Silicon Valley is drunk on the phrase "rational frenzy."
The current consensus among tech analysts, venture capitalists, and enterprise buyers is that the massive capital expenditure on artificial intelligence is entirely justified because it mirrors the early days of the internet. They look at the hundreds of billions of dollars flowing into data centers, Nvidia chips, and energy grids, and they call it a structural super-cycle. They tell you that even if there is a short-term bubble, the foundational infrastructure will remain valuable, just like the fiber-optic cables laid during the dot-com boom.
They are completely wrong.
The comparison to the dot-com fiber boom is a fundamental misunderstanding of asset depreciation and technological obsolescence. This is not a standard infrastructure build-out. This is a hyper-depreciating capital trap. Enterprise buyers and investors who are hoarding current-generation hardware under the assumption that it will retain value are in for a brutal awakening.
The Myth of the Reusable Data Center
The core argument for the "rational frenzy" rests on the idea of durable infrastructure. The narrative suggests that even if software applications fail to monetize immediately, the underlying physical assets—the chips, the servers, the real estate—will serve as a bedrock for the next two decades of computing.
This view ignores the brutal reality of hardware lifecycle mechanics.
When telecom companies laid millions of miles of fiber-optic cable in the late 1990s, the physical glass stayed in the ground. When those companies went bankrupt, the next wave of internet pioneers bought that same fiber for pennies on the dollar. Crucially, that dark fiber was still highly functional; it just needed better electronics at the endpoints to light it up. The asset did not rot.
Silicon is different.
An H100 or B200 GPU cluster is not a passive glass cable. It is a highly specialized computing unit with a functional obsolescence clock that ticks faster than any enterprise hardware we have ever seen. The pace of architectural innovation in AI compute means that hardware bought twenty-four months ago is already nearing economic irrelevance for training frontier models.
Imagine a logistics company buying a massive fleet of delivery trucks, but every single year, a new engine design comes out that is four times faster and uses a quarter of the fuel. The resale value of the old fleet does not just drop; it completely evaporates.
I have watched enterprise technology officers commit tens of millions of dollars to build internal clusters, only to realize by the time the cooling systems are fully operational, the hardware is vastly outperformed by cloud providers offering newer architectures on a consumption basis. They are left holding depreciating silicon that costs more to power and cool than the value of the compute it generates.
The Compute Efficiency Paradox
The current investment thesis assumes that demand for raw compute will scale linearly forever. The logic goes: bigger models require more parameters, which require more chips, which require more power. Therefore, buying chips today is a safe bet on tomorrow's demand.
This ignores a foundational principle of technology shifts: engineering always optimizes away from raw brute force. We are already seeing this shift occur through several distinct mechanisms.
Algorithmic Efficiency Gains
The software layer is rapidly learning how to do more with less. Research from major labs demonstrates that optimization techniques can reduce the compute required to train or run a model of equivalent capability by orders of magnitude.
- Quantization: Reducing the precision of weights (e.g., from FP16 to INT8 or INT4) allows models to run on significantly cheaper, less power-hungry hardware without a meaningful drop in accuracy.
- Small Language Models (SLMs): Highly optimized models trained on curated, high-quality datasets are routinely outperforming older, massive models while requiring a fraction of the parameters.
- Architectural Shifts: The dominance of dense Transformer architectures is being challenged by MoE (Mixture of Experts) and alternative state-space models that route tokens dynamically, drastically cutting the active compute required per inference token.
When a software tweak can suddenly make a model four times cheaper to run on existing hardware, or allow it to run on standard consumer devices, the desperate enterprise demand for massive, centralized cluster space softens significantly.
The Inference vs. Training Delusion
The "super-cycle" narrative conflates training compute with inference compute. Training requires massive, tightly coupled clusters with ultra-low latency interconnects like InfiniBand. This is where the extreme capital expenditure lives.
Inference—the actual running of the model for users—is highly parallelizable and can be distributed across cheaper, less specialized hardware, including edge devices and standard CPUs.
