Tesla AI5 (In-House Automotive and Robotics AI Chip)
Product Overview
The AI5 is Tesla's in-house next-generation AI compute chip, serving both the vehicle autonomous driving (FSD) and Optimus humanoid robot product lines — the first time Tesla has a single chip carrying inference workloads for both vehicles and robots.
The supply chain is the most closely watched part of the AI5: Samsung's Taylor (Texas) fab started 2nm-class mass production in July 2026, after completing the AI5 tape-out; meanwhile, Tesla is placing part of the capacity at TSMC on 3nm, forming dual sourcing to avoid single-point dependency.
Musk's performance framing is that the AI5's energy efficiency is about 3x that of NVIDIA Blackwell at less than 10% of the cost. It should be noted that this is a product claim targeted at Tesla's own inference workloads, not a general-purpose benchmark result.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Tesla proprietary NPU (AI5) |
| Process Node | Samsung 2nm (Taylor fab) + TSMC 3nm (dual sourcing) |
| Positioning | Vehicle autonomous driving + humanoid robot inference |
| Compute | Not disclosed |
| Memory Capacity | Not disclosed |
| Memory Bandwidth | Not disclosed |
| TDP | Not disclosed |
| Mass Production Start | Samsung Taylor fab, 2026-07 |
| Expected In-Vehicle Deployment | Mid-2027 (production switchover) |
⚠️ Data Source Note: Tesla has not disclosed the AI5's compute, memory, or power parameters. The "3x energy efficiency / 10% cost" claims are Tesla product claims targeted at its own workloads and should not be treated as general-purpose benchmarks.
Why the AI5's Cadence Is Held Back by "Stockpiling"
Musk has stated explicitly that Tesla needs to stockpile hundreds of thousands of finished AI5 boards beside the production line before switching vehicle production — meaning there is roughly a one-year gap between chip mass production starting (2026-07) and actual in-vehicle deployment (mid-2027).
| Phase | Time | Notes |
|---|---|---|
| Tape-out complete | Before 2026-07 | Samsung completed the AI5 tape-out |
| Mass production start | 2026-07 | Samsung Taylor fab begins 2nm-class production |
| Stockpiling period | 2026-07 ~ mid-2027 | Hundreds of thousands of boards need to be accumulated |
| In-vehicle switchover | Mid-2027 | Vehicle production lines switch to the AI5 |
Tesla AI Chip Roadmap
| Chip | Generation | Foundry | Use | Status |
|---|---|---|---|---|
| Dojo D1 | Training | TSMC 7nm | Training cluster (since scaled back) | Released |
| Dojo v2 | Training | TSMC | Training cluster | Not mass-produced |
| AI5 | Inference | Samsung 2nm + TSMC 3nm | FSD + Optimus | Taped out, mass production started |
| AI6 | Inference | Samsung (Texas) | Next-generation vehicles and robots | Planned |
Key Insight: Tesla's chip strategy makes an interesting contrast with cloud providers — Google, AWS, Microsoft, and Meta all point their in-house chips at the data center, while Tesla's point at the edge and robots. The AI5 using an advanced node like 2nm/3nm for edge inference represents the new trend of "advanced processes moving down to the edge"; this also explains why Tesla locked in capacity at both Samsung and TSMC simultaneously.
Vendor Information
| Parameter | Value |
|---|---|
| Company | Tesla, Inc. |
| Headquarters | Austin, Texas, USA |
| In-House Chip Team | Tesla AI chip design team |
| Foundry | Samsung (Taylor fab, 2nm-class), TSMC (3nm) |
| Application Products | FSD (Full Self-Driving), Optimus (humanoid robot) |
Use Cases
- ✅ Vehicle autonomous driving inference (FSD)
- ✅ Humanoid robot inference (Optimus)
- ✅ Low-latency edge AI (no cloud round trip)
- ❌ Data center large-model training
- ❌ External sales (Tesla internal use only)
Related Cards
- Tesla Dojo D1 — Tesla training chip
- Tesla Dojo v2 — Training chip successor
- NVIDIA Drive Thor — Automotive AI compute competitor
- NVIDIA Jetson Thor — Robotics AI compute platform
- Full Comparison Table