Three things happened in the AI world this week. Any one of them is a headline. All three together? A structural earthquake that rewrites who holds the power in AI infrastructure.
First: South Korean memory giant SK Hynix listed on the Nasdaq, raising $26.5 billion. Not $2.65 billion — $26.5 billion. That’s the largest foreign IPO in US history, topping Alibaba’s $25 billion from 2014. The book was oversubscribed more than 7x. Shares popped 14% on the opening bell.
Second: NVIDIA’s stock is down 15% from its May peak. Meanwhile, Micron — the memory company — has tripled over the same period. On the spot market, H100 GPU rental prices have slid from $3.20/hour to under $2.60/hour. At the same time, DRAM spot prices have risen tenfold since summer 2025.
Third: Meta announced its custom MTIA AI chips will enter mass production in September 2026. OpenAI is building an inference chip codenamed “Jalapeño” with Broadcom. Anthropic is in talks with Samsung for custom silicon. Google has TPU v8. Amazon has Trainium. Microsoft has its own chips.
These three events converge on a single conclusion: AI’s bottleneck has moved. It used to be GPUs. Now it’s something else entirely — something that almost nobody is talking about.
You might push back: isn’t NVIDIA’s revenue still exploding? Aren’t GPUs still in short supply? Yes and yes. But price signals are more honest than earnings reports. When H100 spot prices are falling while DRAM prices are going vertical, the market is telling you something. GPU scarcity is easing. A different scarcity is tightening its grip.
This shift — from compute scarcity to memory scarcity — will shape the AI industry for the next decade, more profoundly than any single model release. Models come and go. Supply chain structures, once they harden, don’t.
Fig 1: One side falling, one side soaring. The bottleneck has moved.
Let’s talk about that $26.5 billion IPO.
SK Hynix makes memory. Specifically, it makes HBM — High Bandwidth Memory — the stacked, ultra-fast memory chips that sit right next to every high-end AI GPU. Each NVIDIA H100 or B200 ships with 8 to 12 HBM stacks glued to its side. Without HBM, a GPU is a very expensive paperweight.
Only three companies on Earth can manufacture HBM: SK Hynix, Samsung, and Micron. SK Hynix currently commands the largest share. So when US investors heard SK Hynix was listing on the Nasdaq, the order book went past 7x oversubscribed. People were clawing for shares they couldn’t get.
There’s a term in finance: the “Korea discount.” Korean companies typically trade at lower valuations than global peers — corporate governance concerns, geopolitical risk, chaebol complexity. SK Hynix didn’t just avoid the discount. It priced above its Korean listing by 2.7%. Investors didn’t care which country the company came from. They cared about one thing: this company holds the world’s most critical hard-to-replace component in the AI supply chain.
And almost in the same breath, US Commerce Secretary Howard Lutnick appeared at a Micron event and publicly told Samsung and SK Hynix: build fabs in America. You cannot leave this industry concentrated in one country.
Micron immediately pledged $250 billion for US manufacturing, promising 90,000 jobs. Samsung and SK Hynix had just two weeks earlier committed $550 billion to Korean domestic investment. Now the Americans are saying: share some of that.
Three companies. Trillion-dollar investment commitments. In any other industry, this would sound like science fiction. In memory, it’s the table stakes.
Fig 2: Custom silicon from above, HBM lock-in from below — NVIDIA is caught in a vise.
But here’s the truly counterintuitive part: GPUs are getting cheaper while memory is getting more expensive. That’s not supposed to happen. The standard story is “AI is compute-hungry, therefore GPU prices go up.” The standard story is wrong.
TechCrunch ran a headline that stung: “Nvidia is a victim of the compute marketplace it created.” NVIDIA built this ecosystem. Now the ecosystem is eating its creator.
Why? Wayne Nelms, CTO of Ornn, put it with surgical precision: “Everybody is making their own silicon right now, but nobody is making their own DRAM.”
Building a chip takes money and determination. Broadcom handles the design. TSMC handles the manufacturing. OpenAI paired with Broadcom for Jalapeño. Meta used Broadcom for MTIA. Anthropic is talking to Samsung. Even China’s Sugon built a 100,000-card cluster using domestic Hygon accelerators.
Memory is different. HBM is built on decades of process engineering — layer stacking, through-silicon vias, thermal management at microscopic scales. You can’t just raise money, hire Broadcom, and draw a DRAM. The barrier isn’t funding. It’s physics, accumulated over thirty years of clean-room experience.
