Etched
4.5Paid80
Editorial matchup · June 2026
Etched vs SambaNova: Which AI Tool Is Better in 2026?
Side-by-side comparison of Etched and SambaNova — pricing, features, and use cases. Reviewed by our editorial team in Jun 2026.
Use-case score 0–2Updated Jun 2026
AI Infrastructure
SambaNova
4.8Paid350
Etched's Sohu has not shipped to customers as of March 2026 and no independent benchmarks exist, making it a high-conviction bet rather than a proven platform.
Etched's Sohu is optimized to the millimeter for transformer operations—attention, projections, and feed-forward layers—achieving much higher performance per watt by eliminating the overhead of general programmability. Sohu claims 500K+ tokens/sec on Llama 70B, built on TSMC 4nm with 144GB HBM3E.
The core risk is architectural lock-in: DeepSeek V4 is the most downloaded model on Hugging Face as of early 2026 and it is a 671B MoE architecture that Sohu cannot serve. By contrast, SambaNova's SN50 RDU is built for this future and will start shipping to customers in the second half of 2026.
The SN50 RDU is SambaNova's fifth-generation AI inference processor designed specifically for large-scale and agentic workloads.
It uses its unique Dataflow technology and three-tiered memory architecture to reduce data movement, enabling faster inference, lower latency, and improved energy efficiency compared to traditional accelerator designs. SambaNova Composer is the proprietary layer that compiles your model graph for the RDU.
Unlike CUDA, which accepts arbitrary kernel code, Composer only supports architectures it can map to the RDU dataflow. This means no vLLM, no SGLang, no custom attention kernels. In February 2026, SambaNova raised a funding round led by Vista Equity, with Intel participating as a co-investor.
The raise came after reported acquisition talks with Intel failed and followed a period in which the company had struggled to close a new funding round amid intensifying competition with Nvidia.
Etched targets inference-only transformer-dominated workloads at hyperscale; SambaNova targets enterprise agentic AI and heterogeneous workloads that mix training and inference.
Etched's Sohu makes a permanent architectural bet; SambaNova's RDUs reconfigure per model, accepting broader model support at the cost of proprietary compilation.
For organizations serving homogeneous transformer inference at massive scale, Etched's specialization wins on throughput-per-watt if the transformer thesis holds.
For enterprises deploying diverse models, agents, and mixed training-inference, SambaNova's flexibility and current availability make it the lower-risk choice.
Pure transformer inference at hyperscale
Etched
Etched claims the total speedup reaches 20x over the H100 for Llama 70B inference, targeting inference-only deployments where flexibility is not required.
Agentic AI and multi-model inference
SambaNova
SambaNova RDUs take on high-speed decoding and can handle the prefill and decode phases of agentic inference, with support for MoE and diverse architectures.
Available hardware today
SambaNova
SambaNova raised a funding round in February 2026 and has existing deployments, while Etched is still ramping production.
Section 01
Best for what
4 use cases scored. Etched wins 0, SambaNova wins 2.
- EvenPricing value
Neither tool publishes a starting price.
- EvenFree tier
Neither tool offers a free tier or trial.
- SambaNovaUser ratings
SambaNova averages 4.8 / 5 vs 4.5 / 5 on the other side.
- SambaNovaReview volume
SambaNova has 161 ratings vs 90 on the other.
Section 02
Pros & cons
Where each tool earns its rating — and where it falls short.
Etched
AI InfrastructurePros
- Etched claims 90%+ FLOPS use versus 30-40% on GPUs, theoretically delivering 2-3x more useful compute from the same transistors.
- Sohu's circuitry is optimized for transformer operations—attention, projections, and feed-forward layers—achieving much higher performance per watt.
- The chip is manufactured on TSMC's 4nm process at reticle-limit die size and features 144GB of HBM3E memory.
- Etched raised a large Series B round bringing total funding to over 600 million and achieving unicorn valuation.
- Sohu targets inference-only deployments where training happens elsewhere, particularly fitting hyperscalers and inference-only service providers.
Cons
- Sohu hasn't shipped to customers as of March 2026 and no third-party benchmarks exist.
- DeepSeek V4 and Qwen3 are two of the most widely deployed open-weight models as of April 2026 and both are MoE architectures incompatible with Sohu, meaning a significant fraction of current production inference workloads cannot run on Sohu at all.
- If a different AI architecture emerges tomorrow, the chip will become obsolete.
- Etched has not published pricing or per-rack costs, and the toolchain migration cost is unknown but likely significant for any team with a mature vLLM deployment.
- The trade-off is flexibility versus efficiency: if transformers stay dominant for 5+ years, Sohu's cost-per-token advantage compounds; if a new architecture displaces transformers, Sohu is obsolete.
Etched
AI InfrastructurePros
- Etched claims 90%+ FLOPS use versus 30-40% on GPUs, theoretically delivering 2-3x more useful compute from the same transistors.
- Sohu's circuitry is optimized for transformer operations—attention, projections, and feed-forward layers—achieving much higher performance per watt.
- The chip is manufactured on TSMC's 4nm process at reticle-limit die size and features 144GB of HBM3E memory.
- Etched raised a large Series B round bringing total funding to over 600 million and achieving unicorn valuation.
- Sohu targets inference-only deployments where training happens elsewhere, particularly fitting hyperscalers and inference-only service providers.
