Lead Analysis
Nvidia Shifts Soaring Memory Costs to Customers as AI Server Prices Jump Over 15%
Nvidia has told some of its largest customers that prices for servers built around its AI chips will rise by more than 15% in many cases, Bloomberg reported Saturday, via Fortune. The increases hit systems built on the current Grace Blackwell platform and the incoming Vera Rubin platform shipping early next year, and contract manufacturers building racks for Microsoft, Google and Oracle have already passed the word to their own customers. The driver is memory, not silicon: DRAM contract prices rose 58%-63% quarter-over-quarter in the second quarter after a 90%-95% surge in the first, per Tom's Hardware, citing TrendForce, and Nvidia's Rubin GPU ships with up to 288GB of HBM4 per package, putting more than 20TB of HBM in a single NVL72 rack. Nvidia closed Friday at $214.72, down 0.98%.
"We see overall demand signals as strong as ever." — Timothy Arcuri, Analyst, UBS, via TipRanksArcuri's read — issued alongside a price-target raise to $205 ahead of Wednesday's earnings — is the bet underlying the whole trade: that Nvidia's roughly 75% non-GAAP gross margin gives it room to pass memory inflation straight through to hyperscalers without denting orders. Monday's premarket action tested that thesis from the supply side instead. Samsung shares fell 9% in Seoul after its 90-110 trillion won ($67-$80 billion) shareholder-return plan underwhelmed investors, and SK Hynix and Micron each slipped roughly 3% in early U.S. trading, a reminder that the same memory scarcity forcing Nvidia to raise prices is also making the stocks that supply it unusually volatile heading into a week when both Nvidia and Marvell report.
Markets
Chips Steady Ahead Of Earnings
★ highlighted = today's news-driven movers.
Companies & Financials
Nvidia Chases Perplexity as Marvell Heads Into Earnings With Google Deal Questions
- Nvidia is in talks to invest in AI search startup Perplexity as part of a funding round that would value the company at more than $30 billion, per Reuters, citing The Information. Why it matters: the talks follow Nvidia's $6 billion Poolside licensing deal disclosed last week, extending a pattern of Nvidia funding the AI application layer even as its own hardware margins face pressure from rising memory costs ahead of Wednesday's earnings.
- Marvell heads into Thursday's fiscal second-quarter report with Wall Street's attention fixed on its $120 billion Google custom-chip agreement, and Wells Fargo's Aaron Rakers raised his price target to $310 from $240 on Monday, per CNBC. Why it matters: Marvell shares fell 5.57% Friday on profit-taking even after the Google warrant news, leaving the stock's reaction to Thursday's revenue-timeline disclosures as one of the week's clearest tests of how much of the custom-silicon story is already priced in.
Memory Desk
Samsung's Payout Disappoints as YMTC's IPO Adds a New Supply Wildcard
- Samsung Electronics shares fell 9% in Seoul trading Monday after its 90-110 trillion won ($67 billion to $80 billion) 2026 shareholder-return plan fell short of expectations, dragging SK Hynix and Micron down roughly 3% each in early U.S. trading, per TechStock². Why it matters: JPMorgan flagged the absence of an instant buyback-and-cancellation, the structure SK Hynix used last week, as the source of the letdown — a sign the memory trade is now priced for near-flawless capital-return execution, not just record margins.
- The parent of Yangtze Memory Technologies is pressing ahead with a plan to raise 33 billion yuan ($4.9 billion) in a Shanghai listing to fund NAND capacity upgrades and research, per TechStock². Why it matters: YMTC already ranks third globally in NAND shipments behind Samsung and SK Hynix, and fresh IPO capital aimed at capacity would be the first concrete supply-side check on a NAND market that has priced in scarcity lasting into 2027 and beyond.
Supply Chain, Equipment & Fabs
TSMC Wins Xiaomi's New Chips as HBM Supply Stays Locked Up Through 2027
- Xiaomi unveiled a new Xring O3 smartphone processor and contracted TSMC to manufacture it on the 3-nanometer node, alongside two more Xring chips — a 6-nanometer neural processing unit and a 3-nanometer autonomous-driving chip, per Reuters. Why it matters: a Chinese fabless customer routing three separate leading-edge designs through TSMC, rather than a domestic or trailing-edge alternative, underscores how little AI-driven capacity tightness has dented TSMC's pull with non-AI customers.
- TrendForce projects server DRAM contract prices will climb another 13%-18% quarter-over-quarter in the third quarter, and Nvidia is weighing 8-stack HBM4E and HBM4 options for its Rubin Ultra chip alongside the originally planned 12-stack configuration, per The Herald Business. Why it matters: designing flexibility into next-generation packaging is a tacit admission from Nvidia that the memory shortage now driving its own price hikes could extend through 2027, forcing product-design compromises rather than a simple wait-it-out approach.
Policy & Geopolitics
Taiwan Charges Nine in Nvidia-Linked AI Server Smuggling Probe
- Prosecutors in Keelung, Taiwan indicted nine people, including employees of Nvidia and Super Micro, on charges of illegally exporting AI servers equipped with export-controlled Nvidia chips to China and other destinations, per the Taipei Times, citing Reuters. Why it matters: prosecutors said the defendants "colluded with one another at various levels for enormous profit," a criminal finding that raises compliance-cost and reputational stakes for both companies just as Nvidia heads into its most closely watched earnings report of the year.
- Prosecutors said the scheme "increased corporate compliance costs and severely damaged" Taiwan's international standing, and neither Nvidia nor Super Micro immediately responded to requests for comment, per the Taipei Times. Why it matters: the case lands as Taiwan tightens export enforcement more broadly, adding legal and diplomatic friction to the island's chip-export relationship with Washington at a moment when both governments are already negotiating over where AI chip supply chains are allowed to run.
Analyst Corner
SemiAnalysis Tests the CUDA Moat as Wells Fargo Lifts Marvell's Target
- SemiAnalysis (published August 24): Cam Quilici, Bryan Shan and colleagues released AgentX 1.0, a $3 million, fully open-sourced agentic-coding inference benchmark spanning Nvidia's GB300 NVL72 against AMD's MI355X and B200 systems, finding both vendors post strong results on real-world multi-turn workloads — a more contested read on Nvidia's software moat than the narrative heading into Wednesday's earnings.
- Wells Fargo's Aaron Rakers (via CNBC, August 24): Rakers raised his Marvell price target to $310 from $240, arguing the stock — already up nearly threefold in 2026 — has further to run as its custom-silicon backlog builds toward Thursday's earnings report.
Calendar / Week Ahead
Nvidia and Marvell Earnings Anchor a Pivotal Week for AI Chips
- Wednesday, Aug. 26 — Nvidia reports fiscal second-quarter results after the closing bell, its first print since the weekend price-hike report and Monday's Taiwan indictment, per Investing.com.
- Thursday, Aug. 27 — Marvell reports fiscal second-quarter results, the first look at how its $120 billion Google agreement and Wells Fargo's raised price target hold up against actual guidance, per TipRanks.
|
