Intel Jaguar Shores (Rack-Scale System)
Overview
Intel Jaguar Shores is a rack-scale AI system project restarted by Intel after cancelling Falcon Shores in May 2026, expected to launch in 2027-2028. Unlike Falcon Shores' "single GPU chip" strategy, Jaguar Shores is a rack-scale integrated system combining:
- Gaudi-series IP AI accelerators (replacing Gaudi 3)
- Xeon 6 / Xeon 7 CPUs (Clearwater Forest / Diamond Rapids)
- 800G networking NICs
- Intel Foundry 18A / 14A process
Jaguar Shores is the flagship of Intel's "AI systems vendor" strategy — no longer pursuing single-chip competition with NVIDIA, but delivering full rack-scale solutions.
Core Specifications (Speculative)
| Item | Spec |
|---|
| Form Factor | Rack-scale system (not a single chip) |
| AI Accelerators per Rack | 64-128 units (speculative) |
| CPUs per Rack | 32-64 Xeon units |
| AI Accelerator IP | Gaudi v4 architecture (evolved from Gaudi 3) |
| Process | Intel Foundry 18A (accelerators) + TSMC 3nm (partial IP) |
| HBM Capacity (per accelerator) | 144 GB HBM3e |
| HBM Bandwidth (per accelerator) | ~5 TB/s |
| FP8 Compute (per accelerator) | ~2,500 TFLOPS (speculative) |
| FP8 Compute (per rack) | ~160-320 PFLOPS |
| Networking | 800G integrated NIC + topology optimization |
| TDP (per rack) | ~80-120 kW |
| Launch | 2027-2028 |
⚠️ Not officially announced: All specs above are speculative, Intel has only disclosed roadmap-level information. All figures subject to future Intel announcements.
Falcon Shores Cancellation Timeline
| Event | Date | Details |
|---|
| Falcon Shores first announced | 2023 | 200 PFLOPS single chip targeting B100 |
| Roadmap adjustment | 2024-12 | 200 PF target cancelled, shifted to "system-level" |
| Officially cancelled | 2026-05-14 | Intel earnings disclosed Falcon Shores cancellation |
| Pivot to Jaguar Shores | 2026-05 | New rack-scale system project confirmed |
| New direction | 2027-2028 | No longer competing with single-chip NVIDIA |
Reasons for cancellation:
- Third consecutive quarterly loss (2026-Q1) tightening R&D budget
- 200 PF single chip unrealistic (Rubin R200 = 50 PF already at limits)
- HBM supply constraints (Micron / SK Hynix prioritizing NVIDIA)
- Market shift to rack-scale (NVIDIA NVL72, AMD Helios dominant)
Jaguar Shores vs Falcon Shores
| Metric | Falcon Shores (Cancelled) | Jaguar Shores |
|---|
| Form Factor | Single-chip GPU | Rack-scale system |
| Target | Compete with B100 / B200 single chip | Compete with NVL72 / Helios racks |
| Process | Intel 18A | Intel 18A / 14A |
| HBM Capacity | 288 GB HBM3e | Rack-scale ~10-15 TB |
| FP8 Compute | 200 PFLOPS (pre-cancellation target) | Rack-scale ~160-320 PFLOPS |
| Integrated CPU | Standalone | Xeon 6/7 tightly coupled |
| Integrated NIC | None | 800G integrated |
| Launch | 2025 (cancelled) | 2027-2028 |
| Status | ❌ Cancelled | 🟡 Roadmap stage |
Intel AI Strategy Restructuring (May 2026)
| Strategy | Details |
|---|
| Gaudi product line | Continue Gaudi 3 / Gaudi 3E (sustained through 2026) |
| Falcon Shores | ❌ Cancelled |
| Jaguar Shores | ✅ Rack-scale AI system restarted |
| Foundry services | Intel Foundry 18A fab for NVIDIA / AMD / AWS |
| x86 leadership | Strengthen Xeon 6/7 (AI server CPU dominance) |
| Habana brand | Retained, Jaguar Shores integrates Gaudi IP |
Intel no longer directly competes with NVIDIA on AI GPUs:
- Short-term: Gaudi 3 sustained (cost-effective competition)
- Mid-term: Jaguar Shores system-level (rack-scale competition)
- Long-term: Intel Foundry 18A fab for AI vendors (Intel as "AI foundry")
Recommended Deployment Configurations
| Scenario | Recommended Config |
|---|
| Mid-to-large LLM training | Jaguar Shores rack + Xeon 6 hosts |
| Enterprise AI inference | Jaguar Shores single rack |
| Telco/Government localization | Intel ecosystem + Xeon + Gaudi IP |
| x86 ecosystem preservation | Suitable for existing Intel customers |
Use Cases
- ✅ Intel ecosystem retention (existing x86 customers)
- ✅ Rack-scale AI systems (2027+ era)
- ✅ Intel Foundry customers (18A process synergy)
- ❌ Current mainstream training (2026 still NVIDIA-dominated)
- ❌ Ultra-low latency inference (Groq LPX superior)
- ❌ High-density clusters (NVL72 / Helios already GA in 2026)
| Item | Details |
|---|
| Vendor | Intel Corporation |
| First Announced | 2026-05 (earnings disclosure) |
| Codename | Jaguar Shores |
| Form Factor | Rack-scale system (not a single chip) |
| Estimated Launch | 2027-2028 |
| Process | Intel Foundry 18A / 14A |
| Related Products | Gaudi 3 (IP source), Xeon 6/7 (CPU synergy) |