The Economics of Decentralized Compute

Date: 2026-03-12 · Scope: Cloud GPU pricing, consumer hardware operating costs, Bittensor miner economics, break-even analysis, compute marketplace pricing

1. Cloud GPU Pricing: What Centralized Compute Actually Costs

Training-class GPU compute from major cloud providers runs from $1.50 to $100+ per hour depending on the GPU, the provider, and the commitment level. These are the prices that a decentralized compute protocol must undercut to be economically relevant.

Data Center GPUs (Training Grade)

Sources: CoreWeave pricing page (March 2026), AWS EC2 pricing (public documentation), GCP compute GPU pricing. AWS and GCP prices are approximate — their pricing pages load dynamically and exact numbers shift with region and configuration.

Key Observations

CoreWeave is roughly half the price of AWS/GCP for equivalent hardware. An A100 on CoreWeave runs ~$2.70/GPU/hr vs. ~$4.10 on AWS. This price gap has persisted for years because hyperscalers bundle GPU compute with their broader cloud ecosystem (networking, storage, IAM, compliance) and charge accordingly.

Reserved instances cut prices 40-60%. A 1-year commitment on AWS drops p4d pricing to roughly $20/hr. CoreWeave advertises up to 60% discounts for committed usage. But reserved pricing requires capital commitment and utilization guarantees — exactly the constraints that decentralized compute markets aim to eliminate.

The H100 is the current training workhorse. At $6-12/GPU/hr depending on provider, an 8xH100 node costs $50-100/hr. A typical 7B model fine-tuning run takes 4-8 hours; a 70B pre-training run might consume thousands of GPU-hours. Training costs scale linearly with GPU-hours and the bill adds up fast.

2. Consumer GPU Marketplaces: The Decentralized Alternative

A parallel market for GPU compute has emerged where individuals rent out their hardware. Prices are 5-20x cheaper than hyperscalers, but with tradeoffs in reliability, networking, and support.

Vast.ai Marketplace Pricing (March 2026, On-Demand)

Source: Vast.ai API, live marketplace data, single-GPU on-demand listings sorted by price.

Provider Earnings on Vast.ai

Vast.ai takes approximately 15% commission. What hosts actually receive:

The catch: These earnings assume 100% utilization. Real-world utilization on marketplace platforms typically runs 30-60% for consumer GPUs. A more realistic monthly earning for an RTX 4090 host on Vast.ai is $30-60.

CoreWeave Per-GPU Comparison

The comparison breaks down because these are different markets serving different needs. CoreWeave sells reliable, networked, enterprise-grade clusters. Vast.ai sells raw compute from individual machines. An RTX 4090 at $0.17/hr cannot replace an H100 at $6.16/hr for most training workloads — the H100 has 3x the memory bandwidth, 80 GB HBM3, and NVLink for multi-GPU scaling. But for smaller models (0.5B-7B) and fine-tuning tasks, consumer GPUs are dramatically more cost-effective.

Other Decentralized Compute Platforms

io.net — Aggregates GPU supply from data centers, crypto miners, and consumers. Referenced A100 pricing at $3.33+/hr (comparable to cloud providers). Claims to offer lower prices through aggregation but specific consumer GPU pricing is opaque.

Akash Network — Cosmos-based decentralized cloud. Advertises GPU compute but pricing is set through reverse auction. Transparent pricing data is not publicly available outside the live marketplace.

RunPod — Hybrid model with both "Secure Cloud" (data center) and "Community Cloud" (individual hosts). Community Cloud pricing for RTX 4090 runs approximately $0.34-0.44/hr — roughly 2x Vast.ai for the same hardware, reflecting RunPod's emphasis on reliability and user experience.

3. Consumer Hardware: What It Actually Costs to Run

The economics of contributing home compute come down to two numbers: electricity cost per hour and hardware amortization. The electricity math is straightforward.

Power Draw at Training Load

These are sustained power numbers during continuous ML training workloads — not peak/burst or idle.

