Developers Ran AI Models on Retired Crypto Hardware
Tech enthusiasts repurposed discarded FPGA mining cards to execute Qwen3.5 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?
Developers recently demonstrated the use of discontinued FPGA hardware to run Qwen3.5 language models. By leveraging secondary market boards, the projects achieved inference speeds as high as 8.18 tokens per second on September 9, 2026.
Why it matters
Supply scarcity and the high cost of modern GPUs have forced developers to seek alternative, affordable silicon for AI research. The collapse of the crypto mining market provided an unexpected inventory of low-cost hardware for these inference experiments.
The SQRL FK33 FPGA card features 8 gigabytes of HBM2 memory with 400GB/s bandwidth. The llm.vhdl project operates at a 75MHz clock speed to manage the 9B model using INT4 quantization.
The players
Micron
This global semiconductor company reported record fiscal fourth-quarter revenue of $54.2 billion and monitors memory supply constraints.
eBay
This online marketplace serves as the secondary hub where developers source inexpensive, retired cryptocurrency mining FPGA cards.
The details
Engineers developed custom VHDL inference engines and instruction sets to interface with legacy FPGA fabric. By utilizing INT4 quantization, the teams compressed Qwen3.5 9B and 27B models to fit within the limited 8 gigabytes of onboard memory available on boards like the SQRL FK33.
Timeline
September 9, 2026: The fable5_llm project reached 7.29 tokens per second.
September 30, 2026: Micron reported record fiscal fourth-quarter revenue.
The Tech Race
These experiments mark a departure from the industry standard of relying exclusively on high-end GPUs for AI workloads. This shift demonstrates how legacy systems can potentially bridge the gap as the sector waits for memory supply to catch up with global AI demand by 2028.
This development highlights that high-performance AI inference may become accessible to researchers without needing expensive flagship hardware. Enthusiasts can now find capable boards on the secondary market for approximately $280 to $350.
The takeaway
Using discarded FPGA boards offers a cost-effective path for those looking to experiment with language models without securing modern GPU hardware. Readers should prioritize cards with sufficient HBM2 memory capacity if they intend to replicate these inference setups.
Further reading
For broader trends in chip development, see our coverage on Semiconductors.
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?
