Public product program · 2024
Sohu
Public name for the transformer-inference processor program announced on 25 June 2024 and retired later in 2024. Etched presents the current 2026 system separately.
Company briefing
Etched is building rack-scale systems for AI inference.
Inference is the work of running a trained AI model to produce an answer. Etched designs the custom processor, memory system, boards, servers, racks and operating software used to perform that work.
The current system combines the first manufactured revision of Etched's processor with stacked high-bandwidth memory (HBM), links between processors, server control, power and liquid cooling. TSMC makes the processor using its N4P chip-manufacturing process.
Etched calls its low-voltage compute approach LVI: running the math blocks at lower voltage to reduce energy use. Its CSM design shares a combined fast-memory and HBM pool across linked chips. The CEO describes the first generation as one pre-assembled rack with 32 Etched chips, linked so each chip can use memory attached to other chips in the system.
Etched says it has more than $1bn in signed customer contracts. The contract orders cover complete systems containing processors, custom racks and software.
- System
- Rack-scale inference cluster
- Silicon
- First manufactured revision (called A0), made by TSMC using its N4P 4-nanometre process
- Physical form
- Accelerator modules, servers and liquid-cooled racks
- Operating footprint
- San Jose, Milpitas and Taiwan
- Seed financing
- Sohu program
- First manufactured chips
- Customer-system validation
01 / System
From a model request to a rack
Etched designs the compute, memory and connectivity together. Servers and racks provide the control, power and cooling around that silicon.
- 01 Model request Input and context
- 02 Inference software Model placement and execution
- 03 Etched accelerator Custom processor with stacked memory beside it and fast memory built in
- 04 Links between processors Move model data across accelerators
- 05 Server tray Boards, internal connectors and control computer
- 06 Liquid-cooled rack Power, cooling and networking
| Layer | What it does | System elements |
|---|---|---|
| Accelerator | Runs model operations beside stacked memory that feeds it data quickly | Custom processor, high-bandwidth memory, fast on-chip memory and physical signal interfaces |
| Server | Connects, starts and manages accelerators | Accelerator boards, a main board, internal connectors, a management controller, startup firmware and high-speed networking |
| Rack | Supplies power, cooling and cluster connectivity | 48 V conversion, power shelves, liquid loop, manifolds and networking |
| Stage | Publicly described position | What it means in practice | Dated account |
|---|---|---|---|
| System form | A pre-assembled rack with 32 Etched processors linked through Cluster Scale Memory, plus custom boards, power, liquid cooling and operating software. | The commercial unit is a complete inference appliance comprising processors and the surrounding system. | Etched CEO, Jun 2026 |
| First silicon and cluster | The first manufactured version, called A0, returned from TSMC's N4P 4-nanometre-class process in early 2026. Sequoia says the first cluster ran a range of models within 40 days. | The programme progressed from chip design to working manufactured processors and a linked multi-chip system. | Etched and Sequoia, Jun–Jul 2026 |
| Model workloads | Claims cover prompt processing, next-token generation, dense transformers, sparse mixture-of-experts models that activate only part of the network, and Mamba-class sequence models. | Etched describes support for both reading a prompt and generating the answer across several current model structures. | Etched and Sequoia, Jun–Jul 2026 |
| Customer evaluation | Etched describes direct rack validation with customers and an invite-only remote Evaluation Service that can use pre-release hardware, prototype systems and development software. | Customers can run workloads before general rack delivery through remote access and direct validation work. | 2026 service terms and updates |
| Production and delivery | Etched says it began fabricating hundreds of millions of dollars of clusters. On 30 Jun, Oceans investor Joshua Rahn described first-generation racks as shipping; Etched's same-day update described customer validation and scheduled first racks for summer 2026. | The public accounts place the programme between customer validation and initial rack delivery while manufacturing ramps. | Etched / Oceans, 30 Jun–23 Jul 2026 |
Chip program names
Chip-test name in a 2026 role
Caelius
Name used for chip-level manufacturing-test regressions in an Etched role published 13 April 2026. Etched filed Sohu and Caelius trademark applications together on 3 July 2025 for hardware and microchip goods.
Historical software and device lineage
2023–2026 development record
From graph prototype to manufacturing tests
pci.ids maps device 20a1:0001 to Etched / Sohu. In 2023, Etched converted PyTorch model graphs into SohuGraph. In 2025, Etched added a software command wrapper named SohuSendCmd. In 2026, Etched added code for testing chip settings and status while chips were still on the wafer.
02 / Company
Legal records and operating footprint
The business appears in US records under Etched, Inc. and Etched.AI, Inc., and through a newly registered Taiwan company. Its operating footprint spans San Jose, Milpitas and Taiwan.
| Record or site | Identifier and dated record | People or counterparties | Operating connection |
|---|---|---|---|
| Etched, Inc.US legal-name record | Active Delaware foreign-profit corporation in Florida; document F26000004323, filed 20 Jul 2026. Principal address: 3155 Olsen Drive, San Jose. | Robert Wachen, president; Telos Legal Corp., Florida registered agent. Mailing address: 56 Broad Street, Suite 14030, Boston. | Exact name used in current online-service terms and as the buyer in supplier purchasing terms effective 1 Jul 2026. |
| Etched.AI, Inc.US operating-record name | Delaware foreign-stock corporation qualified in California; entity 5730681, filed 22 May 2023. US labour filings use FEIN 92-0793021. | A Mar 2026 registry-derived record lists Gavin Uberti as CEO, acting CFO and secretary, with Mark Ross and Robert Wachen as other officers. | Name used in patent, lender, investment and Milpitas tenant-work records. Labour records trace the operating address from Cupertino in FY2024 to San Jose in FY2025–26. |
| 恆久科技股份有限公司Taiwan company | Unified Business Number 62107843, approved 21 Jul 2026; NT$500,000 registered and paid-in capital; 50,000 issued shares. | Gavin Uberti, responsible person and chair; Robert Wachen, director. Both represent 美商 Etched. ai, Inc.; the same 50,000-share juristic-person holding appears in both director rows. | Registered office: 3F, No. 168, Sec. 4, Chengde Road, Shilin District, Taipei. |
| Operating leadershipCompany roster and board record | Current company roster in Jul 2026; Primary Venture Partners board statement dated 30 Jun 2026. | Gavin Uberti, co-founder and CEO; Robert Wachen, co-founder and president; Chris Zhu, co-founder; Mark Ross, CTO; Brian Schechter, board member. | Leadership spans company direction, silicon and system execution; Primary identifies Schechter as a board member. |
| San Jose3155 Olsen Drive | Two SEC-filed lease records: 49,526 sq ft from 19 Mar 2025 for five years at $68.04 per sq ft gross; 15,197 sq ft from Jan 2026 for 4.6 years at $53.16 per sq ft triple-net. | The records identify the tenant as Etched AI / Etched.AI. Triple-net means the tenant also bears property taxes, insurance and maintenance. | Etched describes a 2-megawatt data centre, chip and system test house, and new-product prototyping facility running hardware around the clock. |
| Milpitas195 S Milpitas Boulevard | May 2026 lease; Jul 2026 demolition permits B-DM26-0032 and B-DM26-0034, valued at $150,000 and $75,000. | Etched AI, tenant; Peninsula Land & Capital, landlord/owner; Topa Architecture, applicant; Iron Construction, contractor. | Tenant preparation includes removal of rooftop equipment, three chillers, three pumps and piping. Etched describes an 80,000-square-foot, 10-megawatt site for new-product introduction, circuit-board assembly and deployment preparation. |
| Taiwan operating programmeFactory, lab and partner interface | YIIE lists an ongoing 2026 design-and-build project for “美商 Etched Inc.” covering 1,500 m². | YIIE, design-and-build firm; Etched's local hardware team; Taiwan design and manufacturing partners. | Etched describes factory, lab, hardware-engineering and partner-manufacturing work used for around-the-clock testing and production support. |
03 / Capital
Capital
Equity rounds, investor entry dates, direct fund positions, venture-credit terms and Etched-named investment vehicles.
