We don't need more users; we need more stewards.
A few weeks ago, I sat through a governance call for a prominent Layer-2 protocol that had just earmarked $30 million from its treasury to purchase high-end GPUs for a decentralized AI inference network. The enthusiasm was palpable—the founding team spoke of bringing the benefits of artificial intelligence to on-chain applications, of democratizing compute, of building the backbone of a new internet. But as I listened, a cold realization settled in: we were repeating the same mistakes we saw in the 2017 ICO era, now dressed in the language of AI and proof-of-work for models rather than for tokens.
This is not a commentary on AI's potential. It is a forensic analysis of a capital allocation crisis that is quietly unfolding across the crypto-AI landscape—driven by the same flawed assumptions that have led traditional tech giants like Google to face a reckoning over their own AI investment returns. If we fail to learn from these signals, we risk not just wasted funds, but the collapse of the very trust that defines our industry.
The context is familiar to anyone who has watched the bear market relentlessly grind down false promises. Since 2024, a wave of crypto projects—from Layer-1s to specialized middleware—have raised, borrowed, or drained treasuries to fund AI-related infrastructure. The narrative is seductive: AI needs decentralized compute, data sovereignty, and transparent inference; crypto provides the economic incentives and trust layer. Yet the numbers tell a different story.
Drawing on a granular analysis of on-chain treasury data and governance proposals from twelve major crypto-AI protocols (representing a combined treasury of over $2 billion at peak), I’ve identified a pattern that mirrors the tension in Big Tech’s AI capex. In traditional markets, Alphabet’s Q2 2024 earnings preview highlighted a structural contradiction: massive capital expenditures in AI infrastructure with uncertain incremental revenue, particularly from Cloud services showing backlog growth slowdowns. For crypto, the dynamic is even starker—our projects do not have advertising or subscription revenue to fall back on. Their only sources of yield are speculative token emissions, lending protocols, and, in a few cases, actual usage fees.
Based on my audit experience during the 2024 bear—when I helped a DAO restructure its overspending on compute resources—I’ve seen the warning signs firsthand: idle GPUs, underutilized data center contracts, and governance tokens whose price action no longer correlates with network activity but with the hype of AI announcements. The underlying question is existential: Are we building for the peak of the narrative, or for the valley where survival matters?
The Core Analysis: Seven Dimensions of Misallocation
Technology Route: The Hollow Promise of Proprietary Models
The first dimension is low relevance but foundational. Most crypto-AI projects do not innovate on model architecture. Instead, they wrap existing open-weight models (LLaMA, Mistral) with a token incentive layer. I reviewed the technical roadmaps of five projects that claimed to be building "decentralized training" — only one had a viable plan for distributed model training across heterogeneous nodes, and even that plan relied on centralized orchestration. We are not advancing the state of AI; we are renting buzzwords. We built not for the peak, but for the valley.
Commercialization: The Token Subsidy Trap
Commercialization is where the tension becomes acute. Many protocols offer subsidized compute to attract users, burning through tokens to simulate demand. For example, Protocol X — a decentralized inference network — spends 60% of its treasury on GPU leases but generates less than 5% of that value in user fees. The gap is plugged by continuous token issuance, which dilutes holders and eventually crushes the incentive structure. I’ve seen this cycle in DeFi before: liquidity mining creates fake volumes; here, compute mining creates fake AI usage. The core insight is that without a sustainable revenue model tied to actual human need (not to speculative AI mania), these projects are Ponzi-like structures that will collapse when token prices fall below the cost of compute.
Trust is the only protocol that cannot be coded. When a project's tokenomics reward waste, trust dissolves.
Industry Impact: The Signal of a Capital Contraction
If one major project — say, the largest Layer-2 by TVL — were to announce a 50% cut in its AI compute budget, the ripple effect would be as devastating as Alphabet cutting its AI capex. It would be interpreted by the market as the leading indicator of the end of the crypto-AI hype cycle. I call this the "Google moment" for our industry. And I suspect it is coming sooner than most realize. The data are already there: on-chain usage of AI oracle services has plateaued since Q1 2025, and the number of active AI agents processing data per day has actually declined over the past two months. The industry impact would cascade into deflation for GPU tokens (like Render, Akash), falling value for data marketplaces, and a sudden re-evaluation of all projects that have tied their identity to AI.
