AI Chip GPUs
Data Center GPU/Accelerator Shipments (Estimate)
Sterling estimate: NVIDIA, AMD and Intel do not publish unit shipments, and no tracker prints these counts. Quarterly shipments of AI accelerators: the specialized chips (GPUs, TPUs, custom ASICs) that train and run AI models in data centers. NVIDIA captures over 80% of unit volume; AMD, Intel, and Google's TPUs trail well behind. Figures in thousands of units.
AI Accelerator Architecture Comparison
Side-by-side comparison of AI training and inference accelerators: specs, memory, performance, and pricing
- Node
- 4nm (TSMC)
- Transistors
- 80B
- Memory
- 80GB HBM3
- Mem BW
- 3.35 TB/s
- FP16 TFLOPS
- 1,979
- FP8 TFLOPS
- 3,958
- TDP
- 700W
- Launched
- 2023-Q1
- Interconnect
- NVLink 4.0 (900 GB/s)
- Price Est.
- $25,000–$40,000
- Node
- 4nm (TSMC)
- Transistors
- 208B
- Memory
- 192GB HBM3E
- Mem BW
- 8 TB/s
- FP16 TFLOPS
- 5,000
- FP8 TFLOPS
- 10,000
- TDP
- 1000W
- Launched
- 2024-Q4
- Interconnect
- NVLink 5.0 (1.8 TB/s)
- Price Est.
- $30,000–$50,000
- Node
- 4nm (TSMC)
- Transistors
- 208B (x72)
- Memory
- 13.5TB HBM3E
- Mem BW
- 576 TB/s
- FP16 TFLOPS
- 162K
- FP8 TFLOPS
- 1440K
- TDP
- 120kW
- Launched
- 2025-Q1
- Interconnect
- NVLink 5.0 (full rack)
- Price Est.
- $2M–$3M (rack)
- Node
- 5nm/6nm (TSMC)
- Transistors
- 153B
- Memory
- 192GB HBM3
- Mem BW
- 5.3 TB/s
- FP16 TFLOPS
- 1,307
- FP8 TFLOPS
- 2,614
- TDP
- 750W
- Launched
- 2023-Q4
- Interconnect
- Infinity Fabric (896 GB/s)
- Price Est.
- $10,000–$15,000
- Node
- 5nm/6nm (TSMC)
- Transistors
- 153B
- Memory
- 256GB HBM3E
- Mem BW
- 6 TB/s
- FP16 TFLOPS
- 1,307
- FP8 TFLOPS
- 2,614
- TDP
- 750W
- Launched
- 2024-Q4
- Interconnect
- Infinity Fabric (896 GB/s)
- Price Est.
- $15,000–$20,000
- Node
- 3nm (TSMC)
- Transistors
- 185B
- Memory
- 288GB HBM3E
- Mem BW
- 8 TB/s
- FP16 TFLOPS
- 5,000
- FP8 TFLOPS
- 10,000
- TDP
- 1000W
- Launched
- 2025-Q2
- Interconnect
- UALink / Infinity Fabric 4.0
- Price Est.
- TBD
- Node
- 7nm (TSMC)
- Transistors
- ~25B
- Memory
- 96GB HBM2E
- Mem BW
- 2.46 TB/s
- FP16 TFLOPS
- 432
- FP8 TFLOPS
- 864
- TDP
- 600W
- Launched
- 2023-Q2
- Interconnect
- 24x 100GbE RDMA
- Price Est.
- $12,000–$15,000
- Node
- 5nm (TSMC)
- Transistors
- ~50B
- Memory
- 128GB HBM2E
- Mem BW
- 3.7 TB/s
- FP16 TFLOPS
- 1,835
- FP8 TFLOPS
- 3,670
- TDP
- 900W
- Launched
- 2024-Q4
- Interconnect
- 24x 200GbE RDMA
- Price Est.
- $15,000–$20,000
- Node
- 7nm (TSMC)
- Transistors
- ~20B
- Memory
- 16GB HBM2E
- Mem BW
- 820 GB/s
- FP16 TFLOPS
- 197
- FP8 TFLOPS
- 394
- TDP
- 200W
- Launched
- 2023-Q3
- Interconnect
- ICI (Inter-Chip Interconnect)
- Price Est.
