Moonshot's Capacity, the AI Meter, and Amending Aggregation Theory
Generative AI's non-zero marginal costs break Aggregation Theory, shifting value capture from consumer platforms to metered infrastructure providers.
By Marcus Vale
Sparked by Moonshot AI suspends new subscriptions due to Kimi K3 demand · discussion
Last week, Chinese AI startup Moonshot issued a rather remarkable statement explaining why its Kimi chatbot had unexpectedly crashed: "Moonshot AI is experiencing an abnormal spike in traffic, prompting it to expand its computing resources."
The conventional read of this event is entirely optimistic. A young company releases a product so compelling that it simply cannot keep up with user demand; the servers melt down, the engineers scramble, and the venture capitalists applaud. We have spent two decades watching this exact sequence play out across the consumer internet, treating downtime as the ultimate proof of product-market fit. That, though, is the rub: this downtime represents a fundamental structural limit.
To understand why Moonshot's outage is a harbinger rather than a growing pain, we must walk through the actual economics of the modern web. For the entirety of the mobile and cloud era, the defining characteristic of consumer technology has been the negligible cost of serving the marginal user. When a billionth person queries Google or loads a Facebook feed, the physical cost of retrieving and transmitting that data is fractions of a cent. The entire advertising and software-as-a-service paradigm rests on this exact foundation of near-infinite, virtually free scalability.
Generative AI operates on a completely different set of physical laws. Every single inference—every prompt typed into Kimi, every image requested from Midjourney, every document summarized by a copilot—requires a forward pass through a massive neural network. This process burns expensive GPU cycles and draws immense amounts of power. The cash conversion cycle is radically inverted. A traditional web platform builds a data center once, amortizes the capital expenditure over five years, and captures near-100% gross margins on incremental usage. An AI provider buys the exact same servers, but instead of enjoying free marginal distribution, it pays a direct, variable toll in compute for every single word generated. Your business model is your destiny, and an architecture that bleeds cash with every marginal user cannot be subsidized indefinitely.
This structural reality forces us to revisit the foundational thesis of the previous era. As Ben Thompson wrote in 2015's Aggregation Theory, the internet shifted the balance of power by altering how value chains are organized.
A subtext to Thompson's article was a pivotal framing question: what has a zero marginal cost?
In the pre-internet epoch, the answer was nothing. Both supply and distribution were constrained by physical scarcity; newspapers had to buy physical delivery trucks, and television networks had to secure finite broadcast spectrum. In the Web 2.0 era, the answer became distribution. Because it cost essentially nothing to send a digital good across the world, companies that aggregated consumer demand could attract suppliers on a massive scale, which improved the user experience, which in turn attracted more consumers in a virtuous cycle.
The Aggregation Theory value chain operates as a straightforward feedback loop. You have three nodes: suppliers, distributors, and consumers. Because distribution costs nothing, the most effective distributors capture the market by delivering a frictionless interface, locking in massive consumer demand that organically pulls in every relevant supplier, spinning up an inescapable flywheel. The suppliers are subsequently commoditized, forced to interface with the aggregator on the aggregator's terms. Value flows unimpeded away from the edges and pools entirely in the center, captured by the entity that owns the demand.
The technology industry is currently attempting to graft that exact demand-aggregation model onto an AI paradigm where those physics are fundamentally broken.
Consider the history of early power generation. When early power companies sold electricity at a flat rate per lamp, the model worked smoothly for a brief period. Utilities calculated a rough average of how much a typical residential household would burn a single bulb in a month, set a recurring subscription fee, and priced accordingly. The system entirely collapsed when industrial users entered the picture. A factory owner realizing they could leave dozens of lamps burning twenty-four hours a day mathematically bankrupted the supplier. The utility was absorbing a distinctly non-zero marginal cost for the coal, generation, and transmission, while collecting a capped, flat-rate revenue from the user.
The unit economics were structurally upside down.
This is precisely the predicament facing generative AI companies today. They are selling flat-rate monthly subscriptions to a consumer base using an infrastructure that demands brutal, variable, non-zero marginal costs.
We must explicitly amend Aggregation Theory to account for this reality. The original framework assumes that gaining a new consumer passively improves the system at zero cost. In the Amended AI Value Chain, however, we must introduce a fourth, load-bearing node: compute infrastructure. When the cost of serving the marginal user is non-zero and tightly constrained by physical GPU availability, the aggregator's flywheel breaks down. A new heavy user does not improve the ecosystem; they actively degrade the system's capacity for everyone else, forcing the company to throttle speeds or, as Moonshot's engineers just experienced, crash entirely.
In this amended model, the value no longer flows inexorably to the consumer aggregator. Instead, the margin is captured by the infrastructure layer that controls the scarce, expensive compute. If you do not own the silicon and the data centers, you are effectively a reseller of someone else's margin, exposed to the same existential risk as a 19th-century utility selling unlimited flat-rate coal power.
How did those early electric utilities survive? They abandoned the flat rate. The grid was saved by the transition to Edison's chemical meter, which measured the precise amount of current consumed by a specific building and billed the user accordingly. Consumption was paired directly with production, forcing efficiency and aligning the incentives of the buyer and the seller.
To put it another way, the flat-rate consumer AI subscription is a ZIRP-era anomaly; the meter is coming, and it will permanently alter where value accrues in the tech stack.
The inevitable endpoint of an ecosystem burdened by non-zero marginal costs is metered consumption. We are already seeing the precursors of this shift in the developer ecosystem, where API access is priced strictly per thousand tokens. The consumer layer is currently shielded from this reality by billions of dollars in venture capital, which is subsidizing the illusion of zero marginal costs in a desperate bid to build a classic demand monopoly. But as agentic AI systems arrive—bots that will loop autonomously, browsing the web, testing code, and burning inference cycles without human intervention—a flat-rate subscription will become mathematically suicidal.
Ultimately, the brute physics of silicon and electricity dictate the transition from aggregation to infrastructure dominance. My expectation is that infrastructure providers will inevitably impose metered consumption on the entire ecosystem. This transition may take a decade or longer as well-funded startups continue to subsidize flat-rate consumer compute, but the underlying structural direction is clear. When exactly the venture capital runs out, though, is a different question entirely: as always, it depends.