What is Etched's Sohu chip and how does it differ from a GPU?

Sohu is a transformer-specific ASIC (Application-Specific Integrated Circuit) that hardwires the transformer attention mechanism directly into silicon. Unlike GPUs, which are general-purpose processors adapted for AI, Sohu runs only transformer-based models but achieves dramatically higher throughput at lower cost for inference workloads.

What performance claims has Etched made about Sohu?

Etched claims an 8-chip Sohu server delivers 20x the throughput of 8 H100 GPUs on Llama 70B, reaching approximately 500,000 tokens per second, with 140x better price-performance. These claims have not been independently verified by third parties as of publication.

When will Etched Sohu chips be available to enterprise customers?

Etched announced first racks would begin shipping to select clients in summer 2026. The company has booked over $1 billion in signed customer contracts and is currently testing systems with those early clients.

What is the architecture lock-in risk of Sohu for enterprises?

Sohu is physically hardwired for transformer models only. It cannot run CNNs, LSTMs, state-space models, or any non-transformer architecture. Enterprises deploying Sohu are making a hardware bet that transformers remain the dominant architecture for the useful life of that hardware, typically 3 to 5 years.

Who are Etched's investors and what does their backing signal?

The Series C is led by Sequoia, the highest-valuation Series C the firm has ever backed. Other investors include Andreessen Horowitz, SK Hynix, Jane Street, Peter Thiel, Andrej Karpathy, Dylan Field (Figma CEO), and Amjad Masad (Replit CEO). The roster reflects broad conviction that transformer inference will be a large, long-running cost center at enterprise scale.

Etched Raises $300M at $10.3B: The Transformer Chip Bet

On July 23, 2026, Etched announced a $300 million Series C at a $10.3 billion valuation, led by Sequoia. The company’s product is a chip called Sohu: an ASIC designed to do exactly one thing at high efficiency, which is run transformer model inference. If Etched’s performance claims hold under production conditions, the economics of running large language models and agentic AI systems at enterprise scale change substantially.

What Etched Built: The Sohu Transformer ASIC

The core bottleneck in GPU-based transformer inference is memory bandwidth. The attention mechanism requires constant reads of key-value (KV) cache data, and GPU architecture was not designed to optimize that specific pattern. Sohu hardwires the attention mechanism directly into silicon, separating weight and KV cache read paths to bypass the bandwidth wall that constrains GPU throughput.

The chip runs on TSMC’s N4P 4nm process, carries 144GB of HBM3E memory per chip, and is designed to be deployed in 8-chip rack systems. The company claims those systems deliver roughly 500,000 tokens per second on Llama 70B inference. For context, 8 Nvidia H100 GPUs deliver approximately 25,000 tokens per second on the same workload under similar conditions.

Metric	Sohu (8-chip server)	8x H100 GPUs
Process node	TSMC N4P (4nm)	TSMC N4 (4nm)
Memory per chip	144GB HBM3E	80GB HBM3
Throughput (Llama 70B)	~500K tokens/sec (claimed)	~25K tokens/sec
Price-performance ratio	~140x vs H100 (claimed)	Baseline
Architecture scope	Transformer models only	General purpose

These figures are Etched’s own claims, provided to TechCrunch and corroborated by technical breakdowns published in early July after the company’s exit from stealth. Independent benchmark validation from third parties has not yet been published.

The Round and What It Signals

Etched was valued at $5 billion in December 2025 when it raised a $500 million Series B led by Stripes. The $10.3 billion Series C doubles that figure in approximately seven months. Sequoia has described this as the highest valuation it has ever led at the Series C stage.

The investor composition is worth examining. Alongside Sequoia:

Andreessen Horowitz brings infrastructure conviction and broad AI portfolio context.
SK Hynix is the HBM memory supplier to Nvidia and AMD; its investment in a chip that uses HBM3E at scale signals supply chain confidence and strategic interest in HBM demand diversification.
Jane Street has reportedly committed over $100 million across Etched rounds, a level of quant fund involvement that suggests proprietary inference cost modeling supports the thesis.
Andrej Karpathy (former head of AI at Tesla, founding member of OpenAI) provides architecture-level credibility.
Dylan Field (CEO of Figma) and Amjad Masad (CEO of Replit) represent the developer and design tool communities that run high-volume, latency-sensitive inference for code generation and creative AI features.

Earlier investors include Geoffrey Hinton and Fei-Fei Li. Total funding now exceeds $1.1 billion. The company reports over $1 billion in signed customer contracts, though no customers have been named publicly.

Why Inference Hardware Is an Enterprise Strategy Question

The practical case for Sohu rests on two dynamics that are accelerating simultaneously: inference cost pressure and agentic workload density.

