mstack

INTC

9 min read

INTC — Intel

TL;DR

Intel is a two-businesses-in-one-stock situation that the corpus reads as bear on products, conditionally bull on foundry. Intel Products missed the 2025–26 AI-driven datacenter CPU demand surge — Q4’25 showed supply unprepared and margins still weak1, Diamond Rapids is expected 40% faster than Granite Rapids but with 192 cores vs AMD Venice’s equivalent per-core lead, and Intel cancelled the 8-channel Diamond Rapids-SP mainstream platform outright, ceding the enterprise stronghold to AMD through 20282. Intel Foundry, meanwhile, is the only plausible non-TSMC leading-edge option and is being treated as a national-strategic asset — the U.S. government is now an explicit stakeholder3, 18A-P is rumored to be a “fixed” 18A with early Apple base-M-series interest ($630M annual revenue at 20% share)4, and Lip-Bu Tan’s public threat to abandon the leading edge surfaced the external support that 18A couldn’t win on merit5. Near-term catalyst: whether 14A design-win customer announcements land in the 2H26–1H27 decision window1.

Business

Intel is an IDM split across Intel Products (Client Computing Group for consumer CPUs, Datacenter and AI for Xeon server CPUs and Gaudi accelerators, NEX for networking silicon) and Intel Foundry (external wafer business fabbing for third parties on Intel 18A / 18A-P / 14A plus legacy Intel 16 in Ireland, plus an advanced-packaging business: EMIB, Foveros, Foveros Direct hybrid bonding)24. Revenue is overwhelmingly Products today; Foundry is a strategic option on leading-edge redundancy that the US government, hyperscalers, and potentially Apple are being asked to underwrite34.

Thesis

  • Intel Foundry is the last leading-edge alternative to TSMC, and the U.S. government has finally committed as sponsor. Fabricated Knowledge argues the endgame is now visible: the Trump administration’s pattern of “invest in America, keep overseas revenue” positions Intel Foundry for a series of hyperscaler investment announcements, and Intel’s public threat to end nodes past 14A was a calculated lever to force that help3.
  • Lip-Bu Tan’s “queen sacrifice” worked. Threatening to abandon leading-edge gave Intel the air cover to delay into 18A-P, which rumors inside the industry now treat as a fixed 18A with viable PDK — the node 18A was supposed to be. Advanced packaging group has been “whipped into shape” under LBT and can show revenue traction faster than 18A-P’s 2027 external ramp5.
  • Apple opened a door at 18A-P. Intel 18A-P (shipping late 2026) is the first theoretically viable non-TSMC leading-edge alternative since Apple left Samsung in 2016. A 20% base M-series shift would imply ~18k wafer ASP, with 70%+ yield at ~150–170mm² die size. Apple could qualify Intel on base M-series, WiFi/Bluetooth, display drivers, or power management without risking the flagship A/M lineup4.
  • Intel has zero externally-usable leading-edge capacity until 2027 — which is also bullish for Foundry demand as TSMC stays booked. TSMC has no capacity available; AI is 60% of 2026 N3 output. When 18A-P ramps in 2027, it does so into a structural wafer shortage, not an open market5.
  • Chipstrat’s view on 14A customer announcements: the cadence is a 2H26–1H27 decision window, so positive news if it comes will be a 2027 event, not a 2026 one. Anyone expecting foundry wins on near-term earnings calls is front-running the design-win cycle1.

