Etched.ai
Designs custom ASICs for AI workloads, including the Sohu chip for transformer models.
Website: https://www.etched.com/
Cover Block
PUBLIC
| Field | Value |
|---|---|
| Name | Etched.ai |
| Tagline | Designs custom ASICs for AI workloads, including the Sohu chip for transformer models |
| Headquarters | United States |
| Founded | 2022 |
| Stage | Series B |
| Business Model | B2B |
| Industry | Deeptech / Semiconductors |
| Technology Type | AI / Machine Learning hardware (ASIC) |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (3): Chris Zhu, Gavin Uberti, Robert Wachen |
| Funding Label | $100M+ |
| Total Disclosed | ~$620,000,000 |
Links
PUBLIC
- Website: https://www.etched.com/
- LinkedIn: https://www.linkedin.com/company/etched-ai
- Crunchbase: https://www.crunchbase.com/organization/etched-ai
Executive Summary
PUBLIC
Etched.ai is a semiconductor startup building application-specific integrated circuits (ASICs) optimized for transformer model inference, the workload class that powers most modern large language models [Wikipedia] [Crunchbase]. Founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, the company has bet that the transformer architecture is durable enough to justify burning it directly into silicon, an approach that trades flexibility for raw throughput on the model family that today dominates production AI [TechFundingNews] [Rambus]. Its first chip, Sohu, is fabricated on TSMC's 4nm process and pairs each die with 144 GB of HBM3 memory, with the company claiming roughly 20x the inference throughput of Nvidia's H100 on transformer workloads [Crunchbase News] [Wikipedia]. Investor conviction has scaled quickly: a $120 million round disclosed in June 2024 was followed by a $500 million financing led by Stripes that valued the company at $5 billion [Crunchbase News] [Yahoo Finance] [TechFundingNews]. Total disclosed capital now sits near $620 million, an unusually deep war chest for a hardware company at this age, reflecting both the cost structure of bringing a new ASIC to volume and the scale of the inference market the company is targeting [Yahoo Finance]. The team has recruited senior silicon and go-to-market talent including Mark Ross, former CTO of Cypress Semiconductor, and operators with backgrounds at Databricks, MosaicML, and Symphony, signaling intent to build a full-stack commercial company rather than a research outfit [TechFundingNews]. Over the next 12-18 months, the questions that matter for investors are tape-out timing for Sohu, named hyperscaler or model-lab customers willing to put silicon into production, and whether the published performance claims hold under independent benchmarking against Nvidia's Hopper and Blackwell generations.
Data Accuracy: GREEN -- Confirmed by Crunchbase, Yahoo Finance, TechFundingNews, and Wikipedia.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series B |
| Business Model | B2B |
| Industry / Vertical | Deeptech / Semiconductors |
| Technology Type | AI / Machine Learning hardware (ASIC) |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (3+) |
| Funding | $100M+ (~$620M disclosed) |
Company Overview
PUBLIC
Etched.ai began in 2022 as a dorm-room project between three Harvard undergraduates who concluded that the transformer architecture had become entrenched enough in production AI to justify a chip designed exclusively for it [TechFundingNews] [Rambus]. Co-founders Gavin Uberti (CEO), Chris Zhu, and Robert Wachen left school to pursue the company, with Uberti bringing prior engineering exposure from OctoML and Xnor.ai and Wachen taking on the COO role after earlier work scaling an AI accelerator program [TechFundingNews]. The bet is narrow by design: rather than build a general-purpose GPU or a flexible AI accelerator, Etched is hard-wiring the transformer's compute graph into the silicon itself, a tradeoff that surrenders programmability for inference efficiency [Wikipedia].
The company's milestone arc has been compressed. After incorporation in 2022, Etched announced its $120 million round in June 2024, with reporting that named early customers who had reportedly reserved capacity for Sohu hardware [Crunchbase News]. A larger $500 million financing followed, led by growth investor Stripes at a reported $5 billion valuation, marking one of the more aggressive AI-hardware repricings of the cycle [Yahoo Finance] [TechFundingNews]. The company has confirmed a manufacturing relationship with TSMC for 4nm fabrication and a memory-subsystem collaboration with Rambus, two anchor partnerships for any merchant silicon vendor [Rambus] [Crunchbase News].
Etched's public-facing positioning frames the product as inference infrastructure for transformer models, with company materials describing the work as "building the hardware for superintelligence" [LinkedIn]. The legal entity is Etched.ai Inc., headquartered in the United States [Wikipedia].
Data Accuracy: GREEN -- Confirmed by Wikipedia, Crunchbase News, TechFundingNews, Yahoo Finance, and Rambus.