| Date | Round | Amount / valuation | Lead | Public participants and timing |
|---|---|---|---|---|
| Seed | $5.4m$34m reported valuation | Primary Venture Partners | MAX Ventures; Devin Wenig. Primary says it invested in every round through Dec 2025. | |
| Series A | $120m | Primary Venture Partners; Positive Sum Ventures | Two Sigma Ventures; Earthshot Ventures; Soma Capital; Peter Thiel; Amjad Masad; Thomas Dohmke; Balaji Srinivasan; Kyle Vogt; Charlie Cheever. | |
| Series B | $500m$5bn reported post-money valuation | Stripes (reported lead) | Peter Thiel; Positive Sum Ventures; Ribbit; Jane Street (publicly named by Feb 2025); Hudson River Trading (backer); Two Sigma Ventures (since Series A); VentureTech Alliance (strategic investor); Radical Ventures; Argo; Andrej Karpathy; Fei-Fei Li; Primary Venture Partners (since seed). | |
| Series C | $300m$10.3bn valuation | Sequoia | Andreessen Horowitz; Jane Street; Diffusion; Argo; SK hynix. New investors in this round: Sequoia, Andreessen Horowitz, Diffusion and SK hynix. |
| Basis | Amount | Construction | Dated record |
|---|---|---|---|
| Named operating-company rounds | $925.4m | $5.4m seed + $120m Series A + $500m Series B + $300m Series C. | Publicly described rounds from Mar 2023 through 23 Jul 2026. |
| Issuer cumulative basis after Series C | About $1.1bn | $800m across four financings stated on 30 Jun 2026, plus the $300m Series C announced 23 Jul 2026. | Etched's two cumulative financing statements. |
| Unitemized portion of issuer basis | $174.6m | Approximate $1.1bn issuer basis less $925.4m of named rounds. | Difference carried by the issuer's cumulative financing basis. |
| March 2025 private-stock reference | $110.84m | Forge groups Etched AI, AI-1 and AI-2 private-stock class amounts dated 25 Mar 2025. | Forge structured private-stock entry dated 25 Mar 2025. |
30 Jun 2026: Etched reported $800m across four financings and also named Jump Trading, Geoffrey Hinton, Stanley Druckenmiller, Arthur Mensch and Scott Wu.
Jane Street: publicly named as a backer by 24 Feb 2025; an earlier lead role was disclosed on 30 Jun 2026.
Anti Fund: portfolio entry labelled Series B and “Since 2025.”
Hummingbird: says it first partnered with Etched in 2024 at the early stage.
| Instrument / channel | Holder / vehicle | Dated record | Reported position |
|---|---|---|---|
| Consolidated preferred position | StepStone Private Venture and Growth Fund / SPRING Subsidiary LLC | First purchase 30 Dec 2025; reported 31 Mar 2026 | 67,513 preferred shares of Etched.ai, Inc.; $1,039,078 cost and fair value. All or part held through SPRING Subsidiary LLC. |
| Preferred shares and warrants | TriplePoint Private Venture Credit Inc. (TPVC) | Shares 9 Apr 2024; preferred warrant 19 Apr 2024; common warrant 18 Jul 2025 | 307,570 preferred shares; separate 307,570-share preferred warrant; 12,110-share common warrant at 31 Mar 2026. |
| Venture credit | TriplePoint Private Venture Credit Inc. (TPVC) | Acquired 18 Jul 2025; balances at 31 Mar 2026 | $4.5m outstanding; $5.5m unfunded; Prime + 4.00% with a 10.75% floor; maturity 30 Apr 2028. |
| Venture credit — separate registrant | TriplePoint Venture Growth BDC Corp. (TPVG) | Acquired 18 Jul 2025; balances at 31 Mar 2026 | $4.5m outstanding; $25.5m unfunded; 10.75% floor; 12,110-share warrant. Earlier warrant schedules show 5,154 shares at Sep 2025 and 1,211 at Dec 2025. |
| Patent, trademark and copyright security | TriplePoint Capital LLC; TPVC | 19 Apr 2024; amended and restated 18 Jul 2025 | TriplePoint Capital LLC was collateral agent and lender; TPVC was lender. Security covers present and future patents, trademarks, copyrights and proceeds. |
| Investment-vehicle fund interests | CGF2021 / Sydecar Etched-named series (SEC Form D) | Vehicle first sales from 16 Apr 2024 to 9 Jul 2026 | Vehicle fund-interest amount sold: Etched AI Mar 2024 $1.633m; F&F Mar 2025 $1.659m; Series B F&F $4.060m; Angels III $8.760m; Angels IV $0.480m. |
| Investment-vehicle fund interests | Oceans Ventures Etched I–IV (SEC Form D) | First sales / filings from 2 May 2024 to 6 Feb 2026 | Vehicle fund-interest amount sold: Etched I $4.841m; Etched II and III $0 at their 29 Jan 2026 filings with first sale pending; Etched IV $10.000m. |
| Other Etched-named vehicles | Etched SPV, LLC; Innova Etched LLC | Innova position from 2 Jan 2026; Etched SPV first sale 5 Jan 2026 | Etched SPV fund-interest offering: $3.75m Form D amount sold; Benjamin Lincoln named manager. Innova: broker-reported $500,000 restricted Class A membership-interest position in a single-asset SPV holding Etched.AI, Inc. equity. |
04 / Peers
Inference-chip peers
The tables compare independent, merchant and operator-controlled inference systems.