Risks

  • 18A yield is still bad, and Intel IR is now admitting it. The rotating defense that “defect density <0.4” has given way to open acknowledgement that 18A yield is “not where it needs to be.” Intel’s pivot to “18A is internal only” is a retreat, not a strategy6.
  • Intel Products fumbled the AI-CPU demand surge. Q4’25 showed server CPU supply unprepared for the agentic AI demand surge despite multi-quarter telegraphing. Stock ran 120% in five months on the “Trump bump,” then crashed 17% (-$46B market cap) on the Q4’25 disclosure. The signals were there; management wasn’t staffed to catch them1.
  • Clearwater Forest delayed H2’25 → H1’26 on Foveros Direct integration issues. The 288-core Xeon 6+ is Intel’s lead vehicle for hybrid bonding, and it slipped. The vertical-disaggregation interconnect is also bandwidth-constrained at 35 GB/s per 4-core cluster2.
  • Diamond Rapids vs AMD Venice: structural gap widens. Diamond Rapids (192c/192t) is expected only ~40% faster than current Granite Rapids, while AMD’s 96c Turin already matches 128c Granite Rapids on per-core. Venice will extend AMD’s lead. Worse, Intel cancelled the mainstream 8-channel Diamond Rapids-SP platform, conceding the high-volume enterprise segment to AMD’s new 8-channel Venice SP8 through 20282.
  • Chiplet engineering discipline. Unlike AMD’s design — where the same CCD leading-edge die is fungible across consumer and server SKUs — Intel’s Panther Lake (consumer) chiplets cannot be reused in datacenter products. This is a strategic error that persists across multiple generations; margin recovery is harder without die-level fungibility6.
  • Dedicated-short trade is live and publicly sized. Ridire Research is long GFS / short INTC at 1.14% / -0.86% gross vol-adjusted weights — explicitly framing Intel as “policy-driven momentum” while GFS has “proven strategic relevance” with already-secured long-term government and industrial cash flows7.
  • Gross margin issue. Despite a datacenter supply shortage (see Micron retiring Crucial to free up datacenter wafers) Intel margins remain depressed; the sell-side is openly skeptical of the Zinsner answer6.

Recent catalysts

  • 2026-04-14 — FundaAI: Google TPU v8e (Humufish) confirmed on EMIB-T (Intel advanced packaging), not CoWoS as market rumored. Google prepaid Intel to secure capacity. 300k TPU v8e units projected in 2027Q4; 2.2M for full-year 2028. Intel actively securing substrate capacity from Unimicron and Ibiden.8
  • 2026-04-15 — ISSCC 2026: Intel presents UCIe-S chiplet interconnect paper (Paper 8.1). Signal of continued die-to-die IP investment but not a commercial catalyst9.
  • 2026-03-28 — Chipstrat: Trainium3 moving from Intel Sapphire Rapids (x86) head nodes to Graviton4 (ARM Neoverse V2); Intel loses ~125k CPU sockets per 1M Trn2 chips as AWS shifts to ARM in next-gen cluster.10
  • 2026-02-09 — SemiAnalysis datacenter CPU landscape: Clearwater Forest delayed to H1’26; Diamond Rapids 8-channel mainstream platform cancelled; AMD Venice set to widen per-socket gap2.
  • 2026-01-29 — WSJ via Chipstrat: stock down 17% (-$46B) post-Q4’25 on server CPU supply mismatch; analyst Stacy Rasgon characterizes the run-up as “vibes and tweets”1.
  • 2026-01-23 — Irrational Analysis Q4’25 review: Q4 earnings call forced Intel IR to admit 18A yield issues, margin pressure continues despite datacenter shortage; chiplet strategy faulted for lack of die fungibility6.
  • 2026-01-08 — SemiAnalysis Apple-TSMC piece: establishes Intel 18A-P as Apple’s first viable non-TSMC option since 2016; models $630M revenue at 20% base M-series share4.
  • 2026-01-04 — Irrational Analysis: Lip-Bu Tan queen sacrifice assessed as “worked”; 18A-P rumors positive; advanced packaging revenue can ramp ahead of 18A-P external revenue (2027)5.
  • 2025-10-10 — Ridire Research initiates long GFS / short INTC trade at 2% gross vol allocation7.
  • 2025-08-15 — Fabricated Knowledge: rumors of hyperscaler investment in Intel Foundry as part of Trump-admin “invest in America” pattern3.