Cons
- Sohu hasn't shipped to customers as of March 2026 and no third-party benchmarks exist.
- DeepSeek V4 and Qwen3 are two of the most widely deployed open-weight models as of April 2026 and both are MoE architectures incompatible with Sohu, meaning a significant fraction of current production inference workloads cannot run on Sohu at all.
- If a different AI architecture emerges tomorrow, the chip will become obsolete.
- Etched has not published pricing or per-rack costs, and the toolchain migration cost is unknown but likely significant for any team with a mature vLLM deployment.
- The trade-off is flexibility versus efficiency: if transformers stay dominant for 5+ years, Sohu's cost-per-token advantage compounds; if a new architecture displaces transformers, Sohu is obsolete.
SambaNova
AI/ML ModelsPros
- The SN50 RDU is designed specifically for large-scale and agentic workloads using Dataflow technology and three-tiered memory architecture to reduce data movement, enabling faster inference, lower latency, and improved energy efficiency.
- SambaNova's Dataflow Architecture allows data to flow from one AI operation to the next as an assembly pipeline, eliminating frequent, energy-intensive memory bottlenecks and enabling faster AI inference with better energy efficiency.
- The SambaRack can hot-swap between models in milliseconds and manage many models on the same infrastructure without the latency spikes common in shared cloud queues and GPU clusters.
- The SN50 delivers 5X more compute and 4X more network bandwidth than the fourth-generation SN40.
- SambaNova partnered with Intel to deliver high-performance inference architecture available in H2 2026, with GPUs handling prefill, RDUs handling decode, and Xeon 6 CPUs orchestrating tasks.
Cons
- SambaNova struggled to close a new funding round amid intensifying competition with Nvidia before securing recent backing.
- SambaNova Composer is proprietary and only supports architectures it can map to the RDU dataflow, meaning no vLLM, no SGLang, no custom attention kernels.
- RDUs must prove competitive on cost and ecosystem maturity, and enterprises must see measurable efficiency gains, with the software layer required to make the system usable, not just possible.
- SambaNova has pivoted to model-as-a-service after limited adoption of its hardware offerings.
- The SN40L features a novel three-tier memory system with 520 MiB of on-chip SRAM, 64 GiB of HBM, and up to 1.5 TiB of DDR DRAM, adding deployment complexity.
Section 03
At a glance
Every spec on one page. Live-pulled from each tool's detail page.
Frequently asked
Etched vs SambaNova FAQs
Quick answers to the questions readers ask before picking between these two.
Can Sohu run MoE models like DeepSeek?
No. DeepSeek V4 is the most downloaded model on Hugging Face as of early 2026 and it is a 671B MoE architecture that Sohu cannot serve. Sohu only supports dense transformer architectures.
When will Sohu ship to production customers?
As of April 2026 Sohu is not publicly available for purchase or rental. Etched is in customer engagement, but no public availability timeline has been confirmed.
Which platform is available to buy today?
SambaNova. The SN50 RDU will start shipping to customers in the second half of 2026, and earlier-generation SN40L systems are already deployed at enterprises and sovereign AI centers.
How do the memory architectures differ?
Sohu features 144GB of HBM3E memory per chip, while the SN40L RDU features a novel three-tier memory system with 520 MiB of on-chip SRAM, 64 GiB of on-package HBM, and up to 1.5 TiB of off-package DDR DRAM. Sohu optimizes for high bandwidth; SambaNova optimizes for capacity and flexible tiering.
What is the primary architectural difference?
Sohu's circuitry is optimized to the millimeter for the key operations of Transformers: attention, projections, and feed-forward layers, making it fixed-function. SambaNova's Dataflow Architecture allows data to flow from one AI operation to the next as an assembly pipeline, eliminating frequent, energy-intensive memory bottlenecks, and reconfigures per model.
Which is better for enterprise agentic AI?
SambaNova. The SambaRack can hot-swap between models in milliseconds and manage many models on the same infrastructure without the latency spikes common in shared cloud queues and GPU clusters, essential for agent-driven workflows.
Bottom line
Etched Sohu and SambaNova RDUs target overlapping but distinct personas within enterprise AI infrastructure.
Etched is an all-in bet on transformer dominance: if your workload is pure transformer inference at massive scale—serving models like Llama 70B with minimal model diversity—and you can wait for production availability and tolerate the risk of architectural obsolescence, Sohu offers unmatched throughput-per-watt and cost-per-token claims.
It suits hyperscalers and inference-only service providers who amortize fixed-function silicon over enormous inference volumes. SambaNova is a hedge against architectural fragmentation: its reconfigurable RDUs run transformers, MoE models like DeepSeek, and other workloads on the same silicon.
The dataflow architecture minimizes memory movement, making it efficient for agentic AI where models switch frequently and latency is critical.
With recent funding, Intel partnership, and fifth-generation hardware shipping in H2 2026, SambaNova offers lower execution risk for enterprises building production AI agents and sovereign AI. Choose Etched if you believe transformers are permanent and can commit to specialized ASIC hardware. Choose SambaNova if you need flexibility, agentic inference support, and hardware available now.
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 Tenstorrent
Read comparison →vs
Etched vs FriendliAI
Read comparison →vs
Cerebras vs SambaNova
Read comparison →vsGroq vs SambaNova
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