# Etched Sohu

> Source: https://aiwiki.ai/wiki/etched_sohu
> Summary: Sohu is a transformer-specialized application-specific integrated circuit (ASIC) built by Etched, a Silicon Valley AI hardware startup founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen.
> Updated: 2026-10-15
> Fact-checked: 2026-10-15
> Categories: AI Compute, AI Hardware, AI Inference
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Etched Sohu." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/etched_sohu
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Etched Sohu | |
| --- | --- |
| General information | |
| **Manufacturer** | [Etched](https://aiwiki.ai/wiki/etched) |
| **Country of origin** | United States |
| **Announced** | June 25, 2024 |
| **Status** | A0 silicon back from TSMC in 2026; first rack shipped to Jane Street in August 2026 (Etched) |
| **Architecture** | Transformer-only ASIC |
| **Process node** | TSMC 4nm (Etched's 2026 A0 silicon used TSMC N4P) |
| **Memory** | 144 GB HBM3E per chip (2024 specification) |
| **Memory bandwidth** | Not disclosed (The Register estimated about 4 TB/s) |
| **Server configuration** | 8x Sohu chips per server (2024 specification) |
| **Claimed throughput** | 500,000+ tokens/sec on Llama-3 70B (8-chip server, Etched claim) |
| **Performance claim vs H100** | One 8xSohu server replaces 160 H100s (more than 20x faster on transformer inference, Etched claim) |
| **Software stack** | Drivers, kernels and serving stack promised as open source (2024) |
| **Website** | [etched.com](https://www.etched.com) |

**Sohu** is a transformer-specialized [application-specific integrated circuit](https://aiwiki.ai/wiki/asic) (ASIC) built by [Etched](https://aiwiki.ai/wiki/etched), a Silicon Valley [AI hardware](https://aiwiki.ai/wiki/ai_chip) startup founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen.[^2][^3] Announced on June 25, 2024, Sohu was billed by Etched as "the world's first transformer ASIC," a chip designed to run only one [neural network](https://aiwiki.ai/wiki/neural_network) architecture, the [transformer](https://aiwiki.ai/wiki/transformers), and Etched claims that a single server with eight Sohu chips delivers more than 500,000 [tokens per second](https://aiwiki.ai/wiki/tokens_per_second) on [Llama-3 70B](https://aiwiki.ai/wiki/llama_3) and "replaces 160 H100s."[^1][^24] Because it hardwires the transformer into silicon instead of offering general-purpose programmable compute, Etched says Sohu runs transformer [inference](https://aiwiki.ai/wiki/inference) more than 20 times faster than an [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) GPU while drawing less energy. These figures remain Etched's own claims.[^1][^2][^30]

The chip is fabricated on [TSMC](https://aiwiki.ai/wiki/tsmc) 4 nanometer process technology and pairs the compute die with 144 GB of [HBM3E](https://aiwiki.ai/wiki/hbm3e) memory.[^1][^31] According to the chart Etched published at launch, one server fitted with eight Sohu chips can sustain more than 500,000 Llama 70B tokens per second, compared with roughly 23,000 tokens per second for an eight-GPU H100 server and roughly 45,000 tokens per second for an eight-GPU [Blackwell B200](https://aiwiki.ai/wiki/nvidia_blackwell) server; Etched labeled only the Sohu figure, and said its Blackwell figures were estimates.[^1] Etched argues that this performance comes from achieving over 90 percent FLOPS utilization, against approximately 30 percent for general-purpose [GPUs](https://aiwiki.ai/wiki/gpu) running NVIDIA's TensorRT-LLM.[^1]

Sohu did not ship to customers in the year after its announcement: a Manifold prediction market on whether it would ship within a year resolved "no" on July 2, 2025.[^19] In June 2026, Etched said its A0 silicon had come back from TSMC's N4P process earlier that year and that it was validating its first rack-scale product with customers to fulfill $1 billion in demand; in August 2026 it said it had shipped its first rack to the trading firm Jane Street. Etched's 2026 announcements describe the product as "frontier inference clusters" and do not use the Sohu name.[^25][^26][^28] The company has attracted significant capital. Etched raised a $120 million [Series A](https://aiwiki.ai/wiki/series_a) in June 2024 co-led by Primary Venture Partners and Positive Sum Ventures, with participation from [Peter Thiel](https://aiwiki.ai/wiki/peter_thiel), [GitHub](https://aiwiki.ai/wiki/github) CEO Thomas Dohmke, Cruise co-founder Kyle Vogt and Quora co-founder Charlie Cheever.[^2][^3] A $500 million round led by the investment firm Stripes closed in December 2025 at a $5 billion post-money valuation, followed by a $300 million Series C at a $10.3 billion valuation in July 2026, led by Sequoia Capital, and a $700 million round at a $21 billion valuation in August 2026, led by Jane Street.[^14][^28][^25][^26]

