Assembled from a long thread of notes and posts I was collecting. Not investment advice.
Today, Intel is the only leading-edge logic foundry in America, and the only real alternative to TSMC anywhere. MC is barely above $200B; TSMC is around $1.4T.
For years, Apple was Intel's defining customer, both as a chip buyer and their anchor tenant in the foundry business. That concentration was a liability: being that reliant on one account is fragile. What's changed is that Intel now has the capability to bring customers back — especially as AI demand explodes and those chips increasingly need to be made on US soil. No one else can offer leading-edge, made-in-America capacity at scale. AMD's old foundry (GlobalFoundries) is less advanced, less capable, and has a weaker customer base. So Intel is, quite literally, the only other game in town. That's the asymmetry I keep coming back to.
Listing all of the cases here:
CPU demand through the roof
"CPU-based RL environments & verifiers. Guess what's going to happen to CPU demand later this year, anon??"
Reinforcement-learning workloads will increasingly run on CPUs — for environments, simulation, and verification — which boosts demand for server CPUs and benefits Intel’s Xeon line. Though, most incremental server-CPU spend has recently favored AMD, so it could be AMD's GPU growth matters more than Intel's CPU-plus-foundry upside.
The bull rebuttal (h/t @zephyr_z9) is that Xeons still have strong adoption, and that scaling RL could even cause CPU shortages — which would amplify Intel's pricing power. On top of that, Intel has additional tailwinds from Apple and from advanced packaging, both of which strengthen the case.
A non-AI server replacement cycle
There's also a general server refresh underway that's quietly fueling memory and CPU demand:
"US CSP general server procurement drives memory price demand. Since September, DDR5 16Gb spot prices have surged more than threefold... We expect 2026F US CSP general server shipments to grow by 50% YoY, driven by: (1) CSPs' extended depreciation periods starting in 2022, creating a replacement peak in the next two years; (2) general-server budgets being constrained by AI investments over the past two years; and (3) some 2027 demand being pulled forward to address expected shortages."
As @zephyr_z9 put it: a general server replacement cycle, growing shipments ~50% YoY — good for both AMD and Intel.
Competitive ground regained against AMD
Intel is also re-arming on the client side.
"20 years after Nehalem, Jensen finally won. Intel will be integrating an RTX iGPU with their CPU to take on AMD's Strix Halo APUs." — @zephyr_z9
This is Serpent Lake: reportedly 16× LPDDR6, with Nvidia GPU IP likely based on the Rubin RTX architecture (the GPU tile on TSMC N3P), and SoC tiles from the Titan Lake family. An Intel CPU with a genuine Nvidia-class iGPU is a direct shot at AMD's APU franchise.
The 18A ramp — strength and the asterisk
The bull pillar here is 18A, Intel's leading-edge node and the centerpiece of "Intel Tech Minute: All About Intel 18A." But I want to be honest about the asterisk, because the thread had both sides.
The skeptical read (via @jukan05, citing Bernstein):
"Bernstein noted in a report published today that Intel's 18A ramp-up is seeing yield improvements that are falling short of its initial targets."
And the analyst framing of what actually matters:
"Intel's ramp into 18A will be gradual as its yield progress is behind its original target. Whether to build meaningful capacity hinges on whether 14A successfully secures design-wins from internal and external customers — and we expect a decision on that probably around the end of 2026."
So the node is real, but the pace is behind plan, and the decision that matters (14A capacity) is gated on landing external customers towards the end of ’26.
Production ramp: small scale now, customers by end-2026
The base case from GF Securities (HK) lines up with a measured ramp:
"We still expect Intel to ramp 18A and 18A-P at small scale in 2026 to ensure yield and quality. That said, the external customer is expected to start by end-2026." — via @stfbutnou
And the catalyst that moved the stock (@jukan05, GFHK raising its target price to $54):
"Intel-related update — 18A yield, 14A external customer wins, and EMIB developments. On Nov 28, Intel's share price was up 10% owing to news of orders..."
The roadmap they shared is the real tell on who is showing up: Panther Lake, Wildcat Lake, Nova Lake (on TSMC N2), Clearwater Forest, Diamond Rapids, Razor Lake, Coral Rapids — and on the external side, a USG Defense Program, Apple M-series on 18A-P, a Google TPU v6e on EMIB-T, an Apple smartphone SoC, an Apple/Broadcom ASIC, and a Tier-1 Ethernet switch. That's a customer list, not a wishlist.
The advanced-packaging ramp up
This doesn’t depend on Intel winning the bleeding-edge node war; it only depends on AI demand being too big for one company to package.
AI chip demand is so high that TSMC can't handle all advanced packaging in-house, so the work is being split across partners like Intel and ASE (I also have a small stake in). Intel "winning AP orders" means a customer's chips may be fabricated at TSMC (including Arizona) but then sent to Intel for advanced packaging — effectively inserting Intel into the AI supply chain regardless of whose transistors are inside.