| Company / category | Product and access | Corporate / capital | Upstream supply | System BOM and delivery evidence |
|---|---|---|---|---|
| GroqPrivate · operated cloud | Compiler-scheduled LPU using on-chip SRAM and direct chip links. GroqCloud and the Aramco/Dammam route provide operated capacity. | Groq, Inc.; $650m growth capital announced 22 Jun 2026, led by Disruptive and Infinitum. Groq remained independent after a non-exclusive NVIDIA licence. | GlobalFoundries was named for the planned deployment of more than 108,000 LPUs by the end of Q1 2025. | Groq reported 13 datacenters in Jun 2026. |
| CerebrasNasdaq CBRS · appliance / cloud | WSE-3 wafer-scale processor with 44 GB on-chip SRAM; CS-3 rack appliance and Cerebras Cloud. | Nasdaq-listed Cerebras Systems Inc.; IPO closed 15 May 2026. At 31 Mar 2026 it reported a $982.9m OpenAI working-capital loan and $2.7bn cash, cash equivalents and restricted cash; its $850m revolver was upsized 17 Jun 2026, and Amazon received a commercial warrant on 19 Jun. | TSMC fabricates the 5 nm wafers; ASE deposits redistribution layers; Cerebras performs final wafer package, assembly, test and burn-in in Sunnyvale. | Cerebras integrates PCB, I/O, cooling and power assemblies in the CS-3. Cerebras announced AWS CS-3 installations and a Trainium/CS-3 service in Mar 2026. |
| d-MatrixPrivate · production ramp | Corsair uses SRAM digital in-memory-compute chiplets, JetStream I/O and Aviator software. Full production was announced in Jun 2026; volume shipping was stated to begin that summer. | d-Matrix Corp.; $275m Series C at a stated $2bn valuation announced 12 Nov 2025. It acquired GigaIO's datacenter business on 2 Apr 2026; GigaIO remained independent. | TSMC N6 supports production ramp and Alchip worked from early design into production. Corsair uses organic substrate and low-power DDR5 memory (LPDDR5). | Arista, Broadcom and Supermicro are SquadRack reference-design partners. |
| SambaNovaPrivate · rack / cloud | SN40L reconfigurable dataflow unit (RDU); 16-RDU air-cooled SambaRack, on-premises SambaStack and SambaCloud. | SambaNova Systems, Inc.; $676m Series D led by SoftBank Vision Fund 2 in Apr 2021. Issuer equity notices span 2019–2021; the 14 May 2020 notice recorded $158.1m sold from a $250m offering. | SambaNova says TSMC manufactures SN40L. Each RDU combines HBM, attached DDR and SRAM. | SambaRack has two 64-core CPUs, 2 TB host DDR4, six 7.6-TB solid-state drives (NVMe), 400/200GbE, redundant power distribution units (PDUs) and air cooling. |
| Positron AIPrivate · customer evidence | Archer transformer-inference accelerators; Atlas is available as a 7U server. | Positron AI, Inc., a Delaware corporation registered in Florida on 20 Jun 2025 and active; $230m Series B announced 4 Feb 2026. | Reported manufacturing routes include TSMC Arizona and US card assembly. | Atlas is a 7U, eight-card, 256-GB-HBM, 2-kW server; Positron says it is shipping. Jump Trading is a named customer and evaluator. |
| FuriosaAIPrivate · card / service | RNGD is a passive PCI Express (PCIe) Gen5 inference card with 48 GB HBM3, 256 MB SRAM and a 180-W envelope; also offered in an eight-card server. | $125m Series-C bridge in Jul 2025; issuer-stated total funding of $246m. KDB, IBK, Keistone, PI Partners and Kakao Investment participated. | The company states TSMC 5 nm. Its first 4,000-unit receipt names TSMC and ASUS jointly. | Samsung SDS launched an RNGD NPU service on Samsung Cloud Platform on 20 Jul 2026 with one-, two-, four- and eight-card configurations. Management and device-plugin software support the card. |
| RebellionsPrivate · server and rack programme | Rebel100 has four chiplets connected by chip-to-chip links (UCIe), 144 GB HBM3E and 512 MB SRAM. RebelServer is an eight-card system; current navigation labels RebelRack and RebelPOD “Coming Soon.” | Rebellions and Sapeon Korea completed their merger under the Rebellions name on 2 Dec 2024. $250m Series C at $1.4bn in Sep 2025 and $400m pre-IPO at $2.34bn in Mar 2026; the SqueezeBits share exchange is announced for 1 Sep 2026. | A joint company release names Samsung Foundry 4 nm and Samsung HBM3E for Rebel. | The undated v1.7 guide documents a 5U, eight-card RebelServer with AMD EPYC 9355, 1.5 TB DDR5, NVMe, 400/100/25GbE and six 2.7-kW power supplies (PSUs); it says hundreds of racks are deployed with enterprise and government customers. |
| TenstorrentPrivate · merchant cards / server | Blackhole combines Tensix tensor and RISC-V cores with GDDR6. Public card, firmware and PCI records connect it to the 6U, 32-processor Galaxy system. | Tenstorrent Inc. is the Canadian federal Form-D issuer; its 10 Jun 2021 amendment recorded $164.0m offered and sold across equity and debt. Separate Delaware issuer Tenstorrent Holdings Inc. recorded $65.8m equity sold on 10 Jul 2025; a CIPO trademark-file security agreement names Stifel Bank on 8 Dec 2025. | Tenstorrent describes Blackhole as 6 nm. | Orderable P100/P150 cards and PCI ID 1e52:b140 connect Blackhole to Galaxy TG-00003: a 6U, 32-processor system with AMD EPYC 9354P, Broadcom N2200G, ASPEED AST2600, DDR5, NVMe and high-speed network ports (QSFP-DD). |
| Company / category | Product and access | Corporate / capital | Upstream supply | System BOM and delivery evidence |
|---|---|---|---|---|
| QualcommPublic parent · merchant card | Cloud AI 100 Ultra is a 150-W PCIe card with 128 GB LPDDR4x and 576 MB SRAM; offered for on-premises and cloud software deployment. | A product of Qualcomm Technologies, a wholly owned Qualcomm Inc. subsidiary; capital is carried at the public parent. | The platform uses a 7-nm process and LPDDR4x memory. | Aetina appliance and Core42 Compass service provide named routes for Cloud AI 100 Ultra. |
| Neuchips Viper-ALTaiwan · PCIe inference card | Viper-AL is a half-height, half-length (HHHL) PCIe 5.0 x8 large language model (LLM) inference card using the Raptor N3000 processor and 64 GB LPDDR5; neuchips-smi provides management. | Taiwan record UBN 54968983 shows US Neuchips Inc., represented by Lin Youn-Long, holding 73,545,111 shares, equal to reported total shares as of 24 Jun 2025; paid-in capital was NT$2.145bn. | Viper-AL combines the Raptor N3000 processor with 64 GB LPDDR5 memory. | Vendor documentation identifies PCI 1fd9:1001, field-replaceable-unit memory record (FRU EEPROM) at 0x50, active/passive cooling and 25/45/75-W power settings. |