## What is Sohu?

Sohu is a piece of silicon that does exactly one thing: run the forward pass of transformer neural networks for inference. A conventional GPU is a programmable parallel processor that can execute any computation a compiler can map onto it, which is why the same NVIDIA chip can train a [large language model](https://aiwiki.ai/wiki/large_language_model), render graphics, or run a physics simulation. Sohu gives up all of that generality. Etched says it cannot run the DLRM recommendation models behind ad systems, protein-folding models like AlphaFold 2, or convolutional, recurrent and LSTM networks, and The Register noted that it cannot run [state space models](https://aiwiki.ai/wiki/state_space_model) either.[^1][^4] In exchange, Etched claims it can devote far more of its transistor budget to the arithmetic of transformer layers, which the company says is what enables the headline throughput. The bet is that the transformer, introduced in [Google](https://aiwiki.ai/wiki/google)'s 2017 paper "[Attention Is All You Need](https://aiwiki.ai/wiki/attention_is_all_you_need)," is so dominant across large language models, [diffusion models](https://aiwiki.ai/wiki/diffusion_model), and multimodal generation that a chip running nothing else has a large enough market to be worth building.[^1]

## Who founded Etched and why?

Etched was founded in 2022 by Gavin Uberti, Chris Zhu, and Robert Wachen, three undergraduates who left [Harvard University](https://aiwiki.ai/wiki/harvard_university) to build a dedicated transformer accelerator; all three are [Thiel Fellowship](https://aiwiki.ai/wiki/thiel_fellowship) alumni.[^2][^3][^15] Uberti, the company's chief executive officer, had previously worked at [OctoML](https://aiwiki.ai/wiki/octoml) and Xnor.ai, and Etched describes him as an expert in AI compilers who developed the Cortex-M backend for the TVM compiler.[^2][^27] Zhu, described as Etched's chief technology officer in 2024 coverage, is listed by the company as a co-founder and a mathematics and high-performance computing researcher.[^8][^27] Wachen, Uberti's college roommate, served as chief operating officer and later as president; before Etched he co-founded Prod and Mentor Labs.[^3][^27]

Former [Cypress Semiconductor](https://aiwiki.ai/wiki/cypress_semiconductor) chief technology officer Mark Ross joined the founders early, and by 2026 Etched listed him as its CTO.[^2][^27] In June 2024, the company had approximately 35 employees.[^2] By June 2026 it described a team of more than 400 engineers drawn from NVIDIA, Google's TPU group, [Broadcom](https://aiwiki.ai/wiki/broadcom), [SK Hynix](https://aiwiki.ai/wiki/sk_hynix), TSMC and other companies.[^25]

The core thesis behind Sohu is that the transformer architecture, introduced by Google researchers in the 2017 paper "Attention Is All You Need," has become so dominant for large language models, diffusion models, and increasingly for vision and video generation, that there is enough scale to justify a chip that runs nothing else.[^1] Uberti framed the bet bluntly in an interview at the time of the Series A: "We're making the biggest bet in AI. If transformers go away, we'll die. But if they stick around, we're the biggest company of all time."[^8][^3] Etched's argument is that general-purpose accelerators spend most of their transistor budget on flexibility: it estimated that only 3.3 percent of an H100's transistors are used for matrix multiplication.[^1][^20]

Etched was headquartered in Cupertino, California, at the time of its 2024 funding and later moved to San Jose.[^3][^14][^27] The company has worked with TSMC's Emerging Businesses Group for fabrication and with [Rambus](https://aiwiki.ai/wiki/rambus) on HBM memory controller IP, which Rambus says significantly reduced the complexity of integrating the memory subsystem.[^13][^15] Sohu was unveiled on June 25, 2024, alongside Etched's $120 million Series A, with the marketing line "Meet Sohu, the fastest AI chip of all time."[^24] The announcement was covered by TechCrunch, CNBC, Tom's Hardware and [The Register](https://aiwiki.ai/wiki/the_register).[^2][^3][^4][^20]

## How does a transformer ASIC work?

A conventional GPU such as the H100 or B200 is a programmable parallel processor with thousands of generic compute units (CUDA cores, [tensor core](https://aiwiki.ai/wiki/tensor_core)s) plus a fixed-function memory hierarchy. The compiler maps any neural network onto those units at runtime. Etched has not published Sohu's microarchitecture, but says that because the chip runs only one algorithm, "the vast majority of control flow logic can be removed, allowing it to have many more math blocks."[^1] A chip built only for transformers has to handle the stages of the transformer forward pass:

- Token embedding and [positional encoding](https://aiwiki.ai/wiki/positional_encoding)
- Multi-head attention with QKV projection, scaled dot product, and output projection
- [Softmax](https://aiwiki.ai/wiki/softmax) over attention scores
- [KV-cache](https://aiwiki.ai/wiki/kv_cache) read and write paths
- Feed-forward network with [GELU](https://aiwiki.ai/wiki/gelu) or [SiLU](https://aiwiki.ai/wiki/activation_function) activation
- [Layer normalization](https://aiwiki.ai/wiki/layer_normalization) or RMSNorm
- [Residual connection](https://aiwiki.ai/wiki/residual_connection)s

