From the ashes of 2022, we planted seeds for 2030.
Last week, macro strategist Jordi Visser dropped a thesis that made even hardened crypto natives pause: AI compute demand will surge 20–30x, wiping out half the S&P 500 within a decade. His prescription? Pile into Nvidia, Marvell, Caterpillar, and… digital assets. As a Web3 community founder who has lived through DeFi summers and bear market winters, I read it twice. Not because the direction was wrong—AI is eating the world—but because the map was drawn with a single, centralised compass. Visser’s analysis, while bold, ignores the very infrastructure that could make his exponential future sustainable: decentralised compute. Let me unpack why.
Context Jordi Visser, chief macro strategist at 22V Research, recently published a note (widely circulated in crypto circles) arguing that consumer AI agents, full self-driving, and humanoid robots will create an insatiable demand for compute. He points to cloud providers’ $2 trillion in remaining performance obligations (RPO) as evidence that capacity is already spoken for. Traditional moats—brand, cost advantage—will crumble, he says, making legacy S&P 500 stocks obsolete. His solution: allocate 10–20% of portfolios to “frontier AI” and digital assets. On the surface, it’s a compelling narrative. But having audited dozens of DeFi protocols and Layer2 scaling solutions over the past six years, I see a glaring blind spot: the physical and economic constraints of centralised infrastructure, and the untapped potential of decentralised networks.
Core: The Infrastructure Mirage Visser’s core thesis relies on three assumptions: (1) compute demand will multiply 20–30x, (2) cloud providers can scale linearly, and (3) AI gains will flow exclusively to centralised giants like Nvidia and hyperscalers. Let’s test each against reality.
1. The 20–30x multiplier has no engineering basis. Visser conflates training and inference. Training demand may saturate as models cross diminishing returns—OpenAI’s Orion reportedly hit performance plateaus with more data. Inference, on the other hand, will indeed grow, but not uniformly. Consumer AI agents (think Siri on steroids) require low-latency, always-on compute. Yet today’s models hallucinate, cost pennies per query, and still lack long-term memory. Scaling inference to billions of users demands not just raw chips, but an intelligent, resilient distribution layer—something centralised data centres struggle with due to latency and single points of failure. Based on my experience modeling gas costs on Ethereum after the Dencun upgrade, I’ve seen how congestion ripples through a shared resource pool. Centralised compute is no different: a spike in AI demand will cascade into resource contention and price volatility.
2. The $2 trillion RPO is misleading. Visser cites cloud providers’ RPO as proof of insatiable demand. But RPO includes years of non-AI services—storage, databases, networking. Worse, RPO can be cancelled or renegotiated. In 2023, AWS’s RPO grew 20% year-over-year, yet its compute utilisation hovered around 60–70%. Hyperscalers are building, but they are also overbuilding. The real bottleneck isn’t demand—it’s supply chain: advanced packaging (CoWoS), HBM memory, and megawatt-scale power. Taiwan Semiconductor (TSMC) can only produce so many chips. Even if Nvidia triples output, the physical limits of fab capacity and energy grids will cap growth far below 20x. I recall a similar over-optimism during the 2017 ICO mania, when every project claimed “infinite scalability” on Ethereum. Reality hit when CryptoKitties clogged the network. Centralised compute faces the same physical constraints.
3. Visser’s “winning” stocks face hidden risks. Nvidia’s forward PE may look attractive relative to its growth, but competition is real: AMD’s MI400, Intel’s Gaudi, and a wave of custom ASICs (Google TPU, Amazon Trainium) are eroding Nvidia’s moat. Marvell’s networking chips are essential, but customers (AWS, Azure) are vertically integrating their own networking silicon. Caterpillar and Modine benefit from data centre construction, but real estate, permitting, and energy transmission timelines stretch years. Meanwhile, Visser ignores the model layer entirely—OpenAI, Google, Meta—where the real value may accrue. From my vantage point running a community that bridges DeFi and AI, I’ve watched the model war shift toward open-source: Meta’s Llama 3 is closing the gap with GPT-4, and decentralised inference platforms like Bittensor and Akash are already routing compute to where it’s cheapest. That’s a paradigm Visser’s top-down macro lens misses.
From the ashes of 2022, we planted seeds for 2030.
Contrarian: Why Decentralised Compute Is the Missing Piece Here’s the contrarian angle that Visser—and most Wall Street strategists—overlook: centralised AI infrastructure is fragile, opaque, and ultimately unsustainable for the scale he envisions. A single hyperscaler outage (AWS went down for 5 hours in 2024) could stall millions of AI agents. Energy costs are soaring; data centres already consume 2% of global electricity. And the regulatory hammer is swinging—the EU AI Act, China’s model approval regime, and US export controls could fragment markets overnight.
Decentralised compute networks offer a resilient alternative. Protocols like Render Network (RNDR) and Akash Network (AKT) allow anyone to offer spare GPU cycles, forming a distributed cloud that scales elastically without central bottlenecks. Bittensor’s subnets let AI models compete for compute in a permissionless marketplace. These networks mirror what DeFi did for finance: remove gatekeepers, reduce counterparty risk, and align incentives through tokens. During the 2022 bear market, I watched the Render Network’s utilisation grow 4x despite falling token prices—a signal that real demand was forming. Today, with AI inference projected to outpace training, decentralised compute could absorb the spillover demand that hyperscalers can’t satisfy profitably.
Moreover, Visser’s thesis that traditional moats will crumble applies equally to his own recommended assets. If open-source AI models become just as capable as closed ones (a trend we’re already seeing with Llama 3.1 and Mistral), the value of proprietary inference infrastructure diminishes. Meanwhile, decentralised networks offer something centralised giants cannot: censorship resistance, data sovereignty, and lower costs for long-tail applications. In the Philippines, where I live, AI-powered services are often too expensive for local startups. Decentralised compute could democratise access, just as mobile money (GCash) leapfrogged traditional banking.
Takeaway: The Future Is Hybrid Visser is right about one thing: AI will reshape entire industries. But his binary view—centralised compute wins, legacy stocks die—ignores the messy, fractal reality. The next decade won’t belong solely to Nvidia or hyperscalers. It will belong to networks that combine the efficiency of centralised systems with the resilience of decentralised ones. From the ashes of 2022, we planted seeds for 2030.
As a Web3 community founder, I’ve learned that the most powerful narratives are the ones that acknowledge their own blind spots. Visser’s vision is a useful wake-up call for traditional investors, but for those of us building in crypto, the real opportunity lies not in chasing the same centralised incumbents—but in constructing the decentralised backbone that makes AI’s exponential future truly inclusive, secure, and sustainable.
What will your portfolio look like when the silicon bottleneck hits? The answer may not be a ticker symbol—but a protocol.
