AI Value Shifts From Models to Undervalued Raw Compute Infrastructure
By Nikhil Raghavan · Reporting from San Francisco ·
Open source adoption proves that accelerating compute demand is driving value into the foundational hardware layer.
The idea that a token—a simple unit of computation—requires the same amount of physical resources whether it originates from an expensive proprietary model or a cheap open-source alternative is perhaps the most profound and overlooked insight from this week’s episode of "Invest Like the Best." It challenges the core economic assumption that premium models justify their cost through inherent technological superiority.
On the podcast, the discussion painted a picture of AI development fueled by accelerating quantitative metrics: GPU availability, DRAM spot prices, and token growth are all moving up without deceleration. The central thesis was bullish: despite market pessimism (with AI names down 40-60% in a month), these selloffs are overreactions because the underlying fundamentals for compute are rapidly improving. Several key arguments emerged regarding infrastructure dynamics. For instance, one speaker pointed out that nobody expected old GPU prices to go vertical in 2026; consequently, contracted compute capacity is being acquired at a "massive discount to the current spot market." Furthermore, the argument was made that open source taking market share isn't negative—it merely shifts margin dollars from the proprietary "frontier model layer" straight into the underlying "AI infrastructure layer," thereby driving demand for raw compute.
The Infrastructure Layer is Undervalued
The narrative surrounding AI investment often fixates on the flashy, high-margin front end: the frontier models developed by OpenAI or Anthropic. However, the discussion provided a crucial counterweight by emphasizing that open source acts as "dark matter to the public markets"—a force driving clear and accelerating demand that traditional metrics struggle to measure. This shift fundamentally changes the economic calculus of AI adoption.
The implication for policy is massive: if the value proposition shifts from access to proprietary APIs toward raw, affordable compute power, then regulatory focus must pivot away from content governance and towards ensuring equitable access to physical infrastructure. The fact that a startup can rent thousands of B200 GPUs at mid-$2/hour and expect to pay under $4 in seven months is not just an investment signal; it’s a structural warning about the speed of compute deflation, regardless of who owns the models running on it.
Rethinking Compute Ownership and Value
The conversation also provided some fascinating insights into business model innovation. The suggestion that Nvidia could generate royalties by implementing a credit wrapper—requiring upfront cash for a cut of ongoing revenues—is a powerful strategic tool. It moves the industry toward an extension of Long-Term Agreements (LTAs), where companies trade short-term upside for guaranteed durability and recurring revenue streams.
However, this focus on pure capital flow overlooks the policy challenge inherent in such concentration. While the argument that accelerating operating cash flows for hyperscalers (Microsoft, Meta, Amazon) are improving is materially reassuring, it also highlights a dangerous trend: increasing reliance on private agreements and pre-purchasing capacity. This makes the entire ecosystem highly susceptible to single points of failure—be they geopolitical disruptions or sudden changes in regulatory oversight concerning compute resource allocation.
The overall picture painted by the speakers is one of relentless technological acceleration, driven by fundamentals that are improving faster than public sentiment acknowledges. While the sheer technical momentum is undeniable—and the potential for blue-collar benefits (plumbers and HVAC contractors) through data center construction is a positive policy point often overlooked—the current discussion remains heavily concentrated on capital flow mechanics rather than systemic risk management.
The market's focus, therefore, must be less on whether AI adoption is happening and more on who gets to dictate the terms of compute access. As the industry becomes defined by increasingly complex orchestration layers (the "160 IQ model"), the true bottleneck shifts from technological capability to regulatory friction and equitable distribution of power.