ELON STACK

THE VERTICAL
VS
THE PLATFORM

Dojo is not a NVIDIA competitor. It's a purpose-built chip for one problem: training FSD on 10 million miles of daily driving video. The question is whether vertical integration beats ecosystem scale. The answer is: both are winning, at different things.

● Tesla Dojo● NVIDIA

D1 VS H100 — CHIP SPECS

SPEC
TESLA DOJO
NVIDIA H100
Chip
Dojo D1
H100 SXM5
TFLOPS (FP16)
362
1,979
Mem Bandwidth
10 TB/s / tile
3.35 TB/s
Mem Size
32 GB / tile
80 GB
Unit Cost
~$1,300 / chip
~$35,000 / chip
Purpose
FSD video training
General AI
External sales
No (Tesla only)
Yes (anyone)
Ecosystem
Closed (Tesla)
CUDA — universal
Software stack
Custom compiler
CUDA + cuDNN + NCCL
Use in production
FSD training only
Training + inference

SYSTEM-LEVEL COMPARISON

Chip → full cluster. TFLOPS = FP16 system-level estimate.

Dojo Exapod
Tesla
1.1EFLOP
3,000 chips

3,000 D1 chips. 1.1 EFLOP. Purpose-built for video processing.

DGX H100
NVIDIA
32PFLOP
8 chips

8x H100 SXM. 32 PFLOP. General-purpose, ships to anyone.

DGX H200
NVIDIA
32PFLOP
8 chips

8x H200 SXM. Same FLOPS as H100 but 141GB HBM3e memory each.

Dojo v2 (est.)
Tesla
4.0EFLOP
specs TBD

Next-gen Dojo. Announced but not yet shipped at scale.

REALITY CHECKS

"Dojo will replace NVIDIA for Tesla"
PARTIAL

Dojo handles training. Tesla still buys H100s for inference and as training overflow. Dojo is complementary, not a replacement. Tesla is one of NVIDIA's largest customers.

"Dojo is cheaper per FLOP than H100"
TRUE

$1,300 per D1 chip vs $35,000 per H100. But the D1 has 5.5x fewer TFLOPS and much less memory. The per-FLOP math gets complicated quickly — and Dojo can only be used for specific video workloads.

"Dojo is a real threat to NVIDIA's business"
FALSE

Dojo is a vertical integration play for Tesla's specific problem — FSD video training. It's not sold externally, has no CUDA ecosystem, and cannot run general LLMs. NVIDIA has zero customer risk from Dojo.

"Tesla has a data advantage with Dojo"
TRUE

10 million miles of driving video per day from 6M+ vehicles. Dojo is the only chip designed specifically to process that kind of data at speed. No one else has this volume of real-world physical AI training data.

"FSD will reach autonomy without NVIDIA chips"
PARTIAL

Dojo trains the FSD model. But Tesla's vehicles run inference on NVIDIA chips (Drive Orin/Thor). The training vs inference distinction is critical — Tesla designed Dojo for training, not for the car.

FSD PROGRESS TIMELINE

2016 · SETBACK
Tesla Autopilot NHTSA crash investigation — hardware limitations exposed
2019 · SETBACK
Elon predicts 1M robotaxis by 2020. Doesn't happen.
2021 · PROGRESS
FSD beta launches — city streets, not just highway. Real progress.
2022 · PROGRESS
Dojo supercomputer construction begins in New York
Aug 2023 · PROGRESS
Dojo Exapod begins training runs. 1.1 EFLOP cluster online.
2024 · PROGRESS
FSD v12 — neural network end-to-end. Hands-free on city streets.
2025 · PROGRESS
Cybercab announced. Autonomous rides begin in select US cities.
2026 · PROGRESS
Dojo v2 in development. Training scale increasing. FSD globally expanding.
THE REAL QUESTION

Dojo's success isn't measured by whether it beats NVIDIA at general AI compute. It's measured by whether it trains FSD fast enough to reach full autonomy before a competitor does.

If FSD achieves full autonomy, Tesla's 6-million-vehicle fleet becomes a robotaxi network generating Uber-scale revenue with zero driver cost. That's the trillion-dollar outcome Dojo is actually pointing at — not beating NVIDIA in the data center market.