
Infinite Intelligence, Chip Shortage: An Observational Analysis of Supply Mismatch in Compute Pricing
Author: Kerman Kohli
Compiled by: TechFlow
TechFlow Guide: While everyone else is fixated on falling token prices and shouting bubbles, Kerman Kohli offers a completely opposing framework: infinite demand, constrained supply, and hardware turning into a strategic asset. This article helps you understand why used graphics cards are appreciating, why compute power will not depreciate, and why "owning your own compute capacity" is shifting from a technical question to an existential necessity.

I haven't written much for Substack in over a month, primarily because my learning curve has been so steep that I can't find the time to put pen to paper. This post serves as a digest of what I've recently learned and outlines some new mental models taking shape in my mind. I'll roughly divide my insights into three parts:
The first part covers hardware dynamics and why I believe most people's understanding of hardware is flawed. The second part discusses how models are evolving and what that means for the future. The final part examines the relationship between local compute and frontier compute.
Before diving in, I want to share some updates on my own hardware explorations, as they directly relate to the theme of this article.
Earlier this year, I wrote about building an inference machine for $15,000. I recently upgraded it by adding another GPU, bringing the total investment to $35,000. Interestingly, the GPU I bought for $10,000 at the start of the year now costs $14,000, and memory prices have also jumped by $1,000. In other words, while the build cost $35,000 in today's dollars, my actual baseline cost sits at $30,000. This kind of scenario is highly abnormal for most hardware.
Additionally, I spent another $35,000 on six DGX Spark units, paired with a 16-port 100Gb switch, requiring fiber optic cables to interconnect them all. Going through this process deepened my understanding of GPUs far beyond anything available online. Actually buying hardware and tinkering with it is truly the only correct way to enter this space.
This year, my personal CapEx on AI hardware has grown exponentially, and we at RouteMesh are advancing along the same trajectory; a GPU cluster is the next logical step for scaling. As time goes on, I'll share more insights from our data center explorations.
i) Hardware Dynamics
How can a machine less than a year old actually appreciate in value? Even crazier, a used RTX 3090 (a 2020 GPU) can still sell for its original price in 2026? You can't just dismiss this as a "bubble" and move on; we need to dig deeper.
It all starts with one thing: Memory! And of course, Nvidia's silicon. I will elaborate on this section. When we talk about hardware, we are really talking about two things: Compute and Memory.
Compute does get stronger over time, with new generations utterly obliterating older ones. But this only matters when compute is the bottleneck (by compute, I mean CPU/GPU/XPU). In inference scenarios, however, the bottleneck is memory. Given that memory will be structurally short for at least the next decade (yes, I've upgraded my previous five-year forecast to ten), any depreciation in compute power will be offset by the appreciation in memory value.
This is the first trend. But the second factor, significantly undervalued by the market, is that Nvidia chips derive much of their worth from backward compatibility. Older chips remain useful, backed by a vast ecosystem capable of running on them. A 2020 RTX 3090 or A100 can still execute economically viable inference tasks today, providing every reason to keep them running in data centers!

This is a massive narrative, yet very few pay real attention to it. You cannot short hardware, emerging cloud providers, compute, or memory. The dynamics of hardware have permanently changed.
I've also observed that certain workloads will never require upgrading to a newer generation of chips. Workloads seem to be starting to bind tightly to specific hardware, which hasn't been seen in tech for a very long time.
I am experiencing this firsthand myself. The RTX 6000 Pro Max-Q I bought for $10,000 is now worth $16,000, sometimes even priced at $18,000! This is a trend, not just a short-term supply and demand mismatch.
ii) Model Evolution
Recently, many have pointed out that token prices are dropping and believe this will cause the "bubble" to burst. This is equally misguided, as they only see the price side of the equation, missing the other: demand.
Below is a chart from Silicon Data showing token prices have been continuously falling since June. By that logic, both AI pricing and demand appear to be slowing.