Sources: Consumer hardware guide benchmarks, Apple technical specifications, TechPowerUp reviews, Tim Dettmers GPU guide.

Electricity Cost Per Hour

Using $0.16/kWh (current US residential average — the EIA reports the 2025-2026 national average at approximately 16 cents/kWh, up from ~12 cents in 2020).

The power gap is 10-20x. A Mac Mini M4 Pro costs $0.006/hr to run. An RTX 4090 system costs $0.072/hr. The NVIDIA card is ~3-5x faster for training throughput, but only at 10x the power cost. Per-token-per-watt, Apple Silicon wins decisively.

Regional variation matters. California averages ~$0.30/kWh (nearly double the national average). Texas averages ~$0.13/kWh. Idaho runs ~$0.09/kWh. A Vast.ai host in California with an RTX 4090 pays $0.135/hr in electricity — eating most of the $0.17/hr median marketplace rate. The same host in Idaho pays $0.040/hr, keeping $0.13/hr margin.

4. Bittensor Miner Economics: The Cautionary Tale

Bittensor (TAO) is the largest existing decentralized AI compute network by market cap (~$2.1B, March 2026). Its economics are instructive — and sobering.

Network Economics

Sources: CoinGecko (TAO price/supply), Taostats.io (network metrics, halving schedule), Bittensor whitepaper (emission mechanics).

Reward Distribution

Bittensor splits emission across ~129 subnets, each running different AI tasks. Within each subnet, rewards split approximately:

• Miners: ~41% of subnet emission
• Validators: ~41% of subnet emission
• Subnet owners: ~18% of subnet emission

Average emission per subnet: ~27.9 TAO/day ($6,000/day). With 256 miner slots per subnet, the average miner earns ~0.045 TAO/day ($9.61/day) if rewards were distributed evenly.

But rewards are not distributed evenly. Bittensor uses Yuma Consensus, which allocates emission proportional to performance rankings. Top-ranked miners in a subnet can earn 10-100x the average. Bottom-ranked miners earn effectively nothing and risk deregistration.

Real Miner Costs

Competitive Bittensor mining requires serious hardware:

The Profitability Reality

For the average miner:

• Daily revenue: $9.61 (average)
• Daily GPU costs: $50-200 (cloud) or $5-15 (amortized hardware + electricity)
• Result: Most miners lose money

The miners who profit are those who:

1. Found early, less competitive subnets with favorable emission-to-cost ratios
2. Run proprietary models that score exceptionally well (earning outsized emission share)
3. Already own the hardware (so marginal cost is only electricity)
4. Hold TAO expecting appreciation (treating mining as a token acquisition strategy, not a cash-flow business)

The fundamental issue: Bittensor's economics are driven by token speculation, not by real demand for compute. The $774,000/day in emission is funded by inflation, not by customers paying for AI services. When TAO price drops, mining becomes unprofitable even for top performers. This is structurally identical to proof-of-work mining — revenue depends on token price, not on the value of the work performed.

5. Break-Even Analysis: When Does Home Compute Pay?

Electricity-Only Break-Even

If you already own the hardware and the only marginal cost is electricity, the bar is remarkably low.

At BTC price = $70,000 (1 sat = $0.0007):

5 sats per hour covers the electricity for a MacBook Air doing background training. That's $0.0035. For context, a Vast.ai listing for a comparable GPU (if one existed) would charge orders of magnitude more. The electricity floor for Apple Silicon is nearly free.

Full Cost Break-Even (Electricity + Hardware Amortization)

Hardware amortization assumes 3-year useful life and 50% utilization (12 hours/day average, accounting for sleep, user activity, downtime).

When hardware cost is included, the Mac Mini M4 Pro has the best economics. At $800, it amortizes cheaply. Its 40W power draw keeps electricity costs negligible. Break-even is 96 sats/hr ($0.067/hr). Compare this to Vast.ai marketplace rates for equivalent compute class — there is no direct comparison because Apple Silicon doesn't appear on these marketplaces, but the performance tier is roughly comparable to an RTX 3060 for ML training on small models.

