Our checks suggest smooth AI server shipment ramps in 2H23 so far. The supply chain is seeing increased AI developments from hypercalers, enterprise and alternative AI semi design for 2024 and beyond

AI downstream supply chain: 
• Our checks suggest that Microsoft's AI server suppliers haven't seen any order changes for 2H23. 
• Our checks suggest NVDA L40S orders have increased to 600-700k (up from 500-600k) in 2H23. 
• Our checks suggest that shipment from Meta's AI server will kick in from September - positive for Quanta. 
• Giga-Byte should be on track to ship another 800-900 units of AI servers in Sept, up from 300-400 units in Aug (link here). 
• Gold Circuit guided 4Q revenue to be up single to double-digit q/q with GM to expand 1% q/q to 24-28%, driven by AI server CPU motherboards, switch boards, ASIC baseboards, and GPU module PCB (link here). 
• Our checks suggest AMD MI300 designs with 3D VC for heat dissipation; NVDA GH200 (280W CPU+720W GPU) also goes with VC, and B100 has still not finalized the passive cooling solution between VC and cold plate

AI semi supply chain: 

Alternative AI semis – more aggressive custom silicon and cost saving 

• In a Reuters report, Google said it expects no change in its relationship with AI chip supplier Broadcom. This came after reports it was considering doing so. 

• In our ASIC design service update, we pointed out Asian design service vendors usually charge only 40% gross margin for chip design (or NRE), with a 20% mark-up for chip production (or turnkey service). We therefore see opportunities for Asian vendors, such as MediaTek, as early as in 2025, if some hyperscalers want pursue the cost savings.