# A deployment benchmark is useful when its boundary is visible. FuriosaAI's LG adoption offers specific workload evidence, with performance-per-watt claims that cannot be generalized into universal savings. Canonical: https://kgcf.dsmlholdings.com/insights/furiosa-lg-rngd-enterprise-adoption/ Published: 2026-10-08 Author: [DSML Holdings LLC](https://www.dsmlholdings.com/) Company: FuriosaAI Event: 2025-07-22 Issuer announcement of LG AI Research adoption and completed accelerator installation. ## Reported metrics - Reported performance per watt: 2.25x. Issuer-reported LG EXAONE comparison against its tested GPU solution; not universal or independently audited. [Source 1](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) ## Reported evidence FuriosaAI reported that LG AI Research had adopted RNGD and installed hardware at its Koreit Tower data center. The issuer states 2.25 times better LLM inference performance per watt against the tested GPU solution, using EXAONE workloads. It also described planned supply of RNGD servers to other enterprises. The release does not disclose contract value or establish the same result for all models. ## Investment interpretation An installed enterprise reference is economically stronger than a specification alone. It can demonstrate that hardware, compiler and serving software operate together under customer requirements. The remaining commercial question is whether another customer can reproduce that outcome without disproportionate engineering cost. A successful design win is evidence of a route, not a disclosed order book. ## Economic assessment Compare complete deployment cost: equipment, power, integration, model conversion, utilization and support. Performance per watt can reduce one resource constraint while leaving acquisition and operating costs unresolved. Savings to LG are not automatically Furiosa margin. The supplier must price the accepted solution above its full delivery and continuing support burden. ## The Benchmark's Economic Boundary The reported 2.25-times performance-per-watt comparison concerns LG's tested EXAONE workloads and GPU solution. The release also reports EXAONE 3.5 32B at batch size one on a four-card server: 60 tokens per second with a 4K context window and 50 with 32K. That difference makes context handling commercially relevant even within one model; neither throughput observation is the denominator of a universal GPU-saving ratio. A buyer should fix context, batching, precision, latency and acceptable output before reproducing the comparison. Performance per watt also differs from cost per verified task: host equipment, integration and support still require payment. Supplier evidence can justify a procurement test without establishing independent universal superiority. Preserving these conditions strengthens the design win by identifying the workloads that may benefit and the engineering required to move beyond them. It also prevents the rack's nominal accelerator count from being mistaken for useful capacity sold under the customer's actual service requirements. 1. [FuriosaAI / LG adoption and test description](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) ## Integration as the Product The release describes scaling across accelerator configurations and work on software paths. That suggests the deployed proposition is not just a chip placed in a server. A customer needs models to compile, serve and remain observable within its infrastructure. The supplier's software effort can create defensibility if it supports repeatable deployment, but may become a cost burden if every account requires bespoke work. Analyze how much engineering is reusable and who pays for changes after installation. Compatible interfaces can lower migration friction without eliminating testing and operational risk. Model updates can introduce unsupported operations or different performance requirements. A capital plan should therefore include continuing software development throughout the hardware's useful service life. The deployment establishes that one customer route works; sustainable economics require a process that serves additional accounts without multiplying engineering headcount at the same rate. This is where the company's ability to standardize an accepted workload matters more than the apparent peak capacity of the accelerator. 1. [FuriosaAI / LG adoption and test description](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) ## Power Constraint and Capital Choice Power efficiency is valuable where the customer faces a real electricity, cooling or rack-capacity constraint. If the binding constraint is low utilization or insufficient demand, more efficient hardware may not recover its purchase cost quickly. The appraisal should compare both operating savings and the investment required to obtain them. An installed alternative can provide negotiating leverage even if it does not replace every GPU workload. That option has value, but its magnitude depends on compatibility and the cost of maintaining multiple stacks. For Furiosa, the business model must turn customer benefit into sufficient contribution after manufacturing, inventory and support. A low-power specification alone does not establish that margin. The most persuasive capital evidence would show repeat orders under commercially viable pricing, measured utilization and a support cost that becomes more predictable over time. A design win can justify further investment while leaving the scale and timing of the next funding requirement unresolved. 1. [FuriosaAI / LG adoption and test description](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) 2. [FuriosaAI / subsequent funding and design-win follow-through](https://furiosa.ai/blog/announcing-furiosaais-125m-series-c-bridge-funding-to-scale-sustainable-ai-compute) ## From Reference to Repeat Revenue LG is an important technical and enterprise reference, but the cited pages do not report the contract amount or customer cash receipts. Future supply to additional enterprises is described as a plan. Distinguish completed installation, current customer use and prospective expansion. A lender needs acceptance, payment and renewal evidence before treating a reference as a recurring cash stream. The supplier also must assess whether new customers buy the same configuration or require different hardware and software commitments. Inventory produced for one workload can become less useful if market preferences shift. Commercial expansion should therefore connect production orders with demand confidence and cancellation protection. The subsequent financing provides resources for scale, not proof that scale is already profitable. The investment case becomes stronger as technical acceptance, repeat procurement and collected contribution converge. Until then, the reported deployment is a meaningful reduction in execution uncertainty rather than a quantified financial return. 1. [FuriosaAI / LG adoption and test description](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) 2. [FuriosaAI / subsequent funding and design-win follow-through](https://furiosa.ai/blog/announcing-furiosaais-125m-series-c-bridge-funding-to-scale-sustainable-ai-compute) ## China - DSML comparison Chinese deployments would need independent workload and supply-route checks; this LG reference establishes none. ## Japan - DSML comparison Japanese buyers could evaluate the architecture, but no Japanese contract is disclosed. ## Other Asia - Reported connection A Korean accelerator developer supplies a Korean enterprise AI organization; this is the reported operating link. ## United States - DSML comparison US cloud alternatives are a commercial comparison, not reported US customer revenue in these pages. ## Europe - DSML comparison European energy constraints may make efficiency relevant, but local workload and service evidence remain necessary. ## Counterpoint A narrowly optimized architecture can be more economical than general-purpose hardware for suitable workloads. Its limitation is that an advantage in one workload may require substantial additional software work before it transfers to another. ## Underwriting questions 1. Can a prospective customer reproduce the reported result on its own workload? 2. What integration and support work is reusable across installations? 3. Which acceptance and payment records substantiate repeat commercial demand? ## Primary sources 1. [FuriosaAI / LG adoption and test description](https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus) (2025-07-22) 2. [FuriosaAI / subsequent funding and design-win follow-through](https://furiosa.ai/blog/announcing-furiosaais-125m-series-c-bridge-funding-to-scale-sustainable-ai-compute) (2025-07-30)