Product and Technology
MIXED
Etched's lead product is Sohu, described in public materials as the first ASIC purpose-built for transformer inference [PUBLIC] [Wikipedia] [Crunchbase News]. The chip is fabricated on TSMC's 4nm node and integrates 144 GB of HBM3 memory per die, a memory configuration aimed squarely at the bandwidth-bound nature of large language model inference, where keeping the KV cache and model weights resident on package determines effective throughput [PUBLIC] [Crunchbase News]. By committing the transformer compute graph (attention, feed-forward layers, normalization patterns) directly into fixed-function silicon, Sohu is designed to dedicate a far higher share of die area to the operations that matter for transformer serving than a general-purpose GPU can [PUBLIC] [Wikipedia].
The company has publicly claimed that Sohu delivers approximately 20x the inference throughput of Nvidia's H100 on transformer workloads [PUBLIC] [Crunchbase News]. That figure has not been independently benchmarked in public reporting reviewed for this note, and the comparison set (model size, batch size, sequence length, precision) is not specified in the cited sources, so the claim should be treated as a vendor figure pending third-party validation. Etched's collaboration with Rambus on the memory subsystem is one of the more concrete technical disclosures the company has made and is consistent with the HBM3-heavy design [PUBLIC] [Rambus].
The go-to-market posture, inferred from the website and from open engineering and sales roles, is sale of Sohu-based servers and inference systems to AI labs, hyperscalers, and large enterprise customers running their own models [MIXED] [Etched.ai] (inferred from job postings). Crunchbase reporting at the time of the $120 million round referenced early customers who had reserved funds against future hardware deliveries, though those customers have not been named publicly [PUBLIC] [Crunchbase News]. Roadmap beyond Sohu has not been publicly disclosed.
Data Accuracy: YELLOW -- Core product specs confirmed across Wikipedia, Crunchbase News, and Rambus; performance claims remain vendor-sourced and not independently benchmarked in the cited record.
Market Research and Opportunity
PUBLIC
The AI inference silicon market is the single largest infrastructure category in technology right now, and the question is no longer whether demand exists but whether anyone other than Nvidia will capture a meaningful share of it. Etched is targeting the inference half of that market specifically, the workload that runs every time a user queries a deployed model, which industry observers expect to grow as a share of total AI compute as model deployment outpaces new model training.
Named third-party TAM figures specific to transformer-inference ASICs are not cited in the structured research available for this note. Public reporting around Etched's financings has consistently framed the opportunity by reference to Nvidia's data-center revenue trajectory and to the broader AI accelerator category, where Nvidia's data-center segment alone has scaled into the tens of billions of dollars per quarter [Yahoo Finance] [TechFundingNews]. The relevant question for a specialist like Etched is not the size of the entire accelerator market but the share of inference workloads that are (a) transformer-architected and (b) run at sufficient scale to justify dedicated silicon rather than general-purpose GPUs. Both conditions favor the bet so long as transformers remain the dominant production architecture.
Demand drivers cited in the reporting around Etched include the rapid growth of generative AI deployment, the cost pressure on model serving as token volume scales, and a renewed appetite from large customers to diversify away from a single GPU vendor [TechFundingNews] [Forbes]. Forbes coverage following the DeepSeek release in early 2025 noted that AI chip startups specifically benefited from a narrative shift toward inference efficiency and away from raw training scale [Forbes, January 2025]. Adjacent and substitute markets that bound the opportunity include Nvidia's own inference-optimized GPU lineup, AMD's MI300 series, custom hyperscaler silicon (Google TPU, AWS Trainium and Inferentia, Microsoft Maia), and a peer set of transformer-focused startups including Groq, Cerebras, SambaNova, and Tenstorrent.
Regulatory and macro forces cut both ways. US export controls on advanced AI chips to China constrain the addressable geography for any leading-edge ASIC, while domestic industrial policy (CHIPS Act incentives, defense and federal AI procurement) is broadly supportive of new merchant silicon entrants. Capital intensity is the dominant macro variable: bringing a 4nm ASIC to volume requires hundreds of millions of dollars before first revenue, which is the structural reason Etched's $620 million in disclosed capital matters as much as its product claims [Yahoo Finance].
| Funding Event | Disclosed Amount | Source |
|---|---|---|
| Earlier round | $120,000,000 | [Crunchbase News] |
| Series A (June 2024) | $120,000,000 | [Crunchbase News] |
| Stripes-led round | $500,000,000 | [Yahoo Finance] |
| Total disclosed | ~$620,000,000 | [Yahoo Finance] |
The table reflects what is publicly reported; the two $120 million entries may refer to the same June 2024 round under different source treatments, and readers should treat the ~$620 million total as the conservative cumulative figure cited by Yahoo Finance and TechFundingNews.
