and seven times 2026 earnings, representing a PEG ratio of 0.07. The trailing earnings multiple is around 21. Thirty-five of the thirty-seven analysts covering the Seoul shares rate it a buy, with an average target about 45% above the recent close. Micron, after its own re-rate, trades near 13 times forward earnings. Again, SK hynix is at a 6-7 forward P/E and Micron is 13. Closing this gap is probably what this Nasdaq listing exists to do.

On book value, the picture changes. Memory investors spent thirty years valuing these companies on price-to-book, because at every cycle turn the earnings evaporated and book value was all you could trust. (Who’s old enough to remember MU stuck at $40 for years?) SK hynix trades near nine times trailing book, maybe five times year-end 2026 book if the consensus profit shows up. That is much higher than the historical 3-4x in the best of prior cycles. Even the bullish Korean brokerage notes justifying higher targets lean on the observation that a broad consensus has formed on the durability of this upcycle.

So the P/E on this debut still shows the market pricing SK hynix as a cyclical that will eventually give the earnings back. The P/B is higher than ever on a historical basis. Which market view of the stock wins depends on whether long-term contracts have genuinely broken the memory cycle. Micron has signed 16 multi-year agreements with binding commitments it values at $22 billion, and the market rerated it. SK hynix has its entire 2026 sold.

Momentum: A Debut with a Calendar Attached

The stock in Seoul is up about 675% over the past year. It set a record high in late June, then pulled back roughly 25% into this month’s chip selloff—the same selloff that knocked Micron 22% off its high in two weeks. SK hynix listed directly into that weakness and still opened 14% above pricing. New issues that enter in that environment and hold the bid are telling you something about demand.

This debut has a lot going on in the next few months. On July 13, regular trading under the permanent ticker SKHY begins (the temporary SKHYV symbol retires) the same day a pair of leveraged ETFs on the name launches. Then, second-quarter earnings land on July 23, just ten days into the stock’s U.S. life. On July 29, the newly issued shares list on the Korea Exchange. Looking further out, early October brings the lock-up expiration on the company and its affiliates—an unusually short 90 days that is worth a calendar entry, since that’s when additional supply could legally hit the market. Finally, December brings eligibility for Nasdaq-100 inclusion, the kind of mechanical index demand that doesn’t read anyone’s cycle views.

Revisions: The Estimates Can’t Keep Up

This is quietly the strongest factor and the one that made me act rather than just watch. The analyst revision log on the Korean shares reads like a ratchet: 2026 earnings estimates lifted from around ₩200,000 per share to ₩281,000 in successive rounds, full-year revenue forecasts raised from $162 billion to $209 billion, price targets that roughly quadrupled in months.

One major Korean broker raised its target 43% in a single note this spring, and the stock ran through the new number anyway, and then they raised it again. (Korean analysts are just as useful as American ones, good to know). When a business beats, guides higher, and forces the models up quarter after quarter, that revision pressure is the mechanical engine of a re-rate.

What Can Break It

Three things can trip this up, in order of how much they’d hurt.

China is the structural one. About 40% of SK hynix’s DRAM output comes from its fab in Wuxi, and roughly a quarter of its NAND from the Dalian plant it bought from Intel. Washington revoked the blanket equipment waivers for those plants effective the end of last year and replaced them with annual licenses that allow operation and maintenance but explicitly no capacity expansion or technology upgrades. The Wuxi fab is effectively frozen at older process nodes, and every December is now a renewal event that a trade dispute could complicate.

The most effective mitigation measure is the one mentioned in the funding section: the HBM lines, the transition to leading-edge DRAM, and all the new capital expenditures are located in Korea and Indiana. The frozen assets represent the legacy infrastructure, not the growth engine. However, a license denial would still significantly reduce output and necessitate the costly replacement of capacity years before the planned schedule.

Global HBM Market Shares by Revenue

Competition is the live one, and the prospectus includes a chart I’d frame if I were bearish. SK hynix’s share of HBM revenue was 63.2% in 2025. In the first quarter of 2026 it was 56.4%. In the same stretch, Micron went from 17.4% to 23.1%, passing Samsung to become the #2 HBM supplier in the world, while Samsung stabilized around 20% and won a fifth of Nvidia’s next-generation orders plus the lead slot on AMD’s new accelerator.

