Etched
4.5Paid80
Editorial matchup · June 2026
Etched vs FriendliAI: Which AI Tool Is Better in 2026?
Side-by-side comparison of Etched and FriendliAI — pricing, features, and use cases. Reviewed by our editorial team in Jun 2026.
Use-case score 1–1Updated Jun 2026
AI Infrastructure
FriendliAI
4.5Paid110
Etched and FriendliAI address the same pain point—expensive transformer inference—but from opposite ends of the stack.
Etched is a pre-production hardware ASIC bet that theoretically offers 20x throughput per transformer workload by embedding attention circuits into fixed silicon, while FriendliAI is a production-ready software inference platform claiming 50-90% cost reduction through algorithmic optimizations and GPU utilization improvements.
The choice between them hinges on timeline, risk tolerance, and workload profile.
Etched Sohu has not shipped to customers as of April 2026 and no independent benchmarks exist to validate its 500,000 tokens/sec claims on Llama 70B, while the platform has no published pricing and requires Etched's proprietary compiler, creating high execution and vendor risk for early adopters.
FriendliAI, by contrast, is shipping today with Series A completed in August 2025, powering inference for enterprises including LG Electronics and AI startups like Upstage, with support for 570k+ Hugging Face models and proven cost reductions measured in live production.
However, FriendliAI's architecture remains bounded by GPU throughput constraints—it delivers 3x faster inference than vLLM through algorithmic techniques like continuous batching and speculative decoding, not architectural specialization.
If Etched's Sohu reaches production and validates its performance claims, it could become a generation-defining shift for dense transformer deployments at hyperscale. For teams making decisions today and running production inference workloads at scale, FriendliAI offers immediate cost relief and reliability.
For organizations betting on a 2-3 year horizon where Sohu becomes available, and where transformer-only workloads dominate, Etched represents a potential long-term CapEx advantage—but only if the startup clears manufacturing, software, and supply-chain hurdles that have historically derailed specialty silicon companies.
Production inference today
FriendliAI
FriendliAI is shipping with Friendli Engine delivering 3x faster throughput than vLLM and claimed 50-90% cost reductions, backed by customer deployments at LG and Upstage. Etched Sohu is not publicly available for purchase or rental as of April 2026.
Transformer-only workloads at hyperscale
Etched
Etched's hardwired attention circuits claim 500,000 tokens/sec on Llama 70B with 8-chip server versus 23,000 tokens/sec on 8x H100, but performance is unverified and requires full infrastructure rebuild with proprietary compiler.
Model and architecture flexibility
FriendliAI
FriendliAI supports 570,000+ Hugging Face models and works with vision, multimodal, MoE, and non-transformer architectures. Etched can only run dense transformers and cannot support MoE with dynamic expert routing, diffusion, or architecture changes.
Section 01
Best for what
4 use cases scored. Etched wins 1, FriendliAI wins 1.
- EvenPricing value
Neither tool publishes a starting price.
- EvenFree tier
Neither tool offers a free tier or trial.
- EtchedUser ratings
Etched averages 4.5 / 5 vs 4.5 / 5 on the other side.
- FriendliAIReview volume
FriendliAI has 125 ratings vs 90 on the other.
Section 02
Pros & cons
Where each tool earns its rating — and where it falls short.
Etched
AI InfrastructurePros
- Transformer-specific ASIC architecture with hardwired attention circuits claims 20x throughput advantage over H100 by eliminating instruction fetch and scheduling overhead inherent to general-purpose GPUs.
- 144GB HBM3E memory per chip on TSMC 4nm reticle-limit die provides ample context window and KV-cache capacity for frontier models like Llama 70B.
- If production timeline is met and performance validated, represents potential order-of-magnitude CapEx reduction for inference-only operations at hyperscale deployment.
- Foundational bet on transformer stability could unlock edge-deployment and humanoid-robot real-time inference where latency and power efficiency are critical.
Cons
- Not available for purchase or rental as of April 2026—more than 20 months after announcement. No independent benchmarks exist; all performance claims come from Etched's own marketing materials and controlled demos.
