TSMC Packaging Crunch, Intel's 18A Progress & LPDDR6 Mobile Memory Race

Jamie Call (0:00)

AI Hardware & Chips Daily News.

I'm Jamie Call. Thanks for joining me. Today, the packaging crunch, why GPU capacity isn't enough. The AI infrastructure story has shifted. The constraint isn't GPU supply anymore. It's packaging. TSMC's co-OS and SOIC advanced packaging capacity can't keep up with demand for AI accelerators. Data center servers are sitting delayed, not because Nvidia can't make GPUs, but because TSMC can't package them fast enough. That's a meaningful inversion of where the bottleneck was assumed to be, and it changes how you think about AI infrastructure timelines.

Co-OS, which stands for chip on wafer on substrate, is the integration layer that bonds high bandwidth memory to GPU logic in a single package. Without it, the chip doesn't ship as a functional accelerator. TSMC hasn't disclosed when its packaging capacity expansion will close the gap. That uncertainty alone could push data center delivery delays into 2027

The signal here is that CapEx commitments across the hyperscalers are now gated on packaging allocation, not GPU output. That's a supply chain dynamic that doesn't resolve quickly. Intel had a strong week on a different front. The stock climbed roughly 7% to around $124.50, driven by analyst upgrades and visible progress on its 18 Angrin process node. Tigress raised its price target to $145, citing recoverings beyond demand and Foundry momentum. The CEO added to his personal position, which is a tangible signal of internal confidence. The important distinction is that the upgrade cycle reflects a shift in analyst sentiment, not yet a shift in fundamentals. Intel's x86 market share fell below 70% in the second quarter. AMD continues to take ground. And a capital raise earlier this year at $95 a share added over 210 million new shares, which means margin gains need revenue behind them to matter for shareholders. The Foundry business still has very limited external customer reduction. Progress on 18 Antregum is real, but the execution gap between process milestone and commercial revenue remains wide. Here's the thing, while the data center is constrained by packaging, the mobile side of AI hardware is moving fast on memory. Samsung completed validation of its 16GB LPDDR6 chip, running at 10.7Gbps, on Snapdragon 8 Elite Silicon.

That's a 21% power improvement over LPDDR5X, which matters when the workload is running locally on a phone, rather than offloading to a server.

CH Hynix confirmed its own LPDDR6 development is complete on its 1K process node, with mass production preparation beginning in the first half of 2026, and supply targeted for the second half of that year. Both companies are aiming at smartphones and tablets. Samsung's likely insertion point is the Galaxy S27, which points to 2027 adoption at scale. The underlying driver is on-device AI. Text, image and voice workloads that currently rely on cloud inference are moving on to the device. Faster, lower power memory is what makes that viable.

One development worth flagging from the research side. A team at the University of Colorado Boulder created a Surface Acoustic Wave Phonon Laser, operating at around 1GHz. The implication is a single-chip architecture that could integrate an entire radio front-end, replacing the multi-chip RF modules that smartphones currently use. That reduces board area, component count and potentially cost. Consider this, the catch is the usual one for lab breakthroughs. Manufacturing scalability and cost at volume are unaddressed. The path from university demonstration to smartphone bill of materials is long and uncertain.

The packaging constraint at TSMC is the thread that runs through most of what matters right now. Until that capacity expands, the AI infrastructure buildout is moving slower than GPU production alone would suggest. Watch for any disclosure from TSMC on Co-OS expansion timelines. That's the real unlock. On Intel, the metric that matters is Foundry customer wins, not process node announcements. On LPDDR6, the timeline is clearer now. The question is whether on-device AI workloads grow fast enough to justify the memory upgrade cycle at the premium end.

Thanks for listening. This podcast was built using AI technology, a YesWee production.