Last Thursday, a quiet but seismic shift rippled through the investment community—a signal that the old labels we used to understand the AI revolution are crumbling. Citi strategists, in a note that circulated through institutional terminals, recommended decoupling the AI investment theme from the so-called "Magnificent Seven" (Apple, Microsoft, Google, Amazon, Meta, Tesla, NVIDIA) and instead redirecting focus toward chip makers. The move was framed as a simple reclassification of market narratives, but for those of us who have spent years tracing the flow of value through decentralized networks, it felt like an echo of something deeper. It was the market finally admitting that infrastructure—the layer beneath the applications—is where the true power lies.
I remember sitting in a cramped co-working space in Nairobi back in 2017, auditing smart contracts for the ZEIP-20 standardization working group. We spent six months arguing over edge cases in token transfer logic, trying to ensure that the protocol could not be gamed by centralized validators. One of my mentors, a quiet engineer from Brazil, said something I have never forgotten: "The value of any network is not in the dApps that run on top of it, but in the protocol that secures their state. Build a great protocol, and the applications will come. Build a great application on a weak protocol, and someone will fork you." That lesson has guided my understanding of blockchain ever since. And now, watching Citi's strategists apply the same logic to artificial intelligence—shifting attention from the application layer (Big Tech's AI chatbots) to the infrastructure layer (chip manufacturers)—I cannot help but feel a sense of vindication. The investment world is finally catching up to what the crypto-native mind has known for years: the base layer is where value accrues.
But let us not get ahead of ourselves. The "Magnificent Seven" have dominated headlines for years, their market caps ballooning on the promise of AI dominance. Microsoft integrated ChatGPT into Office, Google launched Gemini, Meta poured billions into open-source models, and Amazon built Bedrock. The narrative was simple: these companies owned the data, the distribution, and the talent to win the AI race. Yet beneath the surface, a subtle tension was building. The AI models themselves began to converge—GPT-4o, Gemini, Claude, and Llama started producing similar outputs for common tasks. The competitive moat was not the algorithm; it was the compute power required to train and serve those algorithms. And who controlled that compute? Not the Seven, but the chip makers: NVIDIA, AMD, TSMC, and the ecosystem of suppliers around them. Every dollar Microsoft spent on a server rack was a dollar that flowed upstream to TSMC's factories and NVIDIA's R&D labs. The application giants were effectively becoming renters in a world built by infrastructure landlords.
This is the context in which Citi's note lands. It is not a sudden revelation but the formal recognition of a trend that has been accelerating since the launch of ChatGPT. The strategists argue that the "AI investment theme" should be separated from the mega-cap stocks because the latter's AI efforts are becoming commoditized, while the former's (the chip makers) are building the only asset with true scarcity: advanced semiconductor fabrication capacity. For a blockchain writer like me, the parallel is almost painful in its clarity. In crypto, we have watched the same dynamic play out. The Ethereum protocol, with its robust smart contract layer, has captured billions in value through ETH itself, while many of the dApps built on top have struggled to retain value. Uniswap, for all its trading volume, has a token that trades at a fraction of its peak, reflecting that the liquidity protocol is valuable but the application layer is easily forked. Similarly, Bitcoin's security budget is sustained by its proof-of-work infrastructure (miners and ASICs), while layer-2 solutions come and go. The lesson is consistent: in any stack, the most defensible layer is the one that cannot be easily replicated or replaced.
Now, let us dive into the core of Citi's analysis. The strategists do not simply say "buy chip stocks." They propose a reclassification that has profound implications for portfolio construction and industry strategy. By detaching the AI theme from the Magnificent Seven, they are effectively arguing that the market has been mispricing two distinct asset classes: the application-dominant firms (with high valuation multiples but uncertain AI ROI) and the infrastructure-dominant firms (with high multiples but extremely predictable demand growth). The logic is rooted in the economics of the AI stack. Training large language models requires tens of thousands of GPUs running for weeks. Every new model from OpenAI, Google, or Meta requires another cluster purchase. The chip makers' revenue visibility extends 12 to 18 months out, supported by pre-orders and capacity bookings. In contrast, the monetization of AI features—through subscription tiers, API calls, or ad improvements—remains unproven at scale. Microsoft's Copilot, for example, costs the company an estimated $4 per user per month in compute, but it only charges $30 for the enterprise version. The margins are thin, and the competitive pressure to keep prices low is intense. The chip makers, on the other hand, enjoy gross margins above 70% and have pricing power because their products are the bottleneck. This is the essence of the unbundling: the market is finally pricing infrastructure scarcity over application promise.
As someone who has spent years building educational content for blockchain developers in Africa, I have seen firsthand how infrastructure scarcity shapes behavior. In 2020, when I launched "The Open Ledger" educational initiative in Kenya, we partnered with three local universities to translate DeFi mechanics into Swahili and English. One of the biggest challenges we faced was not understanding the code, but accessing reliable, low-cost nodes to interact with the Ethereum network. We were bottlenecked by infrastructure—not just internet connectivity, but the computational resources to run validators or even light clients. The students who succeeded were the ones who invested in hardware, not just software. The same pattern repeats in AI. The researchers who make breakthroughs are the ones who have access to clusters, not just ideas. And the companies that provide that access—the chip makers—are the ones capturing the economic rent.