| CambriconSTAR Market · merchant accelerator | MLU370 is a 7-nm chiplet accelerator family using LPDDR5 and MLU-Link; single- and dual-chip cards can form an eight-chip system. | Public Shanghai STAR Market company; annual reports carry holders, cash and debt. | MLU370 uses a 7-nm chiplet design and LPDDR5 memory. | Single- and dual-chip cards scale into an eight-chip MLU370 system. |
| MetaX688802.SH · merchant GPU | C500/C550 general GPUs and N-series inference products. C500 is a 64-GB PCIe card; C550 is described in a 128-card liquid-cooled cabinet. | Public STAR Market issuer since Dec 2025. Final listing and annual-report records carry capital, holders and debt. | SSE filings show purchases of wafers, HBM, package/test, card processing and external IP. | The company buys server, storage and network equipment; developer records show multi-card vLLM use. |
| KunlunxinBaidu-origin · platform-linked | R200 is a 7-nm GDDR6 PCIe card or accelerator module (OAM). Baidu Cloud separately describes eight-card P800 appliances and a 10,000-card cluster. | Originated in Baidu's chip department and completed separate financing in Apr 2021. | R200 combines a 7-nm processor with GDDR6 in PCIe and OAM forms. | R480-X8 uses eight OAM modules on an OAM baseboard; Baidu Cloud describes eight-card P800 appliances and a 10,000-card cluster. |
| Iluvatar CoreXHKEX 09903 · merchant inference card | ZhiKai 100 is a dual-slot PCIe inference card with 32 GB HBM2; Gen3 is the company's later roadmap programme. | Public in Hong Kong; dealings began 8 Jan 2026. Prospectus and annual-report records carry IPO proceeds, holders and debt. | ZhiKai 100 combines an inference processor with 32 GB HBM2 in a dual-slot PCIe card. | The 2022 product record documents the ZhiKai 100 card form. |
| Biren BR10X familyHKEX 6082 · public merchant family | BR106 entered mass production in Jan 2023; BR110 in Oct 2024; BR166 combines two BR106 dies. Biren markets PCIe cards, OAM modules, servers and clusters; BIRENSUPA is its software stack. | HKEX-listed PRC issuer; RMB5.6314bn listing proceeds were received in early 2026. | Issuer filings describe procurement and contract-manufacturing activity. | RMB1.027655bn 2025 product-sales revenue; 2025 volume production and commercialisation covered BR106 and BR166, with multi-thousand-card cluster deliveries at BR10X-family/company scope. |
| Enflame S60 / Suisi 320STAR IPO registrant · third-generation inference card | S60 is Enflame's 2024 third-generation Suisi 320 inference card: dual-chiplet, 300-W air-cooled PCIe and PCIe 5.0 x16. TopsRider V2.3.51 and V3.0 are registered software releases; the platform added native PyTorch ATen operators. | Shanghai Enflame Technology Co., Ltd.; Zhao Lidong and Zhang Yalin jointly control 28.14% of votes. STAR IPO registration was approved 9 Jul 2026; the Dec 2024 E round raised RMB2.71996bn. | Third-generation production used Supplier C for wafers and Supplier E as principal packaging/test provider. Tongfu Chaowei supplied packaging/test; Shennan Circuits and Tiantong Jingdian supplied board processing. GDDR6 is the disclosed memory type. | Third-generation revenue was RMB293.7909m in 2024 and RMB846.5605m in 2025. Enflame's company-wide sales were approximately 66,000 AI cards/modules (6.6万 as disclosed) in 2025; a Cloud Blazer cluster (POD) normally contains four to eight AI servers plus switches, with typically eight cards/modules per server, or 32–64 cards/modules. |
| Moore Threads MTT S5000 / PH100688795.SH · issuer-documented hardware/software | Moore Threads maps S5000 to fourth-generation Pinghu; vendor documentation maps it to PH100 and MUSA 5.2.0, with PyTorch, Megatron-LM, vLLM and SGLang support. | STAR-listed issuer since 5 Dec 2025; Zhang Jianzhong held 36.36% controlled voting rights at the prospectus date. A Dec 2024 pre-IPO issue raised RMB5.224529091bn and the IPO produced RMB7.57605227915bn net. | Issuer filings describe fabless wafer, package/test and board-processing procurement. | Vendor documentation records an FP8 test configuration using six servers with eight cards per server. |
| Huawei AscendIntegrated operator platform | Atlas 800I A3 is a 10U, eight-Ascend-910 inference server that scales into Huawei-controlled clusters, software and support. | Ascend is a Huawei group programme; hardware is delivered through Huawei's integrated platform and partners. | Huawei integrates Ascend into its controlled hardware, cloud and partner platform. | Four Kunpeng 920 CPUs, DDR5, eight 400GbE links, NVMe/SATA and driver/firmware tooling support the system. |
| Company / category | Product and access | Corporate / capital | Upstream supply | System BOM and delivery evidence |
|---|---|---|---|---|
| Google TPU7xAlphabet · cloud-operated | Ironwood / TPU7x is the latest Cloud TPU, accessed through Google Cloud, Google Kubernetes Engine (GKE) and Compute Engine for training, sampling and decode-heavy inference. | Alphabet / Google funds and operates the programme; customers buy Cloud TPU capacity. | Google designs TPU7x and integrates it into Google Cloud infrastructure. | Google documents pods up to 9,216 chips and cloud provisioning for Ironwood / TPU7x. |
| AWS Inf2 / Trn3Amazon · cloud-operated | Inferentia2 is AWS's inference-specific EC2 family; Trainium3 provides newer training and general-serving capacity. | Amazon funds and AWS operates the programme through EC2 and related managed capacity. | AWS integrates Inferentia2 and Trainium3 into EC2 and managed capacity. | Inf2 provides 32 GiB per accelerator and up to 12 accelerators per instance; the live host matrix names AMD EPYC 7R13 CPUs. |
| Microsoft Maia 200Azure · cloud infrastructure | Inference-first Azure accelerator: four directly linked devices per tray and Ethernet scaling to 6,144 accelerators. | Microsoft funds infrastructure used by Azure, Microsoft Foundry, Microsoft 365 Copilot and OpenAI workloads. | TSMC 3 nm, 216 GB HBM3E and 272 MB SRAM form the Maia 200 accelerator. | Maia 200 deployed in US Central in Jan 2026 with Ethernet scale-out and rack liquid cooling. |
| Meta MTIAMeta · internal fleet | Internal accelerators for ranking, recommendation and emerging generative-AI workloads in Meta's fleet. | Funded and operated inside Meta. The deployed v2 / MTIA 300 record is separate from the 400/450/500 GenAI-inference roadmap. | Meta designs MTIA and integrates it into its internal fleet. | Meta said in Mar 2026 that it was deploying hundreds of thousands of MTIA chips for inference. |
