TSMC's CoW Opens Up, Anthropic's Custom Chip & AMD's 107% Surge

**Jamie Cole** (0:00)
A Hardware & Chips, Daily News.
I'm Jamie Cole, thanks for joining me. Today, Packaging Wars, TSMC outsources CoWAS to break bottleneck.
TSMC is opening up its chip-on-way for packaging process to outside manufacturers, and that single decision may do more to ease the AI chip supply crunch than anything else in the pipeline right now. For context, CoWAS is the two-step packaging architecture that connects GPU and AI processes to high bandwidth memory via a silicon interposer. The first step, the chip-on-way for attachment, has been TSMC's exclusive domain. No one else has been doing it at scale. That's been a structural bottleneck for the entire AI hardware industry, limiting how fast chips can be packaged and shipped, regardless of how many wafers foundries actually produce.
TSMC outsourcing that step to ASE signals the end of that single supplier constraint. The important distinction here is that this isn't a capacity tweak. It's a structural shift in how the packaging supply chain is organized. When a choke point this significant opens up, the effects ripple outward fast. The first place those effects land is with the equipment suppliers. South Korean firms, Hanmi Semiconductor, Avaco and Wonik, are positioned directly in front of this wave, as ASE scales up Co1 capacity, it needs dicing and bonding equipment, and orders in this space run above $1 million per unit.
These aren't speculative plays, they're companies supplying tools for a process that now has a confirmed industrial buyer scaling aggressively. The execution risk is real though, Co has been TSMC's process for a reason. Yield management and quality control at this level are hard. Whether ASE can match TSMC's output quality at scale is the proof point that matters. Bottleneck relief depends entirely on that answer. While the packaging story reshapes supply, Anthropic is making a move that reshapes demand. The company confirmed an in-house chip design team with a co-design strategy aimed at cutting clawed inference costs by 50%.
Discussions with Samsung around the two-nanometer process are reportedly underway. Here's the thing. The signal here is strategic, not just technical.
Anthropic joins Google, Meta and OpenAI in building hardware designed around model requirements, rather than adapting models to general purpose GPUs. Each new entrant fragments Nvidia's monolithic hold on inference compute. The timeline is unclear.
No production date, no confirmed manufacturing partner, no architecture details. This is a multi-year program at an early stage. But the direction is set and the incentives are obvious. Cost control, performance optimization, and reduced dependency on a single supplier. IMD's quarterly numbers deserve attention in this context. Data center revenue came in at $6.7 billion, up 107% year over year. Q3 guidance landed at $13 billion, well above consensus expectations. The first Helios rack system is shipping to Meta, OpenAI, and Oracle. That last detail matters. Helios is AMD's integrated rack offering, and it puts AMD in direct competition with Nvidia's end-to-end system ambitions, not just at the GPU level. EPIC CPUs are gaining hyperscaler share, Instinct GPUs are scaling. The question that doesn't go away is whether AMD can sustain this trajectory against Nvidia's software ecosystem and installed base, particularly in the second half of next year. Intel's story this cycle is simpler to state and harder to evaluate. The 18A process is in full production. The 18A-P variant is in risk production. Panther Lake launches as the first Intel product in nearly a decade to close the manufacturing no gap with TSMC. The stock reclaimed $100 on the back of that, with Q2 revenue up 25% year over year to $16.1 billion. Consider this. Here's the catch. Foundry external revenue sits at $293 million against the CapEx commitment in the hundreds of billions. Intel has validated the process. It hasn't yet validated the business model. External customer adoption is the only metric that resolves that question.
Three things to keep close. Whether ASE's CoRAMP delivers yield at scale, whether A&D's $13 billion Q3 guidance holds as Nvidia fights back at the system level, and whether Intel signs a meaningful external foundry customer before the CapEx math starts drawing harder questions. Thanks for listening. This podcast was built using AI technology, a YesWe production.