Data Accuracy: YELLOW -- Funding figures confirmed across Yahoo Finance, Crunchbase News, and TechFundingNews; market sizing draws on contextual reporting rather than a named third-party TAM study.
Competitive Landscape
MIXED
Etched is competing in the most heavily capitalized hardware category in technology, against an incumbent (Nvidia) whose share of AI training silicon is widely estimated above 80% and whose CUDA software moat has compounded for fifteen years.
| Company | Positioning | Stage / Funding | Notable Differentiator | Source |
|---|---|---|---|---|
| Etched.ai | Transformer-only inference ASIC (Sohu) | Series B, ~$620M disclosed | Architecture burned into silicon; 144 GB HBM3 per chip | [PUBLIC] [Yahoo Finance] [Crunchbase News] |
| Nvidia | General-purpose AI GPUs (H100, H200, Blackwell) | Public, multi-trillion market cap | CUDA software ecosystem, full training+inference coverage | [PUBLIC] [TechFundingNews] |
The segment-by-segment map breaks into four groups. The incumbent layer is Nvidia, with AMD's MI300 line as the principal merchant challenger; both sell flexible accelerators that handle training and inference across architectures. The hyperscaler-internal layer (Google TPU, AWS Trainium and Inferentia, Microsoft Maia) is not sold externally but absorbs a growing share of the workloads Etched would otherwise address, since the same hyperscalers are Etched's most logical customers. The transformer-specialist startup layer (Groq, Cerebras, SambaNova, Tenstorrent) is where Etched's most direct architectural competition sits, with Groq in particular having staked a public claim on inference latency. Substitute pressure comes from algorithmic efficiency gains (quantization, speculative decoding, mixture-of-experts routing) that reduce the compute required per token and therefore the urgency of buying new silicon at all.
Where Etched has a defensible edge today: a clean-sheet architecture targeted at the single workload that matters most for production AI economics, a manufacturing relationship with TSMC at 4nm, a memory subsystem co-developed with Rambus, and roughly $620 million in capital, which is enough to fund tape-out, software stack, and an initial commercial deployment without an immediate return to the market [Yahoo Finance] [Rambus]. The talent base, including Mark Ross from Cypress Semiconductor and operators recruited from Databricks, MosaicML, and Symphony, is consistent with a company that intends to ship product rather than license IP [TechFundingNews]. The durability of that edge depends on transformer architecture remaining the production default; if a fundamentally different architecture displaces transformers within Sohu's commercial life, the specialization that today is an advantage becomes a constraint.
Where Etched is most exposed: Nvidia's software stack is the single largest switching cost in AI infrastructure, and any customer adopting Sohu must port and validate inference pipelines built against CUDA. Nvidia's Blackwell generation also narrows the headline performance gap that motivates the switch in the first place, and Nvidia's ability to bundle networking, systems, and financing is something a single-product startup cannot match. Among startup peers, Groq has a head start on public benchmarking and developer mindshare for inference, which is a category Etched will need to contest directly.
The most plausible 18-month scenario splits two ways. Winner if a named hyperscaler or top-tier model lab publicly commits to Sohu deployment with independently verified throughput on a current-generation model, which would convert Etched's vendor performance claim into procurement-grade evidence and unlock the next financing at a step-up. Loser if tape-out slips materially, if independent benchmarks land closer to parity than to the claimed 20x advantage, or if architectural drift away from pure transformers (toward state-space models, hybrid architectures, or aggressive sparsity) erodes the case for hard-wired silicon before Sohu reaches volume.
Data Accuracy: YELLOW -- Subject-row facts confirmed across multiple sources; competitor positioning drawn from widely reported industry context rather than a single citable comparison.
Opportunity
PUBLIC
If the transformer remains the default production architecture for AI for the next five years, the company that owns the most efficient inference silicon for that architecture has a credible path to becoming a multi-billion-dollar infrastructure business.
The headline opportunity. Etched's single largest plausible outcome is becoming the default inference silicon for transformer-based model serving at the hyperscaler and large-enterprise tier. The structural logic is straightforward: inference is becoming the larger and more cost-sensitive half of AI compute spend, transformers dominate the deployed model base, and a chip whose die area is committed entirely to that workload has a theoretical efficiency ceiling that a general-purpose GPU cannot match. The cited evidence that this outcome is reachable rather than aspirational includes a $5 billion valuation backed by a growth investor (Stripes) that typically underwrites later-stage commercial traction, a confirmed TSMC 4nm manufacturing relationship, and reporting of early customers reserving capacity ahead of delivery [Yahoo Finance] [TechFundingNews] [Crunchbase News].
