An Earthquake Is Coming
By John Schroeter and George Gilder
Artificial intelligence is creating huge opportunities for investors. And some of the most powerful opportunities have nothing to do with microchips or accelerators or silicon or semiconductors or any of that stuff. The opportunities are in infrastructure, real estate, and heavy industry. In the world of computation, a huge disruption—a seismic event—is happening.
The seismic map of investment opportunities has four stages. The first are the fault lines, the tensions below the surface. In geology, fault lines form when tectonic plates meet and push against each other with tremendous force. For years or decades there may be no readily observable sign of the risk. People walking on the San Andreas Fault may have no idea that they’re standing above it because nothing appears to be happening.
Something similar happens in the economics of technology. Underlying tensions build up. They may be invisible to casual observers, but they’re creating all the conditions for a disruption.
The second phase is the pre-earthquake tremors, warning that something is about to happen. Then there is the quake itself, which can be massively destructive to entire industries or companies. Finally, there are the aftershocks, the follow-up consequences of the quake. The aftershocks help settle the ground to an apparent equilibrium but also set up new fault lines.
We are just now in the early stages of a huge disruption in AI. Yet few investors seem to have noticed. We read tons of investment advisories and newsletters, we go to conferences, and we have not heard anyone other than ourselves talk about the fault lines and tremors that we’re already beginning to feel.
In business, fault lines are generally caused by the drastically rising cost of some product or service that seems indispensable. In the entire history of the microchip industry from the 50s until recently, we saw just the reverse: chips got smaller and computation got cheaper. The industry progressed from a handful of transistors on a chip to billions. And at every stage the result was that we could do more computing for less cost. That beautiful evolution has been stopped. What’s been created is a fault line of tremendous pressure between the imperative to do more computation for AI, and the suddenly rising price of computation. That rising price comes not only from the chips themselves but the infrastructure supporting them and the relentless increase in energy they demand and heat they generate.
Most current data centers were originally designed for traditional pre-AI computing loads and built to run on public electricity grids. These data centers now strain under the weight of these massive AI models that they were never built to support.
That’s a fault line, a conflict between a “must have” and its savagely rising price. Something’s got to give.
As the pressure increases, we begin to feel the tremors, the first hint of huge disruptions and profits—and losses—to come. Consider the heat generated by a data center packed with Nvidia GPUs. As power consumed by a “rack” of servers pushes past 30-50 kilowatts, air cooling, once the mainstay, is becoming irrelevant. Nvidia will not allow its hardware to be sold to an air-cooled data center. Coming on strong are new liquid cooling companies and their innovative technologies. Many of the most innovative are still privately held. We have covered one we particularly like in Gilder’s Private Reserve. Others we cover in the Gilder Technology Report. You need to know about all of them.
Another tremor reverberates from companies building custom-built data centers designed to handle orders of magnitude more power than legacy centers and increasingly located for the availability of electric power. Even these increasingly refuse to rely solely on the public grid. Companies building and powering local grids for backup are growing fast, and we cover them in the GTR.
All these developments are part of what we call the Great AI Inversion. For the first time since the invention of the microchip, value is flowing downward to the infrastructure, not upward to the software or the apps. That is an enormous and consequential change. And it is inherently unstable. The micro electronics industry will stall unless the prior condition of more computation for fewer dollars is restored.
The earthquake that will bring that about is wafer scale electronics.
Peter Drucker used to say it took 10x advantage to replace an existing entrenched technology. Wafer scale is at least a thousand-to-one improvement.
We already have the proof in vendor benchmarks for very large AI models. A wafer scale system has achieved inference performance 10 times faster than the top-leading Nvidia GPUs.
In wafer scale the physics of computing flips completely, or rather, flips back to its natural state. Instead of paying the energy tax and the latency tax to move data across boards and through back planes and over networks, compute and memory now live in one continuous low-latency fabric. Currently, most of the energy used in computing is spent not in computation but in moving data back and forth between the chips and the boards and the racks. Wafer scale eliminates most of that to-and-fro, consumes dramatically less power, and produces dramatically improved performance. It will invert the inversion.
Data centers that once occupied millions of square feet will fit in one small air-conditioned office. That’s the quake that is coming.
In the aftershocks, software stacks will be rewritten to exploit wafer scale. Entire supply chains will have to reorient around wafer-first packaging. New methods of power delivery and cooling will come on. Even the economics of the cloud will invert because workloads that once required centralized hyperscale infrastructure will run locally, in an office or in your car, or your private jet. The latency collapses because data no longer needs to navigate a maze of discrete components.
Crippling operational complexities get simplified because thousands of potential failure points are collapsed into a single unified system. And wafer scale computation has built-in resilience: if a part of the circuitry fails, it finds a way around that failure.
The whole phenomenon adheres to George Gilder’s law of the microcosm: computing power always migrates to smaller, more efficient, and decentralized forms, from room size mainframes to racks, then to blades, and now to the wafer itself.
Good earthquake, I say.