New Hardware: A 46-Core RISC-V Server, a $59 AI Coding Kit, and a DGX Spark Rival

This week, RISC-V turned up at three very different scales: a 46-core server CPU, a $59 maker kit you can program by talking to an AI agent, and a $40 USB stick small enough to lose in a pocket. Two of them run RISC-V cores only. The third puts RISC-V cores next to Arm ones on Raspberry Pi silicon, so you can pick either at boot.

Here is what landed, what it costs, and what we still don't know.

SpacemiT launched the V100 at RISC-V Summit China on October 20. It is a 46-core server processor, and SpacemiT says it is the first chip fully compliant with the RISC-V server specifications. In practice that means UEFI, ACPI and SMBIOS firmware, the same plumbing that lets an x86 or Arm server boot a stock distro.

That is the part worth watching. Most RISC-V boards to date have needed vendor kernels and custom images. SpacemiT says mainstream distributions install and boot on the V100 unmodified, and that KVM works from the upstream kernel. Ubuntu 26.10, released October 15, is listed as officially supporting it.

The six XiangShan cores come from China's open-source XiangShan project, so part of this chip is built on openly published CPU designs. SpacemiT is also offering three reference servers (compute, accelerator and storage) plus motherboard designs for OEMs.

What we don't know yet: pricing, ship dates, and real performance. The "first fully compliant" claim is SpacemiT's own. Reported clock speeds also differ: Phoronix lists 2.2 to 2.4 GHz, while LinuxGizmos says the XiangShan cores top out at 2.1 GHz. SpacemiT first teased the V100 in January 2025, so this is a formal launch rather than a surprise.

Last week's customer call has the details your proposal needs. Right now they're buried in notes nobody can find.

Wispr Flow Notetaker captures your meetings with names, numbers, and product terms spelled right, then brings them into Claude or ChatGPT through a built-in connector. Ask Claude to draft the proposal from your last three calls with a prospect, or pull together what the deal team said about pricing this quarter.

Need something from a call in March? Ask one question across every meeting you've had and get the answer with citations. Get Wispr Flow Notetaker, free to try on Mac and Windows.

Get Wispr Flow Notetaker

Espressif's ESP-Mosaico is now sold directly for $59, after a stretch where it was only available through Taobao and AliExpress resellers at a markup. It is one of the first affordable ways to get hands-on with the ESP32-S31, Espressif's new dual-core RISC-V microcontroller.

The kit is modular. A core unit with a 2.16-inch 480 x 480 touchscreen snaps onto camera and sensor modules with magnetic headers. The pitch is "chat coding" and "vibe coding": you describe what you want, and Espressif's ESP-Claw and ESP-ZeroCode tools generate and flash the firmware.

The hardware alone makes it a solid ESP32-S31 dev board, with docs and enclosure STL files published. The bigger question is whether AI-driven firmware generation works well enough for beginners, or mostly produces demos.

What we don't know yet: the store mentions pre-orders, so it may not ship right away. It is unclear whether the BaseBoard with the 40-pin headers is in the $59 kit, and the camera sensor changes between PCB revisions. Nobody has published a hands-on test of the AI coding workflow yet.

Axera's AX8860P is a 6nm chip that pairs 16 Arm Cortex-A55 cores with an AI accelerator rated at up to 256 TOPS (FP4) and up to 128GB of LPDDR5x memory. Sipeed, the company behind the Maix and LicheePi boards, is building a small desktop box around it called the MaixStation. Radxa has also teased a "pocket supercomputer" on the same chip.

The target is Nvidia's DGX Spark: a compact box with lots of memory for running large language models locally. Sipeed says a 128GB MaixStation can run a 125-billion-parameter Qwen model, at a target price under $4,000. CNX Software puts the DGX Spark's current price at $6,949.99.

Axera says its software stack supports vLLM, SGLang, PyTorch, ONNX and OpenAI-compatible APIs, and its developer docs are already public. Sipeed also plans a quad-node MaixStation Pro with 512GB of memory, plus PCIe cards carrying up to four chips.

What we don't know yet: final prices and real-world speed. Sipeed's own benchmark shows the 128GB MaixStation only "slightly faster" than a DGX Spark, with multi-token prediction turned off. Axera's claim that its 256 TOPS beats Nvidia's 1,000 TOPS rests on Nvidia's figure being sparse FP4, which is vendor framing. Readers have also questioned whether the A55 cores and the 10GbE link will become bottlenecks, and Radxa's device has no specs yet.

All three products push the same idea: run standard software, and increasingly AI, on hardware outside the x86 mainstream. The V100 boots Ubuntu without a custom kernel, ESP-Mosaico writes its own firmware from a chat prompt, and the AX8860P targets local LLMs well below DGX Spark prices. None is shipping in volume yet, so treat the specs as announced, not tested.