That's exactly what Digitimes Taiwan reported (via @jukan05): Intel and TSMC stitching together a "front-end / back-end" hand-off, with Lo Wei-jen (羅唯仁) — who recently moved from TSMC to Intel — said to be responsible for bridging the two.
Intel is also scaling packaging outside its own walls for the first time:
"Intel to conduct AI packaging at Amkor's Songdo fab in Incheon, Korea — a technology Intel had previously only conducted in its own fabs." — @jukan05
Intel is turning EMIB into a scalable, semi-outsourced platform that can serve both its own chips and external foundry customers like NVIDIA and Apple. EMIB is cheaper and faster than silicon interposers, and the move signals Intel is prioritizing speed, volume, and supply-chain resilience over vertical control. Longer term it sets up mass production of EMIB-T and positions Intel as a serious advanced-packaging provider in the AI ecosystem. And the partner momentum is building — per @zephyr_z9, "even MediaTek may use Intel's EMIB for advanced packaging, and a Qualcomm deal is certainly coming."
Orders on the board
It's not all theoretical. The wins (all via @zephyr_z9) are starting to land:
- Maia 2 (Microsoft): "Good news for Intel — they won the orders for Maia 2."
- Tesla: "Tesla will likely use Intel's advanced packaging for their chips."
- ASICs: "I think they'll use Intel for advanced packaging for their ASICs."
- Google TPUs: "18A isn't going perfectly and the bottleneck is packaging — Google will use Intel's EMIB for some TPUs."
And on capacity, look at where the industry is pointing: TSMC's own CoWoS shipment forecasts get revised sharply upward into 2027, which means it'll massively increase CoWoS equipment purchases and build new plants — with the 2027 ramp spread across Intel, ASE/Amkor, and Powertech. The whole pie is growing fast enough that Intel gets a slice.
The Apple relationship, again
The customer that defined Intel's past may quietly return. Per Ming-Chi Kuo (郭明錤, @mingchikuo):
"Intel is expected to begin shipping Apple's lowest-end M processor as early as 2027. There have long been market rumors that Intel could become an advanced-node foundry supplier to Apple, but visibility had remained low. My latest industry surveys, however, indicate that [this is becoming real]."
Contrary to the 2010’s, Apple is now a return customer on top of a diversifying base.
Talent
You can't ramp a node or a packaging line without the people who've done it before. From a Japanese-language post (@paurooteri), translated:
"A key 5nm person is moving from TSMC to Intel — TSMC is preparing litigation. Lo Wei-jen (羅唯仁), the mastermind behind 'Project Nighthawk,' has moved to Intel."
As @zephyr_z9 framed it: "Lip Bu Tan is probably one of the few people in the world who can make such proclamations and have the wherewithal to achieve them."
Counterpoints
China. Intel has near-term China risk: revenue from China was ~$14B (nearly 30% of the total), and a formal ban on Xeon chips could put roughly $5B in sales at risk (@zephyr_z9).
Hyperscaler design wins. "Intel will make money — but will they win hyperscaler contracts? NO." Making money and winning cloud silicon are two different bars.
Workload shifts toward Arm. Google's Axion is its own datacenter CPU (like Amazon's Graviton), built on Neoverse V2 and TSMC N4. It could be "bad news for Intel and AMD." Every hyperscaler custom Arm CPU is x86 share gone away.
Does EMIB actually scale? This is the most technical risk, and a good question from @Arronwei3n: "If $TSM CoPoS is on track to MP in 2027, or EMIB falls short, will those ASIC proposals to $INTC EMIB switch back to $TSM CoPoS?"
Maybe Intel's EMIB doesn’t scale cleanly for very large AI chips. Embedding many EMIB bridges increases manufacturing complexity and raises the risk that a single defect ruins an otherwise-good, very expensive HBM die. It's a probability problem — more embedded silicon pieces mean more chances for planarization (CMP) to fail. By contrast, TSMC's CoWoS-L grows silicon islands directly on the interposer with CMP at each step, which should keep yields more stable as designs get larger. The implication is that CoWoS-L may be inherently more scalable and lower-risk than EMIB for next-gen AI packages.
TSMC's pricing discipline cuts both ways. TSMC reportedly makes only ~$700 on every H100 it produces, and faces real pushback whenever it tries to raise wafer prices even 5–10% (@jukan05, @zephyr_z9). This could turn into a headwind for Intel and an opportunity: "Samsung and Intel will start winning orders if TSMC raises wafer prices by a lot." Samsung SF3 and Intel 18A "aren't that bad" — they just need to be good enough at the right price.
It's an asymmetric bet on the only alternative to TSMC, at a fraction of TSMC's valuation, right as AI makes domestic leading-edge capacity invaluable.