Since it's difficult to find a unified dataset for token demand, I decided to use Nvidia's data center revenue as a proxy. If lower token prices were sufficient to meet demand, people should theoretically need fewer chips. But that is exactly the opposite of what we're seeing...
As shown in the chart, growth is consistently trending upward, with the only constraint being: Supply.

One thing people fail to grasp about AI and this entire construction wave is that demand is essentially infinite (as I wrote in my last piece), while supply is the binding constraint. This is hard to comprehend intuitively because traditional economics teaches us that supply expands to meet demand. But the demand for intelligence is infinite, and application boundaries are limitless, meaning we will never truly satisfy supply. I know this is a tough pill to swallow, and I recognize it implies a radically different future, but you must internalize this reality, or you will find yourself in a severely disadvantaged position down the line.
Having your own compute is crucial because it is becoming a strategic resource. Those who do not own compute will have their ideas stolen (we have already seen the recent case where the Navier-Stokes problem was solved, earning a $1 million prize).

Everything is unfolding according to plan.
However, the decline in token prices brings another fascinating trend: we can embed more intelligence into everything, unlocking even more workflows. I currently run a vast array of personal software managing my life. With each model advancement, I can embed greater intelligence, which in turn requires more intelligence to coordinate it all. The recursive and explosive nature of agent demand still leaves me in awe.
This circles back to the core question from my last post: Do you believe the demand for intelligence is finite or infinite? Hint: It's infinite.
iii) Local Compute vs Frontier Compute
I've seen a lot of takes on this topic, most of which I believe are completely wrong. Let's start with the facts: Over the past month, incredible breakthroughs have emerged in the realm of local intelligence.
Qwen 3.8 DFlash2 and GLM 5.3 Flash Next DFlash2 have brought transformative leaps in the orchestration speed and intelligence levels achievable on your own hardware. My dual-RTX setup and 6-node Spark cluster have brought me nearly to compute self-sufficiency. I still maintain a frontier model subscription ($100/month for ChatGPT), but I only use it for iOS-specific tasks, computer operations, and some genuinely difficult ad-hoc problems. I gladly pay a heavy premium for that access.
I believe most people will eventually arrive at a similar configuration (assuming they are truly willing to invest in hardware). So does this signal the end of frontier labs?
Quite the opposite. People frequently forget two key traits of frontier labs and frontier models:
The price you pay for cutting-edge intelligence isn't just 10% higher; it could be 10 times higher. You wouldn't just pay top talent 10% more; you are willing to pay multiples for them. The exact same logic applies to intelligence.
The labs don't care if you can download open-source models, because running them still requires compute! Guess who controls vast amounts of compute: They do!
These two factors are the only ones that matter; everything else is noise. Furthermore, they will carry the "American brand" label, granting them a premium in the enterprise market.
OpenAI and Anthropic possess massive compute reserves, and no amount of debate changes that unless you own your own compute (which most people do not). Hardware prices will continue to climb. Every time you evaluate the ROI of buying hardware, you'd rather say "$200 a month is nothing compared to a $20,000 capital expenditure." As programming subscription services wind down (the $200/month ChatGPT tier has already been paused) and hardware becomes even more absurdly priced and scarce, this reality will bite back at people.
Long Compute
If there are two solid anchors in my belief system, they are these two:
This is the worst state these models will ever reach. It only gets better from here. Optimism is paramount.
Price declines drive usage increases. Demand is infinite, and the world is a positive-sum game.
I believe the latter point is the hardest for people to grasp. Infinite demand shatters numerous assumptions, forcing us to view how markets operate, competitive landscapes shift, and human collectives organize from a completely different perspective.
If you are not long compute, you will be left behind. You can go long compute in various ways, including career choices, portfolio allocations, or hardware ownership (perhaps all of the above). But the one thing you absolutely cannot do is stay short compute.
Long Compute.













