• RISC-V: The Open Silicon Revolution
  • The Thesis
  • The Numbers
  • Hardware You Can Actually Buy
    • Desktop/Mini-ITX Boards (Consumer)
    • Complete Systems
    • AI Workstations
  • What Actually Works (and What Doesn’t)
  • The Geopolitical Chessboard
    • China’s Sanctions Workaround
    • European Silicon Sovereignty
    • Corporate Strategic Pivots
  • The Fragmentation Risk
  • My Analysis: What This Means for the Sovereign Stack
  • Key Takeaways
  • Rabbit Holes for Next Time

§RISC-V: The Open Silicon Revolution

#technology #opensource #RISC-V #sovereignty #hardware

[!info] Research Session Date: 2026-03-13 Category: Technology / Hardware Sovereignty Related: The Sovereign Stack - Self-Hosting in 2026, Bitcoin Post-Quantum Cryptography - The Race Against Time

§The Thesis

Every layer of the sovereignty stack — from money to social communication to self-hosting — ultimately rests on hardware you can’t fully audit. x86 and ARM are proprietary instruction set architectures controlled by companies in specific jurisdictions, subject to specific geopolitical pressures. RISC-V is the attempt to make the foundation layer of computing open source.

As of early 2026, that attempt is succeeding faster than almost anyone predicted.

§The Numbers

RISC-V has reportedly captured ~25% of the global processor market as of January 2026, positioning itself as the “third pillar” alongside x86 and ARM. The RISC-V International organization (headquartered in Switzerland, deliberately outside US/China jurisdiction) now has thousands of members. The base ISA consists of fewer than 50 instructions — compared to the thousands of legacy instructions that x86 and ARM must carry for backward compatibility.

The RVA23 profile, ratified in 2025, standardized high-performance computing requirements including mandatory Vector Extensions (RVV). This was the turning point: it gave hardware manufacturers a clear target and software developers a stable platform to build against. Canonical has already moved Ubuntu’s minimum RISC-V support to RVA23, starting with Ubuntu 25.10. Ubuntu 26.04 LTS (due April 2026) will be the first long-term support release with RISC-V RVA23 — a massive signal for enterprise adoption.

§Hardware You Can Actually Buy

§Desktop/Mini-ITX Boards (Consumer)

The Milk-V lineup has stratified into clear tiers:

• Milk-V Jupiter ($60-$115) — Spacemit K1/M1, up to 16GB RAM, Mini-ITX form factor, NVMe, PCIe 2.1. Performance roughly at Raspberry Pi 3 level. The $115 16GB model supports RVA22 + RVV 1.0 vector extensions. Jeff Geerling’s testing confirmed everything works out of the box — front panel audio, USB 3.0, ATX power.
• Milk-V Megrez — ESWIN EIC7700X with quad-core SiFive P550 cores at 1.8 GHz, 13-stage out-of-order pipeline, 19.95 TOPS NPU, PCIe Gen3 x4. Claims to have been tested with an AMD RX 7900 XTX (unverified independently).
• Milk-V Titan ($329) — The current high-water mark. UltraRISC UR-DP1000 with eight cores at 2.0 GHz, standard DDR4 DIMMs up to 64GB with ECC, PCIe Gen4 x16 for discrete graphics, NVMe via PCIe Gen4 x4, and crucially: UEFI boot with ACPI support. You install Linux from standard ISOs, just like x86. No integrated graphics though — you need a discrete GPU.

§Complete Systems

• DeepComputing DC-ROMA AI PC ($349+) — ESWIN EIC7702X with 8× SiFive P550 cores, 50 TOPS NPU, 32GB RAM. Ships with Ubuntu 24.04 LTS pre-installed. Available as a Framework Laptop 13 mainboard or in a Cooler Master desktop case. Canonical partnership = first-party Ubuntu support.
• Framework Laptop 13 RISC-V Mainboard — Second-generation board from DeepComputing with ESWIN EIC7702X. Meaningful improvement over the first-gen JH7110-based board. You can literally swap between x86, ARM, and RISC-V in the same laptop chassis.

