TSMC's HPC-overtakes-mobile crossover and the custom-ASIC supply rebalancing
TSMC's HPC segment passing mobile as the company's largest revenue source for the first time is a structural inflection. Combined with Broadcom's $73B AI-ASIC backlog growing 44.6% YoY (nearly 3x merchant GPU growth), the H2 2026 compute-supply landscape is rebalancing toward custom-silicon faster than 2025-vintage forecasts predicted.
TSMC's HPC segment crossing mobile as the company's #1 revenue source is one of those single-data-point disclosures that captures a structural shift the whole supply chain was waiting for.
What the segment-crossover signals
TSMC sees end-demand commitments 18-36 months ahead of product launches; the foundry's revenue mix is the single best leading indicator of where global semiconductor capacity is heading. The HPC-overtakes-mobile crossover happening ~12 months ahead of most 2025 analyst frameworks is the structural input — AI-compute capacity demand is scaling faster than smartphone-segment decline, and the implication is that other segments (auto, IoT, industrial) face capacity-squeeze pressure as TSMC reallocates incremental capacity expansion toward AI-customer commitments through 2027.
The Broadcom growth-rate anchor
Broadcom's $73B AI-ASIC backlog with 44.6% YoY growth vs 16.1% for merchant GPUs is the matching demand-side data point. Custom ASICs growing 3x faster than merchant GPUs confirms hyperscaler procurement is restructuring around inference cost optimization — and the restructuring is happening at billions-in-revenue scale, not as a speculative direction.
The three-way capacity dynamic
NVIDIA holds the training-tier dominance with merchant GPUs. Broadcom + AMD + custom-ASIC partners are taking inference-tier share. TSMC sits at the bottleneck for both — every advanced-process AI chip from any vendor goes through TSMC. The H2 2026 supply story is no longer about whether NVIDIA can produce enough Vera Rubin (AM cycle confirmed full production); it's about whether TSMC's capacity expansion can absorb both NVIDIA's continued training-tier demand AND Broadcom's accelerating custom-ASIC demand simultaneously.
The institutional-capital reinforcement
The BlackRock/MGX $40B Aligned Data Centers acquisition from this morning's cycle plus the CoreWeave $9B Core Scientific deal plus today's PM Broadcom backlog disclosure together produce the most coordinated compute-supply quarter of the past three years. Institutional capital, GPU supply (Vera Rubin), custom-ASIC growth, datacenter conversion velocity, and foundry-tier capacity reallocation all align in the same H2 2026 window.
What stays uncertain
Whether TSMC can maintain foundry-tier pricing discipline through 2027 as AI-customer demand intensity grows. The HPC-overtakes-mobile crossover gives TSMC pricing leverage; the question is whether that translates to per-wafer pricing increases or stays at margin-mix improvement. Hyperscaler procurement costs through H2 2027 will be heavily dependent on TSMC's pricing behavior under these structural conditions.
Heygotrade — AI Semiconductor Stocks 2026 → · Tom's Hardware — Custom AI ASICs Examined →