The third wave just hit. And this time, it's not about cheaper chips — it's about who gets to set the price of intelligence.
On July 16, 2026, a Chinese AI lab called Moonshot AI dropped Kimi K3 — a 2.8 trillion parameter open-weight model. Elon Musk called it "Impressive." The Nasdaq fell 1.47%. The Philadelphia Semiconductor Index cratered 4.29%. By the next trading day, the Nasdaq 100 futures were down nearly 2%, and the Philly Semi index had extended its retreat from June highs past 20% — bear market territory.
This wasn't just another model release. It was the third shock to a system that had built its entire valuation architecture on three assumptions:
All three are now under active demolition. And the collateral damage is spreading to a part of the supply chain most investors haven't looked at closely enough: the HBM memory triopoly.
DeepSeek R1 was the bombshell that cracked the compute narrative. Trained for roughly $6 million, it matched models that cost $100M+. On the day markets absorbed what DeepSeek meant, NVIDIA lost 17% — $593 billion in market cap vaporized in a single session. The Nasdaq fell 3.1%.
The core question DeepSeek forced: If intelligence no longer maps 1:1 with GPU count, is NVIDIA's growth curve overestimated?
The answer, in hindsight: sort of. Reasoning models ended up consuming more compute per query, not less. But the psychological damage was done. The market now treats every Chinese model release as a potential compute-demand reset.
Manus didn't crash US markets. It didn't need to. What it did was subtler — it demonstrated that Chinese AI products could compete on the agent layer, not just the model layer. The market rotated from "who has the best model" to "who owns the task-completion pipeline."
Manus showed that the next battle wouldn't be about answering questions. It would be about taking actions and delivering outcomes. The A-share AI agent concept stocks surged. The message: product design matters as much as parameter count.
This one is different. K3 is not a "cheap Chinese model." Let me say that again: K3 is not cheap.
This is the uncomfortable implication: The "cabbage-price Chinese model" narrative is being killed by Chinese models themselves.
DeepSeek made investors worry they'd bought too many GPUs. K3 makes them worry they've overestimated American AI companies' pricing power and terminal margins.
As Bernstein analyst Robin Zhu put it: K3 is a "home run" that proves Chinese frontier capability is no fluke.
Here's where we need to talk about what nobody is connecting.
NVIDIA isn't just a chip company. It's the de facto central bank of artificial intelligence. And like any central bank, it controls the monetary base — in this case, the HBM memory that every frontier GPU requires.
Consider the mechanics:
| GPU | HBM Type | Memory per GPU | HBM Stacks |
|---|---|---|---|
| H100 | HBM3 | 80 GB | 5 |
| H200 | HBM3E | 141 GB | 6 |
| B200 | HBM3E | 192 GB | 8 |
The B200 uses 140% more HBM than the H100. Multiply that by the millions of GPUs being deployed, and you have a demand tsunami that manufacturing physically cannot meet.
SK Hynix and Micron have sold out their entire 2026 HBM production. Let that sink in. All of it. Gone.
This gives NVIDIA extraordinary power: it decides which memory supplier gets qualified for which GPU, and that decision is worth billions. One qualification failure — as Samsung is experiencing — can crater a company's AI memory ambitions.
SK Hynix was first to mass-produce HBM3 and HBM3E. It secured exclusive supplier status for the H100. It extended that to the B200. Now it's co-developing HBM4 with TSMC.
The numbers:
The risk for SK Hynix isn't demand — it's concentration. Roughly 50%+ of its HBM revenue depends on NVIDIA's continued dominance. If the "Chinese models need less compute" thesis gains traction, SK Hynix's growth premium evaporates.
Micron skipped HBM3 entirely. Smart move. Instead of playing catch-up on a dying standard, it designed a power-efficient HBM3E solution and got it qualified for NVIDIA's H200.
The numbers:
At 5.6x forward earnings, Micron is being priced like a cyclical commodity stock — not a strategic AI infrastructure supplier with 21% of a market growing at triple-digit rates. That's either a colossal mispricing or a warning about what happens when the HBM supercycle peaks.
Samsung is the world's largest memory maker by total DRAM revenue (38% market share). And yet in HBM — the only DRAM category that matters for AI — it's third place and falling.
The numbers:
Samsung's HBM failure is a cautionary tale about what happens when you lose the NVIDIA qualification game. It's not enough to make great memory. You have to make memory that NVIDIA certifies. And NVIDIA's certification process is a black box that doubles as a competitive weapon.
Here's the scenario nobody's pricing in yet:
If Chinese models achieve frontier capability with export-controlled GPUs (H800-class or alternative domestic silicon), the entire demand model for advanced HBM collapses from the bottom up.
Moonshot AI's own disclosures mention "export-grade NVIDIA silicon and an unnamed alternative GPU vendor." K3 was trained on what the US government considers acceptable exports. Yet it competes with models trained on unrestricted B200 clusters.
This creates a paradox:
Either outcome is bad for the current valuation structure of the AI supply chain.
| Shock | What It Rewrote | Market Signal | Hidden HBM Impact |
|---|---|---|---|
| DeepSeek (Jan 2025) | Compute ≠ Intelligence | NVDA -17%, Nasdaq -3.1% | HBM demand assumptions questioned |
| Manus (Mid-2025) | Agents > Models | A-share AI rotation | Product layer competition intensifies |
| Kimi K3 (Jul 2026) | Chinese models have pricing power | Nasdaq -1.8%, SOX -5.2% | Premium HBM may face substitution pressure |
Each shock attacks a different assumption. Together, they form a pattern: China is not just copying. It's competing on architecture, product design, and now pricing power.
At $202.81, NVIDIA trades at 15.8x forward earnings. That's not extravagant by historical tech standards. But it embeds an assumption: that the company's 75%+ data center gross margins are sustainable.
If Chinese models continue to close the capability gap while running on less or cheaper hardware:
The HBM triopoly would feel this first. When you're sold out of 2026 production, any demand signal change goes straight to your 2027 order book.
The story of AI in 2026 isn't just about who builds the best model. It's about who controls the supply chain that makes models possible — and whether that supply chain's pricing power can survive the arrival of competitors who need less of it.
Kimi K3 didn't cause the Nasdaq selloff. But it forced the market to reprice the probability that American AI dominance is permanent. And when that probability drops, every company in the GPU-HBM supply chain gets cheaper.
The next time a Chinese lab drops a frontier model, don't just check the benchmark scores. Check Samsung's HBM qualification status, Micron's order book, and NVIDIA's gross margin guidance. That's where the real story will be written.
Published: July 19, 2026 Sources: Toutiao/钛媒体 (象先志 original analysis), SiliconAnalysts HBM Dashboard, EnkiAI HBM Supply Crisis Report, Counterpoint Research, Tom's Hardware, SiliconANGLE, The Decoder, Astute Group, Bernstein Research, Bank of America analysis
