Recommended Free Tools
The June 5, 2023 headline referred to Etched, a semiconductor startup founded by Gavin Uberti and Chris Zhu, with Robert Wachen now also listed as a co-founder. Etched raised $5.36 million in seed funding—reported at a $34 million valuation—to develop specialized hardware for large-language-model inference. The round was led by Primary Venture Partners, with participation from MAX Ventures and angels including former eBay CEO Devin Wenig.
That seed round was only the beginning. By August 2026, Etched said it had raised $800 million across four financings, received initial silicon from TSMC’s N4P process, and was validating rack-scale inference systems with customers. Those later figures are company-reported and should not be confused with independently verified performance or revenue.
As an Amazon Associate I earn from qualifying purchases.
What Etched was building
Etched’s original thesis was that transformer models had become important enough to justify a processor designed specifically around them. Instead of building another general-purpose GPU, the startup planned an accelerator focused on running trained models efficiently.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteThe distinction matters:
- Training uses large amounts of compute to create or update a model.
- Inference runs an already-trained model to answer prompts, generate text, or power an application.
- General-purpose accelerators support many workloads and model types.
- Specialized accelerators trade flexibility for potentially better throughput, latency, power efficiency, or cost on a narrower workload.
Etched was targeting companies operating ChatGPT-like services, where inference runs continuously and can become a major operating expense. Its argument was that a purpose-built system could move data and perform transformer operations more efficiently than flexible hardware designed for a much wider range of applications.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
The original EE Times report described the company as an early-stage effort, not a commercial product with public pricing or confirmed deployments.
The founders behind the seed round
The 2023 report centered on Harvard-affiliated engineers Gavin Uberti and Chris Zhu. Uberti was described as having worked on Apache TVM, compiler technology, and microkernels, while Zhu was presented as his technical collaborator. Etched’s current leadership page also lists Robert Wachen as a co-founder and president.
That creates an important distinction. The original article’s “Harvard dropouts” framing described the people at the center of the 2023 funding story. Etched’s current presentation describes a three-person founding team. In either case, leaving university was not the substantive investment thesis. The more relevant combination was expertise in compilers, kernels, systems, and semiconductor development, applied to a rapidly expanding AI workload.
What exactly did Etched raise?
| Item | Reported detail |
|---|---|
| Round | Seed funding |
| Amount | $5.36 million, commonly rounded to $5 million |
| Reported valuation | $34 million |
| Lead investor | Primary Venture Partners |
| Other participants | MAX Ventures and angels including Devin Wenig |
| Original report date | June 5, 2023 |
According to EE Times, the capital was intended for initial hiring, RTL front-end development, and discussions with intellectual-property providers. In semiconductor terms, that is an early bridge from an architectural concept toward a design that could eventually be implemented in physical silicon.
Sohu: Etched’s original accelerator
The early product was called Sohu. The 2023 description indicated that it would be designed around transformer workloads, include substantial memory, support large batch sizes, and use a narrower software stack than a general-purpose GPU. Its intended focus was LLM inference rather than broad AI workloads.
Rank #2
- Delivers 25 TFLOPS of AI inference performance using four MemryX MX3 accelerators in a compact M.2 2280 M-key module.
- Connects through PCI Express Gen 3 and above for straightforward integration into embedded and edge systems with an M.2 M-key slot.
- Includes 42 MB of distributed on-chip weight storage to support low-latency, batch-1 pipelined inference.
- Designed for low-power deployment with a 1 W–12 W operating range and host-managed airflow requirements.
- Supports ONNX, TensorFlow, Keras, and TFLite models through the MemryX software stack on Linux and Windows hosts.
A narrower architecture can simplify parts of the hardware and compiler problem. If the supported models share predictable operations, the chip can devote more of its resources to the memory movement and matrix calculations that dominate those workloads. Large batches may also help an operator keep the hardware busy when serving many users at once.
But Sohu’s reported 2024 availability date was an initial target, not a confirmed shipping date. Etched’s current developer materials still refer to “Sohu preview docs,” while the company now describes its broader direction as frontier inference clusters rather than merely a single accelerator card.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →What the 140× claim meant
The most attention-grabbing early claim was that Sohu could eventually deliver up to 140 times the throughput per dollar of an Nvidia H100 PCIe card on GPT-3-token workloads.
That should not be rewritten as “Sohu was 140 times faster than an H100.” It was a founder-provided projection, not an independently verified benchmark. The comparison also depended on details such as:
- the model and model size;
- prompt and output lengths;
- batch size;
- prefill versus token-by-token decode;
- precision and quantization;
- latency requirements;
- software maturity and utilization;
- power, networking, cooling, and system costs; and
- whether the H100 baseline was fully optimized.
