Axelera AI reveals datacenter AIPU chip Titania, production in Q3 2027
Axelera AI 公布数据中心级 AIPU 芯片 Titania 架构，2027Q3 投产
A startup claiming air-cooled 4U performance comparable to NVIDIA's liquid-cooled NVL72 rack, if validated, reshapes inference hardware economics for anyone running large MoE models. Datacenter operators and enterprises planning 2027 inference capacity should track its benchmarks against NVIDIA alternatives. European chip challengers entering the NVIDIA-dominated datacenter market remain rare, making Axelera's order book a signal of real demand.
Dutch AI chip startup Axelera AI announced the architecture of Titania, its third-generation AIPU chip targeting the datacenter inference market, on 12 November. Mass production is expected in Q3 2027, and the company has already secured over US$300 million in orders. Titania is designed with a 400W power envelope and consists of two compute chiplets built on the third-generation D-IMC architecture.
The chip is optimised for sparse MoE models and supports MXFP4 and MXFP8 data formats, delivering 2.5 PFLOPS of FP8 compute with 128MB of shared on-chip SRAM cache. It embeds 32 RISC-V cores to offload host CPU workloads, includes AV1 and H.265 codecs, and integrates 128GB of LPDDR6 memory in a SiP with 2456GB/s of bandwidth. Titania ships in PCIe AIC and OCP OAM form factors; a single 4U air-cooled eight-card server can run a 1.6T-parameter MoE model locally at FP4 precision.
Axelera AI claims Titania can achieve 7x the throughput per watt of current datacenter-class servers at 1/20 of the total cost of ownership, and says the per-user DeepSeek R1 speed of a 4U Titania air-cooled server can match NVIDIA's liquid-cooled Vera Rubin NVL72 rack. The claims position the startup against NVIDIA in the inference market.
Axelera AI 公布数据中心级 AIPU 芯片 Titania 架构，2027Q3 投产
IT之家 11 月 13 日消息，荷兰芯片初创企业 Axelera AI 当地时间 12 日 公布了其第三代 AIPU 芯片 Titania 的架构 。其瞄准数据中心推理市场，预计 2027 年第 3 季度量产，已获得超 3 亿美元 （IT之家注：现汇率约合 20.06 亿元人民币） 的订单。 Titania 芯片设计功耗为 400W，由 2 个基于第三代 D-IMC 架构的计算芯粒组成。其针对稀疏 MoE 模型优化，支持 MXPF4、MXFP8 数据格式， FP8 算力可达 2.5PFLOPS ；集成 128MB 共享 SRAM 片上缓存。 此外，Titania 内嵌 32 个 RISC-V 内核，可卸载主机 CPU 负载；还拥有 AV1、H.265 等多格式编解码器； SiP 集成 128GB 的 LPDDR6 内存 ，总带宽 2456GB/s。 在更高的层面，Titania 支持 PCIe AIC、OCP OAM 形态； 一台 4U 的八卡 Titania 风冷服务器能以 FP4 的精度本地运行 1.6T MoE 模型 。 Axelera AI 表示，Titania 有望实现当前数据中心级服务器 7 倍的每瓦吞吐量、1/20 的总拥有成本。 4U Titania 风冷服务器的 DeepSeek R1 每用户速度足以匹配 NVIDIA 的 Vera Rubin NVL72 液冷机架 。
This report is categorized under AITOP's "Industry" desk, focusing on Europe's AI governance and technology ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Axelera AI, Titania, NVIDIA, AI chips.
Source attribution: curated from IT之家. Cross-source corroboration: 11 independent sources have covered this development (including CoreWeave, vLLM, a16z), confirming industry-wide signal strength.