Etched argues from first principles that a multiply-add circuit takes about 10,000 transistors and that, by its count, only 2.7 billion of the H100's 80 billion transistors sit in tensor cores; specializing on transformers therefore lets the same silicon area hold far more arithmetic units. Etched claims that this is what allows Sohu to reach more than 90 percent FLOPS utilization on transformer workloads, against roughly 30 percent for GPUs running TensorRT-LLM.[^1]

A consequence Etched emphasizes is the chip's tolerance for very large inference batches. "Instead of being able to run batch size 32 and then go to 64 and have high performance degradation, we can run batch sizes in the thousands without any performance degradation," Wachen told The Register.[^4] Etched's 2024 website also listed [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts) and transformer variants, beam search and Monte Carlo tree search decoding, and multicast [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding) among Sohu's supported features.[^31]

### Precision and physical layout

Etched's headline 500,000 tokens per second figure is reported for Llama-3 70B running in [FP8](https://aiwiki.ai/wiki/fp8) precision, without sparsity, with eight-way model parallelism and 2,048 input tokens and 128 output tokens.[^1][^24] Etched has not published a peak FLOPS figure for Sohu, reporting tokens per second on specific reference models instead; a LessWrong commenter argued that such throughput claims are meaningless without context on model size, context length, latency and cost.[^1][^5]

Sohu is fabricated on a 4 nanometer process from TSMC, and a LessWrong commenter estimated from Etched's images that the die is at the reticle size limit.[^1][^5] Etched specified 144 GB of HBM3E per chip; The Register reported that this memory sits in six stacks.[^31][^4] Etched has not disclosed Sohu's memory bandwidth. The Register estimated it at about 4 TB/s, while a LessWrong commenter put Sohu's capacity and bandwidth at about 0.75 times those of an [NVIDIA B200](https://aiwiki.ai/wiki/nvidia_blackwell) and about 1.8 times those of an H100.[^4][^5]

| Specification | Value |
| --- | --- |
| Process node | TSMC 4nm (A0 silicon on TSMC N4P, 2026) |
| Die size | About reticle size (outside estimate from Etched's images) |
| On-package memory | 144 GB HBM3E (six stacks, per The Register) |
| Memory bandwidth | Not disclosed (The Register estimate: about 4 TB/s) |
| Benchmark numeric format | FP8 |
| Claimed FLOPS utilization | >90% on transformer inference |
| Server configuration | 8x Sohu per server |
| Headline benchmark | 500,000+ tokens/sec on Llama-3 70B (8x Sohu, Etched claim) |

The 144 GB memory budget is enough to fit a 70 billion parameter model at FP8 weights (about 70 GB) with room for the KV cache and activations, which matters for serving long [context windows](https://aiwiki.ai/wiki/context_window). An eight-chip server has about 1.1 TB of memory in total, and The Register noted that the larger KV caches needed at big batch sizes could limit how large a model a single Etched system can serve.[^4]

## What software does Sohu run?

Because Sohu only runs transformers, Etched's software is much narrower than that of a general accelerator. There is no equivalent of [CUDA](https://aiwiki.ai/wiki/cuda) and no general kernel programming model. Etched argued in 2024 that most companies serve models through transformer-specific inference libraries such as TensorRT-LLM, [vLLM](https://aiwiki.ai/wiki/vllm) and Hugging Face's TGI, which allow hyperparameter changes but not new model code, so a fixed transformer target covers the large majority of users.[^1] The company said its software, "from drivers to kernels to the serving stack, will be open source," so that large labs could implement custom transformer layers themselves, and its website listed a "fully open-source software stack."[^1][^31]

Etched said in 2024 that Sohu "supports all of today's models" and "can handle tweaks to future models," noting that tweaks such as SwiGLU activations and RoPE encodings are used across transformer models, but that "if transformers are replaced by SSMs, RWKV, or any new architecture, our chips will be useless"; it added that if a transformer replacement emerged, "we'll build an ASIC for that too."[^1] A LessWrong commenter estimated that adapting to a new architecture would take Etched about three years.[^5]

In parallel with chip development, the company announced the **Sohu Developer Cloud**, which it said would let customers preview Sohu through an online interactive playground.[^2]

## How fast is Sohu?

Etched's headline performance figure is **500,000 tokens per second** for Llama-3 70B running on a single 8x Sohu server in FP8 precision, with 2,048 input tokens and 128 output tokens per request.[^1][^24] Etched's launch chart compared it with eight-GPU servers:

| Platform | Llama-3 70B (FP8) tokens/sec | Ratio to Sohu |
| --- | --- | --- |
| 8x Sohu (server) | >500,000 | 1.0x |
| 8x [NVIDIA B200](https://aiwiki.ai/wiki/nvidia_b200) (Blackwell) | ~45,000 (estimated by Etched) | ~0.09x |
| 8x [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) (Hopper) | ~23,000 (TensorRT-LLM 0.10.08) | ~0.046x |

