MediaTek’s ASIC business is about to go brrrr MediaTek’s ASIC business is projected to explode to NT$1.4T by 2028, becoming 67% of sales. Now UBS sees potential new ASIC projects including Tesla AI6 using Intel 14A. $INTC Intel 𝕏 MediaTek

Biggest takeaway AI's next semiconductor bottleneck may not be GPUs. It may be everything required to manufacture and test them. Some equipment suppliers are preparing their supply chains for a potential $500B/year WFE market, while test-tool production signals point to roughly doubling from 2H26 to 2H28. Detailed takeaway: 1. How strong is the semiconductor equipment cycle? Extremely strong. WFE suppliers are preparing for 30%+ annual growth for the foreseeable future, with parts of the supply chain being asked to support a potential $500B/year WFE market DRAM is a major driver, with industry capacity potentially reaching ~3M wafer starts/month in 2029 and 4M by 2031. Some equipment suppliers are already filling slots into 2030. 2. What's the overlooked Intel angle? Intel may need to spend considerably more on its own manufacturing capacity. TSMC could begin curtailing wafer allocation to Intel as early as next year, while more customers are showing interest in Intel's EMIB-T capacity/offering. That could drive additional Intel capacity buildout and capex in 2027, plus higher equipment spending over the following years. 3. Is semiconductor test becoming another major AI beneficiary? Yes, and the market may be underestimating it. AI accelerators have: - larger transistor counts - more HBM - more mission-mode testing - less mature DFT capabilities Suppliers are receiving signals from Teradyne and Advantest that tool production could roughly double from 2H26 to 2H28. 4. What could become the next semiconductor supply bottleneck? Semiconductor materials. Wafer starts are already approaching previous peak levels, but advanced nodes also require more materials per wafer. The report points to roughly a 25% node-to-node increase in materials content for Entegris 5. What was the biggest message from OCP? AI's bottlenecks are moving beyond the GPU. As AI clusters become bigger and more complex, performance increasingly depends on: GPU → Networking → Optics → Memory → Power → Cooling 6. What's interesting about Marvell? Investors may be underestimating the interconnect opportunity, while its XPU/custom-silicon opportunity already looks strong. NVIDIA's system-level co-design advantage. Hyperscalers trying to compete on rack-level performance may need similar levels of chip co-design around their own XPUs. Marvell can potentially address many of those chips surrounding the XPU. 7. What about Astera Labs? Astera is trying to expand far beyond PCIe retimers into: PCIe + CXL + scale-up + memory offload + optics + rack-level architecture. The potential content-per-rack opportunity could therefore grow considerably, but many of these businesses are still early. 8. Is Intel's new D-Series GPU finally a serious AI GPU story? It's an encouraging start. Customer sampling is underway. Intel's former Crescent Island, now the D-Series GPU, is targeting inference with 1.3 PFLOPS FP8, 1.5TB/s memory bandwidth, up to 480GB LPDDR5X and 350W power.