| Alibaba / T-Head Zhenwu M890 / Panjiu AL128Alibaba-controlled · cloud/system route | M890 has 144 GB of GPU memory and 800 GB/s inter-chip bandwidth; Alibaba says M890 and ICN Switch 1.0 underpin the 128-accelerator Panjiu AL128, available through Bailian/Model Studio in China. | Alibaba-controlled programme; Alibaba's RMB380bn infrastructure commitment is parent-level. | T-Head is Alibaba's internal chip-design subsidiary and supplies M890 into Alibaba's controlled hardware/cloud stack. | ICN Switch 1.0 is specified at 25.6 Tb/s; AL128 provides PB/s-class internal bandwidth. |
| Tencent / Zixiao C200 / PTX2Tencent-controlled · billable cloud capacity | PTX2 resource shapes with one, two, four or eight Zixiao C200 accelerators are available through Tencent Cloud TI-ONE. | Tencent-funded group programme. | Tencent supplies C200 inside its own cloud service and binds it to PTX2 host configurations. | The published PTX2 host shapes run from 32 virtual CPU cores (vCPU)/96 GB RAM for one C200 to two eight-C200 configurations with 256 vCPU/768 GB and 512 vCPU/1,536 GB host resources. |
| Nvidia GB300 NVL72Public · merchant rack baseline | Current Blackwell Ultra rack platform sold through direct, OEM and cloud channels: 72 GPUs, 36 Grace CPUs and 288 GB HBM3E per GPU. | A product of public Nvidia; manufacturing and system expansion are financed at the group level. | Nvidia integrates Blackwell Ultra GPUs, Grace CPUs, NVSwitch and ConnectX-8 into GB300 NVL72. | Eighteen compute trays, nine NVSwitch trays, ConnectX-8, eight 33-kW power shelves and liquid cooling; reference rack demand reaches about 142 kW. |
| AMD MI350Public · merchant card / OAM | MI350-series accelerators provide 288 GB HBM3E and ship as PCIe cards or eight-OAM OCP-UBB platforms through OEM and cloud routes. | A product of public AMD; capital, cash and debt sit at the group level. | AMD names Micron and Samsung Electronics as MI350 HBM3E suppliers. | Infinity Fabric, OCP-UBB and air- or direct-liquid-cooled systems support OEM and cloud deployments. |
| Intel Gaudi3Public · merchant OEM / cloud | Gaudi3 uses 128 GB HBM2E and 24×200GbE ports; available as PCIe or eight-OAM systems through OEM and cloud channels. | Intel acquired Habana in 2019; Gaudi is an Intel product. | Intel integrates Gaudi3 with HBM2E and its 24×200GbE networking interfaces. | Dell and Supermicro document shipping eight-accelerator systems, 800GbE, server-management controller (BMC), high-power PSUs and air/liquid cooling; IBM Cloud is a named service route. |
05 / Supply chain
One system, built in layers
The product moves from chip design through processor and stacked memory manufacturing, assembly into one package, circuit-board production, server integration, rack power and liquid cooling.
- Design and intellectual property
- Silicon and memory
- Package and test
- Board and server
- Power and cooling
- Deployment
The manufacturing records use L6 for a completed board, L10 for a complete server or tray, and L11 for a complete rack.
- Direct relationship
Etched or the counterparty names the work, tool or collaboration.
- Etched-defined role
Etched describes the required manufacturing function or platform.
- Likely supply route
The closest company or product route for the stated requirement.
- Alternative route
Another company with the required component or operating capability.
Map
Attributable build chain
The currently attributable chain runs from design automation to N4P wafers, memory bring-up, package development, external manufacturing and Etched-operated validation.
| Layer | Party or route | Relationship | Public operating connection |
|---|---|---|---|
| Chip specifications and register automation | Agnisys | Direct relationship | Its IDesignSpec workflow turns one machine-readable specification into chip logic, verification models, software headers and documentation. |
| PCI Express 5 verification | Eximietas Design | Direct relationship | A 2025 project tested the high-speed server link from initial connection through data movement, error recovery and full-chip simulation. |
| Stacked-memory controller | Rambus | Direct relationship | Supplied the control logic for high-bandwidth memory and offered integration services between that controller and the memory's physical electrical interface. |
| Logic wafers | TSMC N4P | Direct relationship | TSMC manufactures Etched's processor on its N4P 4-nanometre-class process and worked with Etched during yield experiments and the first wafer-level tests. |
| Memory development and first-system bring-up | SK hynix | Direct relationship | Supplied high-bandwidth-memory parts before silicon arrived and sent engineers to help establish reliable electrical communication during initial hardware testing. |
| Processor-and-memory package | TSMC collaboration and CoWoS-targeted development | Direct relationship | Etched's packaging lead named TSMC as a collaborator. Etched also described pre-silicon work for CoWoS, TSMC's method for placing the processor and memory stacks together on one large package. |
| Wafer, packaged-chip and system test | Advantest V93000; Chroma | Etched-defined role | Etched's manufacturing roles specify these test platforms for chips before wafer cutting, after packaging and under high-power system load. |
| Boards, servers and complete racks | Unnamed Taiwan product-design partner and Tier-1 contract manufacturer | Etched-defined role | The described handoff covers completed boards (L6), complete servers or trays (L10) and complete racks (L11), including assembly, test, rework, capacity and yield. |
| Production records, inventory and returns | Teamcenter; NetSuite | Etched-defined role | Teamcenter carries controlled product and engineering-change records into manufacturing; NetSuite joins purchases, receipts, inventory, finished units, shipments and returns. |
| Later-generation optical links | Etched optical-interconnect programme | Etched-defined role | A role posted 22 May 2026 covers 100- and 200-gigabit-per-second lanes, standard pluggable modules, moving optical engines nearer to or into the processor package, external lasers, fibre attachment and volume-production test. |
A
Design, verification and intellectual property
Turns Etched's architecture into verified chip logic, memory interfaces, boards and mechanical designs before manufacturing begins.
