Hyperscalers will spend $660–690 billion on AI infrastructure in 2026, nearly doubling 2025 levels. Against that backdrop, decentralized GPU compute networks — Akash, Render, Aethir, io.net — collectively represent a $25 billion token market capitalization and claim cost savings of 50–75% versus ...
"AI capacity is being monetized as quickly as it is installed." — Andy Jassy, CEO, Amazon
Hyperscalers will spend $660–690 billion on AI infrastructure in 2026, nearly doubling 2025 levels. Against that backdrop, decentralized GPU compute networks — Akash, Render, Aethir, io.net — collectively represent a $25 billion token market capitalization and claim cost savings of 50–75% versus AWS, Azure, and GCP. The gap between those two numbers defines the sector's core tension: decentralized compute protocols have assembled supply and attracted speculative capital, but their actual utilization metrics tell a more complicated story than token prices suggest.
Akash Network posted record $5 million in compute spend in Q1 2026, yet its GPU utilization fell to 33.7%. Aethir reported $166 million in annualized recurring revenue from 150+ enterprise clients. Render Network burned 1 million RENDER tokens through its Burn-Mint Equilibrium model but has yet to disclose granular throughput data for its AI compute subnet. The sector is real — but it remains a rounding error against the $100 billion-plus centralized AI cloud market.
This report examines whether decentralized GPU networks are building sustainable economic value or simply repackaging idle hardware into token-incentivized supply that outpaces actual demand.
The five largest cloud and AI infrastructure spenders — Amazon, Alphabet, Microsoft, Meta, and Oracle — have committed $660–690 billion in combined capital expenditure for 2026. Amazon alone plans $200 billion, the majority earmarked for data centers. Alphabet revised its capex guidance upward three times, from an initial $71–73 billion to $175–185 billion. Microsoft disclosed an $80 billion backlog of unfulfilled Azure orders constrained by power availability. Oracle increased spending 136% over 2025 to $50 billion.
AWS holds approximately 31% of global cloud infrastructure market share, followed by Azure at 25% and GCP at 11%. Together, these three control over two-thirds of the market. Google Cloud surged 63% year-over-year in a recent quarter, hitting a $20 billion quarterly revenue run rate driven primarily by enterprise AI adoption.
On-demand H100 pricing across hyperscalers remains elevated: $98.32/hour for an 8-GPU instance on AWS (p5.48xlarge), $98.46/hour on Azure, and $98.32/hour on GCP — effectively $12.29 per GPU-hour. Spot pricing reduces this by 60–70%, but availability is unpredictable and unsuitable for continuous workloads.
The Stargate Project, a joint venture targeting $500 billion by 2029, has begun its initial $100 billion deployment across five sites in Texas, New Mexico, and Ohio. This represents a separate, government-aligned compute buildout that will further concentrate AI capacity among centralized operators.
Four protocols dominate decentralized AI compute: Akash Network, Render Network, Aethir, and io.net. Their combined token market capitalization sits within the broader $25 billion AI crypto category, according to CoinGecko data from June 2026.
Akash Network operates a permissionless compute marketplace built on Cosmos SDK. It processed 43,540 new leases in Q1 2026, a 27.1% quarter-over-quarter increase. The network crossed $5 million in cumulative compute spend during Q1 2026 — a record, but a figure that AWS generates in approximately 12 seconds of revenue. On March 23, 2026, Akash activated its Burn-Mint Equilibrium (BME) mechanism, burning 53,520 AKT by quarter-end.
Render Network migrated from Ethereum to Solana and processed over 63 million frames cumulatively, with 22 million frames rendered in 2025 alone. Its AI compute subnet, Dispersed, launched as a dedicated layer for inference and training workloads, pricing GPU-hours at approximately $0.69. The network burned 1 million RENDER tokens through its BME model by December 2025. Salad, a distributed compute partner, projected $4.3 million in first-year revenue from its Render integration.
Aethir reported $39.8 million in Q3 2025 revenue — its highest quarter — reaching $166 million in annualized recurring revenue. The network claims 150+ enterprise clients and has delivered 1.8 billion compute hours across 430,000 GPU containers supporting NVIDIA H100, H200, B200, and B300 hardware. Unlike most DePIN protocols, Aethir's revenue derives primarily from enterprise billing rather than token emissions.
io.net aggregates 100,000+ idle GPUs on Solana, scheduling them via the Ray framework into clusters of up to 10,000 accelerators. The network reported peak AI training usage as of March 2026 and implemented a dynamic tokenomics overhaul linking emissions to utilization with a fee-burn mechanism. Pricing runs 50–70% below AWS for comparable workloads.
The most revealing metric in decentralized compute is not supply — it is utilization. Here, the data is mixed.
Akash's Q1 2026 numbers illustrate the problem. While new leases rose 27.1%, average GPU usage fell 57.4% quarter-over-quarter to just 84 GPUs. Average GPU availability declined 57.5% to 334 units. This places GPU utilization at approximately 33.7% — meaning two-thirds of available GPU capacity sat idle. Meanwhile, lease revenue dropped 45% to $253,250 despite the increase in transaction volume, indicating a shift toward smaller, cheaper workloads.
For context, enterprise Kubernetes clusters average just 5% GPU utilization according to Cast AI's 2026 State of Kubernetes Optimization Report, drawn from telemetry across 23,000 clusters. This means centralized infrastructure faces a similar problem at scale — 95% of provisioned GPU capacity sits idle at any given moment. Decentralized networks are not uniquely inefficient; they mirror a structural industry challenge.
