AI-memory stocks are getting hit hard even though demand, pricing, and profits are still extreme: SK Hynix is sold out, CXMT is scaling fast, and GPU/compute shortages are feeding a bigger capex cycle. The crash is framed less as a bubble burst than as forced selling, missed targets, and leverage liquidation colliding with a supply chain where HBM is the bottleneck. The bigger call is that the AI trade is unbundling from hyperscalers into semis and memory, while China’s memory push and ASML-style lithography moves are being misread by the market. The conclusion stays bullish on long-run tokens, compute, and energy, but says current valuations in parts of AI are still absurd.
US market losses were described as exceeding the 2008 financial crisis over just a couple of days, while the Korean stock market fell 32% and the Korean index was said to be down over 30%, the largest drawdown ever.
SK HYNIX is down almost 25% in the last month even after a quarter where revenue rose 354% and operating profit almost 600%; the quarter allegedly produced more profit than revenue and more money than in all of 2025 last year.
SK HYNIX missed revenue by about $1.7 billion despite being the number one HBM provider and dedicating fab capacity only to HBM, with supply so tight it could not sell more for the quarter.
HBM was framed as pricier and wafer-intensive than DRAM, so SK HYNIX can earn more than Samsung or MU even when Samsung and MU dump less because they also sell cheaper DRAM.
Three forces were blamed for the memory selloff: analysts punishing revenue misses despite huge growth, upside capped because SK HYNIX is fully sold out into 2026 and pre-sold, and a 2x leverage ETF launched July 13 creating cascading liquidation.
JPMorgan was cited as saying the leverage-ETF drawdown/liquidation is about 90% complete, which was used to argue a bottom may come sometime soon.
SanDisk Corporation was said to be down 50%, and a darling meme-name stock was said to have been up about 4,000% at one point and still up 2,200% on the year even after a 50% drop.
CXMT, ticker CXMT, was described as the world’s number four DRAM maker, up from 1% to nearly 10% share over four years, with its IPO day trading up 466% and raising about $9 billion.
China was described as starved of memory because Western demand is overwhelming, Nvidia export restrictions block frontier chips, and SK HYNIX and Samsung memory are going to the West instead.
Moonshot’s Kimi K3 and Jipu’s GLM were tied to CXMT’s ramp, with the claim that Chinese AI labs need local memory as they train models.
ASML’s EUV machines were described as costing $300 million each and requiring a couple hundred a year, while China’s announced DUV replication was framed as only a prototype aiming for 10 machines, not a scaled flood.
The market was said to be overreacting by treating China’s DUV effort as if it would flood the market with EUV-class tools and force lower GPU costs, dumping NVDA, AMD, and memory stocks.
The AI trade was said to be unbundling from hyperscalers into semis because AI capex is near-immediate revenue for chip makers, while Google was said to have posted negative cash flow for the first time since its IPO 21 years ago.
AI capex was estimated at roughly $250 billion for the year, and the downstream effect was said to be huge because companies collecting money are now spending faster than they earned it.
GPU rent contracts at cloud providers were said to expire at year-end and then re-rate 2x, while spot GPU rental prices were already 2x contracted rates at the start of contracts.
Nvidia chips were cited at $40,000 with 84% gross margin, and nearly 50% of GPU costs were said to be HBM, making memory the key shortage if production expands.
Lisa Su was cited saying AI adoption is faster than expected with no end in sight for demand for tokens and compute, while agentic AI was said to have exploded over the last couple of months and to require CPUs and memory.
Open-source model inference was said not to be cheap: ‘Kimikate 3’ was estimated at $1.1 million to $1.2 million to run effectively at scale.
The closing thesis was that current AI valuations are completely unjustified, but the long-run direction still points to more tokens, compute, and energy, with the only uncertainty being whether the demand wave arrives in 6 months, 6 years, or 60 years.
Sponsored: Ledger’s Agent Stack pitched four open-source tools for Ledger Wallet, Ledger Enterprise, and Ledger Enterprise multisig so agents can rebalance crypto portfolios while humans approve and Ledger signers enforce control.
Ledger’s sponsor example was: ‘Rebalance my wallet. ETH to 40%. Do it in the fewest swaps,’ with the claim that nothing moves until approval.