Nvidia's Earnings Test Arrives With Memory Costs Reshaping the AI Pricing Equation
Published on 08/23/2026 at 20:10 | Redaktion boerse-global.de
The semiconductor sector's most anticipated report of the summer lands Wednesday with Nvidia's fiscal second-quarter results, and the numbers will arrive freighted with a complication that barely existed a year ago: the chipmaker's own input costs are climbing faster than its customers' willingness to pay.
Nvidia closed Friday at €183.78, capping six consecutive losing sessions — its longest such streak since 2022. The weekly decline of 5.6 percent leaves the stock hovering just above its 50-day moving average of €180.82, while the 52-week high of €202.50 sits roughly 9 percent out of reach. The slide comes as reports circulate that Nvidia is raising prices on AI server systems by more than 15 percent for certain configurations starting next year, with the Vera Rubin and Grace Blackwell generations carrying the biggest increases depending on chip version and memory configuration.
The price hikes trace directly to the memory market. Server builders serving Microsoft, Google and Oracle have already passed word down the supply chain, because the cost structure hinges on how much DRAM and HBM goes into each system — and that supply is dominated by Samsung, SK Hynix and Micron. SK Hynix alone commands roughly 57 percent of the HBM revenue share, with South Korean manufacturers collectively accounting for about four-fifths of global HBM capacity.
That dynamic has flipped a long-standing assumption in the AI trade. For years, the logic ran that expensive AI chips meant insatiable demand for Nvidia hardware. Now OpenAI is cutting prices, Chinese providers are undercutting with cheaper models, and the unit economics are getting harder to square. Nvidia's valuation reflects the shift: the stock trades at a price-to-earnings multiple of roughly 33, well below its five-year median of 57. Analysts still expect the coming quarter to deliver around $91 billion in revenue, a 95 percent jump year over year — but the market's question is whether margins can hold as memory costs bite.
Should investors sell immediately? Or is it worth buying Nvidia?
The Memory Makers' Moment
The pricing pressure Nvidia faces is, for SK Hynix, a tailwind of rare strength. The South Korean memory giant capped a volatile week with a 15 percent surge Friday to close at 1,730,000 won, extending its year-to-date gain to 166 percent. The rally followed an extensive capital return program, and the company is now weighing a new memory fab in Japan's Miyagi prefecture — a move that would make it the third foreign company, after Micron and TSMC, to operate semiconductor manufacturing in Japan.
The company is officially noncommittal, citing unresolved questions around US pressure for domestic investment and South Korean public opinion. But analysts see the direction clearly: 13 rate the stock a buy, none recommend selling, and the price target for its US listing sits at $245.43.
The memory shortage that benefits SK Hynix is also reshaping competitive dynamics elsewhere. Nvidia's price increases on AI servers are, in effect, a pass-through of memory scarcity — a structural advantage for HBM leaders that extends beyond any single quarter.
Custom Silicon Widens the Field
While Nvidia still commands the largest single share of data center budgets, the custom-chip lane is getting more crowded. Marvell Technology provided this week's most dramatic illustration. Google signed a deal in late July covering AI inference chips, memory controllers and other semiconductors built around its TPU processors. As part of the agreement, Google received warrants on nearly 59 million Marvell shares — roughly 7 percent of the company — at a fixed price of $206.58.
The initial reaction was euphoric: Marvell jumped about 10 percent Wednesday to close near $237. The enthusiasm cooled quickly once investors ran the dilution math. By Friday the stock had fallen 5.6 percent to €203.00, because the warrant strike sits below the current price — meaning the dilution grows larger the more successful the partnership becomes.
The operational picture remains strong regardless. Marvell posted fiscal first-quarter revenue of $2.418 billion, up 27.6 percent year over year, with the data center business contributing $1.83 billion — roughly three-quarters of the total. Management guided to $2.7 billion for the current quarter and raised forecasts for the next two fiscal years. CEO Matt Murphy cited "extraordinary AI-related order momentum." The Google warrants vest in 240 tranches, each tied to $500 million in revenue Marvell generates with Google through January 2033. The average analyst price target of $288.12 implies upside of about 21 percent from Friday's close.
