The CODEW Semiconductor Watch: AMD's Data Center Surge, HBM Memory Boom, TSMC's CoWoS Expansion, and the Rise of Custom AI Silicon
AMD Data Center Revenue Surges 107%, HBM Memory Supercycle Escalates, TSMC CoWoS Ramps to 130k WPM, and Custom ASIC Scaling
Record Q2 Hardware Earnings Confirm Advanced Packaging and HBM Allocation as Primary Semiconductor Bottlenecks
The global semiconductor industry has entered the second half of 2026 with unprecedented demand across high-performance compute (HPC), custom hyper-scaler ASICs, and high-bandwidth memory. Advanced Micro Devices (AMD) headlined this week’s earnings reports, posting record Q2 2026 revenue of $11.54 billion (+50% YoY), driven by a 107% YoY surge in Data Center segment sales to $6.7 billion. The rapid adoption of EPYC server processors and Instinct accelerator scale-outs confirms that non-NVIDIA merchant silicon options are capturing measurable multi-cloud market share.
Simultaneously, the physical constraints gating global AI chip shipments have shifted from front-end silicon wafer starts to back-end 2.5D/3D advanced packaging and HBM availability. TSMC's Chip-on-Wafer-on-Substrate (CoWoS) monthly capacity expansion is targeting 120,000 to 130,000 wafers per month (WPM) by late 2026, while memory leaders SK hynix, Micron, and Samsung report near-total allocation of 2026 HBM3E supply alongside accelerating HBM4 tape-outs. Enterprise technology buyers face extended lead times for tier-one accelerators, making multi-sourcing and custom ASIC co-development critical strategic mandates.
Ecosystem Connections: Explore related cross-functional analysis in AI Watch, and Semiconductor Watch.
Sources: AMD Q2 2026 SEC Filings · TSMC Technology Symposium Reports · SK hynix Executive Briefings · Semiconductor Industry Association (SIA) Data
Top Headlines
AMD Posts Record $11.54 Billion Q2 Revenue as Data Center Business Doubles
AMD announced Q2 2026 financial results exceeding market consensus, delivering $11.54 billion in total net revenue (+50% YoY) and non-GAAP EPS of $1.66. Data Center revenue surged 107% YoY to $6.7 billion, representing 58% of overall enterprise revenue. CEO Dr. Lisa Su highlighted accelerating deployments of 5th Gen EPYC processors and Instinct MI series accelerators, while confirming that the Helios integrated rack platform begins scaling into production in H2 2026.
HBM Market Projected to Reach $54.6B as SK Hynix & Samsung Accelerate HBM4
The global High-Bandwidth Memory (HBM) market is expanding toward a projected $54.6 billion total addressable market in 2026 (+58% YoY). SK hynix and Samsung Electronics reported near-complete allocation of HBM3E manufacturing lines through 2026. Both memory suppliers are preparing 2048-bit interface HBM4 tape-outs utilizing custom foundry base dies (built on TSMC 3nm/5nm processes) to support next-generation architecture platforms.
TSMC Ramps CoWoS Capacity to 120k+ WPM as 2nm (N2/A16) Tape-Outs Hit Milestones
Taiwan Semiconductor Manufacturing Company (TSMC) confirmed that its 2.5D/3D Chip-on-Wafer-on-Substrate (CoWoS) packaging lines remain fully allocated, with monthly wafer targets expanding toward 120,000–130,000 WPM by year-end 2026. TSMC also verified that over 25 major tape-outs have been finalized on its 2nm (N2) gate-all-around (GAA) process node, with backside power delivery (A16) scheduled for H2 2026 volume manufacturing.
Broadcom & Marvell Capture Multi-Gigawatt Custom Hyperscale ASIC Commitments
Custom merchant ASIC designers Broadcom and Marvell reported expanding multi-year backlog commitments from cloud providers including Google, Meta, and AWS. Broadcom’s recent rollout of the 102.4 Tbps Tomahawk 6 switch chip, coupled with custom 3nm AI XPUs, highlights the trend of hyperscalers deploying tailor-made silicon to lower total cost of ownership (TCO) compared to standard commercial off-the-shelf accelerators.
