The Blackwell-to-Rubin transition is the most consequential AI hardware event of 2026. NVIDIA Blackwell unit shipments are projected to fall from approximately 5.2 million in 2025 to approximately 1.8 million in 2026 as the product line transitions. NVIDIA Rubin has a 5.7 million stated target capped near 300,000 by TSMC N3 capacity reality. AMD MI400 and MI450 ramp into Meta\u2019s first gigawatt deployment. This page tracks the unit projections, foundry constraints, and customer allocation.
Key Findings
- NVIDIA Blackwell (B200, GB200, B300) shipments are projected at approximately 1.8 million units in 2026, down from approximately 5.2 million in 2025 as the product line transitions to Rubin.
- NVIDIA Rubin (R100, R200) has a 5.7 million stated unit target but is realistically capped near 200,000 to 300,000 in 2026 by TSMC N3 wafer-out capacity.
- AMD MI400 ramps to approximately 258,000 units in 2026 from an effective zero base in 2025; MI450 is the chip in Meta\u2019s first gigawatt deployment in H2 2026.
- Intel Gaudi 3 shipments remain modest at approximately 80,000 to 120,000 units in 2026; Falcon Shores is delayed to 2027.
- Combined data centre AI accelerator unit shipments across NVIDIA, AMD, Intel, plus hyperscaler custom silicon (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) are estimated at approximately 6.5 million units in 2026, a slight decline from 2025 driven by the NVIDIA generational transition timing.
2026 AI Accelerator Shipment Projections
| Vendor | Product Line | 2025 Units | 2026 Units (projected) |
|---|---|---|---|
| NVIDIA | Blackwell (B200, GB200, B300) | ~5.2M | ~1.8M |
| NVIDIA | Rubin (R100, R200) | 0 | ~250-300k |
| NVIDIA | Hopper (H100, H200) remaining | ~1.5M | ~0.4M (end-of-life) |
| NVIDIA | RTX Pro / DGX Spark / workstation | ~0.6M | ~0.7M |
| AMD | MI300X / MI325X | ~250k | ~280k |
| AMD | MI355X / MI400 / MI450 | ~0k | ~258k |
| Intel | Gaudi 3 | ~60k | ~100k |
| TPU v6 / TPU v7 | ~700k | ~900k | |
| AWS | Trainium 2 / Inferentia 3 | ~400k | ~600k |
| Microsoft | Maia 2 | ~120k | ~250k |
| Meta | MTIA 2 | ~80k | ~180k |
| Cerebras / Groq / Etched / Tenstorrent | Niche | ~5k | ~12k |
Foundry Constraint Analysis
| Constraint | Impact |
|---|---|
| TSMC N3 wafer-out | Limits Rubin ramp to approximately 250-300k units in 2026 |
| CoWoS-L advanced packaging | Approximately 70k wafers/mo target end-2026; constraint for HBM-stacked GPUs |
| HBM3E and HBM4 supply | Samsung HBM3E ramp pushing into Q2 2026; HBM4 Micron / SK Hynix first samples Q3 2026 |
| GDDR7 | Mostly resolved Q1 2026 after Q3-Q4 2025 shortage |
| InFO-PoP advanced packaging | Apple-priority; constrains other consumer SoC ramps |
AMD MI400 / MI450 Detail
AMD\u2019s MI400 began limited customer sampling in late 2025 and ramped to volume shipment in H1 2026. The MI450, the next-generation training-focused part, is the chip in Meta\u2019s first gigawatt deployment scheduled for H2 2026. Meta is the lead customer for MI450; Microsoft and Oracle are reported as second-tier MI400 and MI450 customers. The AMD roadmap is increasingly credible as a parallel scaling track to NVIDIA, with the ROCm software stack improvements in 2025-2026 closing the gap on PyTorch and JAX compatibility.
Customer Allocation Trends
| Customer | Primary 2026 Mix | Strategic Note |
|---|---|---|
| OpenAI / Stargate | NVIDIA Blackwell + Rubin transition | SpaceX 220k GPU deal disclosed May 6 |
| Microsoft | NVIDIA primary + Maia 2 + AMD MI400 | Multi-vendor diversification |
| Meta | AMD MI450 lead + NVIDIA primary + MTIA 2 | First-GW deployment on AMD H2 2026 |
| Amazon | NVIDIA + Trainium 2 internal | Anthropic anchor on Trainium |
| TPU primary + NVIDIA secondary | TPU v7 ramp | |
| xAI Colossus 2 | NVIDIA Blackwell + Rubin | Memphis expansion |
| Anthropic | AWS Trainium + Google TPU + NVIDIA | Triple-vendor |
| Tesla Dojo + xAI | NVIDIA + Tesla custom (D2) | D2 ramp delayed |
Brand Visibility Implications
GPU and AI accelerator coverage drives a sustained AI assistant query stream from technical buyers, investors, and procurement teams. Brands selling adjacent products (advanced packaging, HBM, GDDR, liquid cooling, power delivery, server OEM integration, ROCm/CUDA developer tools) face strong AI-mediated discovery surface for procurement-research prompts.
Methodology
Shipment figures compiled from NVIDIA investor relations, AMD investor relations, hyperscaler capex disclosures, and analyst reports from TrendForce, Mercury Research, and Bernstein. Some figures are estimated where official disclosures are partial. Updated quarterly.
How Presenc AI Helps
Presenc AI tracks brand-mention rates on AI accelerator queries across ChatGPT, Claude, Gemini, and Perplexity. For brands in adjacent semiconductor categories, the platform identifies the queries driving procurement research and the gaps where new content investment unlocks share of voice.