Anthropic Details Custom AI Accelerator Built on Samsung 2nm Node with Broadcom

Anthropic has advanced its in-house custom silicon roadmap, partnering with ASIC design giant Broadcom and Samsung Foundry to develop proprietary 2-nanometer artificial intelligence accelerators. Fabricated on Samsung advanced SF2 Gate-All-Around (GAA) process node, the custom silicon targets approximately 4,000 TOPS of tensor throughput, packs up to 288GB of high-density HBM3E memory, and integrates next-generation PCIe Gen7 interconnects to power large-scale training and low-latency inference for Claude frontier foundation models.

Samsung SF2 2nm Node and Broadcom Custom ASIC Design

Following the establishment of its dedicated in-house silicon engineering team earlier this year, Anthropic structured a tripartite hardware partnership to bring custom Claude accelerators to production. Broadcom is overseeing physical ASIC architecture, interconnect logic, and high-speed SerDes IP, while Samsung Foundry is contracted to fabricate the monolithic compute dies using its 2nm SF2 process utilizing low-NA EUV lithography. Packaging will leverage Samsung 2.5D I-Cube technology to vertically and laterally bond the compute die with high-bandwidth memory stacks.

Architectural & Silicon Specifications

• Process Node: Samsung Electronics SF2 2-nanometer Gate-All-Around (GAA) Multi-Bridge-Channel FET.
• Compute Throughput: Targets 4,000 TOPS of mixed-precision INT8 and FP8 tensor matrix execution.
• Memory Subsystem: Configurable with 192GB to 288GB of high-density HBM3E memory delivering multi-terabyte-per-second bandwidth.
• High-Speed Interconnect: Integrates early PCIe Gen7 switch fabrics operating across 8 to 16 lanes per accelerator.
• Power & Thermal Envelope: Engineered with a 1,300W thermal design power (TDP) tailored for direct-to-chip liquid-cooled server racks.
• Initial Production Run: Slated for an initial manufacturing run of roughly 200,000 accelerator units.

288GB HBM3E, PCIe Gen7, and Bypassing TSMC Supply Bottlenecks

By partnering with Samsung Foundry rather than TSMC, Anthropic strategically sidesteps saturated CoWoS packaging lines and crowded Taiwanese wafer allocations dominated by NVIDIA and Apple. The move delivers critical custom hardware autonomy for Anthropic Claude models, while marking a landmark commercial win for Samsung Foundry as it proves high-volume viability for its 2nm GAA process node against global foundry competition.

Product Profile Sheet