Semiconductor sales have entered the expected seasonal slowdown, yet the AI-driven demand for memory chips remains red hot, according to a new report from Bernstein. While July marked a traditional trough for the industry, the pricing power and demand for next-generation HBM systems and data center-grade DRAM and NAND chips tied to AI infrastructure have shown remarkable resilience. The investment bank's analysis paints a picture of an industry where the AI compute bull market is gaining fresh momentum, powered by surging memory prices and volumes.

Bernstein's report, citing SIA data, shows global semiconductor sales fell 9.5% sequentially in July, slightly weaker than the historical seasonal average decline of 8.5% but still up an impressive 131.4% year-over-year. The standout performer was memory, with sales soaring 451.7% annually. Excluding memory, the rest of the semiconductor industry grew by a still-solid 35% compared to last year. Notably, memory sales slipped 16.3% from June's record levels, but this is significantly better than the typical 26.1% sequential drop seen during July in past years, underscoring that the seasonal performance of the memory sector is far outpacing consensus expectations.

Digging into the specifics for the month, DRAM revenue surged 427.8% year-over-year with bit shipments up 47.5%, while NAND revenue also climbed 427.8%. More tellingly, the average selling price per bit for DRAM jumped 257.9% and for NAND it skyrocketed 344.1% compared to the same period last year. This data confirms that even as shipment volumes normalize quarter-to-quarter, the pricing momentum for both DRAM and NAND continues to accelerate. The combination of over 40% bit growth and even steeper ASP increases highlights the robust revenue elasticity for the three major memory makers and NAND giants.

On a deeper level, memory is now the primary driver of semiconductor revenue growth. The report notes that global semiconductor sales for the first seven months of this year reached approximately $861 billion, up from $408 billion in the same period last year, a year-over-year surge of about 111%. Memory alone contributed roughly $355 billion of that new sales. Of that, Bernstein attributes about $306 billion—or nearly 68% of the entire industry's incremental sales—to memory price increases and shifts in product mix.

The demand picture is being further bolstered by recent AI developments. OpenAI's newly launched Astra model is expanding the scope of professional tasks AI can handle, while NVIDIA CEO Jensen Huang took to social media on Sunday to declare that the arrival of GPT-6 Astra signifies that "AGI is here." With NVIDIA's confirmed strong revenue guidance and robust shipment outlook—alongside AI research moving into phases of "recursive self-improvement" (RSI)—the investment cycle for AI compute is being extended. Astra's potential to widen commercial AI applications and the shift toward AI "building AI" through RSI are expected to increase demand for frontier model experiments, evaluations, and sustained training投入, adding new demand sources to the historic memory-driven semiconductor upcycle.

The investment significance of Astra and RSI lies in the fact that cutting-edge AI models and AI research itself are becoming new, continuous consumers of compute resources. Astra's release has strengthened market expectations for AGI progress, with observable gains in completing more complex workflows. OpenAI's Legora case study shows that Astra reviewed 41 documents in a single agent run, and the benchmark performance for that financial workflow improved by nearly 40% over previous generations. Meanwhile, OpenAI disclosed on September 6 that it has reached a new operational stage of "automated research interns." By mid-August, for every human research day invested, the team was using approximately 3.1 agent workdays. This new paradigm means that "AI researching AI" (the RSI training model) is itself becoming a major consumer of inference, training, and evaluation resources, adding a powerful new demand curve beyond external commercial applications.

Global capital markets are already signaling renewed enthusiasm for memory and semiconductors, particularly in South Korea and the U.S. tech sector. On September 7, Samsung Electronics rose 5.68% and SK Hynix climbed 8.26%, while the KOSPI index—often seen as a bellwether for AI compute—jumped 4.61% to 6,995.39 points. This marks a cumulative rebound of about 25.06% from the July 30 close of 5,593.56, surpassing the traditional threshold for a technical bull market.

Astra represents a mechanism for expanding cutting-edge performance demand: as model capabilities improve, tasks that were previously too difficult to perform reliably enter the commercially viable range. Additionally, Astra could shift the entire demand curve outward—when AI models become smarter, companies can attempt work they couldn't reliably do before, and competitors must also continue investing in R&D and training. This provides strong new support for the AI spending cycle. OpenAI's GPT-6 Astra and the RSI technical path that AI leaders are focusing on are likely to be two core drivers of exponential AI compute demand expansion. OpenAI has disclosed that Astra scored 98% on the FrontierMath Level 4 test and 99.9% on ARC-AGI-3. Huang has used these results to declare that "AGI is here," and noted that model training used over 100,000 NVIDIA GPUs, with another 400,000 GPUs set to come online soon. While the "AGI is here" assessment remains debatable, the announcement of larger AI GPU cluster deployments directly reinforces expectations for strong AI compute demand as frontier models continue to expand training resource investment.

From a technical standpoint, memory demand is benefiting from changes in how models are run. Training requires saving model weights, activations, gradients, and optimizer states. Long-context inference and parallel agents expand the need for KV caches and working state storage. Automated research adds to the demand for experiments, evaluations, training checkpoints, and data read/write operations. These tasks consume GPU-side HBM, server DRAM, and enterprise SSDs respectively. Memory requirements depend on parameter size, context length, concurrency levels, and experimental density. HBM handles model weights, intermediate training states, and active KV caches on the GPU side; server DRAM manages data processing, runtime environments, and cache offloading; and enterprise NAND SSDs store datasets, training checkpoints, and reusable caches. When stronger models handle longer tasks and more agents run simultaneously, and when RSI research processes add parallel experiments and checkpoint saves, the demand for capacity, bandwidth, and read/write throughput expands in tandem.

NVIDIA, the "AI chip super leader," has already translated this strong memory demand into system-level server architecture. The Rubin platform configures each GPU with up to 288GB of HBM4 and up to 22TB/s memory bandwidth, and introduces a flash-based shared context storage layer to handle reusable KV caches. From this, if model capability improvements drive more concurrent tasks, longer runtimes, and denser experimentation, memory demand will expand across the entire tier. The measure of AI economics will also shift further toward the total cost per successful task, rather than just the quoted price per million tokens.

Memory chips in AI data center server clusters remain the clearest supply bottleneck in the AI compute supply chain. Market research firm TrendForce forecasts that server DRAM contract prices will cumulatively rise by about 270% in 2026, with enterprise SSD prices up about 235%. In 2027, HBM contract prices could still rise another 70%–140%, reflecting the combined effect of AI compute expansion and memory price increases. TrendForce also estimates that DRAM and NAND combined will account for 47% of major cloud service providers' capital expenditures in 2026, rising to 68% by 2027, driven by both increased purchasing volumes and higher prices.

Bernstein remains extremely bullish on the memory supercycle, with price targets that include SanDisk soaring to $3,000 and NVIDIA racing toward $400. The震撼 impact of this memory supercycle can be seen directly in the market share and growth contribution figures from the WSTS report cited by Bernstein. According to the WSTS spring forecast, the memory market will grow from $230.042 billion in 2025 to $803.941 billion in 2026, a year-over-year increase of 249.5%, and further to $1,062.085 billion in 2027, an additional 32.1% growth. Based on this, memory's share of global semiconductor sales will rise from 28.9% to 53.2%, and then to 55.5%. Its contribution to the entire industry's incremental sales in 2026 and 2027 would be approximately 80.2% and 64.1%, respectively. In other words, the preliminary scale of the memory chip category alone in 2026 will be slightly larger than the entire semiconductor market of 2025.

These WSTS figures align with Bernstein's bullish logic on memory: AI compute demand expansion provides the volume foundation, limited supply strengthens pricing power, and product upgrades and long-term agreements improve profit structures. AMD, NVIDIA's strongest GPU competitor, reaffirmed at Citi's tech conference on September 8 that the AI data center accelerated computing market has expanded to $2 trillion by 2030. AMD also pointed out that AI inference demand has become the fastest-growing source of AI compute resource demand, with AI agents focused on agentic workflows driving both GPUs and server CPUs. AMD stated at the conference that the future procurement demand forecasts from its three Helios core customers—Meta and two other AI labs—are all higher than initially expected when the strategic partnerships were first established. The company expects server CPU business to grow over 80% year-over-year in the second half of this year and over 70% next year. This latest forecast combination strongly supports continued compute demand expansion, though it is concentrated in significantly upgraded customer demand expectations rather than fully committed non-cancellable infrastructure orders.

Upcoming long-term compute procurements from Anthropic are further increasing the visibility of future AI infrastructure demand around memory. According to media reports, Anthropic has signed approximately $35 billion in cloud computing agreements with Lambda and a roughly $45 billion, six-year compute rental deal with Nscale, totaling around $80 billion across the two agreements. These are multi-year contracts, but their collective direction is clear: frontier labs are locking in future training and inference infrastructure well in advance. The July WSTS data provides evidence of the volume and price growth that has already occurred, while the capacity procurements and model progress in August and September strengthen the case for continued demand.

In this report, Bernstein maintains "Outperform" ratings for Samsung Electronics, SK Hynix, Micron, and SanDisk, which have been on a rally this year, with price targets of 440,000 KRW, 3.3 million KRW, $1,300, and $3,000 respectively, reflecting positive views on the memory cycle. Bernstein notes that based on the latest surveyed volume and price data and downstream procurement, memory leaders benefit from bit growth, high-value product upgrades, and pricing power collectively supporting profitability. However, memory price increases will also raise material costs for GPU and server makers, meaning industry profits will not grow evenly. Bernstein adds that general DRAM price increases have widened the wafer profitability gap with HBM, pushing next year's HBM contract price renegotiations, leaving room for upward earnings estimate revisions. Nevertheless, Bernstein cautions that the most differentiating metrics going forward are memory makers' actual selling prices, shipment volumes, and free cash flow, as well as whether downstream customers can maintain strong returns from AI compute deployments amid higher hardware and financing costs.

Beyond memory giants, Bernstein's favored semiconductor names also cover AI chip leaders, wafer foundry and semiconductor equipment makers, advanced packaging, and high-end semiconductor testing chains. According to Bernstein's latest price targets, the upside for NVIDIA, SK Hynix, SanDisk, and Samsung Electronics corresponds to approximately 77.20%, 77.80%, 72.61%, and 62.66%, respectively. Bernstein's target price of $400 for NVIDIA, the world's most valuable company, is among the most optimistic on Wall Street.