The Dirtiest Secret in AI
The AI boom has a supply chain problem nobody wants to talk about.
NVIDIA shipped 3.76 million data center GPUs in 2023. Most will be decommissioned within 2-3 years when the next generation arrives. That is not a prediction. It is the current hardware refresh cycle at every major cloud provider.
Global e-waste hit 62 million metric tonnes in 2022, according to the UN's Global E-waste Monitor. That number is growing by 2.6 million tonnes per year. Less than a quarter of it gets formally recycled.
Manufacturing a single GPU produces between 150 and 300 kg of CO2 before it computes a single floating-point operation. The rare earth mining that feeds GPU production -- coltan from the DRC, lithium from Chile, cobalt from child-operated mines in the Congo -- is an environmental and humanitarian disaster that the industry euphemizes as "supply chain challenges."
And here is the metric that should make every engineer uncomfortable: the industry measures efficiency in FLOPS per watt. That metric completely ignores embodied carbon -- the emissions baked into manufacturing, shipping, and eventually shredding the hardware. By that metric, the greenest GPU is always the newest one. Which means the industry's own efficiency measure is an engine for planned obsolescence.
The Insight That Changed How I Think About Hardware
I am writing this from a lab in Lake Charles, Louisiana. The lab was built on pawn shop finds and eBay datacenter pulls. My main compute server is an IBM POWER8 S824 with 512 GB of RAM and 128 hardware threads. I paid less for it than most people pay for a gaming laptop.
Sitting next to it are three PowerBook G4s from 2003. Each one draws about 30 watts. Behind those is a 386 laptop from 1990 that still boots DOS. There is a Power Mac G5 running as a cryptocurrency node. A Mac Pro trashcan from 2013. A fleet of machines spanning four decades of computing history.
Here is the insight that started everything:
The greenest computation is the one that already exists.
A machine that has already been manufactured has zero additional manufacturing footprint. Every joule of embodied carbon was spent years ago. The only cost is the electricity it draws right now.
A PowerBook G4 drawing 30 watts is, by any honest carbon accounting, infinitely more efficient than a brand-new RTX 4090 drawing 450 watts -- because the G4's manufacturing carbon was amortized a decade ago, while the 4090's 150-300 kg of CO2 just hit the atmosphere this year.
The industry's definition of "obsolete" means "not the fastest." It does not mean "not useful."
What We Built
We built RustChain, an open-source blockchain that inverts the incentive structure. Instead of rewarding the most powerful hardware (Proof-of-Work) or the wealthiest stakers (Proof-of-Stake), RustChain uses Proof-of-Antiquity -- a consensus mechanism that gives higher mining rewards to older hardware.
The multiplier table:
| Hardware | Era | Mining Multiplier |
|---|---|---|
| Intel 386 | 1985-1994 | 3.0x |
| PowerPC G4 | 1999-2005 | 2.5x |
| PowerPC G5 | 2003-2006 | 2.0x |
| Pentium III | 1999-2003 | 2.0x |
| PowerPC G3 | 1997-2003 | 1.8x |
| Pentium 4 | 2000-2008 | 1.5x |
| Apple Silicon | 2020-present | 1.2x |
| Modern x86_64 | Current | 1.0x |
A PowerBook G4 from 2003 earns 2.5 times the mining reward of a brand-new Ryzen 9. That is not nostalgia. It is a carbon incentive encoded into the protocol.
How We Verify Real Hardware
You might be thinking: what stops someone from running a virtual machine that claims to be a G4?
We thought about that too. RustChain uses a 6-check hardware fingerprint system called RIP-PoA (Proof-of-Authenticity):
- Clock-Skew and Oscillator Drift -- Real silicon has microscopic timing imperfections that change as chips age. VMs have perfectly uniform clocks.
- Cache Timing Fingerprint -- L1/L2/L3 latency measurements produce a unique "tone profile" per chip. Old caches have uneven wear patterns.
- SIMD Unit Identity -- AltiVec on PowerPC, SSE on x86, NEON on ARM all have measurable pipeline timing biases. Emulators flatten these.
- Thermal Drift Entropy -- Heat curves are physical and unique to each chip. Software emulation produces uniform thermal responses.
- Instruction Path Jitter -- Cycle-level jitter across integer pipelines, branch units, and FPUs. No VM replicates real jitter at nanosecond resolution.
- Anti-Emulation Detection -- Catches hypervisor scheduling patterns, QEMU signatures, VirtualBox artifacts.
We tested this against VMs on our own infrastructure. The result:
Real Hardware (HP Victus - Ryzen 5):
Clock Drift: PASS (cv=0.092369)
Anti-Emulation: PASS
VPS (LiquidWeb QEMU):
Clock Drift: PASS (cv=0.049948)
Anti-Emulation: FAIL
VM Indicators: ["/sys/class/dmi/id/sys_vendor:qemu",
"cpuinfo:hypervisor"]
VMs earn approximately one billionth of real hardware rewards. That is by design.