Competition: The War of Wallet Sizes
Competition among Layer-2s and DeFi protocols for the "crypto-AI compute market" is a war of attrition, not innovation. We are seeing a bifurcation: incumbents with large treasuries (like Ethereum L2s with billions in treasury) can afford to buy GPUs in bulk and subsidize AI activity, while smaller chains are borrowing to keep up. This is unsustainable. The competition for AI market share is driving a destructive arms race where the first project to blink will be the one that gets blamed for the sector’s downturn. My own community — The Alignment Circle — conducted a survey of 50 DAO builders in late 2025; 80% believed that the current level of AI infrastructure spending is excessive relative to user demand. Yet they continue to approve budgets for fear of being left behind. This is a collective action problem without a decentralized solution.
Ethics and Security: The Regulatory Harmony We Ignore
Ethics and security are the forgotten dimensions. Many crypto-AI projects boast about "trustless inference" but rely on centralized APIs for model updates. They ignore emerging regulations around AI liability and data privacy. In 2025, when I collaborated with a DeFi protocol to audit its KYC processes (integrating privacy-preserving compliance), we found that the AI module had been accessing user data without explicit consent, essentially eroding the trust we claim to uphold. The tension between regulatory harmony and decentralization is not an obstacle—it is an opportunity. Projects that proactively design for compliance (without sacrificing privacy) will survive the coming capital contraction, because they will not face regulatory shutdowns. But most current spending is on hardware, not on governance or legal frameworks. That is a misallocation of priorities.
Investment and Valuation: The Token as a Share of GPUs
From an investment perspective, the valuation of crypto-AI projects is dangerously tied to the price of GPUs and the narrative momentum. When Google or Microsoft cuts AI capex, the price of GPUs may fall, but for crypto projects that have already purchased hardware at peak prices, their balance sheets are locked into a depreciating asset. I have seen token models where the primary value accrual comes from staking rewards generated by GPU rental yields—those yields are now compressing. The risk of a death spiral is real: if utility token prices fall below the cost of operating the hardware, the network becomes uneconomical, leading to further collapse.
Infrastructure: The Data Center Mirage
Finally, infrastructure. The physical side—data centers, power agreements, GPU clusters—is where the capital is bleeding. I recall a 2024 conversation with a founder who had leased a full warehouse in Southeast Asia for AI inference. His cost per transaction was 10x higher than a centralized solution. He argued that decentralization justified the premium, but his user base never materialized beyond 100 testers. This is the crypto-AI equivalent of "build it and they will come." But they didn't. And now the data center sits half-empty, draining treasury.
We don't need more users; we need more stewards. Stewards would ask: does this infrastructure serve a real need, or does it serve the myth of inevitability?
The Contrarian Angle: The Blind Spot of Centralized-Efficiency Worship
The contrarian view—and one I must confront honestly—is that crypto-AI infrastructure spending might be a necessary long-term bet. Just as early internet infrastructure companies burned cash for years before the dot-com survivors emerged, perhaps today's GPU purchases are seeds for tomorrow's decentralized intelligence. Optimists argue that the cost of compute will drop, and the quality of decentralized inference will rival centralized giants. They point to the resilience of Bitcoin mining: yes, it consumed capital, but it built an asset that stores value.
But this analogy fails. Bitcoin mining produces a commodity (hash power) that directly secures the network and derives value from the coin. Crypto-AI infrastructure produces services that compete with free or near-free centralized alternatives. The value proposition is weaker. Moreover, the cryptographic guarantees of decentralized AI are often negligible for most use cases—most users do not care whether an AI inference ran on a TEE or a standard GPU. The cost premium for decentralization will not be tolerated when the market is rational.
Furthermore, there is an even more contrarian possibility: the entire crypto-AI narrative may be a manufactured crisis to push new token sales, much like I argued about DeFi liquidity fragmentation. Venture capitalists who invested in compute-intensive projects need to see a "GPU shortage" story to exit. The cutting of capital expenditure might actually be a healthy correction that forces the industry to focus on application-layer innovation rather than infrastructure speculation. In that sense, a slowdown is not a disaster but a purification.
We built not for the peak, but for the valley. The valley is where we will discover which projects built on solid governance and real usage, and which built on sand.
Takeaway: A Call for Stewardship, Not More Compute
If I have learned anything from the burnout of 2022 and the cautious community building of 2024, it is that our survival depends on aligning capital expenditure with actual human need and sustainable tokenomics. The protocols that will endure are those that treat infrastructure as a service to be leanly provisioned, not a monopoly to be built on speculation.
So I end with a question: What are we building toward? If the answer is "decentralized AI that serves people," then we must start by serving the people who hold the tokens, the developers who build on the platform, and the regulators who set the rules of engagement. That means spending less on GPUs and more on community governance, regulatory compliance, and user experience. Because trust is the only protocol that cannot be coded, and we are burning capital as if it were infinite.
The valley is coming. Let us not meet it with empty servers.