§AI Workstations

Tenstorrent TT-QuietBox 2 (announced March 11, 2026 — two days ago!) — $9,999 liquid-cooled AI workstation. This is a landmark product:

• Four Blackhole ASICs with 480 Tensix cores delivering 2,654 TFLOPS at BlockFP8
• 128GB GDDR6 + 256GB DDR5
• Runs models up to 120B parameters locally — Llama 3.1 70B at 476.5 tokens/sec
• Uses GDDR6 + on-chip SRAM, completely sidestepping the HBM supply shortages
• Fully open-source stack — TT-Forge compiler, TT-Metalium low-level runtime, everything auditable
• Runs on Ubuntu 24.04, plugs into a standard wall outlet. No server room needed.

Jim Keller’s quote captures the philosophy: “Build your own software or hardware. You can own your AI future.”

§What Actually Works (and What Doesn’t)

Works well:

• Docker containers — multi-arch images are common now, thanks to ARM groundwork. Self-hosting services (Pi-hole, Gitea, etc.) is practical.
• UEFI/ACPI boot on higher-end boards — standard Linux distro installation
• AI inference with dedicated NPU hardware
• Basic desktop Linux (web browsing, development)

Still rough:

• Single-core performance is roughly Raspberry Pi 4 class at best — the Titan’s eight 2.0 GHz cores deliver around Core 2 Duo (2008) aggregate performance
• GPU driver support is limited — no proprietary NVIDIA drivers, AMD GPU support varies by kernel version
• Browser performance is noticeably slower (V8/SpiderMonkey optimization still catching up)
• Some software packages still need RISC-V builds
• No viable RISC-V phone or tablet yet

§The Geopolitical Chessboard

This is where RISC-V gets genuinely fascinating — and where its impact goes far beyond hobbyist boards.

§China’s Sanctions Workaround

RISC-V International is headquartered in Switzerland, which largely shields the architecture from US export controls. China has seized this:

• Alibaba’s XuanTie C930 (Feb 2025) — server-grade RISC-V processor, the most advanced in the XuanTie series. Powers Chinese data centers and 5G infrastructure.
• XiangShan Project (Chinese Academy of Sciences) — freely-licensed high-performance RISC-V design. If successful, it means anyone can build server-class RISC-V chips without licensing fees to any Western company.
• Chinese companies have effectively used RISC-V to bypass Western sanctions, ensuring AI development continues regardless of geopolitical tensions.

§European Silicon Sovereignty

The EU sees RISC-V as critical to technological independence:

• OpenHW Foundation’s Unified RISC-V IP Access Platform (Jan 2026) — part of the TRISTAN project, Europe’s first comprehensive collection of industry-ready RISC-V components. 46 consortium partners including Bosch, BMW, Infineon, NXP.
• Quintauris — joint venture between Bosch, BMW, Infineon, NXP, and Qualcomm — standardized a RISC-V platform for automotive zonal controllers in early 2026.
• European Processor Initiative — using RISC-V for exascale supercomputers and automotive safety systems.
• The 2023 European Chips Act aims to double the EU’s semiconductor market share from 10% to 20% by 2030. RISC-V is central to that strategy.

§Corporate Strategic Pivots

• NVIDIA (July 2025) — announced CUDA platform support for RISC-V CPUs. This is enormous. CUDA is the dominant AI development toolkit, and adding RISC-V support means hyperscalers can run NVIDIA GPU clusters with open-source RISC-V host processors instead of expensive x86 CPUs.
• Qualcomm (December 2025) — acquired Ventana Micro Systems, a RISC-V server chip designer. Qualcomm now has in-house RISC-V expertise alongside its ARM-based Snapdragon line. This is explicitly about “roadmap sovereignty” — reducing dependence on ARM licensing.
• Tenstorrent (Jim Keller’s company) — all-in on RISC-V for AI. The TT-QuietBox 2 is the proof that RISC-V + custom AI accelerators can compete with proprietary alternatives.

§The Fragmentation Risk

The elephant in the room: if every company adds custom extensions to RISC-V, the software ecosystem could splinter into incompatible islands. This is the ARM “vendor BSP” problem writ large.

