AI Boom or Bust?

Earnings seasons announcements are beginning to sound awfully repetitive.

The hyperscalers are spending tens of billions on AI infrastructure. Data center construction is accelerating. Power demand forecasts are rising. NVIDIA’s (NASDAQ: NVDA) valuation implies demand outstripping supply of GPUs (graphical processing units) for years to come.

Naturally, everyone wants to know if this is sustainable or merely the early stages of a capital expenditure bubble?

Many people make the comparison to the late-1990s telecom build-out. Then, too, infrastructure spending raced ahead of revenue. Fiber was laid at an extraordinary pace. The overshoot was real.

The comparison is defective. The challenge of the telecom bubble was not merely the failure to “monetize” the internet. The larger problem was that too much fiber went unused, not for long, but long enough to bust the companies that laid it. Not for the first time were a flood of bankruptcies the harbingers of huge wealth creation.

Unlike that neglected early fiber, those GPUs from NVDA are not going unused. They are deployed and supporting a huge uptake in AI use. Whether the suppliers of AI are charging enough or reaping enough revenue is a separate question. It will sort itself out.

The better question is not simply whether AI spending is “too high.” It is whether the current capital intensity will be a permanent requirement of artificial intelligence. Or is it rather a function of the current, possibly unsustainable architecture?

Today, frontier AI systems are built around thousands — sometimes tens of thousands — of individual processors networked together across complex switching fabrics. A significant portion of the cost and power consumption is not for computation but coordination: moving data between chips, managing memory across distributed nodes. A huge burden is “synchronizing gradients” so that adjustments one GPU makes to the model are coordinated across all the processors in the fabric.

The modern AI data center is not just a room full of processors. As Danny Kaye sang in White Christmas, “they’re doing choreography?”

Choreography is expensive.

If webs of GPU chips continue to define AI infrastructure, then high capital intensity could be forever—or until AI providers go broke.

History suggests differently. Progress in computation has always been driven by cutting costs. Indeed, calculating by operations per second, GPUs and other processors optimized for AI are orders of magnitude cheaper than their predecessors. For AI to fulfill its promise for eight billion humans, however, it must get cheaper still.

It will. As my partner George Gilder says, we are about to enter “Life After Chips.”

Cerebras Systems, rather than connecting thousands of separate GPUs across racks and data center rows, builds wafer-scale processors, entire silicon wafers as a single computational device.

Cerebras integrates hundreds of thousands of cores onto one wafer-scale engine, with high-bandwidth on-chip memory and extremely low-latency communication across the entire surface. What the dominant GPU model achieves by networking thousands of discrete chips, Cerebras attempts within a single piece of silicon.

Cerebras overcomes the central challenge of current AI architecture. In large GPU clusters, interconnect bandwidth becomes a bottleneck. Synchronization overhead grows. Power consumption rises as more hardware is added. Capital intensity increases non-linearly as each additional processor increases complexity and the challenge of coordination.

By collapsing many of those communication layers into a single device, wafer-scale integration will reduce networking overhead, simplify topology, and improve training efficiency per watt and per dollar.

That will not automatically make Cerebras dominant or doom NVDA. The incumbent GPU ecosystem is vast, mature, and deeply embedded in software stacks. CUDA is NVIDIA’s software layer enabling programmers to harness its GPUs for AI. Over time, it has become the industry’s default language for high-performance machine learning. Whole libraries of software and programing tools depend on it.

Nevertheless, if alternative architectures can deliver comparable training performance with fewer components, simpler interconnect, consuming vastly less power the slope of capital intensity will moderate.

That, however, will mean not less spending but more. Cutting costs will not terminate or even slow the AI boom, it will extend and accelerate it.

Computation is the supreme example of Jevon’s Paradox. When a process becomes more efficient, its scope does not contract but increases. Enhanced efficiency makes power, or steel, or cotton goods cheaper and more is consumed.

The era of the microchip extended the global population of computers from dozens to billions.  When storage costs fell, data volumes exploded. When bandwidth prices declined, streaming and cloud applications proliferated.

As AI training becomes more efficient per dollar, new, previously uneconomic applications will become commonplace. More models will be trained more frequently, customized more narrowly, deployed in more domains.

You may build your own on your desktop (laptop?) wafer.

For investors, the right question is not whether hyperscalers are overspending, but what assumptions about architectural permanence are embedded in current valuations?

If investors assume that today’s distributed GPU model will live forever, then valuations tied to relentless capital intensity may appear justified.

History undercuts that assumption.

As for NVIDIA, its current valuation reflects not merely strong current demand, but the assumption that large-scale, distributed GPU clusters will remain the dominant architecture for frontier AI. In effect, the market is capitalizing years of continued hyperscale build-out, sustained networking complexity, and persistent GPU scarcity.

NVIDIA’s position is formidable. But as system design shifts via wafer-scale integration, new memory hierarchies, and more efficient training paradigms, then super-expensive, high-margin GPU fabrics will no longer dominate.

At Gilder’s Technology Report, we focus less on how much capital is being deployed today, and more on how new paradigms reshape markets. Wafer scale promises to move mountains.