Etched AI is live · custom pricing · last verified 9 Nov 2026 · viability monitor
Etched AI
Etched builds custom inference silicon and rack-scale frontier inference clusters for trillion-parameter MoE prefill and decode.
The most credible non-GPU inference bet in years, backed by $1B in booked demand and a team that shipped HGX, DGX, and TPU v1-v5. But it's still a forward-looking bet: LVI and CSM are vendor-stated, you can't buy capacity off a shelf, and CUDA tooling doesn't come along. Fund a pilot next to your GPU fleet if you have frontier MoE traffic and patience.
Verdict generatedVerified 5d ago · liveness 65/100 · cite: rightaichoice.com/tools/etched-ai
- Frontier AI labs serving many-trillion-parameter sparse MoE inference at extreme scale
- Hyperscalers and cloud providers scoping non-GPU inference capacity for power-capped sites
- Teams hitting thermal throttling or decode-latency walls on long-context and agentic traffic
- Organizations able to run a multi-quarter rack-scale pilot while keeping a GPU fallback
- Teams that need production inference hardware delivered and supported today
- Shops with a deeply CUDA-bound stack and no appetite for a software migration
- Small startups without the budget, power, or facilities for a rack-scale deployment
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Etched if you need inference capacity in production this quarter, if your stack is locked to CUDA, or if your workloads include training and fine-tuning on the same cluster.
Rack-scale hardware is a capital purchase, so the sticker is not the run rate, colocation, power, cooling, and spares all land on your facilities budget.
The comparison is against GPU capacity: the claim is better watts-per-token than an NVIDIA fleet on frontier sparse MoE inference, which is what would justify the capital outlay. Nvidia remains the incumbent you are measuring against, and the Sohu ASIC versus GPU debate is the frame buyers are using in 2026.
In short
Etched AI — Etched builds custom inference silicon and rack-scale frontier inference clusters for trillion-parameter MoE prefill and decode. Best for Frontier AI labs serving many-trillion-parameter sparse MoE inference at extreme scale, Hyperscalers and cloud providers scoping non-GPU inference capacity for power-capped sites, Teams hitting thermal throttling or decode-latency walls on long-context and agentic traffic. Pricing is on request (custom quotes).
No head-to-head comparison for Etched AI yet — the alternatives below are the closest matches we list.
What is Etched AI?
Etched AI — Etched builds custom inference silicon and rack-scale frontier inference clusters for trillion-parameter MoE prefill and decode. Best for Frontier AI labs serving many-trillion-parameter sparse MoE inference at extreme scale, Hyperscalers and cloud providers scoping non-GPU inference capacity for power-capped sites, Teams hitting thermal throttling or decode-latency walls on long-context and agentic traffic. Pricing is on request (custom quotes).
Etched is a chip and systems company building what it calls frontier inference clusters — custom silicon, racks, interconnects, cold plates, and software co-designed so very large transformer models run faster and cheaper per token than on general-purpose GPUs. It is inference-only by design; training isn't a use case here. The company reports first racks shipping this summer against more than $1B in customer contracts, with early customer tests showing SOTA throughput, latency, and power efficiency. Two architectural bets anchor the pitch. Low Voltage Inference (LVI) runs the chip's math blocks at under half the voltage of most AI chips, which Etched says lets trillion-parameter sparse MoEs hit 80%+ peak FLOPs without thermal throttling. Cluster Scale Memory (CSM) pools HBM and SRAM across a scale-up domain over a proprietary ultra-low-latency interconnect, aiming for SRAM-like decode latency at HBM-scale capacity. A0 silicon came back from TSMC N4P earlier this year. Etched says it has tested racks in representative data center deployments, run terabytes of production traffic through its simulator, and stood up a Taiwan factory plus a San Jose data center, test house, and NPI prototyping lab to stay vertically integrated toward gigawatt scale. The team is 400+ engineers from NVIDIA, Google TPUs, Broadcom, SK Hynix, and TSMC, including the platform lead who built HGX and DGX and the TPU v1-v5 software lead. $800M raised across four financings, including $700M at a $21B valuation led by Jane Street. Positioning-wise, this is not a drop-in alternative to renting NVIDIA capacity tomorrow. It's a hardware bet for labs and hyperscalers whose frontier MoE, long-context, and agentic inference traffic is already hitting thermal or decode-latency walls.
