Developers Ran AI Models on Discarded Crypto Hardware
Hobbyists are repurposing legacy FPGA mining boards to run large language models at low costs.
Updated on Oct. 5, 2026 in Semiconductors
Live Poll
Is now a good time to repurpose older, cheaper hardware for local AI projects?
As of September 9, 2026, researchers have successfully deployed Qwen3.5 language models on discarded FPGA hardware previously used for cryptocurrency mining. These projects demonstrate that specialized, low-cost silicon can perform functional AI inference.
Why it matters
The scarcity and high cost of modern data-center GPUs have incentivized engineers to find ways to run AI on cheaper, secondary-market hardware. This shift reflects broader efforts to democratize AI compute access by utilizing existing, underused components.
The llm.vhdl project runs models at a 75MHz clock speed on an SQRL FK33 card, which costs $280 to $350 on eBay. With 8 gigabytes of HBM2 memory, the system uses INT4 quantization to fit models within the card's memory constraints.
The players
Micron
A major global semiconductor manufacturer that produces memory and storage solutions and tracks long-term AI hardware supply trends.
The details
Developers achieved this by writing custom VHDL—a hardware description language used to model electronic systems—to create dedicated inference engines and instruction sets on the FPGA fabric. To fit the Qwen3.5 9B and 27B models into the limited 8 gigabytes of onboard memory, they applied INT4 quantization, a compression technique that reduces the precision of a model's weights to 4-bit integers. This allows the logic gates on the FPGA to handle model computations without requiring external, high-performance VRAM.
Timeline
September 9, 2026: The fable5_llm project reached a peak speed of 7.29 tokens per second on an SQRL BCU-1525 card.
September 30, 2026: Micron reported $54.2 billion in record fiscal fourth-quarter revenue.
The Tech Race
The transition to secondary-market hardware highlights the ongoing competition for specialized silicon needed to scale AI inference. This development follows a pattern set by the 2028 memory supply demand projection from Micron, illustrating the pressure of supply constraints on current AI development.
For hobbyists and researchers, these projects provide a path to run locally hosted models using hardware that is widely available on secondary markets for under $350. Users must possess the technical skills to implement custom VHDL code, as these solutions lack the plug-and-play software stacks of modern commercial GPUs.
The takeaway
This trend highlights the capability of legacy hardware to bridge the gap in AI compute availability. Watch for further performance benchmarks as developers optimize quantization methods to squeeze larger models onto limited FPGA memory.
Further reading
Learn more about hardware optimization and chip design trends in the Semiconductors section.
Source note: This article includes information reported by Startup Fortune.
Live Poll
Is now a good time to repurpose older, cheaper hardware for local AI projects?
