OpenAI Reveals Custom SKIP Chip, Early Tests Show H100-Beating Performance
OpenAI has disclosed initial clinical test results of its custom AI chip SKIP, demonstrating single-GPU performance exceeding NVIDIA’s H100. The chip demonstrates significant advantages in mixed-precision inference operations. Key facts:
- Test completion: Clinical testing recently concluded
- Initial deployment: Q3 2026 within selected internal model inference services
- Public availability: Expected Q1 2027 for enterprise partners
- Model support: Limited to OpenAI internal models only; third-party model compatibility not planned
- Design focus: Inference-optimized, not designed for training workloads
Architecture and Testing Details
SKIP chip is fabricated on a 5nm process node, employing a distributed array architecture that replaces traditional GPU streaming multiprocessors with dedicated inference units. Testing used the MilliSpeed benchmark suite covering 12 tasks across three domains: LLM inference, image generation, and speech recognition.
Critical Test Results:
- Lamp-3.1-70B inference: 820 tokens/sec (SKIP) vs. 680 tokens/sec (H100)
- Logic reasoning (Sophon-GT dataset): 78.3% accuracy (SKIP) vs. 69.5% (H100)
- Power consumption: 280W average (SKIP) vs. 700W (H100), ~2.5x efficiency improvement
- P99 latency: 42ms (SKIP) vs. 68ms (H100)
The most surprising finding involve reasoning accuracy against expectations—SKIP achieves superior performance on complex logic and coding problems, challenging the prevailing assumption that “reasoning tasks demand high-bandwidth memory” and demonstrating the effectiveness of specialized architecture.
Chip Comparison
| Dimension | SKIP (OpenAI Custom) | H100 (NVIDIA) | H200 (NVIDIA) |
|---|---|---|---|
| Fabrication | 5nm | 4NP | 4N |
| Memory | 96GB HBM3 | 80GB HBM3 | 141GB HBM3 |
| Power | 280W | 700W | 750W |
| llama-3.1-70B Throughput | 820 tokens/sec | 680 tokens/sec | 750 tokens/sec |
| Financial Doc QA Accuracy | 86.1% | 82.3% | 84.7% |
| Per-GPU Price | Internal only | $30,000 | $45,000 |
| Model Compatibility | OpenAI internal only | Full ROCm生态 support | Full ROCm生态 support |
Note: H200 specifications from NVIDIA official documentation; SKIP pricing not publicly disclosed as internal deployment only.
Deployment Considerations
Adopt immediately if:
- Your service relies heavily on OpenAI enterprise API with strict latency requirements (e.g., real-time customer support)
- Your organization has signed priority deployment agreements with OpenAI
- Deploying large models like llama-3.1-70B to edge devices (low power advantage critical)
Wait before adopting if:
- You develop smaller open-weights models (SKIP lacks standard framework compatibility)
- Your workload involves model training (SKIP inference-only)
- Your team lacks experience in inference optimization (SKIP’s sparse computation requires model distillation for optimal results)
Final Thoughts
SKIP represents a shift toward specialized AI hardware, yet its closed ecosystem may slow industry standardization. When chips serve only proprietary models, ——————————————————– scalability and cross-industry adoption face structuralconstraints—a challenge for OpenAI moving forward.