ASIC Rise, TSMC's $40B Quarter & Nvidia's $1.5B Packaging Bet

Jamie Cole (0:00)

AI Hardware & Chips, Daily News. I'm Jamie Cole. Thanks for joining me. Today, the ASIC inflection, why hyperscalers are dumping GPUs. TrendForce just revised its AI server growth forecast upward again, and the number that stands out isn't the headline figure.

It's this. ASIC servers are on track to represent 27.8% of all AI server shipments by 2026, growing faster than GPU servers.

The infrastructure layer of AI is quietly shifting away from general purpose GPUs toward custom silicon built for specific workloads. The important distinction is that this isn't a prediction anymore. The hyperscalers aren't debating whether to build custom chips. They're already deep into production.

Google and Marvell expanded their partnership this year to cover the full AI infrastructure stack. That means AI inference accelerators, storage controllers, networking and near-memory computing products built around the TPU ecosystem. Marvell's custom chip business is now tracking toward more than $10 billion by fiscal 2029 Here's what that trajectory tells us. Google isn't just designing around Nvidia. It's building an entire custom hardware ecosystem and Marvell is being pulled into every layer of it. The concentration risk is real. Marvell's custom business depends heavily on a single customer relationship.

Tape-out costs on advanced nodes are substantial and any manufacturing delay compresses the ramp timeline for 2028 and beyond. But the direction of travel is clear. Nvidia's response to all of this is telling. The company committed $1.5 billion in advance payments to Amkor to expand advanced packaging capacity in the United States. Amkor has also signed a 10-year agreement with TSMC to build out facilities in Arizona. The key implication here is that Nvidia is directly funding the packaging infrastructure to secure its own supply chain. That's not a standard procurement move. That's a company looking at COWOS and HBM topacity constraints and deciding it can't afford to wait.

TSMC's advance packaging is already running tight and the Amkor Arizona facility is still in ramp up, carrying depreciation and execution risk on equipment transfers and recalcation. Capacity relief is coming, but not immediately. Here's the thing. TSMC's own numbers make the constraint visible. Wevenue hit $40.2 billion in the second quarter of this year, up 33.7% year over year. High performance computing now accounts for 66% of total revenue. 2nm and 3nm nodes together represent 33% of wafer sales. That's a foundry being pulled hard in one direction. The demand is there. The question is whether CoWAS packaging capacity can keep pace with what multiple hyperscalers will need simultaneously in 2027 That remains genuinely unresolved. SK Hynix earned TSMC's Partner of the Year Award for the second consecutive year, with HBM5 and CoWAS Corporation formally deepened. The company controls roughly 50% of global HBM revenue and is already negotiating 2027 volumes and pricing. The buyback progress, 73.74% complete, at nearly $31 trillion spent, signals management confidence in the memory pricing outlook.

The unresolved piece is Solibic Financing and the margin picture if memory pricing corrects. Worth watching. Consider this. Arm gained 5.18% in a single session closing at $307.49.

The market was repricing the stock for AI workloads beyond mobile, edge computing, autonomous vehicles, robotics and adjective AI.

The top five customers still represent 57% of revenue, so diversification is a genuine risk as Arm moves further into production silicon. The metrics that matter most right now are TSMC's Co-OS expansion timeline, Marvell's tapeout execution heading into 2028, and whether SK Hynix can hold its HBM pricing position through 2027 negotiations. The ASIC inflection is real and accelerating. The supply chain bottleneck is packaging, not design. That's where the pressure builds next. Thanks for listening. This podcast was built using AI technology, a YesWe production.