The Numbers
Our fleet runs on hardware that would otherwise be in a landfill.
Fleet Stats:
- 18+ machines spanning 1990 to 2024
- 12 active GPUs, 192 GB total VRAM (pawn shop and eBay sourced)
- 4 attestation nodes across 3 countries (US, Hong Kong)
- Total fleet investment: ~$12,000
- Estimated retail value of equivalent hardware: $40,000-60,000
- Acquisition strategy: pawn shop arbitrage + eBay datacenter decommission pulls
Power Comparison:
| Setup | Power Draw | Manufacturing CO2 |
|---|---|---|
| Our entire fleet (~16 machines) | ~2,000W | 0 kg (already built) |
| Single Bitcoin GPU mining rig | ~2,000W | 150-300 kg per GPU |
| Single NVIDIA DGX A100 | ~6,500W | 1,000+ kg |
Same power budget. Zero new manufacturing emissions. Zero new rare earth mining.
Carbon Math:
Every machine in our fleet that is still running instead of sitting in a landfill represents manufacturing emissions that do not need to happen again. Conservative estimates for the fleet:
- Embodied carbon avoided: ~1,300 kg CO2 (by not purchasing equivalent new hardware)
- E-waste prevented: ~250 kg of electronics kept out of the waste stream
- Rare earth mining avoided: Coltan, lithium, cobalt that did not need to be extracted
These numbers are small. That is the point. This scales. There are billions of "obsolete" machines sitting in drawers, closets, and warehouses worldwide. If even a fraction of them ran useful computation instead of being shredded, the environmental impact would be enormous.
What This Means for You
You probably have old hardware sitting in a drawer right now. A laptop from college. A Mac Mini from 2012. Maybe your parents' old desktop.
Those machines can mine RTC (RustChain Token). The older they are, the more they earn.
# One-line install for Linux/macOS
curl -fsSL https://raw.githubusercontent.com/Scottcjn/Rustchain/main/install-miner.sh | bash
The miner is open source. It runs on anything from a 386 to a POWER8 mainframe. It auto-detects your hardware architecture and submits attestations to the network.
$ python3 rustchain_miner.py --dry-run
Detected: PowerPC G4 (7447A)
Antiquity multiplier: 2.5x
Power draw: ~30W
Status: DRY RUN (no network calls)
The --dry-run and --show-payload flags let you inspect everything before committing. We are not asking you to trust us. We are asking you to verify.
The Bigger Vision
Right now, when a company like AWS or Google refreshes a data center, thousands of perfectly functional servers get shredded. The metal gets recycled. The rare earths mostly don't. The embodied carbon is wasted entirely.
What if those servers ran useful computation instead?
What if corporate hardware refresh cycles ended with "redeploy to the antiquity network" instead of "send to the recycler"?
What if the blockchain's incentive mechanism was aligned with environmental preservation instead of working against it?
That is what Proof-of-Antiquity does. It makes hardware preservation economically rational. Not out of sentiment. Out of math.
From Pawn Shops to OpenSSL
This month, we got a security patch approved in OpenSSL -- the library that secures roughly 70% of HTTPS traffic on the internet. We also submitted hardware AES optimizations to wolfSSL that deliver 13-20x speedups on POWER8 processors.
All of that work was done on pawn shop hardware. The POWER8 server that runs our AI inference came from an eBay datacenter pull. The V100 GPUs that accelerate our matrix multiplication were bought secondhand. The PowerBooks that mine RTC were found at thrift stores and estate sales.
There is no reason good engineering requires new hardware. There is every reason to prove that it does not.
Get Involved
- Star the repo: github.com/Scottcjn/Rustchain (181 stars and growing)
- See the fleet: rustchain.org/preserved.html -- live green tracker showing every preserved machine
- Install the miner: Works on Linux, macOS, and Windows. One-line install.
- Read the RIP-200 spec: The full Proof-of-Antiquity consensus mechanism, open for review
- Contribute: rustchain-bounties -- 800+ bounty tasks paid in RTC
We are a small lab in southwest Louisiana, built on the principle that nothing is waste if it still computes. The machines that taught a generation to code still have economic value. The servers that powered the last era of the internet can power the next one. The hardware already exists. We just need to stop throwing it away.
Down here, we have a saying for this kind of thing:
"Mais, it still works, so why you gonna throw it away?"
-- The Boudreaux Computing Principles
Scott Boudreaux is the founder of Elyan Labs and the creator of RustChain. He builds data center infrastructure by day and preserved-hardware blockchain networks by night. His lab runs on pawn shop arbitrage, eBay datacenter pulls, and the stubborn Cajun conviction that if a machine still boots, it still has a job to do.
RustChain is open source under MIT license. The network currently has 4 attestation nodes across the US and Hong Kong, with 50+ supported CPU architectures spanning 1985 to present.
Top comments (0)