- Requires complete serving stack rebuild with Etched's proprietary compiler. No vLLM, TensorRT-LLM, or CUDA support—migration from established GPU frameworks carries high implementation and technical risk.
- Cannot run MoE with dynamic expert routing, diffusion models, vision transformers, or any non-transformer architecture. A single architecture pivot by the AI community leaves hardware obsolete.
- Startup execution risk: founded 2022 with no prior shipped silicon products. Supply chain constraints on HBM3E memory and TSMC yield on reticle-limit die could delay ramp significantly.
- No published pricing, no disclosed TDP, server architecture unclear, no enterprise support roadmap disclosed.
Etched
AI InfrastructurePros
- Transformer-specific ASIC architecture with hardwired attention circuits claims 20x throughput advantage over H100 by eliminating instruction fetch and scheduling overhead inherent to general-purpose GPUs.
- 144GB HBM3E memory per chip on TSMC 4nm reticle-limit die provides ample context window and KV-cache capacity for frontier models like Llama 70B.
- If production timeline is met and performance validated, represents potential order-of-magnitude CapEx reduction for inference-only operations at hyperscale deployment.
- Foundational bet on transformer stability could unlock edge-deployment and humanoid-robot real-time inference where latency and power efficiency are critical.
Cons
- Not available for purchase or rental as of April 2026—more than 20 months after announcement. No independent benchmarks exist; all performance claims come from Etched's own marketing materials and controlled demos.
- Requires complete serving stack rebuild with Etched's proprietary compiler. No vLLM, TensorRT-LLM, or CUDA support—migration from established GPU frameworks carries high implementation and technical risk.
- Cannot run MoE with dynamic expert routing, diffusion models, vision transformers, or any non-transformer architecture. A single architecture pivot by the AI community leaves hardware obsolete.
- Startup execution risk: founded 2022 with no prior shipped silicon products. Supply chain constraints on HBM3E memory and TSMC yield on reticle-limit die could delay ramp significantly.
- No published pricing, no disclosed TDP, server architecture unclear, no enterprise support roadmap disclosed.
FriendliAI
AI InfrastructurePros
- Production-ready with active deployments: powering inference for LG Electronics (EXAONE models), Scatter Lab, Upstage, and 25-30 enterprise customers as of August 2025.
- Friendli Engine with iteration batching and speculative decoding delivers 3x faster throughput than vLLM and claimed 50-90% cost reduction versus Together AI and Fireworks on equivalent workloads.
- Flexible three-tier deployment: serverless API for easy adoption, dedicated endpoints for predictable workloads, Friendli Container for on-prem data sovereignty—fits enterprises, startups, and hybrid architectures.
- Supports 570,000+ Hugging Face models plus custom fine-tuned and proprietary models; works with vision, multimodal, MoE, LoRA adapters, and quantized models without architectural lock-in.
Cons
- Bounded by GPU hardware throughput—fundamental NVIDIA H100/B200 memory-bandwidth limits mean throughput gains plateau where Etched claims orders of magnitude advantage.
- Cost reduction claims of 50-90% depend on algorithmic optimizations and batching efficiency, not architectural advantage—replicable by competitors with sufficient engineering resources.
- Not yet profitable despite 6-7x revenue growth in 2025. Heavy R&D spend needed to maintain performance lead as GPU efficiency improves and competitors adopt similar techniques.
- Competitive moat tied to proprietary engineering (iteration batching, speculative decoding, TCache, Friendli DNN Library)—advantage degrades if techniques are published or adopted industry-wide.
Section 03
At a glance
Every spec on one page. Live-pulled from each tool's detail page.
- PricingPaidPaid
- Pricing modelPaidPaid
- Free tierNoNo
- Free trialNoNo
- Rating4.5 / 5 (90 ratings)4.5 / 5 (125 ratings)
- Saves80110
- CategoriesAI Infrastructure, Engineering & SimulationAI Infrastructure, AI/ML Models
- VerifiedNoNo
- Top 100 tier——
- Last updatedMay 2026May 2026
Frequently asked
Etched vs FriendliAI FAQs
Quick answers to the questions readers ask before picking between these two.