Source: Etched announcement, June 25, 2024. Only the Sohu bar was labeled; the GPU values are approximate readings from Etched's chart, which was titled "Sohu is >20x faster than H100s."[^1][^24]

Etched concludes that a single 8x Sohu server is equivalent to about 160 H100 GPUs on transformer inference.[^1][^24] In its launch post the company stated plainly: "With over 500,000 tokens per second running Llama 70B, Sohu lets you build products that are impossible on GPUs. One 8xSohu server replaces 160 H100s."[^24] Uberti has described Sohu as "an order of magnitude faster and cheaper than even Nvidia's next generation of Blackwell GB200 GPUs."[^2]

Critics raised several concerns. LessWrong commenters argued that throughput figures are meaningless without latency, cost and model-size context, and that LLM inference is usually limited by memory bandwidth rather than compute; another commenter replied that Etched's higher throughput came from trading latency for throughput by batching prompts and completions together.[^5] Etched itself chose the 2,048-input, 128-output benchmark, which it said NVIDIA and AMD also use, and argued that continuous batching makes the combined workload compute-bound.[^1] The Register estimated Sohu's memory bandwidth at about 4 TB/s, behind the [H200](https://aiwiki.ai/wiki/nvidia_h200) and [MI300X](https://aiwiki.ai/wiki/amd_instinct_mi300x), and a LessWrong commenter found the 20x claim "hard to believe based on the memory bandwidth."[^4][^5] Siddharth Garg, an electrical and computer engineering professor at NYU Tandon, told MIT Technology Review he was "very skeptical about a 10x improvement just from smarter or more specialized design."[^10] On the other side, Carl Feynman, an electrical engineer commenting on [LessWrong](https://aiwiki.ai/wiki/lesswrong), called the claims "entirely plausible," noting that moving from a programmable device to a similarly sized special-purpose ASIC can yield "a factor of 10 to 50 in performance."[^5]

At announcement, The Register reported that Sohu had not taped out and existed only in emulation on [FPGAs](https://aiwiki.ai/wiki/fpga), and that Wachen implied the first chips were less than two years away.[^4] In June 2026 Etched said its A0 silicon had come back from TSMC earlier that year.[^25]

## How does Sohu compare with other AI accelerators?

Sohu enters a crowded field of [AI accelerator](https://aiwiki.ai/wiki/ai_accelerator) startups, each of which has made a different bet about the right architectural specialization point.

| Chip / system | Vendor | Strategy | Specialization |
| --- | --- | --- | --- |
| [H100, B200, GB200](https://aiwiki.ai/wiki/nvidia_blackwell) | [NVIDIA](https://aiwiki.ai/wiki/nvidia) | General-purpose GPU plus [tensor core](https://aiwiki.ai/wiki/tensor_core)s | Wide: training and inference, all model classes |
| [TPU v5p / v6](https://aiwiki.ai/wiki/tpu) | [Google](https://aiwiki.ai/wiki/google) | Systolic-array matrix engine | Training and inference, internal use plus cloud |
| [Trainium 2 / Inferentia 2](https://aiwiki.ai/wiki/aws_trainium) | [AWS](https://aiwiki.ai/wiki/amazon_web_services) | Custom AI ASIC | Training and inference inside AWS |
| [MTIA](https://aiwiki.ai/wiki/meta_mtia) | [Meta](https://aiwiki.ai/wiki/meta_ai) | Custom inference ASIC | Internal recommendation and [language model](https://aiwiki.ai/wiki/language_model) inference |
| [Groq LPU](https://aiwiki.ai/wiki/groq_hardware) | [Groq](https://aiwiki.ai/wiki/groq) | Deterministic streaming compute with on-chip SRAM | Ultra-low-latency LLM inference |
| [Cerebras WSE-3](https://aiwiki.ai/wiki/cerebras_wse_3) | [Cerebras](https://aiwiki.ai/wiki/cerebras) | Wafer-scale chip with on-chip SRAM | Training and inference for very large models |
| [SambaNova SN40L](https://aiwiki.ai/wiki/sambanova_sn40l) | [SambaNova](https://aiwiki.ai/wiki/sambanova) | Reconfigurable dataflow architecture | Training and inference for [foundation models](https://aiwiki.ai/wiki/foundation_models) |
| Taalas hardcoded models | [Taalas](https://aiwiki.ai/wiki/taalas) | Each chip is a single trained model in silicon | One model per chip, extreme specialization |
| **Sohu** | **Etched** | **Hardcoded transformer architecture (not model)** | **Transformer inference only** |

Several distinctions are worth highlighting:

