The Compute 100 — Monday, July 13, 2026

SK Hynix’s CEO told Reuters on Friday, the day of the company’s record $26.5B Nasdaq debut, that 2027 will be “the worst year in the industry’s history” for memory supply, with demand outstripping capacity “even beyond 2030.” Three trading days later, on Monday, SK Hynix shares fell more than 15% in Seoul — the stock’s largest one-day drop on record — dragging Samsung down nearly 11%, and pulling Micron, Sandisk, and Western Digital down roughly 5% in the U.S. alongside ASML, ASMI, and Besi in Europe. The reversal reflects profit-taking after a run that tripled SK Hynix’s shares this year, but the speed and breadth of the selloff is the clearest signal yet that investors are no longer taking the memory supercycle’s staying power for granted, even as the company’s own leadership doubles down on scarcity lasting past 2030.

Private Companies

At the RAISE Summit in Paris on July 10, Vast Data co-founder Jeff Denworth disclosed a $1.17B commercial agreement with Coreweave and said the company has built a single, exabyte-scale AI storage cluster supporting 300,000 GPUs, per SiliconANGLE. Denworth said the “biggest deals in our pipeline are actually KV cache-oriented” — evidence that as inference splits prefill and decode across specialized silicon, the cache holding intermediate attention state is becoming a shared storage problem, not GPU-memory overflow. Vast, still private north of a $30B valuation with Nvidia among its backers, just showed that pipeline is worth billions with one customer alone.

Kandou AI, the EPFL spinout building copper-based chip-to-chip interconnect for AI systems, opened a new engineering headquarters in Hyderabad on July 11 covering chip design, verification, and systems engineering, with Telangana IT minister KT Rama Rao attending alongside CEO Srujan Linga. The move follows Kandou’s $225M Series A for its “Chord signaling” copper interconnect, which the company claims runs at roughly a tenth the power of optical alternatives for breaking AI memory bottlenecks. It is a bet that India’s chip-design talent pool, already anchoring large captive centers for Qualcomm and the EDA majors, can scale a well-funded interconnect challenger’s headcount faster than Europe or the U.S. can.

Public Markets

Samsung fell 10.7% to ₩254,500 in Seoul Monday, its steepest drop in years, as foreign investors pulled roughly $1.1B from KOSPI-listed shares — erasing about $150B of Samsung’s market value and confirming this is a Korean-memory-sector unwind, not a SK Hynix-specific one.

Emerging

A team of researchers published “Who Needs DRAM? We Have Fiber” on arXiv July 10, proposing to replace HBM entirely for storing immutable LLM weights with recirculating, space-division-multiplexed optical fiber paired with co-packaged optics — claiming it eliminates redundant weight storage across 10,000-accelerator clusters and cuts weight-delivery energy more than 70% versus HBM3e. It is a research-stage idea, not a product, but it lands the same week memory scarcity became the sector’s top story, and it is a reminder that the HBM bottleneck is drawing serious architectural workarounds, not just capacity expansion. Watch whether a funded interconnect or photonics startup picks up the approach before a hyperscaler’s own silicon team does.

DigiTimes’s weekly chip roundup, published July 13, reports Samsung’s HBM4 base-die lines have saturated at 4nm, setting up 15–20% price hikes in the back half of 2026, while SK Hynix’s new Nasdaq capital “won’t deliver significant new HBM capacity before 2028.” Separately, TSMC signed a 10-year advanced-packaging supply pact with Amkor, and ASE raised its 2026 capex target from $7B to $8.5B — both bets that packaging capacity, not wafer starts, is the durable constraint on AI silicon supply.

Watch This Week