Chips & Silicon in
Data Centers
Data centers are powered by silicon at every level: CPUs, GPUs, custom AI accelerators, networking ASICs, and DPUs. This page surveys the leading-edge chips driving today’s AI factories, hyperscale clouds, and sovereign data centers.
Hyperscaler In-House Silicon
| Company |
Chip |
Type / Role |
Foundry |
| Amazon / AWS |
Trainium2 • Inferentia2 • Graviton4 |
Training • Inference • Arm CPU |
TSMC 5nm |
| Google |
TPU v5p • Axion CPU |
Training ASIC • Cloud CPU |
TSMC 4nm / 3nm |
| Microsoft |
Maia 100 • Cobalt 100 |
AI accelerator • Arm CPU |
TSMC 5nm |
| Meta |
MTIA v2 |
AI training & inference accelerator |
TSMC 5nm |
| Apple |
M4 Ultra • A18 Pro |
Edge inference (PC, mobile) |
TSMC 3nm (N3/N3E) |
GPU & Accelerator Vendors
| Company |
Chip |
Type / Role |
Foundry |
| NVIDIA |
H100 (Hopper) • B100/B200 (Blackwell) • GB200 NVL72 |
Training GPUs • AI superchips |
TSMC 4N • TSMC 3nm (Blackwell) |
| AMD |
MI300X • MI325X |
AI accelerators |
TSMC 5nm + 6nm (CoWoS packaging) |
| Intel |
Gaudi 3 • Xeon 6 (Granite Rapids, Sierra Forest) |
AI accelerator • x86 CPUs |
Intel 3 / Intel 7 (self-fab) + TSMC packaging |
Alternative Architectures
| Company |
Chip |
Role |
Foundry |
| Cerebras |
Wafer Scale Engine 3 (WSE-3) |
Training accelerator (largest single chip) |
TSMC (7nm/5nm) |
| Graphcore |
Bow IPU |
Inference / training IPU |
TSMC 7nm |
| Tenstorrent |
Grayskull • Wormhole |
RISC-V CPUs + AI accelerators |
Samsung + TSMC |
| Ampere Computing |
AmpereOne |
Cloud-native Arm CPU |
TSMC 5nm |
Networking & DPUs
| Company |
Chip |
Role |
Foundry |
| Broadcom |
Tomahawk 5 • Jericho3 |
Networking ASICs (switch fabrics) |
TSMC 5nm |
| Marvell |
Teralynx • Octeon |
Networking ASICs + DPUs |
TSMC 5nm/7nm |
| NVIDIA |
BlueField DPUs • NVLink/NVSwitch |
Networking + composable infrastructure |
TSMC 7nm/5nm |
| Fungible (Microsoft) |
SmartNIC / DPUs |
Data processing + storage offload |
TSMC |
China & Sovereign Silicon
| Company |
Chip |
Role |
Foundry |
| Huawei |
Ascend 910B / 920 • Kirin 9000S |
AI training / inference • smartphones |
SMIC 7nm (DUV multipatterning) |
| Alibaba (T-Head) |
Yitian 710 • HanGuang |
Arm CPUs • AI accelerators |
TSMC (pre-sanctions) • now SMIC |
| Baidu |
Kunlun v2 |
AI inference accelerators |
Samsung + TSMC (mixed history) |
| Biren / Cambricon |
BR104 • Cambricon AI chips |
Training/inference ASICs |
SMIC (limited node capability) |
China’s Shift: Ban on Nvidia & Domestic Push
In September 2025, China reportedly ordered domestic companies to halt purchasing Nvidia chips and cancel existing orders — a landmark policy forcing migration toward Huawei’s Ascend/Atlas line.
- Immediate effect: Chinese hyperscalers and AI labs will phase out Nvidia H100/B200/GB200 GPUs in favor of Huawei silicon.
- Huawei roadmap: Ascend 950/960/970 and Atlas “supernode” clusters are positioned as drop-in substitutes for frontier training and inference workloads.
- Deployment impact: China’s future AI factories may evolve on parallel architectures, with unique interconnects, software stacks, and cluster topologies.
- Risks: Gaps in HBM supply, advanced packaging, and CUDA-class software ecosystems could constrain performance relative to Nvidia-based systems elsewhere.
