This morning’s Lightbits Inferra note was the prefetch half of the KV ladder after HBM fills. Last night’s Astera Leo X note was the fabric-attached DDR half. Tonight is a different purchase: a shipping inference AIPU with named watts, TOPS, PCIe SKUs, and OEM server validation — not another optics booth and not another memory-tier slide.
On September 15, 2026, Eindhoven-based Axelera AI put Europa on the AI Infra Summit floor as its second-generation AI Processing Unit, extending the Metis edge architecture into enterprise and Physical AI inference. Primary sources: Axelera’s launch press release, StorageReview’s Sep 15 ship write-up, New-TechEurope’s Sep 21 product brief, and Axelera’s Europa / Edge 232p product pages. Specs and tok/s deltas below are vendor claims unless a named third-party pod result is cited.
Spec sheet (named numbers)
| Product | Europa — second-generation AI Processing Unit (AIPU) |
| Vendor | Axelera AI (Eindhoven; Metis gen-1 edge AIPU) |
| Public launch | Sep 15, 2026 — AI Infra Summit Santa Clara (booth 930; CEO keynote same day) |
| Process | Samsung 5 nm; Digital In-Memory Compute (D-IMC) cores |
| Compute | 8 second-generation AI cores → claimed 629 TOPS at INT4 / INT8 / INT16 |
| Host offload | 16 RISC-V vector cores for on-chip pre-/post-processing |
| On-chip memory | 128 MB L2 SRAM |
| External memory BW | 200 GB/s DRAM bandwidth |
| Power | 45 W TDP on the product page; FAQ also cites ~30 W average — keep both rows |
| Video path | Integrated HEVC decoder (keep frames off host DRAM for vision pipelines) |
| Card SKU (1 chip) | Axelera Edge 232p — HHHL PCIe; 32 GB or 64 GB LPDDR5; PCIe Gen 4 x4; active or passive cooling |
| Card SKU (4 chips) | Axelera Server 250p — FHFL PCIe; four Europas → 2500+ TOPS narrative |
| Bare die | Standalone Europa for customers designing their own boards |
| OEM validation called out | Dell XE5 and Supermicro 111AD with Edge 232p; broader Metis-era OEM list includes Advantech, Axiomtek, HPE, Lenovo, Seco, 2CRSi |
| Software | Voyager SDK (YAML pipelines) across Metis + Europa; Voyager Wingman; AxScript model compilation |
| Named language bench (vendor) | Edge 232p: 83.4 tokens/s on QWEN 30B (product page) |
| Efficiency marketing | Up to 6× tokens/s per watt vs “GPU-based solutions”; 3–5× perf/$ and 2–3× perf/W vs unnamed competitors on Llama 3 family charts (Axelera internal + published competitor NIM data) |
| Availability language | Launch PR: “Shipping now”; Edge 232p brief directs orders; main Europa page still invites early access — treat as qualified early ship, not a commodity CSP SKU |
Europa is not “another edge NPU sticker.” It is a claim that inference watts and host-CPU offload matter once you stop buying frontier GPUs for every agent and vision lane that can run on a 45 W card.
What the silicon is supposed to change
Metis was the sub-10 W / vision-first story. Europa doubles the AI core count to eight second-generation D-IMC cores and keeps the RISC-V vector ring for pre- and post-processing so the host CPU stays on application logic. The HEVC decoder is the vision footnote: multi-camera and robotics pipelines stop bouncing decoded frames through system memory just to feed the model.
Form factor is deliberate. Edge 232p drops one Europa into an existing HHHL slot (Dell XE5 / Supermicro 111AD called out). Server 250p lines four Europas on a full-length board for the “Europa 4Chip” column in Axelera’s comparison charts. Same Voyager compile path as Metis — YAML pipelines, not a forked runtime per generation.
That is a different CapEx sentence from Groq 3 LPX LPU interactivity, Jalapeño captive InferenceX, or d-Matrix Raptor decode silicon on NVLink Fusion. Europa is arguing for PCIe cards you add to servers you already own, aimed at agentic, VLM, and computer-vision workloads that do not need an HBM island.
The published efficiency claims (and what they are not)
Axelera’s Edge 232p page puts a named figure on language serving: 83.4 tokens/s for QWEN 30B. The broader charts claim 3–5× performance per dollar and 2–3× performance per watt against unnamed competitors on Llama 3 8B / 70B and Llama 3.2 11B Vision, with the four-chip Server 250p doing most of the winning. The launch PR rounds the family to up to 6× tokens per second per watt versus GPU-based solutions.
