Microsoft Azure Cobalt 200: The 3nm 132-Core ARM Monster That’s Kicking x86 Out of Azure
Tuesday, December 09, 2025Microsoft Azure Cobalt 200: The 3nm 132-Core ARM Monster That’s Kicking x86 Out of Azure
In the cut-throat world of hyperscale cloud, every watt and every dollar counts. At Ignite 2025, Microsoft dropped a bombshell: the Azure Cobalt 200 — a fully in-house designed, TSMC 3nm, 132-core ARM server CPU that’s already rolling out across Azure data centers and is set to replace AMD and Intel x86 processors on a massive scale.
Yes, you read that right. The company that helped make x86 the undisputed king of computing is now aggressively moving its own cloud empire to custom ARM silicon — and the performance-per-watt numbers are brutal.
From Experiment to Empire: The Cobalt Timeline
- 2023 – Cobalt 100 (5nm, Neoverse N2) quietly launches as Microsoft’s first production ARM CPU for Azure
- 2024 – Cobalt 100 proves itself in real workloads, Microsoft starts internal migration
- November 2025 – Cobalt 200 announced: 3nm, Neoverse V3, up to 50% higher performance than Cobalt 100 at the same power
- Now – 2026 – Cobalt 200 instances coming to public preview, full customer availability in 2026
Under the Hood: Cobalt 200 Technical Deep-Dive
The Cobalt 200 is a dual-chiplet design with a total of 132 Arm Neoverse V3 cores (66 per chiplet). Each core gets:
- 3 MB private L2 cache (huge for server workloads)
- Per-core DVFS (independent voltage/frequency scaling)
- Dedicated hardware accelerators for compression, encryption, and data analytics
Shared resources include a massive 192 MB L3 cache and a 12-channel DDR5 memory interface — matching the best x86 offerings while running on a far more efficient architecture.
| Feature | Azure Cobalt 200 | Cobalt 100 | AMD EPYC Genoa/Zen 4 (typical) |
|---|---|---|---|
| Process Node | TSMC 3nm | TSMC 5nm | 5nm / 4nm |
| Cores | 132 × Neoverse V3 | ~64 × Neoverse N2 | Up to 128–192 × Zen 4 |
| L2 Cache per Core | 3 MB | 1–2 MB | 1 MB |
| Shared L3 | 192 MB | Lower | Up to 384 MB |
| Memory Channels | 12 × DDR5 | 8–10 × DDR5 | 12 × DDR5 |
| Perf vs Previous Gen | Up to +50% | — | — |
| Target | Cloud-native scale-out | General cloud | Broad enterprise |
Why This Is a Nightmare for Intel & AMD
x86 dominated servers because it had to support everything from 1980s DOS to modern AI training. Hyperscalers don’t need that baggage. They run highly optimized, containerized, modern codebases — perfect for a clean-slate RISC architecture like Arm.
Microsoft, Amazon (Graviton), Google (Axion), and soon Meta and others are all moving massive portions of their fleets to custom ARM. That’s well over 60% of global server CPU shipments that could flip away from x86 in the next 5–7 years.
What This Means for PC Builders & Enthusiasts
Short term? Nothing changes — your gaming rig and workstation will stay x86 for years. But the writing is on the wall:
- ARM is already crushing in laptops (Apple M-series, Snapdragon X Elite)
- Server economics will push Arm desktop SoCs harder
- Windows on ARM is finally maturing — native apps are pouring in
We’re heading toward a world with real architectural choice again — x86 for legacy and maximum single-thread grunt, ARM for efficiency and future-proofing.
Final Verdict
The Azure Cobalt 200 isn’t just another server CPU. It’s Microsoft firing a cannon across the bow of the x86 empire, built on real telemetry from running the world’s second-largest cloud.
If early internal benchmarks hold up — and Microsoft rarely overpromises on this stuff — we’re looking at one of the most important server chips of the decade.
x86 isn’t dead. But for the first time in 40 years, it’s no longer inevitable.
Let the silicon wars begin.