- **Sohu versus general GPUs:** Sohu trades flexibility for throughput. It cannot run convolutional networks, recommender models, state space models, or non-transformer recurrent networks.[^1][^4] If a non-transformer architecture displaces transformers, Etched has said its chips would be "useless," a risk GPUs do not carry.[^1]
- **Sohu versus Groq LPU:** Both target inference, but [Groq's LPU](https://aiwiki.ai/wiki/groq_hardware) uses a deterministic streaming dataflow with on-chip [SRAM](https://aiwiki.ai/wiki/sram) as primary weight storage to minimize latency, and Groq offers access through its own API service. Etched sells full rack-scale systems that customers run in their own data centers; Jane Street, which received Etched's first rack, runs it in its own data center.[^16][^26]
- **Sohu versus [Cerebras WSE-3](https://aiwiki.ai/wiki/cerebras_wse_3):** Cerebras's wafer scale engine is the largest chip ever built and supports both training and inference, with very high single-job throughput on models that fit in its on-chip memory. Sohu is a conventional reticle-size die packaged with HBM and is inference only.[^16]
- **Sohu versus Taalas:** [Taalas](https://aiwiki.ai/wiki/taalas) goes a step further than Etched by hardcoding individual trained models into silicon, not just the architecture, which leaves its chips with no flexibility to run other models.[^16]
- **Sohu versus TPU and Trainium:** [Google TPU](https://aiwiki.ai/wiki/tpu) and [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium) are general AI ASICs developed by hyperscalers. They are wider than Sohu architecturally (they handle training as well as inference) but less specialized for the transformer attention path specifically.

## How much funding has Etched raised?

Etched said in June 2026 that it had raised $800 million across four previously unannounced financings, including a strategic investment from VentureTech Alliance; it then announced a $300 million Series C in July 2026 and a $700 million round in August 2026, which by its own figures brings the total to about $1.8 billion.[^25][^26]

| Round | Date | Amount | Lead investor | Notable participants |
| --- | --- | --- | --- | --- |
| Seed | March 2023 | ~$5.4 million (at a $34 million valuation) | (undisclosed) | Multiple early-stage investors[^8] |
| Series A | June 25, 2024 | $120 million | Primary Venture Partners, Positive Sum Ventures | [Peter Thiel](https://aiwiki.ai/wiki/peter_thiel), Thomas Dohmke ([GitHub](https://aiwiki.ai/wiki/github)), [Amjad Masad](https://aiwiki.ai/wiki/amjad_masad) ([Replit](https://aiwiki.ai/wiki/replit)), Kyle Vogt (Cruise), Charlie Cheever (Quora)[^2][^3][^8] |
| Unannounced round | Closed December 2025 (reported January 2026) | ~$500 million (at a $5 billion post-money valuation) | Stripes | [Peter Thiel](https://aiwiki.ai/wiki/peter_thiel), Positive Sum, Ribbit Capital[^11][^12][^13][^14][^28] |
| Series C | July 23, 2026 | $300 million (at a $10.3 billion valuation) | [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital) | [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz), Jane Street, Diffusion, Argo, [SK Hynix](https://aiwiki.ai/wiki/sk_hynix)[^25] |
| Growth round | August 18, 2026 | $700 million (at a $21 billion valuation) | Jane Street | [Kleiner Perkins](https://aiwiki.ai/wiki/kleiner_perkins), Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, Blackstone[^26] |

Bloomberg first reported the Stripes-led round on January 13, 2026, putting the valuation at $5 billion and, citing one person, the company's total funding at almost $1 billion; Etched later said the round had closed in December 2025 at a $5 billion post-money valuation and that its total raised through mid-2026 was $800 million.[^14][^28] Jon Peddie Research, writing in January 2026, put the total at $620 million.[^7] Etched's backers also include angel investors such as [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy), [Geoffrey Hinton](https://aiwiki.ai/wiki/geoffrey_hinton), [Fei-Fei Li](https://aiwiki.ai/wiki/fei_fei_li), Arthur Mensch and Scott Wu, and its investor list includes Two Sigma and Hudson River Trading.[^28] Tech Funding News also named Balaji Srinivasan among Etched's earlier angel backers.[^13]

The valuation rose from $5 billion to $21 billion between December 2025 and August 2026.[^29] In October 2026, TechCrunch reported, citing sources, that Etched was reviewing offers valuing it at $40 billion to $50 billion in early-stage talks; Etched declined to comment.[^30]

At the time of the Series A announcement, Uberti said that unnamed customers had reserved "tens of millions of dollars" in hardware.[^2] By June 2026 Etched said it had signed more than $1 billion in customer contracts.[^25]