Expensive hardware has worse ROI. A $5,000 Mac Studio M2 Ultra needs 564 sats/hr to break even. It's faster, but the higher capital cost raises the bar. The marginal economics favor cheap, efficient hardware running at high utilization — not premium hardware running sporadically.

The "Already Own It" Advantage

Most potential participants already own their hardware. They bought a Mac for work. They built a gaming PC for fun. The GPU is sitting idle 80%+ of the time. In this case, the marginal cost is electricity only, and the break-even numbers become trivially low:

• 9 sats/hr to cover electricity on a Mac Mini
• 103 sats/hr to cover electricity on an RTX 4090 gaming rig

Any payment above these floors is pure profit (ignoring wear and tear, which is minimal for solid-state hardware running within thermal specs).

6. Comparison to Bitcoin Mining

This audience knows mining economics. Here's how decentralized AI training compares.

Bitcoin Mining Today (March 2026)

Antminer S21 Pro (current generation ASIC):

• 234 TH/s, 3,531W power draw
• Hardware cost: ~$6,000
• Daily revenue: ~$7.21
• Daily electricity at $0.10/kWh: $8.47
• Daily profit: -$1.26 (unprofitable at residential rates)
• Breakeven electricity rate: $0.085/kWh

Source: WhatToMine calculator (March 2026), Bitcoin network data from mempool.space.

The Mining Comparison

The honest comparison: Bitcoin mining at home is dead for most people. Post-halving hashprice is below the breakeven electricity rate in most US states. Home mining requires industrial-scale power costs ($0.04-0.07/kWh) to be viable. ASICs generate substantial heat and noise, run 24/7 at multi-kilowatt draw, and have zero utility beyond mining.

AI training compute has three structural advantages over mining:

1. Dual-use hardware. You bought the Mac or gaming PC anyway. The marginal cost is electricity only. ASICs have no other use.
2. Power efficiency. A Mac Mini at 40W is 100x more power-efficient per device than an Antminer S21 at 3,531W. Even an RTX 4090 system at 450W is 8x more efficient. This matters because home electrical circuits, cooling, and noise tolerance all have limits.
3. Revenue from work, not inflation. A protocol that pays for useful computation (gradient contributions that improve a model) generates revenue from the value of the output, not solely from token inflation. Bitcoin mining revenue is definitionally inflationary until transaction fees dominate — which hasn't happened in 15 years.

The honest risk: If nobody values the models being trained, the revenue source dries up. Bitcoin mining has demand certainty (block reward exists by consensus rules). AI training payment depends on someone being willing to pay for the training. The protocol's economic sustainability is only as strong as the demand for its compute output.

7. What Rate Makes This Work?

The Target Range

Based on the analysis above, a decentralized training protocol needs to pay:

The "worth my time" column is subjective but important. A protocol that pays 10 sats/hr technically covers a MacBook Air's electricity, but nobody will bother installing software and dedicating compute for $0.007/hr. The payment needs to feel meaningful, even if it's small — similar to how early Bitcoin mining was "free money" even at pennies per block.

Comparison to Vast.ai Provider Earnings

What Vast.ai hosts actually earn provides a market-rate benchmark:

These are real market rates for consumer GPU compute. A decentralized training protocol competing for the same hardware supply should target similar or better rates to attract providers. But it has one advantage Vast.ai doesn't: participants earn Bitcoin, not fiat. For the target audience (Bitcoin/Lightning developers), there's an ideological premium. Stacking sats has value beyond the exchange rate.

