The Bottleneck of All Bottlenecks

If you’ve ever built your own computer, you know the core components. CPU, GPU, RAM, storage, motherboard. Between these, one represents the biggest bottleneck in computing today and it’s not the one you’d expect.

Funny enough, I experienced this firsthand. In 2024 I decided to build myself a new PC and went with 64GB of DDR5 RAM. Paid around $150. I just checked my local retailer. The same kit is now $1,000+.

I could probably pull it out of my computer right now, sell it used, and turn a good profit. Not a bad trade for hardware I’ve been running for a year.

The reason why is the story of this article.

Every major tech company on the planet is racing to build AI infrastructure and the numbers are hard to wrap your head around. Meta just announced they are launching a cloud compute business and spent $145 billion on AI infrastructure this year alone. Google, Microsoft and Amazon combined are spending hundreds of billions more. The hyperscalers collectively committed over $1 trillion in AI capex for 2026. NVIDIA just crossed a $4 trillion market cap on the back of this. Jensen Huang said at GTC 2026 that NVIDIA is targeting $1 trillion in revenue by 2027. The US government declared AI infrastructure a national security priority. Data center construction permits in the US hit all time highs in 2025. Power grid demand from AI data centers is projected to double by 2030.

This is not a maybe. This is not a theme or a narrative. This is the most obvious and largest capital deployment in the history of technology and everyone with serious money knows it.

The question is not whether compute wins. That is settled. The question is what compute actually needs to exist. And the answer to that is memory.

Your computer has two main processors. The CPU does one thing at a time, just very fast. The GPU does millions of things simultaneously, originally built to render every pixel on your screen dozens of times per second.

It turns out AI is essentially massive amounts of simple math done in parallel. The same thing GPUs were already doing for graphics. NVIDIA figured this out around 2012 and pivoted the entire company toward it. The gaming GPU and the AI chip are the same fundamental architecture, just taken to completely different extremes.

That pivot took NVIDIA from a gaming company to the most valuable company in the history of technology.

Now that you understand why GPUs run AI, here is the problem that needed solving. A GPU doing millions of calculations simultaneously needs to be fed data at an equally insane rate. Regular RAM, the same kind in your PC, cannot keep up. The GPU sits idle waiting for data. In a machine costing millions of dollars, that is an unacceptable waste.

This is where HBM comes in. Remember the RAM you bought for your PC? HBM is the same fundamental idea taken to a completely different dimension. Same concept, completely different league.

Regular RAM is laid flat on your motherboard. HBM is stacked vertically, layer on top of layer, like a skyscraper built directly on top of the chip. If regular RAM is a two lane road, HBM is a 100 lane highway. That architecture delivers roughly 10x the bandwidth of anything else available, fast enough to keep the GPU fully fed at all times.

The other thing that makes it irreplaceable is physical. HBM is not plugged in separately like the RAM sticks in your PC. It is bonded directly onto the chip package during manufacturing. You cannot swap it, upgrade it, or design around it. It is baked in at the hardware level. No HBM, no AI chip. The dependency is absolute.

Meet Vera Rubin. NVIDIA’s newest AI rack and the most sought after piece of hardware on the planet right now. A single unit packs 72 GPUs into one liquid cooled system and costs $9 million. Every major tech company is racing to get their hands on it.

Here is where it gets interesting. Those 72 GPUs cost $55,000 each. Do the math and that is roughly $4 million in GPUs per rack. So where does the other $5 million go? Networking, cooling, chassis. And memory.

Memory alone accounts for $2 million of that $9 million rack. Specifically HBM, the same technology we just explained, not the chip everyone talks about. Memory. Two years ago memory was 5% of the cost of an AI rack. Today it is 25%.

Vera Rubin allocations for 2026 are gone. Meta, Google, Microsoft and Oracle locked up everything available before the product even fully ships. These are the largest technology companies in the world with unlimited capital and they cleaned out the entire supply. If you want meaningful capacity today, you are in a queue that stretches into 2027.

GPU compute gets more efficient every generation, which actually works against the GPU’s share of the rack cost over time. The chip gets relatively cheaper. But model sizes keep growing, inference demand keeps scaling, and HBM demand grows faster than those efficiency gains. The memory share of rack cost is not stabilizing at 25%. It is heading higher.

Three companies make HBM. SK Hynix invented it and controls roughly 50% of global supply. You cannot close that gap quickly. Building a new HBM fab costs over $10 billion. The equipment is entirely different from regular DRAM production. Training the workforce takes years. Even if you started today you would not see meaningful output for 3 to 4 years.

Because supply is constrained and there is no substitute, SK Hynix sets the price. They are not competing on cost. They are the only real option. That is the definition of pricing power.

And the margin story makes it even more interesting. HBM gross margins run in the 50 to 60% range. Regular DRAM sits at 20 to 30%. The more the market shifts toward HBM, the more profitable SK Hynix becomes even without growing unit volume. They win on price and they win on margin simultaneously.

