All the Way Down

My partner was building a home server workstation. Effectively what would be running our homestead infrastructure (think: planting details, watering systems, home automation). The first attempt went wrong somewhere in the assembly and he lost his gaming PC in the process, down a machine instead of up one. He reordered parts. The prices were egregiously over market. He got lucky finding what he needed on eBay thanking "Silicon Valley" people who can upgrade even higher that they're selling their already powerful unused parts.

He's building it because he needs more horsepower. The tools he has access to at work, they run better with more compute. So he's spending his own money, on his own time, to do more of what he does at work. We both are, in different ways. I pay for tools on personal accounts to stay current with what I use professionally. We talked about it that night, going back and forth between how much these tools have changed things for us and the quiet strangeness of absorbing the cost ourselves. Empowered yet double-paying.

When the assembly failed, we lost more than the machine. The gaming PC was running a local copy of our shared Obsidian vault, the brain behind our homestead plans. A five-year self-sufficiency build. Seed inventory, propane data, financial projections, a changelog of every real-world update we'd ever fed into it. Gone. I have the only surviving (stale) copy now, on my laptop. One machine. One hard drive. A distributed system with a single point of failure we didn't see until it failed. Lesson learned, as you say about systems that have already broken once.

Neither of us was complaining, exactly. That's the part that stays with me. We were genuinely energized, comparing notes on what had become possible, what we'd figured out, what we were still learning. The enthusiasm was real. The bill was real. So was the workstation half-assembled on the desk, parts too expensive, some never arriving. The tools that make us better at our work have also quietly reorganized our evenings and our bank accounts and our ideas about where work ends. How does that feel?

We meandered through the conversation. The cost of GPUs. The cost of RAM. The chain running all the way from that eBay listing back through distributors, manufacturers, the handful of fabs that produce the chips that make everything else possible. A hobbyist pricing out parts for a gaming rig is touching the same infrastructure as a frontier model lab burning through compute.

I'd just come back from a conference held at McCormick Place. Researchers, builders, industry in the same room, the kind of floor I don't get to stand on often enough. And underneath the energy, something nobody was saying directly: most of the products being built right now are wrappers around the same handful of frontier models. The intelligence is shared. The differentiation is surface. A few providers sit at every layer of it, all the way up.

The fabs that produce AI chips number in the single digits globally. TSMC alone holds 72% of the global foundry market and manufactures over 90% of the world's most advanced chips. Samsung and Intel are the named competitors; neither is close. The GPU my partner couldn't find at market price is the same hardware type running the inference behind every product at that conference. When frontier labs compete for compute, the pressure moves through the chain. Hobbyists feel it. Small businesses feel it. The consumer market for components is downstream of decisions made by a handful of companies building systems most people will never see.

The data centers running these workloads require land, water, and grid capacity that communities didn't negotiate to absorb. A single modern hyperscale data center uses as much electricity as 100,000 homes. Residential electricity prices jumped 7.1% in 2025, more than double the inflation rate, and data center expansion is a significant driver. The carbon footprint of AI systems alone could reach the equivalent of New York City's annual emissions in 2025, while their water footprint may approach the global annual consumption of bottled water. Inference compounds all of this continuously: every query, every completion, every agent action across millions of users.

I watched an earlier version of this pressure at Hologram, where I worked on IoT connectivity. The chip shortage that followed the pandemic didn't announce itself as a supply chain crisis. It showed up as delays. Hardware manufacturers waiting on components. Devices that couldn't ship meant eSIM activations that couldn't happen, which meant cellular revenue that didn't materialize. One constraint at the fab level moved downstream through hardware, through activation, through revenue: four industries registering the same original failure in different languages. The chain connects everything. The seams are just invisible until one pulls.

Most of the products being featured at McCormick Place were wrappers (with a side of services). Same models underneath, different surfaces. The intelligence has been commoditized faster than anything built on top of it could differentiate. A fair calculation to continue through volatile pricing changes. Model behavior changes. A version update that deprecates the patterns your product depends on. Sometimes, you find out when it happens.

The smart people in that room knew this. They were building anyway, because the work is real and the clients are real and the alternative is to watch from the outside. A well-designed dependency looks exactly like a reasonable decision from where you're standing when you make it. It is pretty rare, in fact, that teams stumble on something truly valuable AND novel at the same time.