The Economic Model

For a training coordinator funding gradient contributions via Lightning micropayments:

At 200 sats/hr ($0.14/hr) per peer:

• 100 peers = 20,000 sats/hr = $14/hr = $336/day
• 1,000 peers = 200,000 sats/hr = $140/hr = $3,360/day

For context: Fine-tuning a 3B model on CoreWeave (single A100) costs ~$2.70/hr. A swarm of 100 Mac Minis contributing gradients at 200 sats/hr each costs $14/hr total but delivers ~15-30x the raw node count. The throughput comparison is not apples-to-apples (an A100 is individually faster), but for gradient-parallel training with SparseLoCo compression, many weak peers can match or exceed a single strong peer on models that decompose well.

The training coordinator's cost: $3,360/day for 1,000 peers is expensive relative to cloud compute for a single model run. But the coordinator isn't paying for raw FLOPS — it's paying for geographic distribution, censorship resistance, permissionless participation, and the inability for any single entity to shut down training. These properties don't exist at any price on centralized cloud platforms.

8. What the Economics Don't Support

Honesty requires acknowledging where the numbers fall apart:

Large model pre-training. Training a 70B+ model from scratch requires sustained petaFLOPS-scale compute with fast interconnects. Consumer hardware cannot provide this. The inter-node bandwidth requirements (100+ Gbps NVLink) are 1,000x what home internet delivers. Decentralized compute is viable for fine-tuning and small model training, not for frontier pre-training.

Competing with hyperscaler spot pricing. AWS spot instances for A100 can drop below $1.50/GPU/hr during off-peak. For a price-sensitive ML team that doesn't need persistence, spot instances on centralized cloud are cheaper and more reliable than a swarm of consumer devices. The protocol competes on properties other than price.

Paying fair market rate for H100-class compute. At $6+/GPU/hr for H100s, the payment rates needed to attract data center hardware make the protocol uneconomical for small-scale training runs. The protocol's sweet spot is consumer hardware doing work that consumer hardware can actually perform well — small model fine-tuning, LoRA training, gradient computation on 0.5B-7B models.

Constant high utilization. Vast.ai hosts report 30-60% utilization on consumer GPUs. A decentralized training protocol would likely see similar or worse utilization, since training runs are finite and peers are intermittent. Revenue projections based on 100% utilization are fantasy.

Sources

Cloud GPU Pricing

• CoreWeave pricing page (coreweave.com/pricing), March 2026 — H100, A100, H200, B200 on-demand rates
• AWS EC2 instance types (aws.amazon.com/ec2/instance-types/p4/, p5/) — A100, H100 configurations
• Vast.ai marketplace API (cloud.vast.ai/api/v0/bundles/) — Live consumer GPU rental rates, March 2026
• H100 supply/demand analysis (gpus.llm-utils.org) — DGX H100 purchase pricing, cloud economics

Consumer Hardware

• Apple Mac Studio technical specifications (apple.com/mac-studio/specs/) — 480W max continuous (M3 Ultra model)
• Apple MacBook Pro technical specifications (apple.com/macbook-pro/specs/) — 140W USB-C adapter (M5 Max model)
• Tim Dettmers, "Which GPU for Deep Learning" (timdettmers.com, 2023) — GPU TDP and training power profiles
• l402-train consumer hardware guide (whitepaper/research/consumer-hardware-guide.md) — MLX benchmarks, power draw measurements

Bitcoin Mining

• mempool.space API — Network hashrate (~1,000 EH/s), difficulty (145T), block reward (3.125 BTC)
• WhatToMine ASIC calculator (whattomine.com/asic) — SHA-256 mining profitability, $0.031/TH/s/day hashprice
• minerstat (minerstat.com/coin/BTC) — Network hashrate, difficulty confirmation

Bittensor / Decentralized AI

• CoinGecko (coingecko.com/en/coins/bittensor) — TAO price ($215), market cap ($2.1B), circulating supply (9.6M)
• Taostats.io — Subnet count (129), halving schedule, emission rate (0.5 TAO/block), total staked
• Bittensor whitepaper (bittensor.com/whitepaper) — Yuma Consensus, incentive mechanism design, 50/50 split framework

Electricity

• US EIA Electricity Data Browser (eia.gov/electricity/data/browser/) — National average residential rate (~$0.16/kWh)