And here’s the multiplier: every AI GPU needs 8 to 12 HBM stacks. When GPU supply increases — which it is, massively — HBM demand multiplies with it. But HBM supply can’t multiply at the same rate. So the price goes vertical. Simple economics, devastating consequences.
Fig 3: Spring for memory makers, autumn for GPU makers.
Here’s an even more damning number: 86%.
That’s the percentage of enterprises whose GPU utilization runs at 50% or below, according to a VentureBeat survey of 573 companies. They spent tens of thousands — sometimes millions — on AI hardware, and half the time, it’s sitting idle.
I found this number hard to believe at first. You pay six figures for a GPU cluster and it gathers dust half the day? But enterprise procurement follows a brutal logic: you spec for peak demand, and most of the time, you’re not at peak. Worse, buying GPUs doesn’t mean your team knows how to use them. The software stack isn’t ready. The scheduler isn’t tuned. The model isn’t optimized. You bought a professional kitchen and you’re still ordering takeout.
The survey also found that 44% of enterprises can’t even track how much they’re spending on AI compute or what return they’re getting. AI spending is, for a lot of companies, a leap of faith with no post-flight debrief. In any mature industry, this would be unacceptable. In AI, it’s normal.
So here’s the paradox: Wall Street is debating whether AI infrastructure is overbuilt, while enterprises are running their most expensive hardware at half capacity. Both things are true simultaneously. Companies don’t buy GPUs to run them at 100%. They buy them for insurance — for the day they might need them. Like buying a fire extinguisher, not a toaster.
The “Everyone Builds Their Own Silicon” Club Keeps Growing.
Meta’s MTIA chips have been in development since 2023. September 2026 marks mass production — Broadcom designed them, TSMC will fabricate them, and Samsung will supply the RAM. These chips are targeting Meta’s internal training and recommendation workloads, the kind that currently burn through NVIDIA GPUs. Meta’s 2026 capex is projected at $125 to $145 billion. They plan to deploy 7 gigawatts of compute this year and double that next year. A significant portion will run on their own silicon.
Google’s TPU is on its eighth generation, running at scale on Google Cloud. Apple and Anthropic both train models on Google TPUs. Amazon’s Trainium chips have won training contracts from Anthropic, OpenAI, and Apple. Microsoft has shipped its own inference chips. OpenAI’s Jalapeño with Broadcom is the latest entry — and think about this: OpenAI was one of NVIDIA’s biggest customers. When your biggest customer starts building their own product, it’s not about saving money. It’s about not having a single point of failure.
Line up the timelines. Meta MTIA: September 2026. OpenAI Jalapeño: in development. Anthropic custom silicon with Samsung: discussions underway. Google TPU v8: already running. Amazon Trainium: already running. This isn’t a coincidence. It’s an organized exodus.
It reminds me of a martial arts movie. NVIDIA is the master swordsman. Everyone wants to learn his technique. Years pass. The students graduate. Some open their own schools. Some find new masters. And the old master realizes: the truly irreplaceable thing isn’t his sword. It’s the iron ore used to forge it.
Meanwhile, China Took a Different Path.
The Sugon 8000 cluster in Zhengzhou is China’s first domestically-built 100,000-card AI cluster. It uses Hygon accelerators — Chinese-designed chips — with fully domestic networking, storage, and cooling. The cluster has already been validated across 300 application scenarios spanning 20 fields: materials science, electromagnetics, quantum computing, biopharma. 80,000 cards running protein folding simulations. 88,000 cards on turbulence modeling. 90,000 cards on atomic-scale simulation.
The US approach is “de-NVIDIA-fication” — Meta, OpenAI, Google all building alternatives to NVIDIA. China’s approach is “de-Americanization” — chips are sanctioned, so build domestic. Two paths, one destination: reducing dependency on a single supplier.
These two supply chains don’t intersect. They run in parallel — one anchored around the US-Korea-Taiwan alliance, the other pursuing full-stack indigenization. They’re not competing. They’re diverging.
So What Does This Reshuffling Mean?
First, AI chips are commoditizing. When Meta, Google, Amazon, Microsoft, and OpenAI all have their own AI silicon, NVIDIA’s pricing power erodes. GPUs stop being a “must-buy” and become a “can-choose.” That’s good for the industry — lower prices, more options, faster innovation at the hardware level.