SK hynix still holds an estimated 70% of Nvidia’s HBM4 allocation, and Nvidia was 14.8% of its total revenue last quarter, with a second customer crossing 10% for the first time. But the direction of the share chart is not in dispute; the bull case only requires that the erosion stay slow while the pie grows fast. That chart, more than anything else, is why my position size looks the way it does.

The cycle is the existential one, and while I believe sales are changed with the technology, I don’t know about the margins. No memory company has ever sustained a 70% operating margin. The whole debate—the P/E lens against the P/B lens—compresses into whether contracted, allocated, multi-year AI memory demand behaves differently from every boom in this industry’s boom-and-bust history. My framework doesn’t require me to answer that with confidence. It requires me to size the position as if I might be wrong.

What I Actually Did

Ultimately, we have a super high-margin business that hasn’t yet fully reached a price that reflects its current business. I did buy it. But my overall prospectus reading left me feeling like I wanted to buy, but not sell MU or anything crazy, just to purchase more.

Every stock I own must pass two hurdles. The first is my extensive framework. I won’t go over all of it today, but I ask a long list of questions, such as if Google, Amazon, Microsoft, or Meta decided tomorrow that it wanted to enter this market, does this company still win? Is this company headquartered in, or critically dependent on, high-risk jurisdictions? Are they an indispensable architectural bottleneck, that grows as the technology grows? SK hynix did pass all of my questions and is probably a close second to Nvidia as a reliable supplier in AI infrastructure. The second gate is independent confirmation from a quant system, which grades valuation, growth, profitability, momentum, and revisions without caring about my narrative. Micron sits at Strong Buy there.

SK hynix has no Quant rating at all. It can’t, it’s too new. My quant gate exists for a great reason because I have bought an lost a lot on great stocks that only passed my own screener. This second gate exists precisely for moments like this: when the story is this good, and the temptation to skip the process is strongest.

So I own a starter position in SKHY, and it earns its way to full size as the confirmations arrive: the July 23 earnings date, the first full quarter of regular trading, Quant coverage initiating, and December’s index decision. Micron remains my anchor in the space, long from $212, and I feel great about my MU after seeing the SK hynix prospectus.

Disclosure: I am long MU and SKHY. Everything here describes my personal portfolio and process. It is informational only — not individualized investment advice. Do your own research.

A company spends years being valued as one thing, a memory commodity, a niche hardware maker, a boring HVAC industrial and then the market changes it’s thinking and the stock explodes in value. The multiple expands. The story changes. That’s a re-rate, and if you owned it before the recognition, you didn’t need leverage, options, or timing. You needed to see the future slightly earlier than everyone else, and then do the hardest thing in investing: hold on.

That’s what I look for. Not trades. Not alerts. The small number of companies positioned in front of technology shifts the market hasn’t priced yet.

Every stock that took this path, like Apple or Nvidia, AMD made its money for the people who understood why they owned it and sat still. I built this newsletter around a common sense version of that: finding which companies are sitting in a great position with great management in a sector that will be worth a lot more in the future. I spent a decade as a financial advisor and share my expertise as I would be doing this anyway. I also write for seeking alpha, and you can see my work here: Blake Callahan Seeking Alpha

A track record you can check. In 2025, I published the case for Micron as the most undervalued AI play — HBM memory as a chokepoint of the AI buildout — when the market still saw a commodity cyclical. That call is the template for everything here: find the re-rate before the reprice.

What subscribers get

Free subscribers receive selected research, market commentary, and three fully disclosed positions from my portfolio, thesis, entry logic, and ongoing updates.

Subscribe now

The Re-Rate is my personal portfolio and my personal research, published for informational and educational purposes. It is not individualized investment advice — I don’t know your situation, and nothing here is a recommendation tailored to you. I own the securities I write about and will always disclose my positions. Do your own diligence; markets carry risk, including the loss of capital.