Skip to content
Intel's 50% Demand Gap, TSMC's Packaging Park & Qualcomm's 2nm Bet artwork

Intel's 50% Demand Gap, TSMC's Packaging Park & Qualcomm's 2nm Bet

AI Hardware & Chips: Daily News

September 23, 2026

(00:00:00) Intel's 50% Demand Gap, TSMC's Packaging Park & Qualcomm's 2nm Bet (00:01:03) TSMC Packaging Industrial Park (00:01:44) KLA, Camtek, Nova Equipment Play (00:02:21) Qualcomm Snapdragon 8 Elite Gen 6 (00:02:58) Samsung 4nm HBM4 Logic Expansion (00:03:29) Memristor Breakthrough and...

Speakers Jamie Cole

TopicsNewsBusiness

Jamie Cole (0:00)

A Hardware & Chips Daily News. I'm Jamie Cole. Thanks for joining me.

Today, Intel's unfilled CPU demand, real supply crisis or hype?

Intel is fulfilling only around 50% of its current AI-related CPU demand. That's not a rumor. That's a number now attached to the supply story, and it changes the conversation.

The question worth sitting with is whether this is a genuine supply crisis or a managed narrative that conveniently supports pricing power?

The honest answer is probably both.

When agentic AI workloads scale, they don't just need GPU acceleration. They need substantial general-purpose CPU capacity to coordinate, route, and manage those agents.

Meta's Muse agent architecture is a clean example of this. It validates that CPUs aren't being displaced by AI infrastructure. They're becoming load-bearing within it.

The important distinction is that unfilled demand isn't the same as uncaptured revenue forever. It suggests Intel has room to expand volume and push price. What it doesn't tell us is whether Nvidia's accelerating push into edge AI CPUs starts eating into that advantage before Intel can fully supply into it. That's the risk that deserves watching. Separate from Intel, TSMC just announced a dedicated advanced packaging industrial park. That's a CAPEX signal, not a roadmap slide. When TSMC builds infrastructure at that scale, it's confirming that co-os demand isn't a short cycle. The projection is 810,000 co-os wafers by 2027, roughly double current levels. The signal here is who captures the margin from that scaling. It's not always the chip designer. Advanced packaging integrates multiple dies late in the manufacturing process, which makes defect sensitivity extremely high. A packaging failure destroys chips that already cost a fortune to produce. That makes inspection and metrology tools non-negotiable.

KLA, Camtek, and Nova are positioned at exactly that choke point. Camtek is carrying over $600 million in orders extending into 2027 Nova posted record revenue. These aren't speculative positions. They're order book backed, and the underlying demand driver is structural. Here's the thing. The risk worth naming is that Camtek's backlog is downstream of Nvidia and AMD multi-die roadmaps. No. If AI chip orders slow, that cascade hits the equipment layer quickly. But right now, the packaging equipment trade is one of the cleaner ways to track AI infrastructure spending without taking direct chip design risk. Qualcomm launched the Snapdragon 8 Elite Gen 6 on TSMC's 2nm process. It's targeting on-device AI with a 30 billion parameter capability. That's a meaningful spec for a consumer chip. The catch is memory. Smartphone memory is in shortage, and unit volumes are expected to contract 14% in 2026 Premium tier SOSEs like this one still sell, but the addressable market is narrowing. Qualcomm's bet is that the users who remain are spending more per device. That might be right, but thermal management and memory constraints at this performance tier are real engineering problems, not sold ones. Samsung is expanding 4-manometer production specifically to supply HBM4-based chips. That's the first major foundry capacity allocation tied directly to high bandwidth memory logic. HBM4-based chips handle the data exchange between GPU dies at advanced nodes, so this is infrastructure work, not headline chip design. Consider this, the execution risk is real. If HBM demand softens or SK Hynix maintains its supply dominance, Samsung's utilization case weakens. This is a capacity bet placed before the market settles. One development worth flagging, a USC lab demonstrated a memristor operating at 700 degrees Celsius with a theoretical 92% efficiency gain for natrix multiplication. That's a striking result. The manufacturing path to wafer scale integration is years away, but the underlying efficiency logic is sound enough to track. The near-term signals that matter most watch whether Intel's 50% on-fill demand figure gets third-party confirmation, and whether Nvidia's CPU roadmap accelerates on a timeline that closes that window. Watch TSMC's Packaging Park capex schedule for confirmation that the CoWAS ramp holds. And watch Qualcomm's Memory Supply Guidance as a leading indicator of where premium smartphone AI lands in 2026

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

Thousands of transcripts fetched by people building searchable podcast archives

Fetch the whole transcript

The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000791222953