You think the AI boom is just about NVIDIA and OpenAI? Look at Celestica. A 40-year-old electronics manufacturer just raised guidance by over 50%, all because of AI infrastructure demand. The market cheered. But I see something else: the same centralized supply chain bottleneck that will eventually accelerate the pivot to decentralized compute networks. Alpha hidden in the noise.
Celestica is not a household name. It’s an electronic manufacturing services (EMS) provider—basically, a professional assembler of high-end servers, switches, and storage boxes. Its customers are the hyperscalers: Amazon, Microsoft, Google, Meta. When they need racks of H100s or B200s, Celestica puts the pieces together. Its revenue spike is a direct measure of how many AI servers are being physically built. That’s not narrative. That’s hardware in the real world.
But here’s the context the headlines miss. Celestica’s 50% growth is not about innovation. It’s about capacity. The AI industry has hit a manufacturing bottleneck. Every hyperscaler is placing orders years in advance, buying every available GPU and networking chip. Celestica is one of the few qualified assemblers for these complex, high-power systems. Its guidance raise tells me that the order books are overflowing. The “sell shovels” play is working—but only for a handful of centralized factories.
Now let’s dig into what this actually means. The core insight is about capital expenditure (CapEx) concentration. Celestica’s growth depends entirely on a small group of mega-corporations. Its 10-K will likely show that 70-80% of revenue comes from less than five clients. That’s a single point of failure. Code doesn’t lie, but narratives do. The narrative is “AI infrastructure boom.” The gritty truth is that this boom is fragile, opaque, and prone to shocks. If one hyperscaler cuts its orders—say, due to regulatory pressure or a shift to in-house manufacturing—Celestica’s guidance collapses.
Compare this to crypto mining. In 2017, I watched the same pattern: Bitmain and Canaan monopolized ASIC production, and every miner was at their mercy. When the bear market hit, those centralized suppliers slashed output, and the network hash rate suffered. Today, AI hardware is repeating the same mistake. The entire AI ecosystem is trusting a handful of manufacturers like Celestica, Foxconn, and Flex to deliver the shovels. That concentration creates leverage for the suppliers, not the users.
But there’s a deeper technical angle. Celestica’s growth is a proxy for energy consumption and physical footprint. Each AI server consumes 7-10 kW. Liquid cooling is becoming mandatory. The supply chain for these components is just as concentrated. If a single factory in Thailand or Mexico faces a disruption—flood, trade war, pandemic—the AI industry halts. This vulnerability is exactly why decentralized physical infrastructure networks (DePIN) are not a luxury but a necessity.
Here’s the contrarian take: Celestica’s success is a sell signal for centralized AI infrastructure. The same logic applies to crypto mining rigs. When the hardware supply chain is concentrated, it creates a second-order effect: the cost of compute becomes opaque and political. Decentralized compute networks like io.net, Akash, and Golem offer an alternative. They aggregate idle GPUs from individuals and small data centers, creating a permissionless, resilient supply. They don’t need Celestica. They need code.
I’ll give you a specific example. In 2021, I helped a Thai artist mint NFTs using a decentralized rendering network. The experience was clunky—latency issues, inconsistent quality. But today, those networks have matured. They’re now capable of handling AI inference workloads. Celestica’s customers are building hyperscale data centers. The alternative is a world where anyone with a gaming GPU can contribute compute and earn tokens. That’s not a pipe dream. It’s already happening.
What the market ignores is that Celestica’s 50% revenue growth is a lagging indicator. It shows what we already know: demand for AI compute is exploding. The leading indicator is the number of new decentralized compute nodes coming online. I track this weekly. The growth rate of GPU nodes on Akash has tripled in the past six months. The narrative is still “cloud is king,” but the data is shifting.
The takeaway is not to short Celestica. The takeaway is to recognize that the infrastructure layer of AI is the most centralized part of the entire stack, and that centralization is a risk the market will eventually price in. Trust is the new currency. Right now, trust is placed in a few factories and logistics chains. The future of AI compute—just like the future of finance—will be trustless, permissionless, and distributed. The pieces are already being assembled. You just have to look past the noise.
