io.net Spot Instance Reliability: Availability Data 2026

io.net Spot Instance Reliability: Availability Data 2026

Photo by Kevin Ku on Unsplash

Quick Answer: io.net reliability in 2026 is tier-dependent: verified data-center-class nodes deliver roughly 97-99% uptime, while residential consumer-GPU nodes sit closer to 90-95% with interruption rates of 2-8% per 24-hour job. That's worse than AWS spot's managed two-minute-warning model on paper, but io.net's prices (RTX 4090 from $0.19/hr, H100 from ~$1.45/hr) mean checkpointed, interruption-tolerant workloads still come out 40-60% cheaper. Treat every io.net spot node as preemptible, checkpoint every 10-15 minutes, and it's a genuine bargain; rely on a single node for a 72-hour uncheckpointed run and you will eventually get burned.

On This Page

• How io.net Sources Its GPUs
• Spot vs Reserved on io.net
• Realistic Uptime by GPU Class
• Interruption Reality: What Actually Kills Jobs
• Checkpointing Strategies That Make Spot Safe
• Cluster Provisioning Times
• The $IO Payment Economics
• io.net vs AWS Spot vs RunPod Community
• Frequently Asked Questions

How io.net Sources Its GPUs

io.net doesn't own hardware. It aggregates GPUs from three supplier categories and coordinates them with token incentives:

1. Independent data centers — smaller DCs and former mining operations with racked A100s, H100s, and 4090 farms. The most reliable tier.
2. Render/professional farms — studios monetizing idle capacity between jobs.
3. Individual suppliers — enthusiasts and ex-miners with 1-8 consumer GPUs on residential or small-business connections. The largest pool by count, the weakest by uptime.

Suppliers earn block rewards in $IO for keeping hardware online and passing periodic proof-of-compute benchmarks, plus the actual rental fees when their nodes are hired. The block-reward layer matters for reliability: it pays suppliers to stay online even when idle, which keeps the available pool deep. But it also attracts marginal operators chasing emissions with flaky power, consumer PSUs, and no UPS — which is exactly why the network's verification tiers exist.

Every node gets benchmarked on join and re-verified periodically: compute throughput, bandwidth, and uptime history feed a supplier score. Filter by that score. The difference between "any 4090" and "verified 4090, data-center tier, 99% uptime history" is the difference between a lottery and a tool.

Spot vs Reserved on io.net

io.net offers two consumption models in 2026:

	Spot / On-Demand	Reserved Clusters
Pricing	Market rate, cheapest	10-30% premium over spot, discounts for 1-4 week terms
Preemption	Node owner or scheduler can interrupt	Contractual hold on specific verified nodes
Best for	Batch jobs, sweeps, dev boxes	Multi-day training, demos, steady inference
Availability	Deep pool, variable quality	Thinner pool, verified tiers only

Reserved doesn't mean "AWS-grade SLA" — it means the network commits specific verified nodes to you and penalizes suppliers (slashing rewards) for dropping a reservation. In practice reserved uptime on data-center-tier nodes approaches centralized-cloud territory.

Realistic Uptime by GPU Class

io.net publishes aggregate network stats, but the numbers that matter are per-tier. The table below is an illustrative synthesis of mid-2026 community-reported and dashboard-observed figures — treat them as planning estimates, not guarantees:

GPU class / tier	Typical uptime	Interruptions per 24h job	Notes
H100 80GB (DC tier)	98.5-99.5%	<0.5% of jobs	Nearly all H100 supply is data-center hosted
A100 80GB (DC tier)	98-99.5%	~1%	Ex-mining DCs and small clouds
RTX 4090 (verified DC/farm)	97-99%	1-3%	The sweet spot for price vs stability
RTX 4090/3090 (residential)	90-95%	4-8%	Power cuts, ISP drops, owner reboots
RTX 3090 (unverified)	85-93%	8%+	Cheapest listings; expect churn

Two takeaways: enterprise GPUs are reliable because their owners are data centers, and consumer-GPU reliability is almost entirely a function of the supplier tier you filter for, not the silicon.

"The uptime story on decentralized compute bifurcated in 2025: data-center-tier nodes became boring and dependable, while the residential long tail stayed a lottery." — The Register, Q1 2026

Interruption Reality: What Actually Kills Jobs

Interruptions on io.net don't look like AWS's polite two-minute warning. Common failure modes:

• Hard node drop — supplier loses power/internet; your container just stops. No warning.
• Supplier withdrawal — owner reclaims the GPU (rare on verified tiers, penalized on reservations).
• Failed re-verification — a node that starts failing benchmarks gets pulled from the pool mid-lease.
• Network partition — node is alive but unreachable; the scheduler eventually marks it dead.

Budget for the P95 case, not the average. If your job can't survive an unannounced kill at any moment, it doesn't belong on io.net spot — full stop.