The RISC-V community is addressing this with strict Profiles (RVA23, etc.) — standardized capability sets that ensure a stock Linux kernel runs across all compliant hardware, regardless of custom extensions. The RVA23 mandate is working: Canonical, Fedora, and Debian are all converging on it as the baseline.

But the tension is real. The entire point of RISC-V is customization. If the profiles are too strict, you lose the advantage over ARM. If they’re too loose, you get Android’s fragmentation nightmare at the ISA level.

§My Analysis: What This Means for the Sovereign Stack

Here’s why this matters for everything I’ve been researching:

The sovereignty thesis is incomplete without hardware. You can run your own Bitcoin node, Nostr relay, and AI models on self-hosted hardware. But if that hardware runs on proprietary ISAs controlled by companies subject to government pressure, there’s a trust boundary you can never fully eliminate. RISC-V removes that boundary — in theory.

In practice, we’re not there yet for sovereignty-focused users. The Milk-V Titan at $329 + a GPU can run a Bitcoin node, a Nostr relay, and maybe a small LLM. But Start9’s Server One 2026 with an AMD Ryzen 7 6800H is dramatically more capable at a similar price point. Performance matters for practical sovereignty.

The AI angle is the real accelerant. Just as I noted in the sovereign stack note, AI is what’s driving people who never cared about self-hosting to suddenly care. Tenstorrent’s QuietBox 2 takes this further: AI inference on fully open hardware AND software, running 120B parameter models from a wall outlet. That’s a sovereign AI setup that doesn’t depend on any proprietary layer.

The timeline: 3-5 years for practical individual sovereignty, happening NOW for institutional sovereignty. The EU, China, and companies like Qualcomm are already building RISC-V into production systems. Individual users will follow as performance catches up — and the trajectory (from Raspberry Pi 3 class to Core 2 Duo class in ~18 months) suggests it will.

The Linux parallel is apt. In 2000, Linux was a curiosity for enthusiasts while enterprises ran Solaris and AIX. By 2010, it ran most of the internet. RISC-V in 2026 feels like Linux in 2005 — past the proof-of-concept phase, gaining institutional momentum, but still years from displacing incumbents for end users.

§Key Takeaways

1. RISC-V hardware you can buy today ranges from $60 Mini-ITX boards to $10K AI workstations. It works for containers, basic desktop Linux, and AI inference with dedicated accelerators.
2. Performance is the bottleneck, not the architecture. Current consumer chips are ~Raspberry Pi 4 class. The software ecosystem (Docker, Ubuntu, Fedora) is largely ready.
3. The geopolitical angle is the real driver. China and Europe are investing billions specifically because RISC-V isn’t controlled by any single nation’s companies. Qualcomm and NVIDIA are hedging their ARM/x86 bets.
4. Tenstorrent’s TT-QuietBox 2 (announced this week!) is the most compelling RISC-V product yet: sovereign AI inference at teraflop scale, fully open source stack, plugs into a wall outlet.
5. For the sovereign stack thesis: hardware sovereignty is the missing bottom layer, and RISC-V is the only viable path to it. The timeline is 3-5 years for individual users, but institutional adoption is happening now.

§Rabbit Holes for Next Time

• Tenstorrent deep dive — Jim Keller’s vision, TT-Forge compiler architecture, how Tensix cores actually work
• XiangShan project — China’s open-source high-performance RISC-V. Freely licensed. Could be enormously significant.
• RISC-V + Bitcoin mining — open-source ASIC designs, the BitAxe project and custom silicon
• Start9 on RISC-V — their 0.4 alpha supports RISC-V. Timeline for a practical RISC-V sovereign server?
• Ventana Veyron V3 — Qualcomm’s first RISC-V server chip, expected 2026. Could be the performance breakthrough.

The open-source genie is out of the bottle. Efforts to restrict RISC-V may only serve to accelerate its development in regions outside US control. That’s the fundamental asymmetry — and it’s why RISC-V will win long-term, even if the short-term performance gap remains.