ETCHED
SOHU // Transformer-Only ASIC
The most audacious bet in AI hardware. Founded by Harvard dropouts who saw transformers taking over the world in 2022, Etched burned the architecture directly into silicon. No CNNs. No RNNs. No flexibility. Just transformers—at 20x the speed of NVIDIA.
BURNED INTO SILICON
The Bet
Why Etched believes transformers will dominate forever
In 2022, when Etched was founded, transformers were just one of many neural network architectures. CNNs dominated vision. RNNs handled sequences. The founders made a contrarian bet that transformers would take over everything—and designed a chip that can do nothing else.
✓ If Transformers Win
The upside is extraordinary:
- 20x performance advantage over general-purpose GPUs
- Dramatically lower cost per inference
- Power efficiency enables applications GPUs can't handle
- First-mover advantage in transformer-specific hardware
- Every AI company becomes a potential customer
✗ If Transformers Lose
The downside is total:
- Chips become completely worthless overnight
- No pivot possible—architecture is hardcoded
- 3+ years to design new chip for new architecture
- $120M+ investment written off entirely
- Company likely fails if major architecture shift occurs
❌ What Sohu CANNOT Run
Technical Specifications
Sohu ASIC Performance Claims
Performance Comparison: Llama-70B Inference
Company Timeline
From Harvard dorms to silicon
Strategic Analysis
The highest-stakes bet in AI hardware
✦ Bull Case
- Transformers have won every domain: text, vision, audio, video, robotics
- ChatGPT, Gemini, Claude, DALL-E, Sora—all transformers
- Network effects: more transformers → more training data → better transformers
- 20x performance advantage is genuinely transformative for real-time AI
- Humanoid robots need edge inference Sohu can provide
- If right, Etched becomes essential AI infrastructure
⚠ Bear Case
- Mamba, RWKV, and SSMs showing promise as transformer alternatives
- Chip hasn't taped out—performance claims are theoretical
- Memory bandwidth may still be limiting factor despite claims
- NVIDIA Blackwell narrows the performance gap
- 3-year chip design cycle means can't adapt to architecture changes
- Binary outcome: massive success or total failure, no middle ground
◎ Competitive Position
- Only company with transformer-only ASIC in development
- Groq and others have transformer optimizations but retain flexibility
- NVIDIA's generality is both strength and weakness
- If successful, creates new category of AI accelerator
- First-mover advantage if transformers remain dominant
- Potential acquisition target for hyperscalers wanting inference edge
⟁ Investment View
- Classic venture bet: low probability, extreme outcome
- Peter Thiel investment signals contrarian conviction
- $120M Series A is meaningful but not enough for multiple pivots
- Secondary valuations of ~$1.5B imply market believes in thesis
- Success could mean 10-50x returns; failure means zero
- Not for risk-averse investors—this is pure speculation on AI's future