But here is where the contrarian angle emerges. If the market is so convinced that chip makers are the new AI trade, then perhaps the trade is already crowded. NVIDIA's market cap has surged past $2 trillion, and its PE ratio hovers above 70. AMD and TSMC have similarly lofty valuations. The risk is that the infrastructure narrative itself becomes a bubble, driven by the same herd mentality that inflated the Magnificent Seven. Moreover, the chip makers face their own structural vulnerabilities. The semiconductor industry is cyclical: a downturn in demand, a geopolitical shock (such as export controls on Taiwan), or the emergence of new manufacturing processes could disrupt the current trajectory. Already, the US government has tightened restrictions on AI chip exports to China, which could cap NVIDIA's growth in one of the world's largest markets. Meanwhile, the Magnificent Seven are actively developing custom chips: Google's TPU, Amazon's Trainium and Inferentia, Microsoft's Maia, and even Meta's MTIA. If these in-house alternatives become competitive, the moat of companies like NVIDIA could erode, shifting value back to the application layer (since those companies would control both the chip and the model).
From a crypto perspective, this dynamic is eerily reminiscent of the debate around Ethereum's dominance versus the rise of layer-2 solutions and alternative layer-1s. For years, Ethereum enjoyed a near-monopoly on smart contract value, but as gas fees rose and competitors like Solana and Avalanche emerged, the ecosystem fragmented. Similarly, NVIDIA's CUDA ecosystem is a powerful moat, but AMD's ROCm is improving, and new architectures (like Cerebras and Groq) are challenging the GPU-centric paradigm. The market may be underestimating the speed at which the chip landscape could diversify. The contrarian bet, then, is not against chip makers entirely, but against the assumption that they are a monolithic, risk-free play on AI.
Yet, even with these risks, the Citi reclassification is a powerful signal for blockchain-native projects. It underscores the thesis that decentralized physical infrastructure networks (DePIN) are the crypto answer to the same infrastructure scarcity problem. Projects like Render Network (which connects GPU owners with artists and AI trainers), Akash Network (a decentralized cloud marketplace for compute), and Golem (a peer-to-peer supercomputer) are all attempting to disrupt the centralized chip makers' monopoly by commoditizing compute hardware through token incentives. If the market is now prioritizing infrastructure over application, then these DePIN projects should theoretically benefit as investors seek alternative, uncorrelated exposure to the same trend. However, the reality is more complex. Decentralized compute networks are still nascent—they lack the scale, reliability, and security guarantees of AWS or NVIDIA's DGX Cloud. Trusting a DAO with your million-dollar training job is a leap most enterprises are not ready to take. The infrastructure thesis may be correct, but the crypto implementation of it is years away from capturing meaningful market share.
I recall a conversation I had last year with a Kenyan farmer who was using a blockchain-based weather prediction model for crop planning. He told me, "I don't care if the compute is decentralized or if it's running on a supercomputer in Palo Alto. I just need the price to be low and the forecast to be accurate." That pragmatism is a healthy reminder: the end user does not care about the stack's architecture; they care about value delivered. Citi's strategists are not endorsing blockchain infrastructure; they are simply following the money. The money is flowing to companies that can produce the most advanced chips reliably and at scale. That is a centralized model—TSMC's factories in Taiwan, NVIDIA's headquarters in Santa Clara. The crypto industry's vision of a decentralized, permissionless compute grid is romantic but impractical in the current technological and regulatory climate.
Nevertheless, the trend toward infrastructure appreciation is undeniably bullish for the broader blockchain ecosystem. When capital flows into infrastructure, it validates the idea that the base layer is valuable, which in turn justifies the tokenomics of many layer-1 and DePIN projects. Solana's recent price surge, for instance, can be partially attributed to its narrative as a high-performance infrastructure for AI and gaming. Ethereum's transition to proof-of-stake and its dominance in settlement similarly benefit from the same logic. By decoupling the AI theme from the Magnificent Seven, Citi has implicitly endorsed the principle that the foundation is more important than the facade.
To bring this full circle, I want to share a personal observation from my years of smart contract auditing. During the ZEIP-20 work, we identified 42 critical edge cases in token transfer logic. One of the most contentious issues was the transfer fee mechanism: some proposals allowed token contracts to deduct a fee on each transfer, which could be used to fund protocol development. The centralized validator argument was that this fee would align incentives. The decentralized counterargument was that it would create an exploitable rent-seeking vector. We eventually settled on a standard that gave developers optionality but required explicit transparency about fees. That experience taught me that infrastructure decisions have moral dimensions. A fee structure that extracts value from users to benefit a centralized entity is no different from a chip manufacturer overcharging for GPUs because they have monopoly power. The crypto ethos is to build alternatives that eliminate such extraction. Yet, the market is now rewarding centralized extraction (chip makers) more than decentralized alternatives. That is a bitter irony for those of us who believe in code as law.
In the end, Citi's reclassification is not a panacea. It is a snapshot of where the market sees value today. Tomorrow, the pendulum could swing back if a new application (say, a breakthrough in medical AI) demonstrates that owning the user interface is more profitable than owning the compute. The key is to remain skeptical of any single narrative and to understand the underlying economic drivers. For the blockchain community, this moment is a call to action: if we believe that decentralized infrastructure can provide a more equitable and resilient foundation for AI, we must build faster and better. The market is listening, but it will only reward true innovation, not just ideology.
Tracing the moral code behind every token. Building libraries where others build empires. Community over capital, always.
As I walk away from the hype to find the soul of this market, I hold onto one certainty: the unbundling of AI from the Magnificent Seven is a mirror for our own industry. In crypto, we are still searching for the killer application that justifies the infrastructure we have built. Perhaps, like with AI, the application is not the end goal—the infrastructure itself is. And if that is the case, we need to ensure that our infrastructure is not just powerful, but also just, transparent, and accessible to all.