Second-order reads

  • 2026-03-24 — SemiAnalysis, Nvidia: Inference Kingdom Expands — Microsoft’s Maia 200 rackscale system deploys Intel Granite Rapids CPUs as head nodes (Cobalt 200 is general-purpose only). Small positive: Intel retains AI-head-node share even when the hyperscaler has an internal ARM option.2
  • 2026-03-12 — SemiAnalysis, The Great AI Silicon Shortage — AI is 60% of 2026 N3 output; TSMC kingmaker with no uncommitted capacity. Raises the value of any viable leading-edge second source → Intel Foundry option value increases.5
  • 2026-02-25 — SemiAnalysis, Vera Rubin: Extreme Co-Design / 2026-03-24 NVDA GTC — NVL72 Vera CPU doubles to 1,536 GB DDR per rack (3× Grace). Positive for server DRAM; neutral/negative for Intel general-purpose CPU socket if Bluefield-4 Context Memory Storage displaces CPU memory hierarchy over time.2
  • 2025-11-27 — Citrini, Carving Up the TPU — Anthropic + Meta adopting TPUs is broadly negative for merchant accelerator vendors but neutral-positive for Intel Foundry insofar as TPU v8e Humufish uses EMIB-T with Intel securing substrate capacity (see FundaAI 2026-04-14).
  • 2026-04-14 — FundaAI, AI Infra 2026 — Google TPU v8e EMIB-T confirmation is a concrete, revenue-visible Intel Foundry win (advanced packaging); 2.2M units in 2028 at scale → positive for TSM CoWoS alternative thesis (Intel EMIB-T is the competing package for massive multi-die AI chips).
  • 2026-03-28 — Chipstrat, Agentic AI Needs CPUs — Trainium3 shifts AWS head nodes from Intel x86 to Graviton4 (ARM) → loss of ~125k CPU sockets per 1M accelerator chips; negative for Intel x86 share in hyperscaler clusters; positive for ARM/AWS custom silicon momentum → watch AMD EPYC response to ARM encroachment.10

Valuation & positioning

  • Price/book — Intel trades at a “much lower” P/B than TSMC, arguably similar to GlobalFoundries, despite a theoretically leading-edge asset base. Bull case is the book is temporarily impaired by self-inflicted 18A execution; bear case is the book should be impaired because 18A never earns its cost of capital6.
  • Trump-bump crash. Q4’25 earnings erased ~$46B market cap (-17%) in a single week; the 120% run-up into the print was “vibes and tweets” per Rasgon1. Not a floor — the underlying Products miss is structural, not a one-quarter air pocket.
  • Embedded optionality. The call options worth tracking: 18A-P external customer announcement in 2H26–1H27, Apple base-M-series qualification decision, and any hyperscaler “invest in America” Foundry commitment354.

Sources

AMD — direct x86 datacenter CPU competitor; Venice SP8 takes enterprise mainstream share ARM — Neoverse CSS licensed by every hyperscaler; Phoenix direct-to-chip move dilutes Intel further QCOM — returning to datacenter with SD2 (Nuvia team); additional merchant-silicon pressure TSM — the incumbent that Intel Foundry must dislodge; TSMC N3 constraint is also Foundry’s opportunity NVDA — Granite Rapids remains the AI-head-node of choice (Maia 200, others); Bluefield-4 CMX is the long-term threat to general-purpose CPU sockets MU — DRAM shortage tailwind + Micron/Intel historical NAND JV legacy; Intel’s cancelled mainstream platform compounds DRAM demand TSEM — Tower Semiconductor deal collapse was a trigger for 18A PDK delay thesis

Footnotes

  1. Chipstrat — Intel Q425: Back to Reality — 2026-01-29 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. SemiAnalysis — CPUs Are Back: The Datacenter CPU Landscape — 2026-02-09 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  3. Fabricated Knowledge — Intel’s One True Stakeholder is Here — 2025-08-15 ↩ ↩2 ↩3 ↩4 ↩5

  4. SemiAnalysis — Apple-TSMC: The Partnership That Built Modern Computing — 2026-01-08 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  5. Irrational Analysis — Fabulous Failures — 2026-01-04 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  6. Irrational Analysis — Intel’s Agentic Dumpster-Fire — 2026-01-23 ↩ ↩2 ↩3 ↩4 ↩5

  7. Ridire Research — Sovereignty over Storytelling — 2025-10-10 ↩ ↩2

  8. FundaAI — Deep: AI Infra 2026, Shifting from Brain to Body — 2026-04-14 ↩

  9. SemiAnalysis — ISSCC 2026: Nvidia and Broadcom CPO — 2026-04-15 ↩

  10. Chipstrat — Agentic AI Needs CPUs, Whose CPUs? — 2026-03-28 ↩ ↩2

Graph View

Backlinks