What if We're Building too Much AI Infrastructure?

We're spending tens of billions of capex dollars on vast datacenters but what if the AI future is more efficient and runs cooler?

Welcome to the hundreds of new subscribers who have joined over the past few weeks. In today’s free post, I take a look at a big risk that almost no one is discussing about the vast AI infrastructure buildout that is happening. Namely, what happens if algorithmic improvements make AI much more efficient or if new chip technology makes GPUs obsolete?

And if you like what you read here and you’re not yet subscribed, consider subscribing. Most of my posts are free though deeper dives tend to be paid.

If the cognitive substrate of AI shifts—and there are signs it might—then today’s datacenter boom becomes tomorrow’s stranded capital.

Billions are pouring into AI infrastructure: multi-gigawatt data centers, substations, transmission lines, 345kV power corridors. Towns like Abilene, Texas are being rebranded as AI boomtowns. Utilities, REITs, and hyperscalers are racing to lock down land, power, and thermal envelopes.

The underwriting premise is rarely stated, but it is always assumed: That today’s AI substrate—centralized, thermally dense, GPU-driven, and transformer-based—will remain dominant for the next 15–30 years.

That is: LLMs will keep scaling. Nvidia will keep shipping. Training will stay hot, centralized, and resource-intensive. And demand will follow the same thermodynamic curve we’re on now, straight up and to the right.

But what if that premise breaks?

Not tomorrow. Not next quarter. But sometime in the next ten years.

What if the compute substrate evolves?

What if the transformer/GPU stack gives way to something cooler, leaner, and less centralized?

What if we stop needing desert-scale campuses to train models that don’t even need to be trained?

If any of that happens, the capital being deployed right now is misaligned. Duration mismatch becomes a balance sheet problem. And the AI boom leaves behind a trail of dead malls: overcooled, underutilized, and fundamentally unfit for purpose.

The Unspoken Fragility of the Current Stack

Today’s AI infra stack has two fragile assumptions baked in:

1. That transformers are the best path to artificial general intelligence.

2. That GPUs are the best way to run them.

Both are contingent. Both could break.

1. LLMs are compute-maximal, not compute-optimal

Transformers work, but they’re brute-force machines. They predict tokens. They don’t understand. And they scale only by throwing more parameters, more FLOPs, more electricity at the problem.

That’s fine when gains scale linearly. But we’re already seeing diminishing returns. Scaling laws are bending. Each additional dollar buys less capability. GPT 4 wasn’t 10x better than GPT 3. We don’t know if GPT 5 will be a 10x improvement over GPT 4.

Meanwhile, researchers are exploring alternatives:

• Sparse models

• State-space architectures

• Structured memory and planning systems

• Hybrid symbolic-neural methods

If even one of them breaks through, the case for centralized 5–10 GW training clusters starts to collapse.

2. GPUs are legacy baggage

Nvidia’s dominance is a historical artifact. GPUs were built for rendering, not reasoning. They became the AI standard because they were there: cheap, parallel, CUDA-compatible.

But they’re thermally inefficient, power-hungry, and increasingly out of step with what’s coming:

• Photonic compute: low-heat, high-bandwidth

• Analog neural nets: orders-of-magnitude better performance

• Neuromorphic chips: sparse, event-driven, closer to biological computation

If compute gets radically more efficient—even just 10×—then the supporting infrastructure becomes misfit. The rack densities, the thermal envelopes, the substations sized for 1.5 kW per sq. ft. All of it mismatched to a new substrate.

You can’t forklift in analog chips and expect them to run hot. You don’t need 400 tons of chilled water for photonic inference.

The assets break. Not functionally, but financially. They don’t pencil.

Duration Mismatch as Structural Risk

This is the core problem. Infrastructure has duration. Compute doesn’t.

• Servers: 3–5 years

• Cooling and power train: 15–20 years

• Substations and transmission lines: 30+ years

The AI substrate of transformer + GPU is six years old. It might last another five. Maybe seven. But no one rational should be underwriting 30-year capex against a five-year substrate without some form of hedging.

Most aren't. They’re pricing permanence.

What Happens If the Substrate Shifts

If models compress…

If chips become radically more efficient…

If training decentralizes or inference moves to the edge…

Then:

• Multi-gigawatt training campuses become overbuilt

• Cooling plants are oversized and uneconomic

• Zoning, tax credits, and grid upgrades tied to AI demand get stranded

• Power purchase agreements go unrenewed

• Hyperscaler leases quietly expire

• Secondary asset value collapses

This isn’t cloud overbuild. You can’t re-lease an H100-optimized data hall to a logistics company. This is non-fungible infrastructure. It doesn’t adapt. It dies in place.

“But demand is infinite”

Only if the task is still thermally demanding. Only if the workload remains centralized.

Yes, cheaper compute expands demand. But at some point, once token prediction gets good enough, marginal utility flattens. You don’t need 10× the model if 2× gives you 99.99% coherence.

If efficiency gains exceed workload expansion, centralized demand contracts. The grid stays, the chiller loops stay, but the tenant doesn’t. That’s when you get dead malls.

Why No One’s Modeling This

Because no one has an incentive to.

• REITs want long-term leases to justify capex

• Utilities want AI demand to underwrite grid expansion

• Nvidia needs the infinite-scaling story to sustain its multiple

• Hyperscalers are buying land and power as an option, not a promise

• LPs and allocators think they’re getting AI exposure via real assets

Everyone’s upstream of the risk. No one’s pricing the fragility.

What To Do Instead

This isn’t a call to stop building. It’s a call to build smarter.

• Use modular thermal systems. Containerized CDU skids, immersion cooling that can be forklifted and redeployed.

• Avoid substrate-specific capex. Don’t design around H100 density if your lease is 5 years.

• Build in renegotiation triggers. If rack power or die temp shifts by 30%, so should your pricing.

• Prioritize dual-use zoning. Pick sites where other industries, such as hydrogen, biotech, or rendering, can absorb stranded capacity.

• Pace capex. Build substations first. Add white space in 50 MW blocks based on observed, not forecast, load.

And most importantly:

Price in regime change. Treat a substrate flip as base case, not tail risk.

Final Word

If the substrate holds through 2040, today’s builders will look brilliant.

If it shifts by 2028, they’ll be holding multi-billion-dollar liabilities tuned to a workload that no longer exists.

You can’t repurpose a dead mall for the mind. You can’t rezone for cognition. Once the tenant leaves, the building doesn’t reprice. It just hums, quietly, fluorescent and empty, waiting for a future that never arrives.

---

Coda

If you're enjoying this newsletter, I write every day. Some posts are public. Some are paywalled. But all of them follow the same arc: cut through the noise, get to the physical substrate, and ask: who controls the power?

If you want more of that, consider going paid. I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)