Over the past 90 days, three of the largest AI-focused crypto projects—Akash, Render, and Bittensor subnets—collectively allocated over $500 million in their native tokens to GPU and data center expansion. Yet, on-chain usage metrics for these networks have decelerated by nearly 40% since Q1 2026. This disconnect between infrastructure spending and actual compute demand mirrors the very dynamic that is now rattling traditional tech investors: the uncomfortable realization that AI capital expenditures may be outpacing real revenue generation.
Context
We didn't need to look far to see this pattern forming. In the traditional tech world, Alphabet’s looming earnings report has triggered a wave of fear about its own AI capex, with analysts warning that rising spending on data centers and TPUs is not yet translating into proportional cloud or ad revenue growth. The same tension now defines decentralized AI infrastructure projects, which have attempted to bootstrap their supply side by locking tokens into validators and GPU providers. The projects promise a future where compute is democratized, but the financial mechanics tell a different story: supply is growing faster than demand, and the return on each token spent on hardware is shrinking.
Core: The DeFi Winter Returns—for Compute
To understand the risk, we must examine the tokenomics of decentralized compute projects. Akash Network, for example, uses AKT tokens to incentivize providers. As the network grows, more AKT is deployed into staking and hardware incentives. In the last six months, the total value locked in Akash’s provider bonds increased by 120%, while actual compute leases grew only 25%. This imbalance means that each new AKT injected into the system is generating less marginal economic activity. During my audit of similar token models for a DeFi lending protocol last year, I saw the same pattern: growth in TVL without proportional usage leads to a correction in token price and eventually a cut in rewards.
Bittensor faces a similar challenge. Its subnet system, which allocates TAO rewards to miners and validators based on the quality of their machine learning models, has seen a proliferation of subnets that are consuming TAO for basic computation. According to on-chain data, the top five subnets account for 80% of all TAO emissions but only 45% of the network's total validated queries. This inefficiency suggests that the token is being spent as a subsidy for compute rather than a value capture mechanism. If the market begins to question whether these subsidies can be sustained, we may see a sell-off in TAO similar to what happened to FIL during the Filecoin hype cycle of 2021.
But the real danger lies in the second-order effects. When decentralized compute projects cut their capex—by reducing provider rewards or delaying planned data center expansions—the impact ripples through the entire crypto AI stack. GPU-backed tokens (like GPU token) and DePIN coins will see their valuations collapse, as their primary use case (speculative hardware leasing) disappears. Even high-value applications like decentralized AI model training could suffer if the underlying compute layer becomes too expensive or unreliable due to underinvestment.
Contrarian: The Efficiency Hypothesis
Yet, there is a counterargument that deserves attention. Decentralized infrastructure may actually be more resilient to capex cuts than centralized giants. Because crypto projects are inherently modular and permissionless, they can pivot to more efficient resource allocation faster than a corporate behemoth. For example, Akash recently introduced fractional GPU leasing, which allows users to rent small slices of powerful GPUs for inference rather than whole cards for training. This increases utilization rates and reduces the need for constant new supply. Based on the pilot test I helped run on Golem’s network for AI content verification in the Philippines, we saw utilization jump 35% when we moved from whole-node rentals to micro-instances. The same principle applies here.
Moreover, the crypto-native ethos of “trust the code, not the CEO” might actually encourage more rational capital allocation. Unlike Alphabet, which must answer to quarterly earnings calls, decentralized projects can adjust token incentives on the fly through on-chain governance. If a subnet is underperforming, the community can vote to reduce its emissions without waiting for a board meeting. This agility could prevent the kind of overinvestment that Google is now struggling with.
Takeaway
The market is about to reward efficiency over excess. For decentralized AI infrastructure to survive the coming capex correction, projects must demonstrate that every token spent on hardware yields measurable compute usage. We didn’t build this ecosystem to replicate the mistakes of Wall Street. The future belongs to those who prioritize network utilization over raw infrastructure accumulation. The question is not whether we will cut spending—but whether we can learn to spend smarter.