And so the same providers inside the products are inside the tools that built them. GitHub Copilot. Cursor. Claude Code. The companies whose APIs power your product are powering the development environment where your team cranks out the product itself. Again? With the double-dipping?

When developers build on these tools, they're offloading more than time. A 2025 Microsoft and Carnegie Mellon study surveyed 319 knowledge workers and found that the more they trusted AI outputs, the less cognitive effort they applied independently. A separate MIT Media Lab study tracked neural connectivity in real time across 54 participants writing essays under different conditions; the brain-only group showed the most distributed activity, the AI-assisted group the weakest. Researchers called the pattern "cognitive debt" and found it persisted even after the tools were removed.

Cognitive scientists draw a line between memory offloading and judgment offloading. Memory offloading has always made workers more productive. Judgment offloading removes the reasoning step itself. The argument for why this is fine goes: the orchestration layer is yours. The platform, the integrations, the accumulated context. You built it and no competitor can take it.

If your senior engineers assembled your system through AI-assisted tools, the tools scaffolded away the architectural doubt as they went. The "this feels wrong" instinct. The tradeoff held under pressure. The accumulated friction of a hard call made without a suggested completion already waiting. Those are the things that don't get exercised when the tool resolves them first (and the tool always resolves them first).

You own code your team may not be able to fully audit, running on infrastructure you don't control, built by judgment the same providers quietly shaped while building the tools you used to build it. I use these tools. My work runs on some of this infrastructure. Finding them genuinely useful, in ways that have changed how I think and not just how fast I work, is part of what makes the rest of this uncomfortable to sit with. In yet another vector, the providers were moving into the financial infrastructure underneath all of it.

OpenAI expanded its partnership with BBVA in late 2025, rolling ChatGPT Enterprise across all 120,000 employees globally in what OpenAI called one of the largest enterprise AI deployments in financial services. Around the same time, Customers Bank announced a direct collaboration with OpenAI to embed AI across lending, deposits, and payments, building bespoke capabilities around the bank's own processes and institutional knowledge, with OpenAI's technical teams embedded inside the work. Last week, Fiserv, which runs the core banking systems for 6,000 financial institutions, announced a strategic collaboration with OpenAI to build AI agents directly into the platforms banks already rely on: not as a separate tool, but as foundational infrastructure.

The same week, personal finance tools arrived in ChatGPT for Pro subscribers.

The app built on their API, built by developers using their tools, built by judgment those tools scaffolded, runs on capital now flowing through systems they're wiring themselves into. Understanding a system well enough to see its shape while standing inside it changes nothing about the exit. The ouroboros swallows its tail not because it's hungry but because the geometry leaves no other move.

We are faster at some of our work because of all of this, more capable in specific ways, and measurably worse in others. The skills that require sustained attention, aesthetic judgment, the kind of design sense that only comes from years of deliberate practice — those don't sharpen through a prompt. The curiosity is real, and it arrived alongside the necessity, and most days it's hard to tell which one showed up first. You can be genuinely energized by something and quietly resentful of what it costs you to stay inside it. I'd like to think most people in that room at McCormick Place were navigating exactly that.

The impulse to fight for sovereignty, to build local, to own your stack, to run your own brain on your own hardware, puts pressure back onto the larger system. Every person who opts out is a node the mesh has to route around. Enough nodes, and the geometry changes. The fight for independence of heart and mind is itself a form of influence on the larger system. For now, sovereignty is a word that gets thrown at the problem. I think it means something more specific and harder than the infrastructure discourse implies. The rest of the system will adapt or it won't. I'm not sure yet which.

References

TSMC foundry market share and advanced node production data: Motley Fool, March 2026

Hyperscale data center energy equivalence: World Resources Institute, February 2026

Residential electricity price increases and data center demand: Consumer Reports, March 2026

AI carbon and water footprint estimates: Lucivero et al., "The carbon and water footprints of data centers and what this could mean for artificial intelligence," iScience, December 2025

Microsoft and Carnegie Mellon critical thinking study: Lee et al., "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers," CHI 2025, Microsoft Research

MIT Media Lab neural connectivity study (preprint, not yet peer-reviewed): Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," arXiv, June 2025, MIT Media Lab

OpenAI and BBVA strategic partnership: OpenAI, December 2025

Customers Bank and OpenAI collaboration: Fintech Global, April 2026

Fiserv and OpenAI strategic collaboration: Fiserv Investor Relations, May 2026