SK Hynix allocations are sold out well into 2026 and beyond. SK Group's chairman recently stated that customers are requesting 4 to 5 times more HBM supply than current capacity can deliver, while wafer capacity is only expected to double by 2030. The gap between what AI infrastructure needs and what can physically be produced is not a near-term problem. It is a multi-year structural constraint. And here is what most people miss. NVIDIA cannot produce more Vera Rubin racks not because of their own limitations, not because TSMC cannot manufacture the GPUs fast enough, but because SK Hynix cannot supply enough HBM4 to keep up with GPU production. The most valuable company in the world is capacity constrained by a Korean memory manufacturer most people have never heard of.

And it does not stop at NVIDIA. Every major AI chip on the market, AMD, Google, Amazon, all run on HBM. SK Hynix supplies the majority of it. It genuinely does not matter who wins the AI chip war. SK Hynix collects on every side of it.

SK Hynix is still priced by the market as a commodity memory company. That is the entire opportunity in one sentence.

The numbers back it up. The stock trades at roughly 6x forward earnings while the broader semiconductor sector trades at 37x. That gap exists because the market is pricing SK Hynix on what it was, not what it’s becoming. A company that controls 50% of the most critical input in AI infrastructure does not belong at a commodity multiple.

The stock has already moved. On the Korean exchange it went from years of flat trading to a near-vertical run as the HBM story clicked for institutional investors, peaking at 2,723,000 KRW. On July 10, SK Hynix listed on the Nasdaq under the ticker SKHY, raising $26.5 billion in the largest US share sale ever by a foreign company. The ADR opened at $170 and jumped 13% on debut day.

Then on July 13 it posted its worst single day on record, down 30% from the all-time high. A Nasdaq debut hangover, one brokerage note flagging a potential Q3 earnings miss, and a broader AI sentiment selloff hit simultaneously. That is the entry point this article is being written around.

The structural story has not changed. Sold out allocations into 2026 and beyond. No substitute for HBM4. Pricing power intact. The risk is near-term earnings volatility. But the gap between how SK Hynix is priced and what they actually are is where the opportunity sits.

Memory companies have always been cyclical. That is not an opinion, it is thirty years of market history. Oversupply kills margins, prices collapse, the weakest players exit, supply tightens, margins recover, repeat. The market learned to never pay a high multiple for memory earnings because those earnings never lasted.

That conditioning is exactly why SK Hynix still carries a commodity discount. The market is applying a playbook that no longer fits.

HBM is not DRAM. Regular DRAM is a commodity where any manufacturer can flood the market and crush pricing. HBM requires a completely different fab, different equipment, different process technology, and years of yield improvement that cannot be replicated quickly. There are three companies in the world that make it. One of them controls 50% of supply and is the primary qualified supplier to NVIDIA. You cannot overcorrect into oversupply when building a new HBM fab costs over $10 billion and takes four years to come online.

The boom/bust cycle requires someone to oversupply the market. Nobody has the capacity to do that here. The discount the market is applying to SK Hynix is borrowed from a different era. That is the re-rate waiting to happen.

Samsung closing the yield gap on HBM4 is the most important variable in this trade. They have the capacity, the capital, and the motivation. They are currently behind on yields and behind on NVIDIA qualification. If they close that gap, the HBM market goes from one dominant supplier to two, and SK Hynix pricing power gets tested. That is the bear case.

But here is how to read it if Samsung does figure out the process. They become the number two supplier in the most critical bottleneck in AI infrastructure. That is not a threat to the thesis. That is confirmation of the thesis. The addressable market is large enough that a qualified Samsung just means more of the supply chain gets priced correctly, not that SK Hynix loses.

Micron is the third player and the domestic exposure angle for US investors who want the memory supercycle without Korean market risk. They are scaling aggressively and are further behind than Samsung, but the US government has strong incentive to help them close the gap given the concentration risk in Asian memory supply. Watch their HBM qualification progress with NVIDIA as the signal.

One more thing worth understanding. SK Hynix and NVIDIA are joined at the hip on earnings. If SK Hynix ships less HBM than expected in a given quarter, NVIDIA builds fewer Vera Rubin racks than expected, and their data center revenue misses too. The Q3 concern that triggered the July selloff is not an SK Hynix specific problem. It is an AI infrastructure timing question that affects the entire supply chain simultaneously. Which means if it resolves, it resolves for everything at once.

The demand itself is not in question. Every new AI chip announced runs on HBM. Every new data center needs more of it. The order book does not empty. Revenue already committed ships eventually. The only variable is which quarter it lands in.

You felt it in your RAM prices. A 64GB DDR5 kit that cost $150 eighteen months ago now runs close to $1,000 at retail. That is not a coincidence. That is what happens when the dominant memory manufacturer redirects capacity toward a product that earns twice the margin, because the buyers at the other end have no choice but to pay.

The data center version of that story runs at $2 million per rack. NVIDIA designs the chip. TSMC builds it. SK Hynix makes it possible. The market has priced the first two correctly. It has not priced the third.

Memory was supposed to be the boring part of the AI trade. It turned out to be the bottleneck.

This is not financial advice. neym is an independent research newsletter. Do your own research before making any investment decisions. The author may hold positions in securities mentioned.