European semiconductor startup Axelera AI unveiled on November 12, 2026, the architecture details of Titania, its third-generation AIPU (AI Processing Unit). The chiplet design aims to deliver data-center-class performance inside existing air-cooled racks without requiring high-power liquid-cooled infrastructure.
What happened?
Each Titania System-in-Package (SiP) combines two chiplets based on the third-generation D-IMC architecture. The package delivers 2.5 PFLOPS of FP8 compute at 400 W, with 128 GB of LPDDR6 (up to 2,458 GB/s bandwidth), 128 MB of shared SRAM and 32 RISC-V cores. It natively supports MXFP4, FP4, MXFP8, FP8 and BF16, plus on-chip video codecs (AV1, H.265) for 8K decode and 4K encode.
A reference configuration with eight SiPs in a 4U air-cooled server provides roughly 20 PFLOPS FP8, 1 TB of memory and capacity for a 1.6-trillion-parameter MoE model at FP4. The company claims up to 7x throughput per watt and up to 20x better TCO versus high-density data-center servers. More than $300 million in confirmed commitments already exist. Production is scheduled for the third quarter of 2027.
The platform uses the same Voyager SDK as previous generations (Metis and Europa), enabling workload migration without code rewrites. Form factors include PCIe and OCP OAM modules, air- or liquid-cooled.
Why does this matter?
Most high-capacity accelerators rely on HBM and liquid cooling in racks drawing tens of kilowatts. Titania bets on LPDDR6 and D-IMC efficiency to run large models (including sparse MoE and agents) on infrastructure companies already own. According to CEO Fabrizio Del Maffeo, the core challenge of inference is tokenomics: many customers cannot justify CAPEX to rebuild data centers.
Axelera simulations indicate a 4U air-cooled server can match the per-user speed of a liquid-cooled Vera Rubin NVL72 rack on DeepSeek-R1 while maintaining interactive responsiveness.
What changes in practice?
Enterprises, neoclouds, telcos and sovereign clouds gain a high-capacity inference option without renovating their data centers. The PCIe card is a drop-in replacement for the Europa-based Server 250p. The architecture was designed in Europe with support from EuroHPC. Additional technical details will be presented at SC26 in Chicago (November 15-19).
Sources: official Axelera AI announcement and IT Home coverage.
Image credit: Axelera AI | Source: official announcement
By Grok
