Samsung Ships First HBM4E Samples at 3.6 TB/s
Key insights
- Samsung's HBM4E chips deliver 16 Gbps per pin and 3.6 TB/s bandwidth, both industry-first figures for high-bandwidth memory.
- Nvidia Rubin and Google Ironwood TPU are Samsung's primary HBM4E customers, linking its production timeline to major next-gen AI platforms.
- Samsung trails SK Hynix in advanced HBM, and this sample shipment is its most direct move yet to close that gap.
Why this matters
HBM4E's 3.6 TB/s bandwidth directly raises the memory ceiling for Nvidia Rubin and Google Ironwood platforms, meaning training and inference throughput benchmarks for the next hardware generation will exceed what current projections assume. Samsung entering the HBM4E market creates a second credible supplier alongside SK Hynix, reducing single-vendor risk for AI hardware buyers and shifting contract pricing leverage heading into 2027 procurement cycles. The Samsung versus SK Hynix competition at the HBM layer is increasingly the rate-limiting variable for AI scaling plans, because Nvidia and Google compute roadmaps now depend on memory vendors executing on custom partner timelines rather than standard product cycles.
Summary
Samsung shipped the first HBM4E memory samples on May 29, opening a lead in the race for next-generation AI memory supply.
The 12-layer chips deliver 16 Gbps per pin, over 20% faster than HBM4, with 3.6 TB/s bandwidth and 48 GB capacity. Nvidia Rubin and Google Ironwood TPU are the named launch customers, tying Samsung's execution schedule directly to those platform timelines.
Essentially: (Samsung, SK Hynix) are competing for the HBM supply position that sets the ceiling on AI compute throughput.
- 16 Gbps per pin and 3.6 TB/s bandwidth are industry firsts for high-bandwidth memory.
- 32 GB and 64 GB configurations are planned, with mass production timing aligned to individual partner schedules.
- Samsung shares climbed 6.5% on the announcement.
Whether Samsung can hold this lead before Nvidia and Google finalize their next-gen supply chains is what the market just priced in.
Potential risks and opportunities
Risks
- If SK Hynix ships competitive HBM4E samples within 60 to 90 days, Samsung's stock premium from this announcement could reverse quickly as the supply differentiation collapses.
- Nvidia Rubin and Google Ironwood qualification timelines may slip if HBM4E yield rates at Samsung fall below expectations, delaying both platforms and pushing major AI cluster procurement into late 2027.
- Samsung's mass production timing is partner-dependent, leaving the company exposed if Nvidia or Google shift volume allocation to SK Hynix before Samsung achieves manufacturing scale.
Opportunities
- AI data center operators (Microsoft, Amazon, Meta) building next-gen clusters now have a second HBM4E supplier to leverage against SK Hynix in procurement negotiations, potentially compressing per-unit costs.
- Samsung can bundle HBM4E supply with its advanced packaging services to strengthen its pitch to Nvidia and Google as a vertically integrated hardware partner, differentiating from SK Hynix on logistics and co-design.
- Smaller AI accelerator startups (Cerebras, Groq, Tenstorrent) planning next-gen chips around HBM4E bandwidth can now qualify Samsung as a supply alternative, reducing dependence on SK Hynix allocation priority.
What we don't know yet
- Whether Nvidia and Google have signed binding HBM4E supply agreements with Samsung or are still in sample qualification as of May 29 has not been disclosed.
- SK Hynix's HBM4E development and sample timeline has not been publicly announced, leaving the size of Samsung's competitive window unclear.
- No confirmed volume production start date for Samsung's HBM4E has been published; the announcement covers sample shipments only, not a ramp schedule.
Originally reported by cnbc.com
Read the original article →Original headline: Samsung Ships Industry-First HBM4E Samples — 16 Gbps, 3.6 TB/s Bandwidth, Shares Surge 6.5%