backend/app/services/gpu/detection.py and the pricing engine at backend/app/services/pricing_engine.py. Real marketplace data will be published once enough fleets are online to make the numbers honest.
Hardware tips: what earns most
The hero-user view of hardware selection — what actually earns once you've moved past Stage 4's "should I buy hardware?" framing and committed to operating a small fleet for the long haul. This is a pattern guide, not a buyer's guide.
The honest framing
Earnings on Zyra are a function of three things, in order of importance:
- Uptime. A 70th-percentile CPU that runs 24/7 beats a 99th-percentile CPU that crashes every 6 hours.
- Capability score. Calculated server-side from the heartbeat — CPU cores, RAM, GPU model (normalized via
normalize_gpu_modelingpu/detection.py), bandwidth. Higher score → access to higher-paying job tiers. - Power efficiency. Earnings minus electricity. Stage 4 chapter 4 covers the math; this chapter focuses on dollars-per-watt at idle and under load.
[VERIFY: published $/capability_score curve — marketplace data will be released as a "fleet performance report" once we have 90 days of post-launch operation.]
CPU sweet spots (as of mid-2026)
Three honest categories rather than a leaderboard:
The boring-but-reliable tier
Ryzen 7 7700 / 7700X (8 cores, 65-105W). Apple M2/M3 base (via the macOS agent). Intel Core i5-13500. These run cool, draw modest power, rarely throttle. They earn the bulk of general-purpose jobs and they keep doing it for years.
The aggressive earner tier
Ryzen 9 7950X / 9950X (16 cores). Threadripper 7960X (24 cores). High capability score, high power draw. The 9950X has known thermal-throttling reports with stock cooling [VERIFY: marketplace failure-rate report once we have N=100+ devices] — a stable 7950X that holds boost clocks earns more over a year than a 9950X that throttles in the 5th minute of every job.
The Apple Silicon path
M2 Pro / M3 Pro / M3 Max — exceptional perf/watt at idle and under sustained load. Earnings cap is lower than top-end x86 GPU rigs, but electricity cost is dramatically lower. For owners on residential rates above $0.20/kWh, Apple Silicon often wins on net earnings even if gross earnings are lower.
GPU tiering (from gpu/detection.py)
The backend classifies GPUs into three tiers via classify_gpu_tier:
- Datacenter — A100, H100, L40S. Almost never in individual fleets.
- Prosumer — RTX 4090, RTX 4080 Super, A4000/A5000, Mac Studio M2 Ultra GPU partition. The earning sweet spot for serious individual operators.
- Consumer — RTX 4070 / 4060 Ti, RX 7800 XT, integrated GPUs. Earn on lighter jobs, rarely qualify for the highest-paying tiers.
If you're buying with earnings in mind, prosumer is the bracket where the curve bends. A used RTX 4090 typically earns more than a new RTX 4060 Ti for less than 2x the cost, with similar power draw.
[VERIFY: tier-by-tier $/hour averages at GA — the platform pays $0.08/device-hour to the Pro-tier customer (verified in pricing_engine.py); device-owner-side payout splits will be published with the launch fleet report.]
Used hardware: where the value is
The big secret of fleet operators is that almost nothing in the fleet is new. The honest sourcing ladder:
- eBay — reliable for consumer/prosumer GPUs (filter "tested working", check seller feedback, avoid "for parts"). Avoid sealed-box "new" listings from no-feedback sellers; frequently rebadged.
- Refurb dealers — Micro Center, Newegg, Server Monkey for rack gear. Usually 90-day warranty. Pay 10-15% more than eBay for the warranty; it pays off when the GPU dies in month 2.
- Local listings (Craigslist, Facebook Marketplace) — best prices, highest scam rate. Test in person. Bring a USB stick with memtest and a stress-test image.
- Crypto-mining liquidation lots — fine for GPUs (mining is sustained at low temps and rarely damages silicon) but the cards have been on 24/7 for years. Replace thermal pads and fans before deployment.
Scams to avoid
- Locked BIOS / vBIOS on ex-corporate workstations. Confirm BIOS unlocks before buying.
- Stripped servers sold as "complete" — populated chassis but only one CPU and one RAM stick. Read the parts list, not the photos.
- Rebadged GPUs — older chips reflashed to a newer name. Verify the reported chip ID with
nvidia-smipost-arrival. - "Mining-only" cards (CMP series) with no display output — do not work as compute devices on Zyra
[VERIFY: agent rejects display-less cards at registration]. - Pulled laptop GPUs soldered onto desktop carrier boards. Unpredictable performance, worse thermals.
Reliability beats raw speed
Two hypothetical devices over 30 days at 80% uptime target:
- Device A: Ryzen 7950X stable. Capability score 720. Holds 100% earning hours. Earns the full month.
- Device B: Ryzen 9950X with marginal cooling. Capability score 760 nominal, throttles to 7950X performance under sustained load, crashes once every ~18 hours. Loses ~10 earning hours per day to restarts and cooldowns.
Device A wins. The capability-score advantage of Device B is wiped out by lost uptime and the platform's reliability scoring [VERIFY: device_reputation.py — devices with frequent disconnects see reduced job dispatch].
The lesson: buy one tier down from the bleeding edge, pair it with overkill cooling, and aim for boring stability.
Storage and networking
- NVMe over SATA. Many job types are I/O-bound. A $40 NVMe upgrade can move a device up a capability tier
[VERIFY: storage-IO factor in capability_score formula]. - Wired Ethernet. Wi-Fi adds latency and jitter; the heartbeat tolerates it but the highest-paying jobs prefer wired. Worth a $10 powerline adapter if you can't run a cable.
- Symmetric upstream. 1 Gbps down / 50 Mbps up caps you on result upload. Fiber owners earn more on data-heavy jobs.
What does not matter
- RGB and case aesthetics. Earns $0.
- Top-tier motherboards (X670E Extreme, Z790 Apex). A $180 board does the job at stock settings.
- DDR5-7200 vs DDR5-5600. Below 1% earnings difference on most workloads.
- Liquid AIO on a 65W CPU. A $35 tower cooler matches an AIO and lasts longer.
Cross-links
- Stage 4 chapter 2: Advanced earnings strategies — how capability tiers map to job dispatch
- Stage 4 chapter 3: Should I buy hardware to earn more? — the ROI math you should do first
- Stage 4 chapter 4: Power, cooling, and noise — kWh math and thermals
What's next
2. Referrals at scale: ambassador and community →
Last reviewed: 2026-05-21