## What is Oasis, the Decart partnership?

On October 31, 2024, Etched and [Decart](https://aiwiki.ai/wiki/decart), an AI lab, jointly released **Oasis**, an interactive, playable [world model](https://aiwiki.ai/wiki/world_model) that generates a [Minecraft](https://aiwiki.ai/wiki/minecraft)-style 3D environment frame by frame from keyboard and mouse inputs.[^9][^10] Oasis performs next-frame prediction, generating each frame of the game autoregressively from the preceding frames and the player's most recent inputs. There is no game engine and no procedural world; the world exists only in the model's predictions.[^9][^10]

The model was trained on millions of hours of Minecraft gameplay and the corresponding player actions, and Etched said its public demo ran in real time at 20 frames per second at 360p resolution on NVIDIA H100 GPUs.[^10][^17] Its architecture combines an [autoencoder](https://aiwiki.ai/wiki/autoencoder) with [diffusion transformer](https://aiwiki.ai/wiki/dit) layers placed between the encoder and decoder, and the companies released the weights of a scaled-down 500 million parameter version along with inference code.[^9] Demand for the H100-based demo outstripped capacity: Chipstrat's author reported being 444th in line with an estimated 15-minute wait, illustrating the cost-to-scale problem the companies argue Sohu addresses.[^17] Etched described Oasis as a demonstration of the kind of workload Sohu is meant to accelerate, claiming that on Sohu it could run in real time at 4K resolution and serve more than an order of magnitude more concurrent users; it told MIT Technology Review it expected Sohu to improve performance by a factor of 10 and planned to run projects on the chip in 2025.[^10][^17] Chipstrat characterized Oasis as a strategic exercise in market creation: rather than searching for workloads to fit Sohu, Etched was demonstrating a workload class (real-time, interactive generative video) that it argued is uneconomical without Sohu.[^17]

## Has Sohu shipped to customers yet?

Etched did not ship Sohu within its first year. In October 2024, the company told MIT Technology Review it planned to run projects on the chip in 2025, and a [Manifold Markets](https://aiwiki.ai/wiki/manifold_markets) prediction market titled "Will the Sohu AI chip ship to customers within a year?", which asked whether it would ship by June 25, 2025, resolved "no" on July 2, 2025.[^10][^19]

In 2026 the company reported the following milestones:

| Date | Milestone |
| --- | --- |
| Early 2026 | A0 silicon returned from TSMC's N4P process, according to Etched[^25] |
| June 30, 2026 | Etched said it was validating its first rack-scale product with customers to fulfill more than $1 billion in contracts, that its first racks would ship that summer, and that it had opened a factory in Taiwan and built a data center, test house and prototyping lab at its San Jose office[^25][^27][^28] |
| August 18, 2026 | Etched said it had shipped its first rack to Jane Street, which said it had tested the chip, was "pleased with the early results," and now had its own rack running in its datacenter[^26] |
| October 2026 | TechCrunch reported that Jane Street had taken delivery of an early system and that Etched had about 400 employees, a 10-megawatt data center in Silicon Valley and a facility in Taiwan near TSMC[^30] |

Etched's 2026 materials describe its product as "frontier inference clusters," combining its chips with custom racks, cooling, interconnect and software, and do not use the Sohu name; the company has not said publicly whether the A0 silicon is the Sohu design announced in 2024, although analysts such as Jon Peddie Research continued to call Etched's chip Sohu.[^7][^25][^27] Etched says the systems use two techniques: Low Voltage Inference, which runs the chip's math blocks at under half the voltage of most AI chips so that it can run trillion-parameter sparse mixture-of-experts models at more than 80 percent of peak FLOPs without thermal throttling, and Cluster Scale Memory, a hybrid HBM and SRAM design with a proprietary low-latency interconnect that pools memory across chips for faster decoding.[^27] TechCrunch described these as a low-voltage prefill chip and a new memory and interconnect for the decode phase, and reported that Etched's systems can run any frontier model.[^29]

Gaps in the public record as of October 2026 include:

- **Performance:** Etched's throughput and efficiency figures remain its own claims; the only outside assessment published by a customer is Jane Street's statement that it was "pleased with the early results."[^26][^30]
- **Customer list:** Jane Street is the only customer Etched has identified publicly; the other parties to its $1 billion in contracts have not been named.[^25][^26][^30]
- **Power consumption:** Etched's public materials do not give chip, board or rack power figures.[^27]
- **Product specifications:** Etched has not published updated specifications (memory capacity, bandwidth or FLOPS) for the 2026 silicon.[^27]

