HBM4 becomes the bottleneck — SK hynix's volume ramp meets OpenAI's UAE-Stargate site and the new compute calendar reshapes the next training cycle
SK hynix kicked off HBM4 volume production this week the same day OpenAI confirmed the second Stargate site in the UAE. The chip-supply story and the data-center-buildout story used to run on independent calendars. They are now the same calendar — and the constraint isn't fab capacity, it's HBM yield.
Two announcements, one bottleneck
SK hynix moved HBM4 into volume production on schedule, the first vendor at that node. OpenAI confirmed the UAE as the second Stargate site with first compute online H2 2026. The UAE site is sized for Vera Rubin generation deployment — which requires HBM4 stacks at densities nobody else can ship yet.
The yield curve is the calendar
HBM3e shipped late, then ramped. HBM4 yield curves matter more because the per-package density is high enough that single-die defects waste large amounts of silicon. Whoever wins the yield race wins the per-GW-hour-of-compute cost curve through 2027. SK hynix's volume ramp this quarter says they think they've cleared the yield knee earlier than expected.
Why the UAE site is the right test
Stargate-UAE is sized at the scale where HBM supply, power supply, and geopolitical clearance all converge. UAE power is cheap, capacity is high, and the sovereign-wealth co-investment structure clears the export-control overhead on bulk silicon delivery. The first gigawatt going online H2 2026 is the calendar HBM4 has to hit.
The compute buildout calendar is now SK hynix's HBM4 yield curve. Everything downstream — model releases, agent rollouts, training cycles — flexes off that one constraint.
What this does to model release timing
Frontier-model labs have been training on H100/H200 plus interim Blackwell. Vera Rubin generation training requires HBM4 density at scale. If SK hynix slips by a quarter, every frontier model targeting H1 2027 release slips. If they hold the ramp, the H1 2027 release window stays — and the labs that pre-committed to Vera Rubin deployment hit it first.
The forward read
- HBM4 supply allocates by H2 2026 — winners and losers are visible by then. Whoever has committed contracts now ships first. Whoever didn't waits.
- Stargate-UAE first GW comes online late Q3-early Q4 2026. That becomes the public benchmark for the new compute generation.
- Samsung's HBM4 timeline becomes the second-source story by Q4. A two-vendor HBM4 supply chain is the difference between an OpenAI/Anthropic-only compute story and a true open-frontier compute story.
Reuters — SK hynix HBM4 volume production → · Bloomberg — OpenAI Stargate UAE confirmation → · SemiAnalysis — HBM4 yield curve analysis →