RISC-V Ecosystem Gains Ground in Datacenter Control Planes & Edge Inference
The RISC-V International consortium reported a major increase in enterprise data center implementations, where RISC-V application cores are increasingly embedded within SmartNICs, DPUs, and specialized AI micro-controllers. Major vendors are leveraging RISC-V to bypass proprietary ISA licensing overhead while customizing control-plane execution inside heterogeneous compute packages.
Industry Analysis
1. Gating Factors: Packaging and HBM Overshadow Wafer Starts
The primary bottleneck in enterprise AI hardware manufacturing has evolved. While leading-edge 3nm and 2nm silicon wafer supply has expanded, the ability to package multi-chiplet topologies via 2.5D/3D interposers remains tightly constrained. TSMC’s CoWoS capacity allocation controls global accelerator output.
Simultaneously, HBM3E yield rates and Through-Silicon Via (TSV) stacking capacity dictate system-level lead times. As a result, hyperscalers and tier-one enterprises are signing 18-to-24-month advance capacity agreements to secure guaranteed memory-and-packaging pairings.
2. Merchant Silicon vs. Custom Hyperscaler ASICs
A structural bifurcation is taking place across server room deployments. Merchant GPU platforms (NVIDIA Blackwell Ultra, AMD Instinct) maintain dominant positioning for training massive frontier models due to their software ecosystems (CUDA, ROCm). However, hyperscale cloud operators are shifting steady-state inference and domain-tailored training workloads onto internal custom ASICs (e.g., Google TPU v6, AWS Trainium/Inferentia, Meta MTIA). Custom ASICs are optimized for specific model topologies, offering higher performance-per-watt and protecting operating margins against high merchant silicon pricing.
Semiconductor Node & Memory Benchmark Matrix
| Segment / Technology | Lead Process / Standard | Key Industry Metric | Market Bottleneck Level |
|---|---|---|---|
| Enterprise Compute (AMD EPYC / Instinct) | 3nm / 4nm TSMC Chiplets | 107% YoY Data Center Revenue Growth | Moderate-High |
| Advanced Packaging (TSMC CoWoS-S/L) | 5.5-reticle size / 2.5D Interposer | 120k–130k WPM Target (End-2026) | Critical Constraint |
| High-Bandwidth Memory (HBM3E / HBM4) | 12-High Stacks / 2048-bit Interface | $54.6B TAM in 2026 (+58% YoY) | Critical Constraint |
| Custom ASIC Design (Broadcom/Marvell) | 3nm GAA / Co-packaged Optics | 82% YoY ASIC HBM Demand Growth | Balanced-Stable |
Market Signals
- Acceleration of Multi-Vendor AI Chip Sourcing: Enterprise CIOs are actively diversifying compute supply chains, utilizing AMD Instinct accelerators and hyperscaler custom ASICs to mitigate single-supplier allocation limits.
- Transition from HBM3E to HBM4 Custom Base Dies: The shift toward HBM4 introduces custom logic dies manufactured on advanced foundry nodes (3nm/5nm), embedding memory vendors more tightly into foundry advanced packaging ecosystems.
- Backside Power Delivery (BSPDN) as the Next Scaling Frontier: TSMC’s A16 and Intel’s PowerVia implementations mark the move toward decoupling power distribution from signal routing, enabling higher clock frequencies and improved thermal management in dense AI silicon.
- Foundry Geography Diversification: Capital investments in global fab sites across Taiwan, the United States (Arizona, Ohio, Texas), Europe, and Japan (Rapidus) are beginning pilot production to improve supply chain resilience against geopolitical disruptions.
The CODEW Perspective
Over the next 12 to 24 months, competitive leadership in the semiconductor industry will be determined by advanced packaging execution and co-design integration rather than standalone transistor counts. As raw process node shrinks yield diminishing cost returns, performance gains rely heavily on high-density interposers, photonic interconnects, and tightly coupled HBM.
Enterprise infrastructure architects must plan for a heterogeneous chip environment. Organizations that build software pipelines compatible across multiple hardware backends—ranging from commercial GPUs to custom cloud ASICs—will secure superior cost efficiencies and avoid deployment delays caused by single-vendor component shortages.
The CODEW Semiconductor Watch is published weekly. Content compiled from verified SEC filings, earnings disclosures, and primary foundry updates. Published August 5, 2026.
Reviewed by Erwin Castro
on
Wednesday, August 05, 2026
Rating: