Hook: We didn't need another Wall Street strategist to tell us that diversification is prudent. But when JPMorgan's Gabriela Santos explicitly calls for spreading AI bets across regions and sectors, she's not just offering portfolio advice—she's validating a truth the crypto world has been screaming for years: centralized infrastructure concentrates risk, not value. The question is whether her version of diversification goes deep enough to escape the gravitational pull of the same handful of hyperscalers and GPU lords.
Context: Santos's thesis, as reported, is straightforward: AI investment is moving from a single-bet-on-infrastructure phase to a multi-layered, application-driven era. The technology is maturing, and the value chain is fragmenting. Hardware, models, applications, and geographies are no longer moving in lockstep. This is conventional wisdom in any maturing tech cycle. But what's missing from the analysis is the structural reality that AI's current bottlenecks—compute, data, and model access—are themselves centralized in a way that traditional diversification cannot fully hedge. The same cloud providers, the same semiconductor supply chains, the same regulatory frameworks tie together seemingly disparate AI investments. Identity isn't about where you invest; it's about who controls the underlying protocol.
Core: From a blockchain-native lens, Santos's advice is both correct and insufficient. Correct because the AI sector is indeed evolving from a winner-take-most narrative to a broader ecosystem play. The 2023-2024 GPU frenzy is cooling, inference costs are plummeting, and application-layer revenue is finally materializing. But insufficient because the diversification she recommends—geographic, sectoral, and by market cap—still operates within a centralized trust architecture. Every AI company, whether in healthcare, finance, or manufacturing, ultimately relies on the same few cloud APIs (AWS, Azure, GCP) and the same few chip suppliers (NVIDIA, AMD). The correlation is not zero; it's dangerously high under systemic stress.
Here's the crypto-native counterpoint: decentralized compute networks (Akash, Render, io.net) and on-chain AI models (Bittensor, Ritual) offer a fundamentally different form of diversification. They distribute the infrastructure layer, not just the investment ticket. When you own a token that represents compute power from thousands of independent nodes, your exposure is not to AWS's pricing power or NVIDIA's next product cycle. You're exposed to the network's ability to aggregate idle resources—a structurally different risk factor. Similarly, Liquidity isn't just about capital allocation; it's about the ability to exit under any circumstance. On-chain, you can swap a tokenized AI compute contract for a stablecoin in seconds, without a counterparty. Traditional AI diversification, by contrast, locks you into equity markets that can halt trading or suffer from counterparty risk.
The real insight from Santos's recommendation is that the AI industry's value chain is fragmenting, but the foundational layer remains centralized. This creates a unique opportunity for blockchain-based alternatives to capture the "diversification premium" that traditional portfolios cannot access. For example, a portfolio that includes $AKT (Akash) alongside $TAO (Bittensor) and $RNDR (Render) is not just diversifying across AI subsectors; it's diversifying across governance models—each network has different tokenomics, different validator sets, different geographic concentrations. That's a more orthogonal diversification than buying 10 different AI stocks that all trade on the NYSE and all use AWS.
Contrarian: But let's be honest: the crypto-AI narrative is still largely speculative. The vast majority of decentralized compute networks have negligible utilization compared to AWS. Bittensor's subnets are mostly experimental. The idea that on-chain AI can scale to meet enterprise demand is, for now, a hope more than a reality. Freedom isn't the presence of consent; it's the presence of viable alternatives. Today, there is no viable decentralized alternative to OpenAI's GPT-4 or NVIDIA's H100 clusters. So Santos's advice to diversify within the traditional AI stack is the rational move for institutional money. The contrarian view is not that crypto-AI is ready to replace it, but that the very act of diversifying into traditional AI assets may be reinforcing the centralized infrastructure that creates the risk in the first place. Every dollar that flows into a traditional AI ETF indirectly funds the same hyperscalers, the same chip monopolies, the same data brokers. The diversification is cosmetic.
Moreover, the risk of "pseudo-diversification" is real. Many AI stocks on the market, even those in different sectors, share the same risk factors: regulatory risk (EU AI Act, US export controls), compute cost risk, and model obsolescence risk. A sudden shift in the foundational model architecture (e.g., from Transformer to something more efficient) could render all current applications obsolete simultaneously. Traditional diversification fails here because the correlation is driven by a common factor—the underlying AI paradigm. Crypto-AI, by contrast, is built on incentive structures that reward heterogeneous compute resources. A network that aggregates GPUs, CPUs, and even specialized ASICs from different owners is less likely to suffer a single point of failure.
Takeaway: Santos's call for AI diversification is a signal that the industry's first wave of hyper-concentration is ending. But for those who understand the deeper mechanics of decentralization, the real opportunity lies not in spreading bets across the same centralized table, but in building a new table where the foundation itself is distributed. The question every investor should ask is not just "Which AI sectors should I own?" but "Which AI infrastructure can I trust to survive a systemic shock?" The answer, increasingly, points to blockchain-based networks that are permissionless, transparent, and resilient by design. The next bull run in AI won't be about who has the best model—it will be about who has the most sovereign compute.