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Hyperscaler design win | A named cloud provider deploys Sohu in a public inference offering | First production tape-out and a benchmarked customer pilot | Hyperscalers are actively diversifying away from sole-source GPU supply [TechFundingNews] |
| Model-lab anchor customer | A frontier AI lab standardizes on Sohu for serving a flagship model | Independent throughput validation at the claimed multiple of H100 | Inference cost is now a board-level line item at every major lab [Forbes, January 2025] |
| Sovereign and enterprise inference | National AI programs and Fortune 100 enterprises buy Sohu systems for on-prem inference | Export-control-driven demand for non-Nvidia merchant silicon | US industrial policy and export controls are reshaping AI hardware procurement |
What compounding looks like. The flywheel for a merchant silicon company runs through software and reference deployments rather than network effects in the consumer sense. Each shipped customer generates a kernel library, a set of validated model conversions, and operational telemetry that lower the integration cost for the next customer. If Etched can convert two or three lighthouse deployments into published benchmarks and open-source model support, the per-customer onboarding cost falls and the comparison against CUDA shifts from "port everything" to "drop in for these models." The capital base also compounds: $620 million is enough to fund a second-generation chip in parallel with first-gen commercialization, which is the cadence that has historically separated durable silicon vendors from one-product companies [Yahoo Finance].
The size of the win. Public comparables suggest the upside envelope. Nvidia's data-center business is the obvious anchor, though no startup will displace it; more relevant are the valuations the market has assigned to AI silicon specialists in recent rounds, which have routinely cleared $10 billion on pre-revenue or early-revenue traction. Etched's own $5 billion valuation is the floor of that comparison set [Yahoo Finance]. If the hyperscaler design-win scenario plays out and Sohu becomes one of two or three named merchant alternatives to Nvidia for transformer inference, a valuation in the $20 billion to $50 billion range is consistent with how the market has priced peer ASIC challengers (scenario, not a forecast). The downside case, in which inference economics shift toward algorithmic efficiency rather than dedicated silicon, would compress that envelope considerably. Either way, the bet is concentrated and legible, which is unusual in AI infrastructure and is part of why the company has attracted the capital it has.
Data Accuracy: YELLOW -- Scenario plausibility grounded in Yahoo Finance, TechFundingNews, Crunchbase News, and Forbes; valuation comparables are illustrative rather than cited from a named report.
Sources
PUBLIC
[Wikipedia] Etched (company) | https://en.wikipedia.org/wiki/Etched_(company)
[Crunchbase] Etched.ai - Crunchbase Company Profile & Funding | https://www.crunchbase.com/organization/etched-ai
[TechFundingNews] Harvard dropouts' Etched raises $500M at $5B valuation to challenge Nvidia | https://techfundingnews.com/nvidia-rival-ai-chip-maker-etched-founded-by-harvard-dropouts-lands-500m-at-5b-valuation/
[Rambus] From Dorm Room Beginnings to a Pioneer in the AI Chip Revolution: How Etched is Collaborating with Rambus | 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/
[Yahoo Finance] AI Chip Startup Etched Raises $500 Million in New Funding Round | https://finance.yahoo.com/news/ai-chip-startup-etched-raises-160625909.html
[Etched.ai] Etched company website | https://www.etched.com/
[Crunchbase News] AI Chip Startup Etched Captures $120M In Hot Market | https://news.crunchbase.com/ai/chip-startup-gets-120m-etched/
[LinkedIn] Etched company page | https://www.linkedin.com/company/etched-ai
[Bloomberg Markets] Chris Zhu, Etched.ai Inc: Profile and Biography | https://www.bloomberg.com/profile/person/25064727
[Bloomberg Markets] Gavin Uberti, Etched.ai Inc: Profile and Biography | https://www.bloomberg.com/profile/person/24046930
[Bloomberg Markets] Robert Wachen, Etched.ai Inc: Profile and Biography | https://www.bloomberg.com/profile/person/25064661
[Forbes, June 2024] The Prompt: Dario Amodei On Anthropic's Most Intelligent AI Model Yet | https://www.forbes.com/sites/rashishrivastava/2024/06/25/the-prompt-dario-amodei-on-anthropics-most-intelligent-ai-model-yet/
[Forbes, January 2025] Why These AI Chip Startups Are Rejoicing Over The DeepSeek Freakout | https://www.forbes.com/sites/richardnieva/2025/01/30/why-these-ai-chip-startups-are-rejoicing-over-the-deepseek-freakout/
Articles about Etched.ai
- Etched Is Betting One Chip on Transformers and Nothing Else — The Harvard dropouts behind Sohu have raised roughly $620M to make a wager Nvidia structurally cannot make.