The ledger remembers what the mind forgets. In early 2025, a report surfaced from a crypto-focused outlet, Crypto Briefing, claiming that cost—not technical capability—is the primary barrier to enterprise AI adoption. On its surface, this is a mundane operational finding. But when I trace the capital flows beneath it, I see something more structural: a market shifting from narrative-driven valuation to unit economics, and a fragility that echoes the 2020 DeFi liquidity mining boom.
This is not a story about AI models. It is a story about how markets price uncertainty when the hype cycle meets the balance sheet.
Context: The Global Liquidity Map and the AI Capex Supercycle
To understand why cost matters, we must place it within the broader liquidity environment. Since 2023, we have witnessed an unprecedented concentration of capital into AI infrastructure. NVIDIA's data center GPU revenue alone surpassed $100 billion in fiscal 2025, with gross margins exceeding 75%. Hyperscalers—AWS, Azure, Google Cloud—have committed hundreds of billions in capital expenditure to AI compute. This is the classic 'picks and shovels' phase, where upstream suppliers capture the majority of the value chain's profits.
The report's finding, which aligns with Gartner's projection that at least 30% of generative AI projects will be abandoned after proof-of-concept by the end of 2025, signals that the bottleneck has shifted. The technical feasibility of AI is no longer the question. The question is whether enterprise customers can build a positive return on investment (ROI) loop when the total cost of ownership (TCO) includes not just inference costs, but data governance, system integration, talent, and compliance.
Based on my 2020 analysis of MakerDAO's stability fee model, I recognized a parallel. In DeFi, liquidity mining programs subsidized total value locked (TVL) figures, creating an illusion of organic demand. When the incentives stopped, the users vanished. Enterprise AI faces a similar dynamic: the 'cost' is the subsidy that has not yet been removed, and the real users are yet to materialize.
Core Analysis: The Unit Economics of Enterprise AI and the Anthropic Signal
Let me deconstruct the cost structure. For a typical enterprise AI deployment—say, a customer service chatbot handling one million daily interactions—inference costs can reach several million dollars annually. Training costs are a one-time capital expense, but inference is an ongoing operational expenditure. This is the crux: the cost curve is linear to super-linear with scale, while the willingness to pay for AI outputs remains uncertain.
Anthropic, the company highlighted in the report, provides a case study in structural fragility. With an estimated annualized revenue of $1 billion and a valuation between $60-80 billion (following its early 2025 funding round), the price-to-sales multiple stands at 60-80x. This valuation implies a future where revenue grows tenfold and gross margins improve to 70%+. Yet, inference costs alone may consume 60-70% of Anthropic's revenue, leaving a gross margin far below the 80%+ healthy benchmark of the SaaS industry.
This is the 'high-cost, high-valuation' model that the market is beginning to question. The report's linkage of cost barriers to Anthropic's valuation is not coincidental. It reflects a paradigm shift in investor logic: from 'technology potential' to 'unit economics.' Investors are now asking the same questions they asked of DeFi protocols in 2022: Where is the gross margin? What is the customer acquisition cost? What is the retention rate?
If we apply my first-principles framework, the core contradiction is this: AI capability value creation has not yet formed a clear, quantifiable ROI loop, while costs continue to rise. This imbalance, if sustained, will lead to longer procurement cycles, smaller project scopes, and a price restructuring across the AI supply chain.
Contrarian Angle: The 'Cost' Barrier Is a Healthy Correction, Not a Death Knell
The counter-intuitive view is that this cost barrier is not a failure but a necessary correction. In the 2021 NFT energy audit, I argued that the market was ignoring externalities. Today, the cost barrier is the market's way of forcing a reckoning with externalities in AI—specifically, the environmental and capital costs of compute.
But here is the blind spot: the report's framing of 'cost' is monolithic. It fails to distinguish between model inference costs, data costs, talent costs, and system integration costs. My analysis suggests that inference costs are the most dynamic component, and they are declining rapidly. Techniques like speculative sampling, KV cache quantization, prefix caching, and continuous batching can reduce inference costs by 50-80%. NVIDIA's next-generation B200 chip promises a 2-3x improvement in inference performance.
If inference costs are the primary barrier, then the barrier is eroding faster than the market expects. The real question is not whether costs will fall, but whether enterprise AI projects can build robust ROI frameworks before the next wave of cost reduction arrives.
The second blind spot is the geopolitical dimension. US export controls on advanced chips (H100/H800) have made compute costs significantly higher for Chinese enterprises, creating a bifurcated AI market. In this context, 'cost' is not a universal constant but a function of geography and policy. This is where my cross-border payment research background becomes relevant: the flow of AI capital is as constrained by borders as the flow of money.
Takeaway: The Cost Discovery Phase and the Crypto Parallel
We are entering the 'cost discovery' phase of the AI cycle. This is analogous to the transition in crypto from the ICO mania of 2017 to the institutionalization of 2024. The market is shifting from valuing potential to valuing proof. For AI companies like Anthropic, this means the 'safety premium' and 'technology premium' in their valuations will be scrutinized against unit economics.
The signal for crypto investors is clear: the AI and crypto narratives are converging on a shared theme—the search for sustainable value creation. In both markets, the 'cost' is the friction that separates speculation from adoption. The ledger remembers what the mind forgets. As I wrote in my 2024 Bitcoin ETF analysis, institutional entry reshapes liquidity landscapes. The same is true for AI. The cost barrier is the new regulatory framework for AI adoption, and it will determine who survives the next cycle.
My forward-looking judgment is this: by 2026, we will see a divergence between AI companies that have solved their unit economics and those that have not. The former will be the 'OpenAIs' of the world—high revenue, growing margins, and sustainable valuations. The latter will face a 30-50% valuation correction, as the market applies the same rigor to AI that it applied to DeFi after Terra.
The question is not whether AI is overhyped. It is whether the market can learn from its own history. The data points don't lie. The cost is the truth. And the market is about to discover it.