## What are the risks and criticisms of Sohu?

| Risk | Description |
| --- | --- |
| Architectural | Sohu only runs transformer-family models. If the dominant architecture shifts to a [state space model](https://aiwiki.ai/wiki/state_space_model), [Mamba](https://aiwiki.ai/wiki/mamba), or a hybrid successor, Sohu's value collapses; Etched has said its chips would then be "useless."[^1][^4] Etched argues that "transformer killers will need to run on GPUs faster than transformers run on Sohu," and says it would build a new ASIC if that happened.[^1] Uberti has stated explicitly: "If transformers go away, we'll die."[^8] |
| Execution | Etched itself said in 2024 that chip projects "cost $50-100M and take years to bring to production," and it had roughly 35 employees at the time of the Series A.[^1][^2] TechCrunch pointed to the high-profile near-failures of AI chip startups such as Mythic and Graphcore.[^2] Etched's A0 silicon came back in 2026, roughly two years after the announcement.[^25] |
| Benchmark verification | Until independent third parties measure Sohu, 500,000 [tokens per second](https://aiwiki.ai/wiki/tokens_per_second) is a marketing claim. Etched's comparison used TensorRT-LLM for the H100 and estimated figures for Blackwell, and critics argued that throughput figures need latency and cost context.[^1][^5] |
| Bandwidth versus compute | The Register estimated Sohu's memory bandwidth at about 4 TB/s, behind the H200 and MI300X, which complicates the "20x" framing because decode-phase inference is typically bandwidth-bound.[^4] |
| Software porting | Customers must port inference workloads from [CUDA](https://aiwiki.ai/wiki/cuda)-based stacks to Etched's compiler and serving software. |
| Supply chain | Sohu depends on [TSMC](https://aiwiki.ai/wiki/tsmc) 4nm capacity and HBM3E, which puts Etched in competition with NVIDIA for memory makers' output; Jon Peddie Research noted that NVIDIA's Rubin generation will take all the HBM that suppliers can provide.[^7] |

## What is next for Etched and Sohu?

Reception of Sohu has been polarized. Skeptical coverage has focused on Etched selling a chip that had not shipped against benchmarks the company controls, and on whether the transformer architecture, while dominant, is permanent; The Register's 2024 report carried the caveat "assuming it's actually built."[^4][^5][^10] Jon Peddie Research observed that competition for HBM supply with NVIDIA's Rubin generation is a structural challenge for any merchant AI chip startup.[^7]

Etched's 2026 plans center on producing rack-scale "frontier inference clusters" in volume. The company says it is vertically integrated, with a Taiwan factory and its own data center, test house and prototyping lab in San Jose, and that it aims to reach "Gigawatt scale."[^27][^25] With its August 2026 round, it said it had raised the money to accelerate production, and that ramping to "Gigawatt-scale" would require new factories, global supply chains and fleet software.[^26]

## See also

- [Etched](https://aiwiki.ai/wiki/etched)
- [AI accelerator](https://aiwiki.ai/wiki/ai_accelerator)
- [Transformer](https://aiwiki.ai/wiki/transformers)
- [Large language model](https://aiwiki.ai/wiki/large_language_model)
- [Inference](https://aiwiki.ai/wiki/inference)
- [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100)
- [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell)
- [Groq LPU](https://aiwiki.ai/wiki/groq_hardware)
- [Cerebras WSE-3](https://aiwiki.ai/wiki/cerebras_wse_3)
- [SambaNova SN40L](https://aiwiki.ai/wiki/sambanova_sn40l)
- [TPU](https://aiwiki.ai/wiki/tpu)
- [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium)
- [Mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts)

