On February 2026, SemiAnalysis asked whether Etched's Sohu chip will support TinyGrad on Day 0. Etched claims 10x performance over GPUs for transformer inference, but has not confirmed TinyGrad compatibility.
Key facts
- SemiAnalysis tweeted on February 2026
- Etched claims 10x performance over GPUs
- TinyGrad has 8,000 GitHub stars
- TinyGrad codebase is under 5,000 lines
- Etched has not confirmed TinyGrad support
On February 2026, SemiAnalysis tweeted a pointed question: "Does anyone know whether the @Etched chip will support @tinygrad out of the box on Day 0 of General Availability?" The query, tagging both Etched and TinyGrad, underscores a critical software ecosystem gap for the transformer-specialized chip.
Etched's Sohu chip is designed to run transformer models like GPT-4 at 10x the speed of GPUs, a claim the company has made publicly. But the chip's success hinges on software adoption. TinyGrad, an open-source deep learning framework with 8,000 GitHub stars, is known for its simplicity and portability, making it a potential bridge for developers.
The Software Ecosystem Question
Etched has not confirmed TinyGrad support on Day 0. The company's blog post focuses on hardware performance, but the software stack remains opaque. This is a common pitfall for AI chips: without a robust software ecosystem, even the fastest silicon fails to gain traction.
TinyGrad's appeal lies in its minimal codebase—less than 5,000 lines of Python—and its ability to compile to multiple backends. If Etched can integrate TinyGrad, it could lower the barrier for developers to test Sohu. But the lack of a public commitment raises concerns.
The company has not disclosed a specific GA date, but early 2026 was the target. If Sohu ships without TinyGrad support, it risks alienating a community that values openness. However, Etched may prioritize its proprietary compiler, which could offer better performance but less flexibility.
What This Means for Developers
For developers evaluating Sohu, Day-0 TinyGrad support would signal a commitment to open standards. Without it, they must rely on Etched's proprietary tools, which may be less documented or harder to debug. This could slow adoption, especially among researchers who favor TinyGrad's simplicity.
SemiAnalysis's question highlights a broader trend: AI chip startups must win on software, not just hardware. The winner will be the one that reduces friction for developers. Etched's response—or silence—will be telling.
The company has not responded publicly to the tweet, but the clock is ticking. As GA approaches, expect more questions about software readiness.
Key Takeaways
- SemiAnalysis questions Etched Sohu's Day-0 TinyGrad support.
- Etched claims 10x performance but hasn't confirmed software compatibility, a critical adoption factor.
What to watch
Watch for Etched's official response to SemiAnalysis, and the GA announcement. If they confirm TinyGrad support, it signals an open-ecosystem strategy. If not, expect community backlash. Also track TinyGrad's GitHub activity for any Etched-related commits.
