The Physics of Solving Durable Bottlenecks

Last month, Cerebras Systems went public, largely on the idea that it had finally solved one of computing’s oldest problems: the bottleneck that forces data to shuttle between memory and the processor — what engineers call the von Neumann constraint. 

By placing memory directly alongside compute on a single piece of silicon, the design sidesteps the round-trip that the rest of the industry has been engineering around since the 1960s. The chip lives up to the claim. Faster inference, simpler systems, and better economics follow from the architecture. 

While I’m not interested in assessing this claim (I’m told the engineering is real), I am intrigued by an older question that the market, in its enthusiasm, may have failed to consider: does solving a constraint and being paid for solving it amount to the same thing? 

In my estimation, it rarely does. The resolution determines whether a bottleneck-solver becomes an inflection or a footnote. 

The Older Question 

Every capital cycle produces companies that solve real constraints and never get properly paid for it. Sony’s Betamax was the technically superior videocassette. VHS had longer recording times and broader licensing. By 1988, Sony was producing VHS players. 

The pattern is not random. It’s driven by the answers to two questions: 

First, can demand find another path to the same answer? When a constraint can be approached from multiple directions, the solver’s pricing power erodes as alternatives mature, even when its approach is technically superior. 

Second, what holds the user in place once they arrive? A bottleneck solution captures durable rent. When choosing it creates ongoing dependencies that make switching expensive: software ecosystems, embedded processes, accumulated developer mindshare, data graphs. Without that, every renewal is a clean evaluation. 

A solution that answers both with “yes” captures the rent. A solution that answers either with “no” is on a constant treadmill. 

The Engineering Is Real. The Structure Isn’t. 

Run the two questions on Cerebras. 

On the first, the same constraint is being attacked from every direction at once: NVIDIA’s roadmap, AMD’s accelerator line, Groq’s inference chips, every hyperscaler’s proprietary silicon.  

The workloads where wafer-scale is uniquely required keep getting narrower. 

Not every user needs the bottleneck removed. They need it managed. NVIDIA’s stack manages the constraint well enough that the developer never has to think about it. For most inference workloads, that is enough. Cerebras sells to the part of the market that needs the bottleneck removed, not the much larger part content to have it managed. 

On the second, the company has no analog to NVIDIA’s CUDA platform: no ecosystem that becomes more valuable to existing users as the user base grows. The narrowness of Cerebras’ customer base is obvious. In 2025, 62% of revenue came from Mohamed bin Zayed University of Artificial Intelligence and 24% from Group 42, two UAE entities the prospectus itself describes as related parties¹. That is 86% of revenue from a single connected counterparty pool. 

The OpenAI Master Relationship Agreement that anchors the diversification story is the most salient answer to the second question. The market is reading it wrong. OpenAI received an equity stake and extended a one-billion-dollar loan to Cerebras, and if the agreement terminates for any reason other than OpenAI’s own uncured breach, OpenAI can call the loan. The same counterparty owns the revenue, the equity, and the credit. 

The Pattern Beneath the Pattern 

This is not an anomaly. As I noted in Beyond the Binary, the NVIDIA-OpenAI financings drew short sellers crying “Enron.”² I was dismissive of the apocalyptic framing then and remain so. The quieter point it obscured is what these structures mean for the second question. 

Customer-equity entanglement manufactures the appearance of switching cost without producing it. A customer who is also an equity holder cannot leave easily. The reason is capital posture, not workflow, data, or developer dependencies. Prior cycles taught the lesson plainly: customer concentration looked like durability right up until the capital cycle turned, at which point the customer relationships evaporated alongside the customers’ own equity. Real product moats survived. Capital-mediated relationships did not. 

The OpenAI structure does not produce a real answer to the second question. It produces a synthetic one. Synthetic answers last only as long as the equity is worth defending. 

The chip solves a bottleneck. The surrounding system does not reward the solving. Sony made the better videocassette. The market kept buying VHS. 

What the Test Sees Next 

J.P. Morgan this week called the AI bottleneck “the most important theme in the stock market” and listed the cohort (memory, photonics, cooling, energy) with the claim that these companies have “market power and technical barriers that are not easily overcome.” This is the framing I am arguing against: find a bottleneck, find the winner. I add one more step. 

Take memory, the lead example and the most aggressively priced. Micron and SK Hynix both crossed trillion-dollar market caps in May on high-bandwidth memory tightness. Demand must flow through three suppliers, which counts in memory’s favor on the first question. But memory has always failed the second question: no customer lock-in, NVIDIA buys from all three interchangeably, full cyclicality when supply catches up. The barriers JPM cites are real, but they are manufacturing barriers, not switching costs. Memory is a real bottleneck. The market is pricing it as a structural one. It is not. 

Power and grid interconnect is the cleanest pass I can identify. The binding constraint of the AI buildout today is not silicon. It is electrons. Interconnect queues at the major operators stretch into the early 2030s. Can demand find another path? Largely no. There is no substitute for electrons at gigawatt scale. What holds the user in place? Twenty-year purchase agreements, sited capital, permits that took half a decade to obtain. Both point the same way. 

Optical interconnect is the more interesting case because the test does not give a clean answer. The bottleneck is real. As compute scales, the speed at which chips can talk to each other becomes the binding constraint, and Jensen Huang has been explicit that NVIDIA needs substantially more photonics capacity than currently exists. 

The past three months show how NVIDIA intends to solve for that: two billion dollars each into multiple players, plus a multi-year partnership with Corning. The structure of each deal is the one I just indicted at the Cerebras-OpenAI layer: equity stakes paired with multi-year purchase commitments. NVIDIA is doing to the photonics layer what OpenAI did to Cerebras. 

The optical players in NVIDIA’s orbit are receiving a synthetic answer to the second question. NVIDIA’s moats are more durable than most because NVIDIA itself is. They are still not real product moats. In a desperate world, dollars flow to whoever has capacity. The constraint will get solved. The economics will be shaped by who NVIDIA keeps in the loop, not by who has the best product. 

Two Questions, Asked Early 

The market will keep surfacing companies with genuine engineering accomplishments solving real constraints. The discipline is distinguishing the ones whose surrounding structure rewards the solving from the ones whose structure does not, including the ones whose answer to the second question is synthetic. 

Two questions. Asked early, before the narrative gets dressed in valuation. Can demand find another path? What holds the user in place once they arrive? Is the answer real, or is it capital? The chip is real. The pitch is real. The structure is what I underwrite. 

Footnotes 

¹ Cerebras Systems Inc., “Form S-1 Registration Statement”, filed with the U.S. Securities and Exchange Commission, April 17, 2026. 
² Mark Scalzo, “Beyond the Binary”, Validex.co, December 5, 2025.