## References

[^1]: "Etched is Making the Biggest Bet in AI." Etched, June 25, 2024 (archived copy; the original page returned 404 in October 2026). https://web.archive.org/web/20240626182414/https://www.etched.com/announcing-etched
[^2]: "Etched is building an AI chip that only runs transformer models." TechCrunch, June 25, 2024. https://techcrunch.com/2024/06/25/etched-is-building-an-ai-chip-that-only-runs-transformer-models/
[^3]: "Etched raises $120 million to build chip to take on Nvidia in AI." CNBC, June 25, 2024. https://www.cnbc.com/2024/06/25/etched-raises-120-million-to-build-chip-to-take-on-nvidia-in-ai.html
[^4]: "Etched scores $120M for an ASIC built for transformer models." The Register, June 26, 2024. https://www.theregister.com/2024/06/26/etched_asic_ai/
[^5]: "New fast transformer inference ASIC: Sohu by Etched." LessWrong, June 26, 2024. https://www.lesswrong.com/posts/qhpB9NjcCHjdNDsMG/new-fast-transformer-inference-asic-sohu-by-etched
[^6]: "Etched Sohu: Transformer-Only Inference ASIC." Awesome Agents Hardware Directory. https://awesomeagents.ai/hardware/etched-sohu/
[^7]: "Welcome to the club, Etched." Jon Peddie Research, January 16, 2026. https://www.jonpeddie.com/news/welcome-to-the-club-etched/
[^8]: "Transformer model chipmaker Etched.ai raises $120M to challenge Nvidia's market dominance." SiliconAngle, June 25, 2024. https://siliconangle.com/2024/06/25/transformer-model-chipmaker-etched-ai-raises-120m-challenge-nvidias-market-dominance/
[^9]: "Etched and Decart Release Oasis, a New AI Model Transforming Gaming Worlds." InfoQ, November 10, 2024. https://www.infoq.com/news/2024/11/decart-etched-oasis/
[^10]: "This AI-generated Minecraft may represent the future of real-time video generation." MIT Technology Review, October 31, 2024. https://www.technologyreview.com/2024/10/31/1106461/this-ai-generated-minecraft-may-represent-the-future-of-real-time-video-generation/
[^11]: "Etched.ai raises $500m for a $5bn valuation - report." DatacenterDynamics, January 2026. https://www.datacenterdynamics.com/en/news/etchedai-raises-500m-for-a-5bn-valuation-report/
[^12]: "Etched Raises $500M, Valued at $5B for AI Chip Sohu." Asia Business Outlook, January 14, 2026. https://www.asiabusinessoutlook.com/news/etched-raises-500m-valued-at-5b-for-ai-chip-sohu-nwid-11065.html
[^13]: "Harvard dropouts' Etched raises $500M at $5B valuation to challenge Nvidia." Tech Funding News, January 14, 2026. https://techfundingnews.com/nvidia-rival-ai-chip-maker-etched-founded-by-harvard-dropouts-lands-500m-at-5b-valuation/
[^14]: "AI Chip Startup Etched Raises $500 Million to Take on Nvidia." Bloomberg, January 13, 2026. https://www.bloomberg.com/news/articles/2026-01-13/ai-chip-startup-etched-raises-500-million-to-take-on-nvidia
[^15]: "From Dorm Room Beginnings to a Pioneer in the AI Chip Revolution." Rambus, June 17, 2025. https://www.rambus.com/blogs/from-dorm-room-beginnings-to-a-pioneer-in-the-ai-chip-revolution-how-etched-is-collaborating-with-rambus-to-achieve-their-vision/
[^16]: "The AI Inference Wars: Comparing Taalas, Cerebras, Groq, Etched, and NVIDIA." The Menon Lab Blog, February 22, 2026. https://themenonlab.blog/blog/ai-inference-accelerators-compared
[^17]: "Etched's Oasis: Creating a Market For Sohu." Chipstrat, November 2, 2024. https://www.chipstrat.com/p/etcheds-oasis-creating-a-market-for
[^18]: "Etched (company)." Wikipedia. https://en.wikipedia.org/wiki/Etched_(company)
[^19]: "Will the Sohu AI chip ship to customers within a year?" Manifold Markets, July 2024 to July 2025 (resolved no). https://manifold.markets/ahalekelly/will-the-sohu-ai-chip-ship-to-custo
[^20]: "Sohu AI chip claimed to run models 20x faster and cheaper than Nvidia H100 GPUs." Tom's Hardware, June 26, 2024. https://www.tomshardware.com/tech-industry/artificial-intelligence/sohu-ai-chip-claimed-to-run-models-20x-faster-and-cheaper-than-nvidia-h100-gpus
[^21]: "Etched: specialized Sohu chips for AI transformer architecture." Aurorean Horizon, 2024. https://blog.auroreanhorizon.com/p/etched
[^22]: "Etched's $500 Million Raise: A Blueprint for Enterprise AI Inference in 2026." AI 2 Work, 2026. https://ai2.work/technology/etcheds-500-million-raise-a-blueprint-for-enterprise-ai-inference-in-2026/
[^23]: "Etched chip 10x faster than Blackwell." Electronics Weekly, June 2024. https://www.electronicsweekly.com/news/business/etched-chip-an-order-of-magnitude-faster-than-blackwell-2024-06/
[^24]: "Etched: Meet Sohu, the fastest AI chip of all time" (X/Twitter announcement). Etched on X, June 25, 2024. https://x.com/Etched/status/1805625693113663834
[^25]: "Progress" (news posts "Frontier Inference Clusters," June 30, 2026, and "Accelerating Inference," July 23, 2026). Etched, 2026. https://www.etched.com/progress
[^26]: "From Zero to One." Etched, August 18, 2026. https://www.etched.com/progress/from-zero-to-one
[^27]: "Frontier Inference Clusters" (company homepage, including leadership). Etched, June 30, 2026. https://www.etched.com/
[^28]: "Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip." TechCrunch, June 30, 2026. https://techcrunch.com/2026/06/30/nvidia-competitor-etched-hits-5b-valuation-1b-in-sales-for-ai-chip/
[^29]: "Etched's valuation doubles to $21B in a month." TechCrunch, August 18, 2026. https://techcrunch.com/2026/08/18/etcheds-valuation-doubles-to-21b-in-a-month/
[^30]: "Etched fields funding offers at $40B+ valuation, sources say." TechCrunch, October 5, 2026. https://techcrunch.com/2026/10/05/etched-fields-funding-offers-at-40b-valuation-sources-say/
[^31]: "Etched: The World's First Transformer ASIC" (company homepage, archived copy). Etched, July 2024. https://web.archive.org/web/20240701000000/https://www.etched.com/