Key features
- Custom A0 silicon fabricated on TSMC N4P
- Low Voltage Inference (LVI) runs math blocks at under half the voltage of most AI chips
- 80%+ peak FLOPs on trillion-parameter sparse MoEs without thermal throttling
- Cluster Scale Memory (CSM) hybrid HBM/SRAM shared memory pool across a scale-up domain
- Proprietary ultra-low-latency, high-bandwidth scale-up interconnect
- Built for both prefill and decode inference workloads
- Optimized for many-trillion-parameter sparse MoE models
- Built for long-context and agentic inference workloads
- Splittable math arrays with novel tiling and scheduling algorithms
- Co-designed chips, packages, PCBs, cold plates, interconnects, and VRM architectures
- Rack-scale systems tested in representative data center deployments
- Simulator validated against terabytes of production traffic patterns
- Vertical integration: Taiwan factory plus San Jose data center, test house, and NPI lab
- Path to gigawatt-scale deployments
- Inference-only by design — no training workloads
Behind the verdict
Etched is worth tracking for one reason: it's attacking the specific wall that frontier inference teams hit at scale — thermal throttling that drops sustained throughput to under half of peak, and decode latency that HBM can't fix. The LVI pitch is the more interesting of the two. Running math blocks at under half the voltage of most AI chips, if it holds in production, means FLOPs density that general-purpose GPUs can't match without clocks falling. CSM is the complement: an HBM/SRAM hybrid across a scale-up domain that claims SRAM-like decode latency without the capacity and yield tradeoffs of SRAM-only chips. Both are vendor claims right now — no independent benchmarks have been published. Pick Etched when your traffic is many-trillion-parameter sparse MoE, long-context, or agentic, and you're already fighting power or decode-latency ceilings on GPUs. The $1B in contracts and the ex-NVIDIA HGX/DGX and TPU v1-v5 pedigree are why we'd take the meeting seriously. Pass if you need supported hardware today, or if your stack is deeply CUDA-bound with no appetite for a software migration. The closest alternative is simply more NVIDIA capacity — more predictable, more expensive per token at the frontier, but available now. Etched's advantage only compounds if it ships on schedule and the performance claims survive customer validation this summer. In practice, treat this as a multi-quarter pilot, not a procurement. Keep your GPU fallback warm, budget engineering time for the software port, and ask hard questions about rack-level support SLAs before you sign.
Models under the hood
as of 2026-11-08
Limitations
- Etched builds custom inference silicon and rack-scale inference clusters rather than a self-serve AI application; the site offers only a "Get Access" contact path, not a signup or download.
- The homepage states the first rack-scale product is being validated with customers against $1B in demand, so the systems are not yet generally available.
- Published performance and efficiency figures (e.g.
- 80%+ Peak FLOPs, LVI/CSM architecture) are vendor-stated claims, and most secondary pages (pricing, docs, changelog, about) returned 404 at scrape time.
as of 2026-10-09
Researching Etched AI? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Etched AI still active?
Etched AI is live · custom pricing · last verified 9 Nov 2026 · viability monitor.
Viability score
How well maintained and how widely used is Etched AI? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: November 2026
How we score →What do users say about Etched AI?
What people actually say about Etched AI — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
ResearchedDiscussed by users across 3 sources (Hacker News, YouTube, Lemmy) · researched Aug 30, 2026.
Average across the 3 sources that answered — each source counts once, not each post.
- +Unique ASIC design purpose-built for transformer inference, unlike general-purpose GPUs.
- +Low Voltage Inference claims sustained 80%+ FLOPs utilization without thermal throttling.
- +Cluster Scale Memory offers HBM-scale capacity with SRAM-like speeds for long-context workloads.
- +$1B+ in customer contracts booked before launch signals strong early market interest.
- +Vertical integration from chip to data center ensures optimized performance and control.
- −No shipped product or independent benchmarks yet, making performance claims unverifiable.
- −Inference-only focus is a major limitation for teams needing flexibility for training.
- −Lack of ecosystem and software tooling forces heavy customization and expertise.
- −Reported 56k tokens/sec is only from FPGA demo, not the real A0 silicon.
- −Very high cost of entry with contact-based pricing, likely out of reach for SMBs.
- • Custom software development and integration likely required
- • Potential long-term contracts with minimum commitment
How we verified this
Every listed fact about Etched AI — pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 9 Nov 2026. Analysis is drafted with AI models and checked against our editorial standards before publishing; see how we verify.
Verification history
We have re-verified Etched AI 24 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed our verdict, description
- — changed limitations
- — changed who should skip it, who it suits, summary, our verdict, features, description
- — changed summary, description, pricing fit, our verdict, limitations, who should skip it, setup time, hidden costs, switching out, switching in, use cases
- — changed use cases, who it suits, description, our verdict, features, hidden costs, limitations, switching in, switching out, who should skip it, pricing fit, setup time, summary
- — changed SSL certificate
Showing the 6 most recent of 24 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Etched AI cost in 2026?
Etched AI: Pricing is on request (custom quotes). Pricing last verified 9 Nov 2026.
Where the pricing makes sense
The company stage and team size where Etched AI's pricing actually pencils out — and where peers do it cheaper.
The comparison is against GPU capacity: the claim is better watts-per-token than an NVIDIA fleet on frontier sparse MoE inference, which is what would justify the capital outlay. Nvidia remains the incumbent you are measuring against, and the Sohu ASIC versus GPU debate is the frame buyers are using in 2026.
Has Etched AI's pricing changed?
No plan price for Etched AI has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Etched AI?