Photo by Teemu on Unsplash

Checkpointing Strategies That Make Spot Safe

The entire economics of DePIN spot hinge on cheap recovery. The playbook:

1. Checkpoint to external storage every 10-15 minutes — S3-compatible object storage (Cloudflare R2 has no egress fees, which matters when nodes churn). Never checkpoint to node-local disk only.
2. Make jobs resumable by design — --resume_from_checkpoint latest should be the default entry point of your training script, not a manual flag.
3. Shard batch work into small units — for inference/embedding jobs, process in 5-10 minute shards with a work queue (Redis/SQS pattern). A dead node = one re-queued shard.
4. Use a supervisor — a cheap orchestrator (even a $5 VPS) that health-checks nodes and re-provisions replacements automatically.
5. Match checkpoint cadence to interruption math — at a 5% daily interruption rate and 15-minute checkpoints, expected loss is under 1 GPU-minute per day per node. That's noise against a 50-60% price advantage.

Our local LLM fine-tuning guide walks through resumable QLoRA setups that map directly onto this pattern.

Cluster Provisioning Times

io.net's Ray-native cluster deployment is genuinely fast for what it does — assembling heterogeneous, globally distributed hardware — but it's not EC2:

Request	Typical provisioning time
Single 4090 node	2-8 minutes
Single A100/H100 (DC tier)	3-10 minutes
8-node 4090 Ray cluster	15-30 minutes
24+ node mixed cluster	30-60 minutes, occasional partial fills

Partial fills are the quirk to know: a 24-node request may come up with 21 nodes and backfill the rest over the following minutes. Write your launch scripts to tolerate late joiners.

The $IO Payment Economics

Compute on io.net is priced in USD but settled via $IO or USDC. What matters practically in 2026:

• USDC listings dominate, so token volatility no longer poisons budgeting.
• Paying in $IO typically carries a small effective discount (fee rebates), and suppliers earn staking-boosted block rewards — that subsidy is partly why listed prices undercut centralized clouds.
• Tax note (US): paying in appreciated $IO is a crypto disposal. With broker 1099-DA reporting live, your exchange is reporting those disposals — track cost basis on every top-up you spend.
• Card on-ramps exist for teams that want zero crypto exposure; you lose the token discount, keep the low base price.

io.net vs AWS Spot vs RunPod Community

Factor	io.net spot	AWS EC2 Spot (g5/p4d)	RunPod Community
RTX 4090 / equiv price	$0.19-$0.28/hr	n/a consumer (A10G ~$0.35-0.50/hr spot)	$0.44/hr
A100 80GB price	$0.75-$1.10/hr	~$1.20-1.80/hr spot (p4de, region-dependent)	$1.19/hr
Interruption warning	None	2-minute notice	None (host-dependent)
Interruption frequency	1-8%/day by tier	<5-10% typical, capacity-driven	Low-moderate
Provisioning	2-30 min	Seconds-minutes (if capacity)	Seconds-minutes
Quotas/paperwork	None	vCPU quota requests, region hunting	None
Compliance	None on spot	Full AWS compliance stack	Limited

The honest summary: AWS spot is more predictable and gives you a warning; io.net is cheaper and has no quota gatekeeping. RunPod Community splits the difference. If you've ever spent a week getting p4d spot quota approved in three regions, io.net's "click, get 20 GPUs" experience explains its growth despite the rougher edges.

Who should rely on io.net: researchers, indie builders, and teams running checkpointed training, sweeps, batch inference, or rendering — anyone whose cost of interruption is minutes of compute.

Who shouldn't: production APIs with latency SLOs, regulated-data workloads, uncheckpointable long runs, and anyone who can't automate recovery. For choosing across the whole rental landscape, see our cloud GPU comparison guide.

Related Reads

• DePIN GPU vs AWS: Cost Comparison for AI Workloads 2026
• Decentralized GPU Network Reliability: Production Reality Check 2026

Key Takeaways

• Filter by supplier tier: verified data-center-class nodes (H100/A100/4090) deliver 97-99% uptime, while residential consumer GPUs average 90-95% with 2-8% daily interruption rates—price advantage evaporates if you ignore tier scores.
• Checkpoint externally every 10-15 minutes to S3-compatible storage (e.g., Cloudflare R2) and design jobs to resume from the latest checkpoint; local disk checkpoints are lost on node drops, which happen without warning.
• Budget for P95 interruption scenarios: hard power drops, supplier withdrawals, or failed re-verifications can kill jobs instantly—assume zero warning and automate recovery via supervisors (e.g., a $5 VPS health-checking nodes).
• Reserved clusters on verified nodes approach centralized-cloud uptime but cost 10-30% more than spot; use them for multi-day training or steady inference, while spot is ideal for checkpointed batch jobs, sweeps, or dev work.
• Provisioning times vary by scale: single 4090 nodes launch in 2-8 minutes, while 24+ node clusters may take 30-60 minutes and arrive partially filled—write launch scripts to tolerate late joiners and backfills.
• Pay in USDC to avoid crypto tax complexity (token disposals are now 1099-DA-reported in the US) unless the small $IO fee discount justifies tracking cost basis for every compute spend.