Is Etched Sohu available to rent or purchase today?
No. As of April 2026, Sohu has not shipped to customers and is not available for purchase or cloud rental. Etched has demonstrated the chip to investors and shown controlled benchmarks, but there is no public production deployment, no independent verification of performance claims, and no announced availability date.
Which platform has verified, independent benchmarks?
FriendliAI. Third-party comparisons from September 2024 and later studies confirm its 3x throughput improvement over vLLM on Llama benchmarks and lowest TTFT at 0.24 seconds among tested inference providers. Etched publishes only its own benchmarks; no third-party testing has validated its 500,000 tokens/sec claim on Llama 70B.
Can either platform run mixture-of-experts models?
FriendliAI supports MoE models with dynamic expert routing, quantized models, and LoRA adapters natively on standard GPU infrastructure. Etched's Sohu cannot support MoE with dynamic routing; Etched has mentioned a separate variant for fixed-MoE architectures, but that chip is not yet available.
What is the difference between Etched hardware and FriendliAI software?
Etched builds transformer-specific ASIC silicon that hardwires attention, projection, and normalization operations for maximum compute density. FriendliAI runs on standard NVIDIA GPUs and optimizes inference through software techniques like iteration batching and speculative decoding. Etched aims for 20x throughput advantage; FriendliAI aims for 3x improvement over vLLM on the same GPUs.
How much does each platform cost for inference?
FriendliAI's pricing is transparent: usage-based per GPU hours or per token, with claimed reductions of 50-90% versus Together AI and Fireworks for equivalent workloads. Etched has not published pricing and is not available for purchase, so cost-per-token comparisons cannot be made. Once Sohu ships, competitive pricing will depend on manufacturing yield and scale.
Which platform supports more AI models?
FriendliAI supports 570,000+ models from Hugging Face, plus custom fine-tuned, proprietary, and multimodal models. Etched supports only dense transformer models and cannot run diffusion, vision, MoE-routed, or SSM-based models like Mamba.
What happens if AI architectures shift away from transformers?
FriendliAI remains viable because it runs on general-purpose GPU silicon; you can immediately switch to new model families without hardware changes. Etched's entire value proposition depends on transformers remaining dominant. If architectures shift, Etched would require a full hardware redesign, a process the company estimates at 3+ years.
Bottom line
Choose FriendliAI for immediate, production-grade cost and latency reduction on transformer workloads running at scale today.
Teams operating Llama 70B inference across thousands of requests should evaluate FriendliAI's cost-per-token math against existing GPU infrastructure, especially if running vLLM or TensorRT-LLM at batching factors where continuous-batching and speculative-decoding optimizations yield measurable savings.
Choose Etched only if you operate at datacenter scale (5,000+ GPU equivalents), can absorb 18-24 month wait-and-see risk on production availability and performance validation, are willing to rebuild serving infrastructure with proprietary tooling, and your workloads are 100% transformer-only with zero need for multimodal, vision, or model-architecture flexibility.
Etched targets the same pain point—transformer inference CapEx—but from a hardware bet rather than software optimization, making it a long-cycle strategy for organizations with patient capital and high conviction on transformer stability. For enterprises making decisions in 2026, FriendliAI is the operational choice.
For hyperscalers or frontier labs betting 2-3 years ahead, Etched represents a potential category shift—but unproven and unshipped.
Related matchups
Keep comparing
More ai infrastructure head-to-heads.
vs
Cerebras vs Etched
Read comparison →vs
Etched vs Groq
Read comparison →vs
Etched vs SambaNova
Read comparison →vs
Etched vs Tenstorrent
Read comparison →vs
Cerebras vs FriendliAI
Read comparison →vsFriendliAI vs Groq
Read comparison →Sign up for our newsletter
Receive weekly updates so you can stay up-to-date with the world of AI
Sign up for our newsletter
Receive weekly updates so you can stay up-to-date with the world of AI