Pick Etched AI if you are Frontier AI labs serving many-trillion-parameter sparse MoE inference at extreme scale, Hyperscalers and cloud providers scoping non-GPU inference capacity for power-capped sites, Teams hitting thermal throttling or decode-latency walls on long-context and agentic traffic, Organizations able to run a multi-quarter rack-scale pilot while keeping a GPU fallback.
Look elsewhere if you are Teams that need production inference hardware delivered and supported today, Shops with a deeply CUDA-bound stack and no appetite for a software migration, Small startups without the budget, power, or facilities for a rack-scale deployment, Workloads that mix training and inference on the same cluster — Etched is inference-only.
Skip it if Skip Etched if you need inference capacity in production this quarter, if your stack is locked to CUDA, or if your workloads include training and fine-tuning on the same cluster.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Etched AI actually fits — and what changes day-one when you adopt it.
You run a many-trillion-parameter sparse MoE in production and your decode latency and power draw are capping how much traffic you can serve per rack. You arrange a pilot of Etched's first rack-scale product alongside your existing GPU deployment and run your own production traffic patterns through it before committing.
Outcome: You get a side-by-side throughput, latency, and power comparison on your actual workload rather than a vendor benchmark, and you decide with a contractual off-ramp whether to shift a slice of capacity to Etched.
You are modeling gigawatt-scale inference buildout and want a non-GPU option in the plan to avoid single-supplier exposure. You evaluate Etched's vertically integrated supply chain (Taiwan factory, San Jose data center and test house) against your procurement and power constraints.
Outcome: You get a defensible second source in your capacity model and a concrete answer on whether rack-scale custom silicon can be sited and operated inside your existing facilities.
Your sustained inference throughput sits well under peak FLOPs because the GPUs downregulate clock speed under load. You test whether the Low Voltage Inference design actually holds 80%+ peak FLOPs on your sparse MoE traffic without throttling.
Outcome: You either confirm the thermal ceiling is a hardware problem worth switching for, or you rule it out and keep optimizing on GPUs.
Use cases
- Serve many-trillion-parameter sparse MoE models in production at lower cost and power than a GPU fleet
- Run long-context and agentic inference workloads that GPU memory subsystems bottleneck on decode
- Scale customer-facing LLM inference toward gigawatt-scale data center capacity
- Replace or supplement GPU clusters specifically for transformer inference
- Pilot a purpose-built inference rack alongside an existing NVIDIA deployment as a benchmark comparison
Setup time & first value
How long it actually takes to get something useful out of Etched AI — broken out by persona, not the marketing-page minute.
There is no signup flow. For an AI lab or hyperscaler the realistic path is a vendor conversation, then a multi-quarter rack-scale pilot that includes facility planning, integration, and benchmarking against your GPU baseline. Expect months to first validated result, not days. For a small team, this is not a purchasable path at all.
Switching to or from Etched AI
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From NVIDIA GPU inference fleet: pilot an Etched rack-scale product alongside your existing deployment and benchmark your own production traffic patterns before moving a slice of capacity.
- →From CUDA-locked inference stack: expect porting work, since a custom ASIC means your kernels and tooling do not carry over unchanged.
- ↗To NVIDIA GPU inference: keep a GPU fallback running during any Etched pilot so you can shift traffic back if benchmarks or timelines do not hold.
- ↗To general-purpose GPU capacity: because Etched is inference-only, any training or fine-tuning workload has to live on a separate cluster regardless.
Resources & guides
Tutorials & Learning
YouTube returned 6 videos for “Etched AI”, and we withheld 6: 6 could not be judged, because “Etched AI” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Etched AI.
Official links
What are the best Etched AI alternatives?
The closest alternatives to Etched AI we list are Rebellions, Recogni and Groq. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Rebellions | Contact Sales | Custom pricing | — |
| Recogni | Contact Sales | Custom pricing | — |
| Groq | Freemium | Pricing not published | — |
Alternatives to Etched AI
View all
Rebellions
Rebellions builds chiplet-based AI inference hardware — Rebel100 silicon, RebelCard, RebelServer, RebelRack and RebelPOD for enterprise LLM
Tools that pair well with Etched AI
Common stack mates teams adopt alongside Etched AI, with the specific reason each pairing earns its keep.
Rebellions
Rebellions builds chiplet-based AI inference hardware — Rebel100 silicon, RebelCard, RebelServer, RebelRack and RebelPOD for enterprise LLM
Recogni
Rack-scale AI inference system using log-space math to serve multi-trillion-parameter models at 608 PFLOPS per rack.
Groq
Groq delivers fast LLM inference on its own LPU and LPX silicon through an OpenAI-compatible API.
Frequently asked questions
Categories
Best-of guides
Topics
For makers of Etched AI
- Get a Verified Listing for Etched AI ($39): a written decision within 1 working day.
- Free listing for another tool: Free reviews are decided in order, within 3 weeks of the review starting.
Used Etched AI? Help shape our editorial sentiment research.