| Component or work | Company | Relationship | Role and operating detail |
|---|---|---|---|
| Executable specifications and register automation | AgnisysPrivate · US and Noida, India | Direct relationship | IDesignSpec and associated licensed design technology. The workflow generates chip logic, verification register models, software headers and documentation from one specification. Disclosed 19 Dec 2024. |
| High-bandwidth-memory controller | RambusNasdaq RMBS · controller lineage from Northwest Logic | Direct relationship | Provided controller intellectual property and offered services to integrate the controller with the physical memory interface. Disclosed 17 Jun 2025. |
| PCI Express Gen 5 verification | Eximietas DesignBengaluru design-services operation | Direct relationship | January–December 2025 work covered link training, packet and flow control, error recovery, regression coverage and gate-level simulation at both block and full-chip level. |
| Near-manufacturing design recovery | Bengaluru development supplier | Etched-defined role | Etched's CEO described embedding roughly 12 senior engineers at the supplier for about six months to audit code, build tools and make design decisions close to the chip's manufacturing release. |
| Historical full-chip emulation before silicon | Synopsys ZeBuNasdaq SNPS · ZeBu Server 5 | Direct relationship | Historical Sohu-program work ran complete transformer workloads and software against an emulated chip. A March 2026 account describes 16 ZeBu Server 5 systems; Etched's CEO described a multi-year equipment-payment arrangement. |
| Chip implementation and circuit-board design | CadenceNasdaq CDNS · Innovus and Allegro | Etched-defined role | Innovus is used for full-chip implementation fixes; Allegro is used for accelerator and server circuit-board design. Etched also presented package-and-board work at CadenceLIVE 2026. |
| Mechanical and thermal design | Siemens NXSiemens Digital Industries Software | Etched-defined role | Etched calls NX its standard three-dimensional design platform for mechanical and thermal engineering across packages, boards, trays and racks. |
| Electromagnetic modelling | Ansys HFSSSynopsys group | Etched-defined role | Builds and validates three-dimensional electromagnetic models for package and board signal and power behaviour. |
| Physical signal interface to stacked memory | Cadence HBM3E physical interface | Likely supply route | The closest published pairing with Rambus controllers; Cadence has a 9.6-gigabit-per-second HBM3E interface qualified for TSMC N5/N4P. |
| Physical signal interface to stacked memory | Synopsys DesignWare; Global Unichip | Alternative route | Synopsys offers a hardened HBM3 interface for TSMC's N5 process family. Global Unichip offers a 9.2-gigabit-per-second HBM3E-generation controller/interface validated on N4P and two CoWoS package variants. |
B
Silicon, memory, package and test
Makes the processor and memory stacks, joins them on a large package substrate and tests the device at wafer, package and system level.
| Component or work | Company | Relationship | Role and operating detail |
|---|---|---|---|
| Logic wafers | TSMCTaiwan · TWSE 2330 / NYSE TSM | Direct relationship | Manufactures Etched-designed logic using the N4P process. Etched described a successful yield experiment run by TSMC and its first wafer-sort session with TSMC personnel in Taiwan. |
| Development high-bandwidth memory and bring-up | SK hynixSouth Korea · KRX 000660 | Direct relationship | Provided HBM parts for pre-silicon package tuning and sent engineers to Etched for memory signal training during initial hardware testing. SK hynix also joined the 23 Jul 2026 Series C. |
| Production high-bandwidth memory | SK hynix HBM3E | Likely supply route | The first route because SK hynix has a direct memory-hardware relationship with Etched. It began 12-layer HBM3E production in 2024 and describes HBM3E as a major 2026 volume product. |
| Production high-bandwidth memory | Samsung; Micron | Alternative route | Samsung produces 36 GB 12-high HBM3E. Micron manufactures HBM3E with major DRAM and back-end operations in Taiwan. Either route would require package, signal, firmware and thermal qualification. |
| Low-voltage accelerator-package development | TSMC | Direct relationship | Etched's Head of Advanced Packaging publicly acknowledged TSMC's collaboration on the package in July 2026. |
| Join logic and memory on one package | TSMC 3DFabric / CoWoS-S | Likely supply route | CoWoS-S uses a silicon interposer to connect the processor and HBM stacks. TSMC offers integrated package assembly and test, and Etched's CEO described historical pre-silicon CoWoS tuning. |
| External package assembly and test | ASE / SPIL; AmkorTaiwan and wider Asian package/test networks | Alternative route | ASE/SPIL provides flip-chip, through-silicon-via and bridge-based multi-die packaging. Amkor provides side-by-side multi-die package assembly and test across Asian manufacturing sites. |
| Package substrate — the circuit board beneath the silicon | Unimicron / UMTC; Ibiden; ToppanAdditional Taiwan routes: Kinsus and Nan Ya | Likely supply route | The first three are TSMC 3DFabric substrate members. Unimicron and Ibiden worked with TSMC on substrate design files and stress-hotspot checks; Toppan's Niigata high-end line began operating in Jan 2026. Kinsus and Nan Ya add Taiwan capacity for fine-pitch, high-layer-count package carrier boards. |
| Substrate build-up dielectric | Ajinomoto Fine-TechnoJapan · package-substrate dielectric | Likely supply route | Supplies the insulating build-up film used to create fine wiring layers in advanced package substrates; its Gunma varnish plant entered customer qualification in 2025. |
| Low-expansion glass cloth | NittoboJapan with Chiayi and Taoyuan operations | Likely supply route | T-glass and NE-glass fibre and cloth reduce expansion and signal loss in AI and high-performance package substrates. |
| Fine copper for substrate wiring | Mitsui KinzokuJapan / Malaysia two-site route | Likely supply route | MicroThin fine copper supports very small substrate wiring. Its high-frequency VSP foil and FaradFlex materials also map into the server-board stack. |
| Underfill, moulding and thermal-interface materials | Resonac; Shin-Etsu | Likely supply route | Resonac supplies underfill, non-conductive film, mould compound and thermal materials for large logic/HBM packages. Shin-Etsu supplies silicone thermal-interface families. |
| Wafer sort and final chip test | Advantest V93000 | Etched-defined role | Etched's test hiring explicitly requests V93000 programme experience for wafer sort, packaged-device testing and HBM test development. Possible outsourced operating sites include TSMC test services, ASE and King Yuan Electronics. |
| System-level test | ChromaTaiwan test-equipment company | Etched-defined role | Named experience for high-power system test using load boards, handlers, thermal control, production programme bring-up and correlation with the test-equipment maker. |
C
Circuit board, server and rack build
Mounts packaged accelerators on high-layer boards, builds server trays and integrates power, networking, cooling and controls into complete racks.