Future / Next-Gen Chips (Watchlist)
Roadmaps shift fast; treat the entries below as expected/rumored until vendor release notes and foundry disclosures confirm specifics.
| Company |
Candidate |
Segment |
Node / Packaging (expected) |
Status / ETA |
Notes |
| NVIDIA |
“Rubin” (post-Blackwell) |
AI training/inference GPU |
TSMC N3 / advanced CoWoS |
2026 (rumored) |
Successor to Blackwell (B200/GB200); likely higher memory BW and NVLink evolution. |
| AMD |
Instinct MI400 series |
AI accelerators |
TSMC N3/N4 + 2.5D/3D packaging |
2026 (expected) |
Follows MI325X; focus on perf/W, memory capacity, and ecosystem maturity. |
| Intel |
Gaudi “next” / Xeon roadmap |
AI accel + x86 CPUs |
Intel 3 / 20A/18A packaging mix |
2025–2026 (rolling) |
Gaudi line continues; Xeon 6 families iterate for AI offload/DDR5/CXL scale. |
| AWS |
Trainium3 / Inferentia3 |
Training & inference ASICs |
TSMC N4/N3 (likely) |
2026 (speculative) |
Follows Trainium2/Inferentia2; focus on cluster-scale throughput. |
| Google |
TPU v6 / v6p |
Training ASIC |
TSMC N3 + HBM4 era |
2026 (speculative) |
Scale-out mesh, memory BW, and perf/W uplift for frontier model training. |
| Microsoft |
Maia 200 • Cobalt next |
AI accel • Arm CPU |
TSMC N4/N3 + CoWoS |
2025–2026 (expected) |
Second-gen Azure silicon; tighter fabric + memory advances. |
| Meta |
MTIA v3 |
AI accelerator |
TSMC N4/N3 |
2026 (rumored) |
Focus on inference density and Meta-specific operators. |
| Apple |
M5 / A19 family |
Edge inference SoCs |
TSMC N3E ? N2 (later) |
2025–2026 |
On-device inference uplift; ties to hybrid cloud/edge strategies. |
| Huawei |
Ascend 930 / next |
AI accelerators |
SMIC 7nm-class (DUV) • advanced packaging |
2025–2026 (domestic) |
Sanctions-bounded cadence; performance rising within node limits. |
| Cerebras |
WSE-Next |
Wafer-scale training |
TSMC N5?N3 (possible) |
2026+ (speculative) |
Larger on-wafer SRAM and yield tricks for larger context models. |
| Broadcom |
Tomahawk 6 • Jericho3-AI (evol.) |
Switching ASICs |
TSMC N3/N4 |
2025–2026 |
AI fabric scale to 100k+ GPU pods; optical-friendly roadmaps. |
| Marvell |
Teralynx next • Octeon next |
Switch + DPU |
TSMC N5/N3 |
2025–2026 |
AI spine/leaf throughput, DPUs for storage/IO offload. |
| NVIDIA |
BlueField-4 • NVLink-next |
DPU + interconnect |
TSMC N5/N4 |
2025–2026 |
Higher line-rate security/IO offload; tighter GPU coherency fabrics. |
| Ampere |
AmpereOne next |
Cloud Arm CPU |
TSMC N3/N4 |
2026 |
Perf/W gains for stateless cloud workloads; CXL features. |
| Tenstorrent |
Next-gen RISC-V + AI |
CPU/accelerator |
Samsung/TSMC (TBD) |
2026+ |
Modular tiles; licensing model to third parties. |
Key Takeaways
- NVIDIA remains the dominant training silicon provider (Blackwell generation).
- AMD and Intel are challengers, gaining traction with cost, openness, or x86 incumbency.
- Hyperscalers (AWS, Google, Microsoft, Meta) are actively reducing reliance on NVIDIA with custom chips.
- China’s ecosystem is forced onto SMIC (7nm-class), limiting competitiveness but sustaining sovereignty.
- Networking & DPUs (Broadcom, Marvell, NVIDIA) are as critical as compute — they define AI cluster coherence.
- Alternative architectures (Cerebras, Graphcore, Tenstorrent) push boundaries but face adoption hurdles.