- Cloud-only
- Node
- 5nm (TSMC)
- Transistors
- ~35B
- Memory
- 32GB HBM3
- Mem BW
- 1.6 TB/s
- FP16 TFLOPS
- 460
- FP8 TFLOPS
- 920
- TDP
- 300W
- Launched
- 2024-Q4
- Interconnect
- ICI v2
- Price Est.
- Cloud-only
| Product | Architecture | Node | Transistors | Memory | Mem BW | FP16 TFLOPS | FP8 TFLOPS | TDP | Interconnect | Launched | Price Est. |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Hopper | 4nm (TSMC) | 80B | 80GB HBM3 | 3.35 TB/s | 1,979 | 3,958 | 700W | NVLink 4.0 (900 GB/s) | 2023-Q1 | $25,000–$40,000 | |
| Blackwell | 4nm (TSMC) | 208B | 192GB HBM3E | 8 TB/s | 5,000 | 10,000 | 1000W | NVLink 5.0 (1.8 TB/s) | 2024-Q4 | $30,000–$50,000 | |
| Blackwell | 4nm (TSMC) | 208B (x72) | 13.5TB HBM3E | 576 TB/s | 162K | 1440K | 120kW | NVLink 5.0 (full rack) | 2025-Q1 | $2M–$3M (rack) | |
| CDNA 3 | 5nm/6nm (TSMC) | 153B | 192GB HBM3 | 5.3 TB/s | 1,307 | 2,614 | 750W | Infinity Fabric (896 GB/s) | 2023-Q4 | $10,000–$15,000 | |
| CDNA 3.5 | 5nm/6nm (TSMC) | 153B | 256GB HBM3E | 6 TB/s | 1,307 | 2,614 | 750W | Infinity Fabric (896 GB/s) | 2024-Q4 | $15,000–$20,000 | |
| CDNA 4 | 3nm (TSMC) | 185B | 288GB HBM3E | 8 TB/s | 5,000 | 10,000 | 1000W | UALink / Infinity Fabric 4.0 | 2025-Q2 | TBD | |
| Gaudi | 7nm (TSMC) | ~25B | 96GB HBM2E | 2.46 TB/s | 432 | 864 | 600W | 24x 100GbE RDMA | 2023-Q2 | $12,000–$15,000 | |
| Gaudi | 5nm (TSMC) | ~50B | 128GB HBM2E | 3.7 TB/s | 1,835 | 3,670 | 900W | 24x 200GbE RDMA | 2024-Q4 | $15,000–$20,000 | |
| Custom ASIC | 7nm (TSMC) | ~20B | 16GB HBM2E | 820 GB/s | 197 | 394 | 200W | ICI (Inter-Chip Interconnect) | 2023-Q3 | Cloud-only | |
| Custom ASIC | 5nm (TSMC) | ~35B | 32GB HBM3 | 1.6 TB/s | 460 | 920 | 300W | ICI v2 | 2024-Q4 | Cloud-only |
AI Chip Product Roadmap
Product availability timeline for AI accelerators across NVIDIA, AMD, Intel, and Broadcom/Google
- 2020 Q2 to 2023 Q4
A100
7nm
- 2022 Q4 to 2025 Q2
H100
4nm
- 2024 Q1 to 2026 Q2
H200
4nm
- from 2024 Q4
B100/B200
4nm
- from 2025 Q1
GB200 NVL72
4nm
- from 2026 Q1
Rubin (R100)
3nm
- from 2028
Feynman
Node TBD
- 2022 Q1 to 2024 Q2
MI250X
6nm
- from 2023 Q4
MI300X
5nm
- from 2024 Q4
MI325X
5nm
- from 2025 Q2
MI350X
3nm
- from 2026 H2
MI400
2nm
- 2023 Q2 to 2025 Q1
Gaudi 2
7nm
- from 2024 Q3
Gaudi 3
5nm
- from 2026 H1
Jaguar Shores
Intel 18A
- from 2023 Q3
Google TPU v5e
7nm
- from 2024 Q4
Google TPU v6e (Trillium)
5nm
- from 2025 Q4
Google TPU v7 (Ironwood)
Node not disclosed
- from 2024 Q3
Meta MTIA v2
5nm
Data Center GPU + Accelerator Revenue by Vendor (Quarterly)
Quarterly data center AI accelerator revenue by vendor (NVDA, AMD, INTC, AVGO, MRVL). NVIDIA's Data Center line is by far the largest.
NVIDIA Data Center Compute vs Networking
NVIDIA-disclosed Data Center compute versus networking, only for fiscal quarters where both lines were printed. Networking (InfiniBand, Spectrum-X, NVLink fabric) was $14.8B in FY27 Q1, almost 20% of Data Center.
Hyperscaler Custom AI Accelerators
Company-stated custom accelerators (Google TPU, AWS Trainium, Azure Maia, Meta MTIA) plus Intel Foundry external customer revenue. This is a landscape table, not a shipment ranking. Foundry names appear only when the cited page states them.
TPU Ironwood (7th gen)
- Owner
- What the owner has said
- Google unveiled Ironwood, its 7th-generation TPU built for inference and training at scale, on April 9, 2025. TPU7x, the first Ironwood release, entered Preview on Google Cloud on November 24, 2025 and became generally available on March 31, 2026, with 192 GB of HBM per chip
- Foundry (only if disclosed)
- Not disclosed by Google; TSMC is widely assumed and is not treated as a Google disclosure here
Trainium2 / Trn2
- Owner
- Amazon Web Services
- What the owner has said
- AWS documents Trainium2 as the current-generation training chip for Amazon EC2 Trn2 instances and UltraServers, sold as an NVIDIA alternative for AWS-hosted training
- Foundry (only if disclosed)
- Not stated on the cited AWS product page
Azure Maia 200
- Owner
- Microsoft
- What the owner has said
- Microsoft introduced Maia 200, the successor to Maia 100, on January 26, 2026 as an inference accelerator with 216 GB of HBM3e. It is deployed in Azure's US Central region near Des Moines, Iowa, with US West 3 near Phoenix, Arizona, named as next, and serves models including OpenAI's GPT-5.2
- Foundry (only if disclosed)
- TSMC 3nm (stated by Microsoft)
MTIA
- Owner
- Meta
- What the owner has said
- Meta's engineering blog describes MTIA as a family of training and inference accelerators for ranking and recommendation, now in production inside Meta data centers
- Foundry (only if disclosed)
- Meta has publicly worked with Broadcom on custom silicon; the cited MTIA blog is about the chip, not a foundry share figure
Intel Foundry external AI silicon
- Owner
- Intel Foundry customers
- What the owner has said
- Intel's Q2 2026 10-Q: third-party foundry and assembly/test revenue was $293 million in the quarter (vs $22 million a year earlier). That is the disclosed external foundry print, not a named list of AI chips
- Foundry (only if disclosed)
- Intel Foundry
| Product | Owner | What the owner has said | Foundry (only if disclosed) |
|---|---|---|---|
| TPU Ironwood (7th gen) | Google unveiled Ironwood, its 7th-generation TPU built for inference and training at scale, on April 9, 2025. TPU7x, the first Ironwood release, entered Preview on Google Cloud on November 24, 2025 and became generally available on March 31, 2026, with 192 GB of HBM per chip | Not disclosed by Google; TSMC is widely assumed and is not treated as a Google disclosure here | |
| Trainium2 / Trn2 | Amazon Web Services | AWS documents Trainium2 as the current-generation training chip for Amazon EC2 Trn2 instances and UltraServers, sold as an NVIDIA alternative for AWS-hosted training | Not stated on the cited AWS product page |
| Azure Maia 200 | Microsoft | Microsoft introduced Maia 200, the successor to Maia 100, on January 26, 2026 as an inference accelerator with 216 GB of HBM3e. It is deployed in Azure's US Central region near Des Moines, Iowa, with US West 3 near Phoenix, Arizona, named as next, and serves models including OpenAI's GPT-5.2 | TSMC 3nm (stated by Microsoft) |
| MTIA | Meta | Meta's engineering blog describes MTIA as a family of training and inference accelerators for ranking and recommendation, now in production inside Meta data centers | Meta has publicly worked with Broadcom on custom silicon; the cited MTIA blog is about the chip, not a foundry share figure |
| Intel Foundry external AI silicon | Intel Foundry customers | Intel's Q2 2026 10-Q: third-party foundry and assembly/test revenue was $293 million in the quarter (vs $22 million a year earlier). That is the disclosed external foundry print, not a named list of AI chips | Intel Foundry |