AMD is pressing its own case in the custom and standard silicon markets. The company's share of x86 client CPUs reached 30.3 percent in the second quarter, up from 23.9 percent a year earlier, while Intel slipped to 69.7 percent. Desktop share climbed from 33.2 to 34.9 percent; notebooks rose from 28.3 to 28.9 percent, even as Intel expanded its own production output. In servers, AMD's EPYC share is now closing in on Intel's Xeon — a shift with direct implications for data center procurement.
AMD's stock closed at €405.30, down 8.7 percent on the week and 16 percent over the past month, though still up 120 percent year to date. The analyst consensus from S&P Global's survey of 53 experts is "Strong Buy," with an average target of $613.09 — roughly 30 percent above current levels. Argus recently lifted its target from $450 to $625.
The Capital Intensity Question
The broader AI complex is increasingly defined by balance-sheet strain rather than product announcements alone. Broadcom is negotiating a financing package of up to $100 billion to secure additional compute capacity for customers like Anthropic, with much of it structured as secured and subordinated debt. Blackstone and Apollo, which established a financing platform with Broadcom in June, are expected to participate again. The initiative aims to fund more than 20 gigawatts of compute capacity. CEO Hock Tan projects AI chip revenue will exceed $100 billion next year, backed by existing OpenAI orders and a separate multibillion-dollar Apple deal.
Nvidia at a turning point? This analysis reveals what investors need to know now.
The market has yet to reward the ambition. Broadcom closed Friday at €315.45, down 7.0 percent on the week and 9.5 percent over 30 days, leaving the stock 27 percent below its 52-week high of €429.60. The RSI of 36.8 points to oversold conditions.
Oracle presents a similar paradox. Its remaining performance obligations surged 363 percent to $638 billion, fueled by $67 billion in AI infrastructure contracts in a single quarter. Yet the company is planning layoffs this month, some affecting double-digit percentages of individual teams, as it manages the capital demands of a buildout that saw capex reach $55.7 billion last fiscal year — more than double the prior year's figure. The stock closed Friday at €125.86, up 3.5 percent on the day and 14 percent over 30 days, but still down 24 percent year to date and 57 percent below its 52-week high of €294.85.
Divergent Paths in the Same Sector
The contrast within the AI trade could hardly be sharper. Palantir delivered a record quarter with revenue up 93 percent to $1.935 billion and adjusted earnings per share of 41 cents against expectations of 35 cents, adding more than $115 billion in market capitalization since the report. The stock rose 3.4 percent Friday to €154.08, extending a 30-day gain of 41 percent. An RSI of 68.1 signals overbought conditions — which helps explain why Cathie Wood's Ark Invest has been trimming, selling roughly $21 million across several sessions while still maintaining a position that ranks among Ark's largest holdings at 3.5 percent of the portfolio.
Tesla, meanwhile, spent the week managing the largest recall in its history in China — nearly three million vehicles over door handles that could impede emergency exit in severe crashes, plus a separate action covering driver monitoring systems in certain locally built Model 3 and Model Y units from 2019 to 2025. The recall is the biggest single contributor to an industry-wide wave affecting nine automakers and more than four million vehicles. The stock nonetheless rose 5.0 percent Friday to €310.30 after Nevada approved driverless robotaxi operations in Las Vegas, leaving the shares up 4.9 percent on the week but still down 21 percent year to date.
What Wednesday Will Tell Us
Nvidia's report will be parsed not just for revenue growth but for management's commentary on rising memory and server costs — whether they represent a margin squeeze or simply confirmation that AI demand remains robust enough to absorb price increases. The stakes extend beyond Nvidia itself: SK Hynix's Japan decision, if it comes, would redraw the memory map of Northeast Asia and feed back into pricing for every AI server built. AMD must demonstrate whether its share gains can continue into the second half, while Marvell balances dilution from the Google deal against long-term revenue certainty in an increasingly competitive custom-chip market against Broadcom. For Nvidia specifically, the question is whether pricing power can survive a moment when its own suppliers hold the leverage.
Ad
Nvidia Stock: New Analysis - 23 August
Fresh Nvidia information released. What's the impact for investors? Our latest independent report examines recent figures and market trends.
