DeepSeek, Margins, and the Smiling Curve
Frontier AI labs are merely commoditized suppliers sinking into a zero-margin trough between upstream hardware monopolies and downstream aggregators.
By Marcus Vale
Sparked by GLM 5.2 and the coming AI margin collapse · discussion
To say that building frontier artificial intelligence requires an impenetrable, multi-billion-dollar capital moat is the conventional wisdom of the generative era; the actual economics of the AI value chain, though, tell a much starker story. Over the last two years, the prevailing thesis in Silicon Valley has been that scaling laws are the ultimate defense. If you can afford to string together a hundred thousand GPUs in a single datacenter, you possess a structural advantage that no startup or open-source collective can bridge.
And yet, this assumption was violently punctured last month by the revelation that DeepSeek R1 achieved state-of-the-art performance on a verified $5.6 million training run. The immediate geopolitical fallout of a Chinese lab matching the best models from San Francisco is certainly a fascinating storyline. The far more consequential shift, however, is the structural realization quietly brewing among developers—a dynamic highlighted by Martin Alderson's analysis of GLM 5.2 and the ensuing community discourse around an inevitable margin collapse.
The AI Cash Cycle
To understand why the frontier labs are uniquely vulnerable, we have to isolate the fundamental finance primitives of the artificial intelligence market. The creation of a foundation model is an exercise in massive fixed costs. A company must purchase or rent vast clusters of silicon, employ the rarest machine learning talent on the planet, and secure the energy contracts required to process petabytes of data. All of this capital is deployed before a single customer ever interacts with the product.
Once the weights are finalized and the model is deployed, the business shifts entirely to a marginal cost game. This is the cost of inference—generating a response to a specific user prompt. In a healthy software market, the creator uses a proprietary capability to charge a premium on that marginal cost, slowly earning back the initial fixed-cost investment over time. If a foundation model retained a unique, unreplicable capability, this math might work out.
The reality of open-weights proliferation, however, means that developers do not have to pay a premium. Because the underlying architecture is increasingly standardized, drop-in API compatibility has driven switching costs practically to zero. A developer building an application can swap an expensive proprietary model for a nearly free open-source alternative by changing a single line of text in a standardized routing layer. To put it another way, the actual product being sold is a commodity.
The Smiling Curve
This dynamic charts a very specific, and ruthless, physical shape across the market: the Smiling Curve. Originally coined in the early 1990s by Acer founder Stan Shih to describe the personal computer manufacturing industry, the Smiling Curve plots where value is captured in a given supply chain. Picture a U-shaped graph where the y-axis represents profit margin and the x-axis traverses the value chain from upstream components, to middle assembly, to downstream distribution.
In the AI value chain, the upstream edge is anchored by the hardware monopolies. This node is where the absolute scarcity lies. Training runs demand massive parallel processing, and right now, only one company can reliably facilitate that at scale. That is exactly why Nvidia's datacenter segment just posted a record $30.8 billion in quarterly revenue. They are capturing the vast majority of the industry's available margin because every single entity attempting to build a model must pass through them to exist.
As we slide down into the middle of the curve, we hit the model builders. This is the trough. Because inference is rapidly becoming a race to the bottom, the undifferentiated middle is sinking toward a zero-margin reality. You have billions in fixed CapEx required to enter the market, but zero pricing power on the output because the product is instantly undercut by open-source competitors.
The curve then swoops sharply back up at the downstream edge, where the customer relationship is owned. The companies that control the end-user endpoints—the operating systems, the browsers, the massive social networks—capture outsized value because they determine which commoditized model actually reaches a human being.
The Aggregator Audit
This brings us to the core strategic delusion of the current boom: the belief that companies building foundation models are the next great aggregators. To test this hypothesis, we must conduct a strict audit of the frontier labs against the three necessary conditions of Aggregation Theory, as originally defined by Ben Thompson.
The first condition is zero marginal cost of reproduction. Foundation models pass this test easily; duplicating a digital model, or instantiating a new instance of it in memory, costs practically nothing in terms of the underlying intellectual property.
The second condition is zero marginal cost of distribution. Thanks to cloud computing infrastructure, delivering a model's output globally is trivially cheap. The labs pass here as well.
The third condition is where the entire edifice collapses: an aggregator must own demand through superior discovery and user acquisition. The defining characteristic of a true aggregator is that users come to them by default, giving them the leverage to commoditize suppliers.
A strict audit of the model providers yields a fatal result: frontier labs do not own demand; they are merely a hyper-commodified supplier layer. Consumers do not inherently want to query a raw large language model—they want answers embedded seamlessly into the workflows, devices, and applications they already use. ChatGPT was a brilliant user acquisition hack in the early days of the technology, but a web interface is not a durable moat against the incumbents who own the underlying operating systems.
Downstream Leverage
If you want undeniable proof of who actually holds the leverage at the downstream edge of the Smiling Curve, look no further than Cupertino. When Apple announced its integration of ChatGPT into its mobile operating system, the financial arrangement laid bare the fundamental economics of the market. Apple isn't paying OpenAI for the integration.
Contrast this with the search market. Google pays Apple roughly $20 billion a year to be the default search engine on the iPhone. Google can afford to do this because Google is a true aggregator—they monetize that demand intensely through an advertising juggernaut. Apple extracts a massive toll because they control the distribution.
The zero-dollar structure of the OpenAI partnership is the ultimate evidence of who holds the cards in the AI era. Apple owns the end user, which means Apple dictates the terms. OpenAI is effectively giving away its product in exchange for distribution, hoping to up-sell a small fraction of those users to a premium subscription tier. In this transaction, OpenAI is not an aggregator capturing value—they are an outsourced supplier providing a commoditized feature for Apple's platform. Apple captures the real value by making its devices more capable, further cementing its own ecosystem moat and ensuring users never leave iOS to find answers.
The structural reality is clear: leverage has irrevocably moved to the edges of the value chain. Upstream, the silicon designers and fabricators will continue to capture extraordinary margins as long as the models require specialized hardware to train. Downstream, the incumbent aggregators and consumer endpoints will reap the benefits of hyper-commodified intelligence, integrating zero-margin reasoning into their surfaces to enhance their own user experiences.
The entities trapped in the middle are fundamentally mispriced by the narrative. They are spending like natural monopolies while competing like agricultural suppliers. As to exactly when the capital markets will force the undifferentiated middle to reckon with their collapsing margins, well, as always, it depends. Epochs of technological capital deployment can sustain irrational valuations for years before the music stops, but the underlying economics have already made their decision.