| Component or work | Company | Relationship | Role and operating detail |
|---|---|---|---|
| Completed boards, server trays and racks | Taiwan Tier-1 contract manufacturer | Etched-defined role | Runs surface-mount assembly, mechanical assembly, server construction, inspection, test, rework, capacity planning, yield improvement and the move from new-product builds into volume production. |
| Component supply and build readiness | Taiwan product-design and manufacturing partner | Etched-defined role | Coordinates custom and standard components, circuit boards, interposers, electronics and mechanical parts for completed boards and servers through engineering validation, design validation, production validation and ramp. |
| Product lifecycle management | TeamcenterSiemens PLM environment | Etched-defined role | Named environment for managing product records, engineering changes and manufacturing handoffs. |
| Manufacturing, inventory and returns control | NetSuite | Etched-defined role | Initiative for manufacturing, supply-chain, logistics and reverse-logistics operations, joining purchases, receipts, inventory movements, finished units, shipments and returns. |
| Freight, customs and warehousing | Forwarders, customs brokers and warehouse partners | Etched-defined role | The operating chain tracks supplier and factory shipments through invoices, packing lists, delivery terms, customs records, warehouse receipts and final destinations. |
| Board-to-rack manufacturing prime | Ingrasys / FoxconnFoxconn group · Ingrasys based in Taoyuan | Likely supply route | The broadest component-to-system route: server boards, network equipment, mechanical parts, liquid cooling and complete rack design and manufacture. |
| Server and liquid-cooled rack integration | Wiwynn / WistronWiwynn TWSE 6669 · Wistron held 35.46% in Mar 2026 | Likely supply route | Custom cloud servers, storage and L10/L11 rack integration with direct liquid cooling, system test and volume manufacture. Wiwynn also publishes a custom-silicon-to-rack operating model with Global Unichip. |
| Server, switch and complete-rack integration | QCT / QuantaQuanta cloud-technology group · TWSE 2382 | Likely supply route | Designs and manufactures AI servers, network switches and integrated racks. Its public LITEON pairing provides a concrete model for combined rack power and cooling. |
| Server new-product introduction and traceable build | PegatronTWSE 4938 · Guishan, Taoyuan server plants | Likely supply route | Server-board assembly and system test with work-order, material, inspection and per-unit barcode genealogy. Current Taoyuan certificates cover motherboards, circuit-board assembly and servers. |
| Server joint development and box build | USIASE Technology group · Nantou, Taiwan | Alternative route | Publishes an L10 service spanning motherboard, boot firmware, management controller, daughtercards, enclosure, thermal design, system integration, test and certification; its Tsaotuen site has six high-speed assembly lines. |
| Additional server and rack primes | Inventec; MiTAC; Compal | Alternative route | Taiwan server manufacturers with board-to-rack, system-test and liquid-cooling capability. MiTAC also appears in a documented Hyve Solutions liquid-cooled factory model. |
| High-layer accelerator and server circuit board | Gold Circuit ElectronicsTaiwan · TWSE 2368 | Likely supply route | Dedicated multilayer circuit-board maker with stated AI, cloud-server, server and switch exposure—well aligned with Etched's high-layer, low-loss, back-drilled 112-gigabit-per-second-class board requirements. |
| High-layer accelerator and server circuit board | Compeq; Unimicron; Zhen Ding | Alternative route | Large Taiwan circuit-board groups with high-density and server/network board capability; separate from the package-substrate workstream. |
| Low-loss board core and prepreg | Taiwan Union TechnologyThunderClad 5R · TU-953R / TU-953R N; alternatives: Elite Material, ITEQ and Ventec | Likely supply route | Core and prepreg designed for AI servers, networking and high-performance computing, with next-generation production work centred in Taiwan. Elite Material, ITEQ and Ventec supply alternative low-loss laminates for the same high-layer board requirement. |
| 48-volt board power conversion | Monolithic Power SystemsNasdaq MPWR · accelerator power systems | Likely supply route | Multiphase conversion from the 48-volt input to the processor's low-voltage core, memory, PCI Express and auxiliary rails. |
| 48-volt board power conversion | Infineon; Renesas; Texas Instruments; onsemi | Alternative route | Digital multiphase controllers, power stages, hot-swap control and monitoring for high-current accelerator boards. |
| Host processor | AMD EPYC; Intel Xeon; Ampere | Alternative route | Provides the server control plane and PCI Express root complex around accelerator cards. |
| Server management controller and firmware base | ASPEED; Nuvoton; AMI | Likely supply route | Management-controller silicon and firmware components for boot, remote management, root of trust, recovery and platform debugging; Etched appears to retain substantial integration work. |
| PCI Express switches and signal-restoring chips | Broadcom; Credo; Astera Labs | Likely supply route | Gen 5 switches, retimers and smart cable modules that preserve signal quality between accelerator cards and the host server. |
| External Ethernet network | Broadcom; Marvell; NVIDIA Networking | Alternative route | Switch, network-interface, physical-layer and cable silicon for the 100-gigabit Ethernet functions described in Etched's system roles. |
| High-speed connectors and cable harnesses | TE Connectivity; Molex; Luxshare; Amphenol; Samtec; BizLink | Likely supply route | Board, card and rack interconnect. Distinct product families include TE Sliver/STRADA, Molex Impulse/Impress and Luxshare KOOLIO/Intrepid APEX. |
| Enclosures, racks and slide rails | Chenbro; Ablecom; King Slide | Likely supply route | Custom sheet metal, chassis, structural parts, racks, rails and cable-management arms. AIC and InWin are additional Taiwan chassis routes. |
D
Rack power and liquid cooling
Converts facility power, feeds high-current servers and removes processor heat through cold plates, manifolds, quick disconnects and coolant-distribution equipment.