Bittensor, the largest AI token by market capitalization at approximately $3.2 billion, operates 128+ subnets but lacks the granular public dashboards common in DeFi. With ~70% of all TAO staked, the network has high participation rates, but proxy metrics — subnet growth, staking volume — do not translate directly into revenue. Bittensor's first halving in 2026 cut block rewards by 50%, tightening supply, but did not resolve the transparency gap in demand-side metrics.
The AI crypto sector's $3.4 billion daily trading volume as of late May 2026 exceeds most protocols' annualized compute revenue by orders of magnitude. Token speculation remains the primary economic activity, not compute utilization.
The cost advantage is the sector's most defensible claim. The data supports it, within constraints.
| Provider Type | H100 Per GPU-Hour (On-Demand) | |---|---| | AWS / Azure / GCP | $12.29 | | CoreWeave / Lambda Labs | $2.00–$4.25 | | RunPod (Community Cloud) | $1.39 (A100 80GB) | | Render (Dispersed) | ~$0.69 | | io.net | 50–70% below AWS |
Specialized neoclouds (CoreWeave, Lambda Labs) offer H100s at $2.00–$3.00 per GPU-hour — 35–50% below hyperscaler on-demand rates. Decentralized peer-to-peer marketplaces push further: Render's Dispersed subnet charges approximately $0.69 per GPU-hour, and io.net claims 50–70% savings versus AWS.
The tradeoff is well-documented: variable availability (commonly 70–85%), potential instance interruptions, limited enterprise SLAs, and narrower managed-service ecosystems. For batch inference, short-duration training runs, and non-production workloads, the economics favor decentralized. For mission-critical, long-running, SLA-bound workloads, centralized infrastructure retains the advantage.
Alphabet's disclosure that it reduced Gemini serving costs by 78% through model optimization underscores a second dynamic: hyperscalers are optimizing their own cost curves aggressively. The pricing gap may narrow as centralized providers improve efficiency rather than decentralized providers gaining share.
The economic models governing these networks determine whether token value accrual follows compute demand or exists independently of it.
Burn-Mint Equilibrium (BME) has emerged as the dominant tokenomic design. Akash's BME mandates that all on-chain compute spending triggers a market buy and permanent burn of AKT. Render's BME burned 1 million RENDER tokens by late 2025. Both models create a direct link between network usage and token deflation — in theory. In practice, burn rates remain modest relative to total circulating supply. Akash burned 53,520 AKT in nine days of BME activation; annualized, that represents a small fraction of total supply.
Emissions-linked models (io.net, Bittensor) tie token issuance to network participation. io.net's 2025 overhaul split rewards between availability and usage, aligning supplier incentives with actual demand. Bittensor's Dynamic TAO upgrade concentrates emissions among top-performing subnets — effectively a market-driven allocation of new supply.
Aethir's model is the outlier: $166 million in annualized revenue from paying enterprise clients represents real economic activity, not tokenomic engineering. Whether that revenue scales beyond the current 150+ client base will determine if Aethir becomes the sector's first protocol to generate meaningful cash flows.
The Artificial Superintelligence Alliance (ASI), formed from the merger of Fetch.ai, SingularityNET, and Ocean Protocol, represents the sector's largest consolidation attempt. Ocean Protocol's October 2025 withdrawal, citing a desire to control its own tokenomics, illustrates the governance challenges in merging decentralized AI projects.
NEAR Protocol's admission to NVIDIA's Inception Program in May 2026 and its launch of a confidential GPU marketplace built on Trusted Execution Environments represent the most concrete enterprise-facing moves in the sector. The v2.13 upgrade, scheduled for June 2026, introduces dynamic resharding and post-quantum cryptographic signing.
A June 2, 2026 industry panel featuring leaders from Akash, Nosana, Aethir, and io.net addressed whether AI infrastructure will be dominated by centralized hyperscalers or decentralized networks. The consensus, according to multiple sources: hybrid architectures are the likely outcome, not wholesale displacement.
Enterprise adoption constraints remain consistent: compliance requirements, SLA expectations, data sovereignty mandates, and integration complexity. Decentralized networks that solve for these constraints — Aethir with enterprise billing, NEAR with TEE-based confidential compute, io.net with Ray framework compatibility — are better positioned than those optimizing solely for cost.
Decentralized GPU compute networks have assembled real infrastructure, attracted enterprise clients, and established a demonstrable pricing advantage over hyperscalers. Those are facts, not speculation. But the sector's economic value — measured in actual compute revenue — remains microscopic relative to both its own token valuations and the centralized market it claims to displace.
The $660–690 billion hyperscaler capex wave will flood the market with additional centralized capacity through 2027–2028. Decentralized networks must convert their cost advantage into sustained utilization growth before that capacity comes online and compresses the pricing gap. Aethir's enterprise revenue trajectory and Akash's BME mechanism offer two credible paths toward value accrual. Whether either can scale fast enough to matter at the margin of a $100 billion-plus market is the open question.
The honest assessment: decentralized compute is a functioning niche with real users and real cost savings, serving a small but addressable segment of AI workloads. It is not, based on current data, a systemic competitor to AWS, Azure, or GCP. The market prices it as if it were.