Topic
Explore English articles about Data Centers.
Google is pulling DDR4 memory out of retired servers and reconnecting it to new-generation AI servers via a dedicated interface, turning component recycling into a supply-defense strategy amid a memory shortage.
LTO tape shipment capacity dipped in 2025 amid trade uncertainty, but Q1 2026 saw a sharp 57% rebound, fueled by AI data storage needs, power constraints, and new 40TB media.
TSMC is targeting 4,100W-class CoWoS packages and has already demonstrated uniform heat dissipation beyond 5kW in test vehicles. Here's what sealing, warpage, and rack integration reveal about the path to mass production of direct-to-silicon liquid cooling.
Cross-referencing Anthropic's $35 billion compute deal with Hut 8's disclosed 704MW facility, this piece breaks down the financing structure and division of responsibility among the four companies involved—including NVIDIA's unusual role as a data center tenant.
Samsung aims for 2x performance and a 20% cut in thermal resistance with HBM5. Meanwhile, zHBM's claimed "8x" improvement is reported against different baselines—HBM4E in some accounts, HBM5 in others—making disclosure of absolute bandwidth and test conditions the key to evaluating the roadmap.
The spot price of 36GB HBM3E has reportedly reached about $2,100 per unit, four to five times the long-term contract price. Samsung plans to allocate 60-70% of its memory production capacity to long-term contracts running through 2031, and whether a buyer is inside or outside those contracts is starting to determine the terms under which AI memory can be procured.
We break down TSMC's projected "50x" leap by 2029 into the roles played by CoWoS, SoIC, and optical interconnects, examining where effective performance ends and mass-production conditions begin. We also clarify the difference between the 50% silicon photonics market forecast and actual CPO adoption rates.
Broadcom's VMware AI Factory for AMD touts GPU sharing and usage visibility, but pricing, release dates, and measured cost reductions remain undisclosed, leaving key conditions for verifying the savings still missing.
Intel expects revenue from EMIB-T-based advanced packaging to grow starting late 2027 and scale up by 2029. Customer qualification, yield, and financial disclosure will determine whether the technology becomes a real external business.
Chinese DRAM giant CXMT is reportedly producing small quantities of HBM3E, with Alibaba's T-Head and Cambricon testing it for adoption. Before any 2027 deployment, customer qualification, yield, and supply volume still need to be confirmed.