“Throughput per dollar” is also different from a universal reduction in cost. A specialized chip might produce more tokens for a given hardware purchase while still facing substantial expenses for development, deployment, networking, power, support, and underutilized capacity.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
Why specialize in transformers?
The opportunity was straightforward: inference can run continuously at very large scale, and even a modest efficiency improvement can matter when multiplied across thousands of accelerators. A purpose-built system could potentially optimize the entire path from chip and memory to rack design, cooling, software, and deployment.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The risk is that the workload may not remain stable. Model developers continue to change attention mechanisms, quantization methods, mixture-of-experts designs, context lengths, and multimodal architectures. A chip optimized around one generation of transformer assumptions can lose value if customers need unsupported operations or rapidly change their models.
Customers may also prefer the flexibility and software ecosystem of Nvidia GPUs, even at a higher per-token cost. The original EE Times coverage acknowledged that Nvidia could pursue the same opportunity if the market became large enough—giving a specialist both a technical challenge and a powerful incumbent to compete against.
Why the seed round was notable
The significance of the financing was not simply that young founders attracted venture capital. Investors were funding a capital-intensive semiconductor company before a commercial chip was available, around a concentrated bet on transformer models while general-purpose GPUs dominated AI infrastructure.
A seed round must fund far more than a software prototype. The company has to recruit chip designers, define the architecture, write RTL, acquire or license IP, build verification infrastructure, work with a foundry, manage packaging and memory dependencies, and eventually develop software that customers can use reliably.
Rank #4
- 48GB AI graphics accelerator
Failure can occur at any stage: the architecture may not translate into practical silicon; memory bandwidth may become the bottleneck; yields or packaging may disappoint; thermal and power-delivery problems may appear at rack scale; or the target market may evolve before production begins.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What happened after the 2023 funding
Etched’s current website says the company has raised $800 million across four financings, received A0 silicon from TSMC’s N4P process, and is validating its first rack-scale product with customers. It also says the company has more than 400 engineers, is working to fulfill more than $1 billion in customer demand or contracts, and has established a new 10-megawatt lab near its office.
These are company-reported claims. “Demand” and “contracts” are not necessarily equivalent to signed purchase orders, recognized revenue, delivered systems, or operational customer deployments.
On July 23, 2026, Etched announced a $300 million financing at a $10.3 billion valuation, led by Sequoia Capital, with participation from Andreessen Horowitz, Jane Street, Diffusion, Argo, and SK hynix. TechCrunch also reported the financing. A private-company financing valuation is not the same as a public-market capitalization.
Free tools Windows power users keep installed
One-click scans. No signup required.
Etched’s later positioning emphasizes frontier inference clusters: integrated systems involving chips, racks, software, manufacturing, cooling, and interconnects. That is a substantially broader ambition than the original headline’s “LLM accelerator” shorthand.
Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
Has the thesis been proven?
There is evidence that Etched progressed well beyond its seed-stage concept. It raised substantially more capital, reached an A0 silicon milestone, expanded its team, and says it is validating rack-scale systems with customers. The later financing also indicates that major investors assign considerable value to the company’s prospects.
That does not independently prove the original 140× projection or establish that Etched has achieved a durable commercial advantage. A full assessment would require public evidence of:
- independent performance results against current Nvidia systems;
- model, batch, precision, latency, and power details;
- delivered systems operating at sustained utilization;
- production volume and yield;
- named customer deployments;
- revenue and recognized contract value; and
- performance across the changing mix of modern inference workloads.
The crucial test is deployment, not a narrow theoretical comparison. A successful inference platform must manufacture enough units, provide usable software, deliver predictable latency and uptime, and secure enough customer volume to justify the enormous fixed cost of custom silicon and rack-scale infrastructure.
What the business model could be
Etched has not published public retail pricing or a self-serve purchase path for Sohu or its inference racks. Its current site directs interested users toward access and company engagement rather than a consumer checkout process.
Analytically, a company pursuing this strategy could sell accelerator cards, sell complete rack-scale systems, operate an inference cloud, license hardware or software technology, or enter long-term capacity agreements with model providers. Etched’s current materials point most clearly toward integrated infrastructure and customer deployments, but they do not establish which commercial model will generate the majority of revenue.
The bottom line
Etched’s story is not simply that “Harvard dropouts raised $5 million.” The June 2023 seed round was an early, high-risk bet that transformer inference was important and stable enough to support specialized hardware. The company has since reported major financing, silicon, team, and customer-validation milestones, and has evolved from a proposed accelerator into a broader inference-infrastructure effort.
At the same time, the strongest early performance claim remains a projection, and many current scale and customer figures come from Etched itself. The company has advanced far beyond the concept described in the original headline, but the ultimate investment thesis still depends on independently demonstrated performance, reliable production, and sustained commercial deployments.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
