The dream of generative AI as a high-margin software business is officially dead. For the past three years, venture capitalists and tech executives operated under the assumption that proprietary foundation models would command the same lucrative 80% to 90% gross margins that fueled the SaaS boom of the 2010s. The release of GLM 5.2, alongside a wave of open-weights models that perform at frontier levels for a fraction of the operational cost, has shattered this illusion. We are witnessing the rapid transition of artificial intelligence from a premium software product into a highly commoditized public utility.
This shift is not a temporary market correction. It is a structural realignment of the technology sector's economics. When intelligence becomes cheap enough to meter like water or electricity, the value shifts away from the raw capability toward distribution, integration, and physical infrastructure. The companies currently spending billions to train slightly better models are running directly into a historical trap.
The Iron Law of Utility Economics
What is happening to artificial intelligence today has happened to every major industrial input over the last two centuries. In the late 1880s, early electric utilities charged premium rates for incandescent lighting, treating electricity as a luxury service. Once Westinghouse and General Electric standardized alternating current, the cost of generating power plummeted, and electricity became a low-margin commodity. The fortune was not made by selling the raw electrons, but by manufacturing the appliances that consumed them.
We saw the same script play out during the fiber-optic boom of the late 1990s. Telecommunications companies laid millions of miles of glass cables, expecting to charge premium rates for data transmission. Instead, overcapacity triggered a brutal price war, driving the cost of bandwidth down by over 90% in some corridors. The massive capital expenditure did not yield high-margin monopolies; it created the cheap, commoditized foundation upon which Google, Netflix, and Amazon built their empires.
- Phase 1: Scarcity. Early innovators charge monopolistic premiums because they hold the only working technology.
- Phase 2: Standardisation. Competitors replicate the technology, open-source alternatives emerge, and proprietary advantages evaporate.
- Phase 3: Overcapacity. Capital floods the market, supply outstrips demand, and prices crash to the marginal cost of production.
- Phase 4: Utility Consolidation. Only the lowest-cost operators survive, running on thin margins backed by massive volume.
The API Price Race to the Bottom
To understand the speed of this collapse, one only needs to look at the pricing metrics of API calls over the last eighteen months. In early 2023, querying a state-of-the-art LLM cost roughly $0.03 per thousand tokens. Today, comparable or superior performance from models like GLM 5.2 can be acquired for less than $0.0005 per thousand tokens. This represents a pricing collapse of over 98% in less than two years.
Photo by panumas nikhomkhai on Pexels
This pricing pressure is driven by three converging forces. First, algorithmic efficiency has dramatically reduced the compute required to achieve a given level of intelligence. Second, hardware optimization has increased throughput per watt. Third, and most importantly, the proliferation of high-quality open-weights models has removed pricing power from proprietary developers. If a closed-source provider tries to maintain high margins, customers simply migrate to self-hosted open models.
This reality leaves foundation model companies in a precarious position. They are spending hundreds of millions of dollars to train models that will face immediate, aggressive price competition the moment they launch. The capital expenditure required to stay at the frontier is escalating exponentially, while the revenue potential per query is collapsing toward the marginal cost of the electricity required to run the inference.
The Illusion of the Proprietary Moat
Many investors still cling to the belief that proprietary datasets and specialized tuning will protect these high margins. This is a misunderstanding of how enterprise buyers operate. A bank or a healthcare provider does not need a trillion-parameter general intelligence to automate its back-office paperwork. They need a reliable, cheap, and secure system that performs a narrow task.
Once a model reaches a threshold of 95% accuracy on a specific task, further improvements are irrelevant to the buyer. If Model A costs $10.00 per million tokens and achieves 96% accuracy, while Model B costs $0.10 per million tokens and achieves 95% accuracy, the market will overwhelmingly choose Model B. The margin is devoured by the cheaper, optimized utility.
Furthermore, the cost of switching between models has dropped to near zero. Standardization of API formats means that an enterprise can swap its underlying model provider with a single line of code. In a market with zero switching costs and standardized outputs, price is the only variable that matters. This is the definition of a commodity market.
What This Actually Means
The implications of this margin collapse will reshape the technology landscape over the next decade. First, the venture capital model of funding foundation model startups is broken. Startups cannot survive a low-margin utility war against hyper-scalers who own their own datacenters and energy pipelines. We will likely see a wave of consolidation, where independent model developers are absorbed by cloud providers who view LLMs as loss-leaders to sell cloud storage and compute.
Second, the real value of the AI boom will accrue to the application layer and the physical infrastructure. The winners will not be the companies building the models, but the enterprises that integrate these cheap cognitive cycles into proprietary workflows, and the physical infrastructure providers who supply the land, fiber, and gigawatts of power required to keep the datacenters running.
Ultimately, cheap intelligence is a massive net positive for global productivity. Just as cheap electricity transformed manufacturing and cheap bandwidth transformed communication, cheap compute will transform knowledge work. But for the technology sector itself, the transition will be painful. The era of easy software margins is giving way to the hard reality of industrial utility economics.
Quick Answers
Why are AI margins collapsing so quickly compared to traditional software?
Traditional software scales with near-zero marginal cost, allowing high margins. AI requires physical compute and electricity for every single query, meaning it has a real marginal cost that cannot be coded away, turning it into a utility.
Will proprietary models always maintain a performance advantage over open-source?
While proprietary models may hold a temporary lead at the absolute frontier, the gap is closing rapidly. For 90% of commercial use cases, open-weights models like GLM 5.2 offer equivalent performance at a fraction of the cost.
Who benefits most from this commoditization?
Enterprises and application developers benefit enormously. As the cost of the underlying intelligence drops to near zero, they can deploy complex AI systems at scale without incurring unsustainable API bills.