| Component or work | Company | Relationship | Role and operating detail |
|---|---|---|---|
| Processor cold-plate design and validation | Etched | Etched-defined role | Etched assigns itself the direct-to-chip cold-plate design: internal pedestal, nozzle and skived-fin geometry; thermal-interface material; internal manifolds; quick disconnects; coolant-distribution equipment; and heat exchangers. Engineers also built thermal stand-ins matching expected hot spots and pressure-tested plates to failure before working silicon arrived. |
| Rack power shelf | Delta ElectronicsTaiwan · TWSE 2308 | Likely supply route | Spans server power, power shelves, battery backup and liquid cooling. Delta publishes a 90 kW 800-volt-direct-current to 54-volt shelf matching Etched's stated planning envelope of up to 90 kW per rack. |
| Rack power shelf and server power supplies | LITEONTaiwan · TWSE 2301 | Alternative route | Produces AI-server power supplies, shelves, battery-backup units, racks, cold plates and coolant-distribution units; its QCT collaboration includes a 110 kW shelf. |
| Rack power and battery backup | Advanced Energy; AcBel | Alternative route | Open-rack and 48-volt power shelves, server power supplies and battery-backup systems. |
| Cold plates, manifolds and coolant distribution | AurasTaiwan · TWSE 3324 | Likely supply route | Publishes customised accelerator cold plates, universal quick disconnects, inner and rack manifolds, pumps and coolant-distribution units. |
| Cold plates and manifolds | Asia Vital ComponentsTaiwan · TWSE 3017 | Alternative route | Taiwan thermal supplier with cold-plate, manifold, chassis and system-integration capability. |
| Custom liquid loop and field support | CoolIT; Delta; LITEON | Alternative route | Custom cold plates, rack manifolds, coolant-distribution units, validation and field support. CoolIT manufactures in Canada and through Asian partners and maintains Taiwan support. |
| Liquid-cooling quick-disconnect standard | CPCDover operating company · Minnesota | Direct relationship | CPC and Etched co-authored the 26 Sep 2025 UQDv2 beta specification and presented it in Oct 2025. CPC's UQD08/UQDB08 families cover rack-scale handmate and blindmate connectors. |
| Interchangeable quick disconnects | Cejn; CoolIT; Danfoss; Stäubli; Parker | Alternative route | UQDv2-compatible contributor ecosystem. The standard requires visible type, size and version marks and cross-validation between suppliers. |
| Facility-side coolant distribution | Auras; Delta; LITEON; CoolIT; Vertiv | Alternative route | Coolant-distribution equipment, facility heat rejection, controls, redundancy and service for the connection between a liquid-cooled rack and the building plant. |
E
Where shipments can bind
The chain is serial: usable wafers, memory, package, board, rack and test results must all arrive together.
| Gate | How it blocks a rack | Why substitution is slow |
|---|---|---|
| Usable N4P logic die | Functional first silicon must become repeatable wafer starts with enough good die per wafer. | A foundry or process change requires a new physical implementation, masks and qualification. |
| Qualified HBM stacks | A complete package requires enough electrically and thermally qualified memory stacks. | Changing maker or generation changes signal training, package mechanics, firmware and thermal validation. |
| Large-package compound yield | One attach, interposer, warpage, substrate or final-test defect can lose good logic and several expensive memory stacks. | Package platform, site, materials and test flow are co-qualified as one assembly. |
| High-layer board yield | Via, impedance, laminate, power-delivery or assembly defects stop a packaged accelerator from becoming a working board. | Board stack-up, materials and factory process must be electrically, mechanically and thermally requalified. |
| Taiwan board-to-rack ramp | Materials, fixtures, rework, test programmes, line capacity and product revisions must converge at the same time. | Tooling, process recipes and failure knowledge accumulate at the manufacturing prime. |
| Power and liquid cooling | A server can boot yet fail under sustained current, heat, pressure, flow or facility-water conditions. | The processor, board, firmware, cold plate, manifold, power shelf and building interface are validated together. |
| Wafer, package and system test | Weak coverage releases marginal memory, interconnect or thermal parts; excessive test time throttles output. | New programmes need correlation across testers, load boards, handlers, sites and production stages. |
F
How Etched expects suppliers to operate
Etched's standard purchase-order terms set the timing, quality, continuity and field-failure controls around the physical chain.
| Control | Etched term | Operational effect |
|---|---|---|
| Order response | Respond to a purchase order within three business days; production forecasts remain planning inputs. | Etched gets a fast commitment signal while separate capacity or long-term agreements govern firm supply. |
| Payment and delivery | Payment initiation at 90 days; supplier delivery to the destination, including freight and customs, is the default. | Suppliers may carry material, working-capital and cross-border execution responsibility during a ramp. |
| Process and site changes | Etched approval for product or process changes and notice within ten days for material changes to controls, subcontractors or manufacturing sites. | Board, package and rack configurations remain traceable through changes below the prime supplier. |
| Product continuity | Twelve months' notice before manufacturing discontinuance and a six-month last-time-buy window. | Etched has time to secure inventory or qualify a replacement before a part reaches end of life. |
| Warranty and systemic failure | Three-year product warranty; the systemic-failure trigger is 2% of units in any three-month period. | Serial-level genealogy, field returns and common-failure analysis become financially important. |
| Resilience and subcontracting | Annual-tested business-continuity and disaster-recovery plans; subcontracting requires Etched approval. | The prime supplier must control recovery plans and material changes across its sub-tier network. |
06 / Production
From design release to a running cluster
Each step adds physical hardware, test data and system-level control until the rack can run customer workloads.
- 01
Custom processor and memory design
Finish the processor logic, memory system, package and signal interfaces.
- 02
Processor, memory and package
Make the silicon, place stacked memory beside the processor and assemble the complete device.
- 03
Device test
Test each chip before cutting the wafer, after packaging and inside a working system.
- 04
Board and server build
Assemble accelerator boards, main boards and complete servers.
- 05
Rack integration
Install power, networking and liquid cooling, then run the rack under load to find early failures.
- 06
Customer operation
Deploy the cluster, load models and operate inference services.