'Sohu,' developed by Etched, is a dedicated chip designed exclusively for Transformer models, offering faster, cheaper, and more efficient AI performance than general-purpose GPUs.
Imagine this: You wake up in the morning and tell your smartphone AI, “Summarize these three meeting documents and give me the essentials.” Current AI requires complex calculations to perform this task, sometimes requiring several seconds of waiting. But what if the way this AI thinks was built directly into hardware chips, allowing the result to appear in 0.1 seconds the moment you give the command? This is the kind of incredible shift currently unfolding in the AI hardware market.
Why It Matters
Most of the powerful AI we currently use runs on Nvidia’s GPUs (Graphics Processing Units). Recently, however, AI startup Etched sent shockwaves through the market by achieving a valuation of $10.3 billion [Source 14, Source 15]. The reason is simple: they didn’t create a general-purpose GPU that “does everything well,” but rather ‘Sohu,’ a dedicated chip that runs only the ‘Transformer’ models that serve as the engines for AI [Source 5, Source 13].
This change is critical because it can significantly reduce costs and drastically increase the speed of AI. There are claims that a single server equipped with just eight Sohu chips can replace the massive workload that would have required 160 Nvidia GPUs [Source 1, Source 3]. For the average user, this is a clear signal that an era is coming where we can enjoy faster, smarter AI at a lower cost than today.
The Explainer
Let’s use an analogy to make it easier to understand. Existing Nvidia GPUs are like “all-around chefs.” They possess highly flexible skills capable of making any cuisine—Korean, Western, Chinese, Japanese, etc. However, this flexibility means it takes time to prepare, such as fetching kitchen tools and prepping ingredients for whatever dish is ordered. In computer terms, this is expressed as “processing via software” [Source 4, Source 6].
On the other hand, Etched’s Sohu chip is a “kimchi jjigae robot.” It has the method for making kimchi jjigae hardwired into the robot’s very framework and mechanics. There’s no need to fetch separate kitchen tools; just press a button, and the perfect kimchi jjigae comes out. Sohu is the hardware embodiment of a recipe called Transformer (an AI structure that identifies relationships between words in a sentence) [Source 4, Source 5].
Sohu implements ‘Attention,’ the core technology Transformer models use to understand sentences, directly into dedicated circuits [Source 6]. Thanks to this, while general GPUs might barely utilize 30–40% of their performance due to complex software processes, Sohu can pour 80–90% of its chip performance solely into that task [Source 6, Source 7].
Where We Stand
Sohu is a state-of-the-art semiconductor manufactured using a 4-nanometer (nm) process [Source 2, Source 6]. The technical data released so far shows some startling figures. It claims to be able to process 500,000 tokens (the unit of characters AI reads) per second for large language models like Llama 70B [Source 1, Source 14].
Of course, the limitations are also clear. Just as a “kimchi jjigae robot” cannot make pasta, Sohu cannot perform any tasks other than Transformer-based AI models [Source 4, Source 5]. Nvidia GPUs have the powerful weapon of “general-purpose capability,” able to handle everything from scientific research to game graphics processing [Source 13]. Etched clearly acknowledges that their chip cannot be used for anything outside of this Transformer architecture and faces the challenge of overcoming limitations that appear in complex models like Mixture-of-Experts (MoE) [Source 16].
What’s Next
The future of the AI hardware market will be an intense battle between “general-purpose GPUs” and “specialized dedicated chips (ASICs).” Etched has already proven the potential of this technology in the market by securing hundreds of millions of dollars in funding [Source 6, Source 14]. Experts predict that this trend could lower AI inference costs by nearly tenfold [Source 2, Source 3].
As a reader, you can simply watch to see “how many more AI models will enter our lives more naturally.” As efficient chips like Sohu become widespread, sophisticated AI features that were previously unthinkable due to server costs can more easily be integrated into our smartphones or everyday home appliances.
MindTickleBytes AI Reporter Opinion
For hardware to forcibly hard-code a specific algorithm is akin to creating a dedicated translator that only perfectly understands one specific language. This is a symbolic event showing that AI technology has become completely solidified in a certain direction. Watching who becomes the ruler of the broader market—Nvidia’s flexibility or Etched’s efficiency—will be the most interesting spectator point for the tech world in 2026.
References
-
[Etched Sohu vs NVIDIA: Transformer ASIC vs GPU (2026) Spheron Blog](https://www.spheron.network/blog/etched-ai-sohu-vs-nvidia-transformer-asic-inference/) - Etched’s $500M Sohu Chip Takes Aim at Nvidia
- Independent AI Chip Companies Challenging NVIDIA in 2026
- Etched Just Raised $300M at a $10.3B Valuation for a Chip That Can Only Run Transformers — And It’s Beating Nvidia’s Blackwell by 10x
- Etched Sohu: the ASIC born solely to run Transformers
- Transformer Chip Startup Etched Exits Stealth: $800M Raised, $1B in Contracts
-
[AI Startup Etched Unveils Transformer ASIC Claiming 20x Speed-up Over NVIDIA H100 TechPowerUp](https://www.techpowerup.com/323887/ai-startup-etched-unveils-transformer-asic-claiming-20x-speed-up-over-nvidia-h100) - Etched’s Jump From $5B to $20B: What aTransformer-Only AI Chip…
-
[Etched $300M Sohu Chip Rivals Nvidia H100 TechPillow](https://www.techpillow.co/blog/etched-sohu-asic-chip-300m-transformer-inference-2026) - AI Chip Startup Etched Reaches 10.3 Billion Valuation to …
- Etched AI Review 2026: Sohu Chip Benchmarks and Limits
- It is equipped with larger memory
- It has the Transformer structure designed directly into the hardware
- It uses cheaper materials
- All types of computer games
- Transformer-based AI models
- High-definition video editing
- It is slower but cheaper
- Similar speed and power efficiency
- Up to 20 times faster processing speed