Inference cost has become a first-order operational concern for enterprises running AI at scale. Unlike training, which is a one-time or periodic expense, inference is a recurring cost that scales directly with usage volume. As organizations deploy AI to handle customer interactions, generate content, process documents, and run agentic workflows, inference spend grows proportionally. GPU pricing is set by Nvidia, and alternatives at production scale have been limited. Together AI raised $800 million partly on the thesis that open-source inference optimized delivery could compete on cost; Etched’s bet is that purpose-built hardware does the same at a deeper level.

Agentic workloads compound the pressure. Agentic systems typically require multiple inference calls per user action: planning steps, tool invocations, synthesis passes. The token-per-second ceiling of underlying hardware determines how many concurrent agent threads can run within acceptable latency tolerances. A 20x throughput improvement is not only a cost reduction; it is an architectural unlock that makes more complex multi-step agentic pipelines operationally viable. This is the direct infrastructure layer beneath agentic capabilities like those described in Claude Sonnet 5’s enterprise deployment.

Databricks’ Unity AI gateway and similar infrastructure plays show that enterprises are building multi-model, cost-aware AI stacks. Purpose-built inference hardware is the logical next layer of that stack when inference becomes a material cost center.

The Architecture Bet: Durable Advantage or Brittle Lock-In?

Sohu’s design strength is inseparable from its constraint. The chip hardwires transformer attention into silicon. It cannot run CNNs, LSTMs, state-space models, or any non-transformer architecture. This is a physical property of the die, not a configuration choice.

This means any enterprise deploying Sohu is making an explicit bet that transformers remain the dominant production architecture for the useful life of that hardware, typically 3 to 5 years for data center equipment. The current evidence favors that bet. Every major frontier model in production today is transformer-based. No alternative architecture has demonstrated that it replaces, rather than supplements, transformers at frontier capability levels. The trajectory of both leading closed and open models suggests continued transformer scaling.

But “favorable today” is different from “locked in for five years.” State-space models such as Mamba showed early promise for long-context tasks. Hybrid architectures combining transformer attention with linear recurrence are an active research area. Etched’s founders acknowledge the risk and frame it as a calculated trade-off: specialization delivers an efficiency advantage large enough to justify the scope limitation for the foreseeable horizon.

Enterprise buyers evaluating Sohu should model this explicitly rather than treat it as a theoretical concern. The question is not whether transformers will exist in 2030. It is whether the specific models a given enterprise runs in production will remain transformer-only, and whether the inference cost savings over the hardware lifetime offset the switching cost of rebuilding around a different chip if the architecture mix changes.

What Enterprise AI Teams Should Do Now

Three questions frame the evaluation:

Map current inference spend. Identify what share of total AI operational cost is inference, broken down by model and workload type. If inference is already a significant budget line, the Etched case merits serious evaluation. If it is a minor cost relative to integration, tooling, and human oversight, the urgency is lower.

Assess your architecture commitment. Enterprises running exclusively transformer-based LLMs with no near-term intent to adopt non-transformer models carry lower architecture lock-in risk from Sohu. Heterogeneous stacks that include or may include non-transformer models face a more complex trade-off.

Calibrate your timeline. First racks are shipping to select clients in summer 2026. General availability pricing and timeline have not been announced publicly. Enterprises with immediate scale requirements cannot depend on Sohu in 2026 planning cycles. Enterprises building 2027 and 2028 infrastructure roadmaps have time to evaluate Sohu alongside competitive offerings from Nvidia, AMD, and other ASIC entrants.

The strategic conclusion is not necessarily “buy Sohu now.” It is that the inference hardware market is genuinely competitive for the first time, and enterprise AI infrastructure planning should reflect that. Nvidia’s dominance has been the one reliable constant of the AI stack for the past several years. Etched’s $10.3 billion valuation, $1 billion in orders, and its roster of investors provide the clearest signal yet that this assumption is being seriously challenged.

Key Takeaways

Etched raised $300M at $10.3B, doubling its valuation in seven months, in the highest-valuation Series C Sequoia has ever led.
Sohu claims 20x throughput and 140x price-performance over H100 GPUs on transformer inference workloads; independent validation is pending.
The chip cannot run non-transformer architectures, making every deployment decision a hardware bet on transformer durability.
$1 billion in signed contracts and first rack shipments beginning in summer 2026 make Etched the most credible GPU alternative to enter the enterprise inference market.
Enterprise AI teams should incorporate inference hardware diversification into 2027 infrastructure roadmaps and model the architecture lock-in risk explicitly.

Enera helps enterprise teams design and deploy agentic AI infrastructure. Book a call to discuss your inference stack and where dedicated hardware fits your roadmap.