Three caveats belong on any capacity ticket before this becomes a fleet assumption:
- Vendor benches, mixed sources. Comparisons combine Axelera internal runs with published competitor NIM data. New-TechEurope flags incomplete test-condition disclosure — treat as marketing until you reproduce on your model, quant, and concurrency.
- 45 W TDP ≠ 30 W average. Product marketing leads with 45 W; the Edge FAQ also cites ~30 W average. Budget the TDP for power/cooling; use measured average for tokens-per-watt math.
- “Shipping now” vs early-access wording. Launch PR and StorageReview say shipping with Dell/Supermicro validation. The main Europa page still invites early access, and several ecosystem quotes read as eval / future integrate. Ask for lead time, firmware rev, and FITs before you treat Edge 232p as a drop-in NeoCloud SKU.
Where it sits next to Inferra, Leo X, and the GPU island notes
| Purchase | Product | What it buys |
|---|---|---|
| Prefetch + NVMe tiers | Lightbits Inferra | KV readiness across DRAM/NVMe so attention does not wait on recompute |
| Fabric DDR | Astera Leo X | KV / agent state as a scale-up fabric citizen |
| Package HBM | NVHBM | Memory controller / stack economics inside the accelerator envelope |
| Merchant GPU island | Vera Rubin NVL72 MLPerf / MI355X 512-GPU | What the HBM island does when you do buy the big box |
| Efficient PCIe AIPU | Europa | 629 TOPS / 45 W inference cards for agent / VLM / vision lanes that should not burn HBM GPUs |
Five purchase domains. Europa does not move KV blocks and does not invent a fabric. It argues that many inference lanes are still over-provisioned on GPUs, and that a European 5 nm AIPU with onboard RISC-V pre/post and a standard PCIe card is how you reclaim those watts.
Dell + E4 quotes on EU AI Factory work in Italy and Luxembourg are a sovereignty / on-prem signal, not a published multi-tenant SLA — same discipline as the OVH sentence on Inferra.
Caveats I would write into any Europa ticket
- Reproduce the QWEN 30B 83.4 tok/s and Llama charts on your Voyager build, batch, and concurrency before you price a SKU.
- PCIe Gen 4 x4 on Edge 232p is fine for many inference cards; wrong if host↔card bandwidth is the multi-stream vision bottleneck.
- LPDDR5 32/64 GB is the Edge memory envelope — do not confuse it with GPU HBM planning for 70B+ multi-user serving until you measure.
- Do not collapse with Metis 214 TOPS. Edge 130p and Edge 232p share Voyager; different TOPS / power / memory envelopes.
- Do not mix with killed Huawei Ascend / Atlas SuperPoD numbers. That lane stays closed on this desk.
- Availability asterisk. “Shipping” + Dell/Supermicro validation is real; early-access page language means you still qualify lead time.
How I would evaluate the claim if I ran clusters
- Tokens/s and tokens/s/W on models you sell (8B–70B + VLMs) at fixed batch/concurrency — Edge 232p vs Server 250p vs a same-rack GPU baseline.
- Sustained power under your SLA — TDP vs average, passive vs active cooling.
- Host CPU busy fraction with RISC-V pre/post and HEVC on vs off.
- Voyager compile coverage for your model zoo — model support is the silent kill for many AIPUs.
- Lead time and OEM SKU path through Dell XE5 / Supermicro 111AD before you assume drop-in PCIe.
Bottom line
Europa’s public launch is the strongest shipping inference ASIC story next to this week’s KV and optics notes: 629 TOPS at a 45 W TDP, 128 MB L2 / 200 GB/s DRAM, named Edge 232p / Server 250p cards, Dell XE5 + Supermicro 111AD validation, and a Voyager path shared with Metis. The honest read is vendor tok/s and tok/s/W charts plus early-ship OEM validation you must re-prove under your SLA. The dishonest read is “GPUs are obsolete” or “European silicon closed the efficiency gap forever.” Until your own rack reproduces the QWEN / Llama tables at sustained power, I treat Europa as a serious on-prem / AI-factory / NeoCloud inference-card POC — and as the watts foil every HBM-GPU ticket should be measured against when the workload is agent, VLM, or multi-camera vision rather than a frontier training island.
Create-only. Specs and dates from Axelera’s Sep 15 Europa launch, the Edge 232p / Europa product pages, StorageReview’s Sep 15 ship write-up, and New-TechEurope’s Sep 21 product brief. Update when a named customer publishes a reproduced QWEN / Llama ladder, or when Axelera posts same-methodology tok/s on a named third-party pod.
Sources: Axelera launch PR · StorageReview · New-TechEurope · Europa product page · Edge 232p.
Anpoo Sivanadi, Staff Software Engineer