As the industry shifts from the training phase of foundational models to the deployment phase of specific applications, the premium on hyper-expensive training clusters will drop. The revenue model changes from capital expenditure to operational efficiency. Those who built giant training facilities assuming they could rent them out at a premium forever will find themselves competing with localized, hyper-efficient inference nodes.
The Energy Grid Bottleneck is a Mirage
A secondary pillar of the rational frenzy argument is the energy play. Media outlets are filled with stories about tech companies buying nuclear power plants and locking up gigawatts of grid capacity. The thesis is that energy constraints create a moat around existing data center players.
This is a classic top-of-market assumption. It assumes that the energy density of computing will remain static.
Historically, computing has always followed forms of Koomey's Law, where the number of computations per joule of dissipated energy doubles roughly every 1.5 years. While the sheer scale of AI has temporarily outpaced these efficiency gains, the pressure to optimize power consumption is now the primary engineering constraint.
Hardware companies are not stupid. They are pouring engineering resources into performance-per-watt metrics, not just raw compute output. At the same time, localized energy solutions, like on-site modular reactors and advanced thermal management, are decoupling data centers from traditional municipal grids. The idea that a company has a permanent competitive advantage just because they secured an interconnect agreement with a local utility provider in 2025 is a short-sighted view of infrastructure evolution.
The Enterprise ROI Reality Check
Let us look at the actual balance sheets. To justify a multi-trillion-dollar infrastructure build-out, the enterprise software built on top of this hardware must generate equivalent value. Right now, there is a massive disconnect between infrastructure spend and software revenue.
The current enterprise deployment pattern looks like this:
- A company licenses a frontier model or builds a custom wrapper.
- They deploy it to automate internal customer support or assist with code generation.
- They see a modest incremental improvement in efficiency (e.g., 15-20% faster ticket resolution).
This is a productivity gain, certainly. But it is a feature, not a new economy. A 20% increase in customer service efficiency does not justify a 500% increase in the company's enterprise software spend.
Furthermore, the switching costs between models are practically zero. If a company builds an application using one provider's API, they can swap it for a competitor's cheaper, faster API over a weekend. There is no lock-in at the intelligence layer. This commoditization means that the margins for software providers will be squeezed down to the bone, which in turn means they cannot afford to pay high premiums for compute infrastructure indefinitely.
The Risk of the Sovereign CapEx Bubble
Much of the current demand driving the hardware super-cycle is not even commercial; it is sovereign and venture-subsidized.
Governments around the world are buying up hardware to build "national AI capabilities." Venture capital firms are funding startups whose primary expense is transferring those venture dollars straight to cloud providers to rent GPUs. This creates a circular economy where artificial demand distorts the true market value of the infrastructure.
When the venture funding dries up, or when nation-states realize that owning a massive GPU cluster does not automatically translate into economic growth, that demand will vanish overnight. The market will suddenly be flooded with secondary hardware, driving rental prices through the floor.
How to Navigate the Infrastructure Shift
Stop buying the narrative that you need to own the stack or secure massive long-term compute contracts to survive. The winning strategy right now is extreme asset agility.
- Rent, Do Not Buy: Avoid locking capital into physical computing assets that depreciate faster than a new car driven off the lot. Use cloud infrastructure on a flexible, short-term basis, even if the spot prices seem higher upfront than amortized long-term contracts.
- Optimize for Portability: Build your application layer to be completely model-agnostic. Do not tightly couple your software to a specific hardware architecture or a single provider's infrastructure.
- Focus on Proprietary Data, Not Compute Scale: The value is not in the machine that processes the information; it is in the unique operational data that feeds it. Direct capital toward capturing and structuring your own data rather than subsidizing a silicon manufacturer's margins.
The tech industry loves a super-cycle story because it excuses reckless spending and creates a sense of inevitability. But infrastructure is only valuable if it remains useful over the timeline of its amortization. The current generation of AI hardware is built on shifting sand, designed to be replaced before the warranty expires.
Turn off the hype. Protect your capital. Let your competitors fund the depreciating laboratories of the early AI era while you build agile, efficient systems designed for the lean, optimized architecture that is inevitably coming next.
