The Physical Constraints of the AI Era: A Systems Perspective

The next phase of AI will not be defined by software breakthroughs alone, but by a collision with physical reality: the scarcity of electricity and the limits of wafer fabrication. While the market oscillates between euphoria and panic, the structural reality is that we are in the middle of the largest industrial build-out in the history of capitalism. This conversation shows that the AI trade is not a monolith. It is a complex, multi-layered system where delayed payoffs, supply-chain bottlenecks, and geopolitical shifts create competitive advantages for those who look past the immediate noise. Investors and operators who fail to map these downstream dependencies, from the cooling requirements of orbital compute to the disaggregation of inference, risk being overwhelmed by a fast-moving transition they are not prepared to navigate.

The Hidden Dynamics of the Watts and Wafers Bottleneck

The immediate instinct when analyzing AI infrastructure is to focus on software models. However, Gavin Baker argues that the true gating factors are physical: energy (watts) and chip production (wafers). The system is currently routing around these constraints in ways that create long-term moats for those willing to endure short-term friction.

Capitalism is efficient at solving industrial shortages, but it requires patience. The current energy crunch should ease by 2027 or 2028 as infrastructure projects come online. Beyond that, the long-term solution, orbital compute, is often misunderstood. Skeptics view data centers in space as giant, impractical structures, but the reality is more elegant: racks in space, connected by lasers, forming a virtual data center. This is not science fiction; it is an extension of the existing satellite infrastructure that SpaceX already operates at scale.

"The reality is 10 years later no other company is consistently capable of landing and fully reusing an orbital rocket none of this makes sense without reusability that means you have to land it."

-- Gavin Baker

The downstream effect of this infrastructure build-out is a change in the useful life of compute. The disaggregation of pre-fill and inference means that older, less capable GPUs like Ampere or Hopper can be repurposed for pre-fill tasks, extending their usable life from two years to over a decade. This shift creates an under-appreciated advantage for private credit and infrastructure financing, turning what was once considered technical debt into durable, revenue-generating assets.

Why the Frontier is the Only Place to Be

Conventional wisdom suggests that AI models will quickly become commodities, with value migrating to the application layer. Baker’s analysis suggests the opposite: economic returns are currently accruing almost exclusively at the frontier.

This creates a new prisoner’s dilemma. If frontier labs keep their best models behind private APIs, they maintain an edge. If one lab defects and releases an open-source version, the entire system shifts. This explains why the perito frontier, the balance between intelligence and cost, is so volatile. Google’s loss of leadership in this space, driven by conservative design choices, allowed OpenAI, Anthropic, and xAI to dominate.

"The closer someone is to AI the more skeptical they are that this will occur [the violation of the bitter lesson]. The reason I am a little less skeptical is I think we are very close to ASI and who knows if the bitter lesson holds for 400 iq models."

-- Gavin Baker

The Bitter Lesson, the idea that more compute and data will always outperform human algorithmic ingenuity, remains the most critical variable. If an Artificial Superintelligence (ASI) emerges, it may find ways to optimize itself that temporarily violate this lesson, creating a disruption in the demand for traditional silicon.

The Competitive Advantage of Hard Trade-offs

In the semiconductor space, many companies are emerging, but most will fail. The system responds to new chip companies with a predictable pattern: if a startup builds something that is different but not hard, the hyperscalers like Nvidia or Google will simply replicate it.

True competitive advantage in this sector requires making difficult, non-obvious trade-offs in the iron triangle of chip design: attack, defense, and mobility. Cerebras is a prime example. By pursuing wafer-scale computing, they made an architectural decision that was genuinely hard to replicate. This creates a moat that prevents the hyperscalers from simply absorbing their innovation. For founders and investors, the lesson is clear: if your product is obvious and easy to build, you will be overwhelmed by the hyperscalers the moment you reach scale.

Key Action Items

• Audit Your Compute Dependency (Immediate): If your business relies on frontier models, transition to usage-based enterprise plans immediately. The lobotomized versions of models provided in flat-rate plans are insufficient for production-grade reasoning.
• Re-evaluate Hardware Useful Life (Next Quarter): Stop assuming a 2-year depreciation cycle for GPUs. With the disaggregation of inference, plan for 10-15 year hardware lifecycles, which changes the ROI of your infrastructure investments.
• Focus on Hard Differentiation (12-18 Months): If you are building in the AI stack, ensure your technical moat is based on hard trade-offs that are not easily replicable by hyperscalers. If your advantage is just a software wrapper, you are not in the token path and are at high risk of displacement.
• Implement Safe Word Security (Immediate): As AI-driven social engineering becomes indistinguishable from reality, move beyond standard MFA. Establish non-digital, out-of-band verification protocols for all high-stakes financial and operational decisions.
• Monitor the Perito Frontier (Ongoing): Track the intelligence-to-cost ratio of frontier models. If Google or others regain leadership, expect a rapid shift in the competitive landscape of the model layer.
• Shift to PSU-Based Compensation (Next 6 Months): For your own firm or portfolio companies, move away from standard RSUs. Align incentives with performance-based stock units (PSUs) that reward actual economic output rather than just time-in-seat.