The Halving Confession — The Week Nvidia Cut Its Own Vera Rubin Memory in Half, Committed $500B More Demand, and CXMT IPO'd the Answer
The world's biggest AI buyer admitted it can't swallow HBM pricing, spec-cut its own rack, then piled another half-trillion into demand-side deals in the same 72 hours
Three faces in one week
In the last week of July, Nvidia did three things at once. (1) It halved the CPU memory capacity on its next-gen Vera Rubin NVL72 rack. (2) It announced a $500B, 2GW data-center partnership with Korea's SK Group. (3) It layered another $10B with Brookfield into Naver's Sejong 200MW expansion, added new equity into Ilya Sutskever's SSI, and locked in Vera Rubin supply for the venture. These are not sequential. All of it happened inside the same 72 hours.
The pattern doesn't fit any single headline. The world's largest AI buyer wrote a surrender note on its own bill of materials in the same week it committed another half-trillion dollars of demand. Two days later, China's CXMT priced an $8.6B IPO — the supply-side answer.
The cost confession: NVL72 memory got cut in half
At least seven outlets confirmed that Nvidia is halving the memory spec on Vera Rubin NVL72 (Wccftech, Tom's Hardware). The reason is simple — HBM pricing has entered the un-swallowable zone. With DDR5 16Gb spot at $50.83 as of today, memory has quietly become the line item that eats rack margin.
This is not a spec-tune story. It is a discarded standard. Nvidia has cut in half the context length, batch size and KV cache its own flagship platform can hold, and dressed it up as "design optimization." The honest read is: even Nvidia can no longer pay full sticker for HBM.
On the other side of that trade, SK hynix pre-announced record profits across DRAM, NAND and HBM simultaneously (story); brokerages lifted both Samsung and hynix targets in unison. Nvidia's cost confession = the memory oligopoly's excess-profit signal.
The demand side: another $500B booked in the same week
Even as the cost line retreated, the demand pipeline thickened:
- $500B, 2GW SK Group partnership targeting 2027 (DCD)
- $10B Brookfield-Nvidia co-investment to bring Naver Sejong to 200MW (DCD)
- Additional equity in SSI plus a Vera Rubin supply agreement
- Vera CPU adopted internally by Cadence and Synopsys for EDA workloads, cutting verification time 1.5x
- A separate $1B Verizon-Google fiber deal wiring hyperscaler backbones
The week's demand commitments emanating from Nvidia now total roughly $750B by analyst count, and that number is what has re-ignited the circular-financing debate (WSJ). Nvidia invests; the investee buys Nvidia silicon; that revenue underwrites the next round.
The same company writing an involuntary cost note on the supply side is manufacturing its own demand collateral on the buyer side. That is the real Nvidia posture this week revealed.
CXMT IPO'd the answer
On the exact same tape, China's CXMT priced its $8.6B IPO (EE Times). The market reacted immediately — SanDisk -12%, SK Hynix -8%, Micron -5%.
Don't misread the drawdown. Hynix earnings did not deteriorate. The market stapled Nvidia's spec-cut to CXMT's fresh capital and priced a scenario: "if premium memory is that expensive, a lower-spec substitute can steal share, and that substitute just financed itself."
In parallel, China moved indigenous DUV lithography tools into mass production, chipping at ASML's monopoly (story); ASML and US equipment stocks sank in sympathy. The cost axis is now being pushed by Beijing, not by Seoul or Hsinchu.
PM view: three trades that fall out of this
- NVDA — the balance test between margin defense and demand expansion. Spec-cut is gross-margin positive in isolation. But the narrative "even Nvidia can't pay its own supplier" is a crack in the premium multiple. If the $750B demand story starts to feel manufactured rather than organic, that is the trigger.
- MU / hynix / Samsung — the premium-vs-commodity split. Nvidia's spec-cut may compress top-line HBM volume, but it strengthens the oligopoly's pricing power on what remains. CXMT still cannot touch HBM3E/4 — it can only bend the DDR5 spot floor. Risk is sentiment, not earnings.
- AVGO — the quiet winner of this frame. Alphabet capex reinforces custom-silicon TAM (story); the Samsung–Broadcom pact spans memory, foundry and packaging. Every time Nvidia confesses to its own BOM, the relative case for custom silicon strengthens.
Bottom line: which of Nvidia's three faces the market ends up paying for determines the second-half AI cycle. Price the confession and the HBM oligopoly wins. Price the demand expansion and Nvidia wins. Price the CXMT/DUV axis and every US equipment name and premium-memory name gets re-rated at once.
Key Sources: - Nvidia and SK Group announce $500bn AI agreement, includes 2GW of data center capacity (DCD, 2026-07-27) - Naver to expand Sejong data center to 200MW with $10B Brookfield, Nvidia investment (DCD, 2026-07-27) - NVIDIA halves Vera Rubin cluster memory amid HBM cost surge (Wccftech, 2026-07-27) - CXMT IPO: Where China's Largest DRAM Maker Stands (EE Times, 2026-07-27) - Nvidia's $750B AI deals spark bubble concerns amid circular financing questions (WSJ, 2026-07-27) - plus 12 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram