OpenAI unveils GPT-6 Astra across two-day rollout
OpenAI released its newest model, ChatGPT-6 Astra, in two parts on Thursday and Friday of the prior week, with president Greg Brockman describing it as "the world's most intelligent and aligned model." The launch drew an enthusiastic reception that investors interpreted as a signal of continued demand for advanced memory chips used to run frontier AI workloads, including high-bandwidth memory, DRAM and GPUs.
Memory-chip stocks rally from U.S. to Asia
The model's rollout triggered sharp moves across the memory complex on Friday and into Monday's Asian session. On Friday, Micron Technology climbed about 6%, Sandisk jumped roughly 11.9%, the Nasdaq-listed depositary receipts of SK Hynix rose about 8.1%, and the iShares Semiconductor Index ETF (SOXX) added 3.5%, lifting its year-to-date return to 72% despite a roughly 30% correction in July. In Monday's Asian trading, SK Hynix ordinary shares spiked about 8% and Samsung Electronics more than 5%, while Japanese peer Kioxia posted an even larger gain.
Sell-side sees compute- and memory-intensive AI demand persisting
Nomura's CW Chung reiterated bullish price targets of KRW 670,000 for Samsung and KRW 4,700,000 for SK Hynix, citing meaningful upside from current levels near KRW 270,000 and KRW 1,783,000. A Saxo Bank note from chief investment strategist Charu Channa said the Astra launch "supports continued AI spending" and argued that frontier AI models remain extremely compute-intensive, with memory and network bottlenecks broadening — conditions Saxo described as highly favorable to SK Hynix and Samsung. Analysts at Goldman Sachs and Morgan Stanley estimated that of the expected $1.3–1.5 trillion of AI capex planned for 2027, more than half would be devoted to memory.
Structural HBM shortage tied to AI scaling, not to a single launch
Reporting framed the rally as a spark rather than new information. AI accelerators feed compute through high-bandwidth memory stacks wired into cubes beside GPUs, and memory bandwidth has become the binding constraint on inference speed, a "memory wall" that tightens as models scale. Analysts cited HBM demand growth of roughly 40% annually through 2030, about twice the rate of ordinary DRAM, with production effectively sold out for 2026. OpenAI has signed letters of intent with Samsung and SK Hynix for memory supply for its data centers, and its in-house "Jalapeño" inference chip built with Broadcom requires HBM like any other accelerator.
DRAM leg carries cyclical risk alongside the structural HBM story
The same coverage warned that only the HBM leg of the memory trade is structural; producing a bit of HBM consumes roughly three bits of conventional DRAM, and the resulting surge in DRAM prices raises the risk of demand destruction as end users push back on costs. Memory stocks were described as trading at peak multiples built on temporary earnings, exposing investors to historical cycles of overinvestment and collapse. The coverage framed the central risk as whether supply expansions planned for 2027–2030 will outpace demand growth before pricing erodes the demand the build-out is meant to serve.
Next milestone to watch
Investors are looking to upcoming memory-sector earnings and capacity guidance from Samsung Electronics, SK Hynix, Micron Technology and Kioxia for confirmation that HBM allocations remain sold out through 2026 and that 2027 capex plans keep the memory share of AI infrastructure spending above half, per the Goldman Sachs and Morgan Stanley framework cited above.
Share this article