Second, memory is the new center of power. When everyone can design a chip but only three companies can manufacture HBM, the bargaining power shifts from chip designers to memory suppliers. That’s why SK Hynix can raise $26.5 billion. That’s why Micron’s stock tripled. That’s why the US Commerce Secretary is personally making phone calls.
Third, the global AI supply chain is splitting in two. One half: US design + TSMC manufacturing + Korean/American memory. The other half: China’s fully domestic stack from chips to cooling. These aren’t competitive lanes. They’re parallel tracks, each running its own race.
I’ll be honest: I don’t know who wins. But I’m certain of one thing — the most important stories in AI right now aren’t happening in keynote presentations or on model leaderboards. They’re happening in memory fabs where technicians in clean-room suits are stacking HBM layers under micron-level precision. They’re happening in TSMC’s fabrication plants, where lithography machines etch Meta’s and OpenAI’s custom silicon 24 hours a day. They’re happening in Zhengzhou, where 100,000 domestic cards hum through protein-folding simulations. These are the things that determine how far AI can actually go.
We spend too much time watching the waves at the surface — GPT-5.6 beats Fable 5 by a few points, GLM-5.2 catches up. We miss the deep current underneath: who controls chip design, who controls manufacturing, who controls memory supply. The distribution of power across these three layers determines how high the surface waves can rise.
And that distribution, over the past three months, has come into sharp focus. The design layer is fragmenting. The manufacturing layer rests on TSMC. The memory layer is locked in a three-company oligopoly spanning two countries. A disruption at any single point in this chain would give the entire global AI industry a fever.
Honestly, this makes me both uneasy and strangely hopeful. Uneasy because the more concentrated the supply, the more fragile the system. Hopeful because — can you see it? — China’s indigenization route is breaking that concentration apart. When an industry develops two parallel supply chains, neither side has an absolute chokehold anymore. For consumers, that’s good news.
Here’s what all of this means for you, practically. GPU prices are falling — that means AI inference costs will continue to drop. Your ChatGPT subscription probably won’t get more expensive. More chip options mean the entire AI service won’t grind to a halt because one company faces a supply disruption. Memory prices are rising — if you’re building an AI startup, your biggest hardware cost might shift from GPUs to RAM sticks.
The deepest impact, though, might be this: when AI chips stop being “something only NVIDIA can make,” AI compute starts looking less like a scarce luxury good and more like electricity — a ubiquitous utility. I grew up in a place where blackouts were just part of life. Then the grid expanded, and nobody worried about blackouts anymore. AI compute might follow the same arc. From scarcity to abundance. From one company’s monopoly to infrastructure as invisible as the power grid. That process might take five years. It might take ten. But it’s already happening.
When the supply side shifts from monopoly to diversity, interesting things happen. Prices keep coming down. Options keep multiplying. Innovation spreads from the model layer to the hardware layer. You can see it already. Meta says custom chips hit production in September. OpenAI’s Jalapeño is on the way. Google’s TPU v8 is live in the cloud. Once all three are running at scale, NVIDIA’s share of AI training and inference will inevitably shrink. Not because NVIDIA fails — but because it stops being the only answer.
For developers and founders, this is liberation. You can pick the chip that fits your workload instead of being forced to buy the most expensive one. You wouldn’t use an F1 car for food delivery. You shouldn’t need an H100 for a simple classification task. Even more importantly, diversified chip supply will spawn a new generation of infrastructure companies. CoreWeave, Lambda, Crusoe — these specialized AI cloud providers don’t need to marry NVIDIA. They can mix and match accelerators based on customer needs. That kind of flexibility is something AWS, Google Cloud, and Azure — tied to their own silicon — can’t easily offer.
You can see the shift in the $26.5 billion IPO, in Meta’s September production timeline, in the hum of 100,000 domestic cards in Zhengzhou. These aren’t dry financial figures. They’re the foundation of a new continent being built, and we’re standing right at the shoreline where the old one ends and the new one begins.
AI’s future isn’t written in model parameter tables. It’s being etched into silicon in fabrication plants, stacked into memory modules in clean rooms, and wired into clusters in data centers.
Looking back from July 2026, this might be the week history remembers — not for GPT-5.6’s benchmark scores, but for SK Hynix’s $26.5 billion and Sugon’s 100,000 cards.
You don’t see them. But they determine what kind of AI you’ll wake up to tomorrow.
AI’s Biggest Bottleneck Isn’t GPUs Anymore was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.