The AI Market Split in Two, and the Middle Is Empty

Meta is paying a competitor to do what its own models were supposed to do. The company spends hundreds of millions of dollars a year on Microsoft Azure AI, mostly for coding assistance, while its Llama models have slipped below 1% of the traffic developers route through OpenRouter ^[1]. Google sits in the same canyon: 6.2% of US companies paying for AI, a rounding error next to the leaders ^[2]. Neither is cheap enough to win on price, neither capable enough to win on quality. That is the death zone — the space between two tiers nobody planned for. The two walls of the canyon are now well defined. On one side, Chinese open-weight models: more than 60% of OpenRouter traffic, 12.96 trillion tokens a week, DeepSeek priced at one-twelfth to one-nineteenth of what OpenAI and Anthropic charge for equivalent work ^[3][4]. On the other, Anthropic: 12% of token share but roughly half of platform spending, and 43.5% of US enterprise adoption ^[3][2]. The market has not decamped to China; it has split. Cheap and open on one side, governed and premium on the other, and very little in between. The floor of the cheap side was not set by a pricing decision. It was set by the export controls. When Washington cut Chinese labs off from Nvidia's best chips, DeepSeek stopped optimizing for US silicon entirely and handed Huawei and HiSilicon a head start on its V4 model ^[5]. The result was the Ascend 950, a domestic processor that let DeepSeek cut V4-Pro prices 75% — permanently — and run 12 to 19 times cheaper than GPT-5.5 and Claude Opus 4.7 ^[6]. Zhipu AI is now training frontier-scale models on a gigawatt of Chinese-made silicon ^[7]. The cost frontier is set from Shenzhen now, not Silicon Valley. Jensen Huang said this would happen.

if they can’t buy from us, they’ll build their own — and then we’ll have a competitor we wouldn’t have otherwise created. — Jensen Huang

The US labs' answer has been to stop competing on the model and start competing on the work around it. OpenAI launched a $4 billion deployment venture with 150 forward-deployed engineers; Anthropic put up $1.5 billion for an implementation company; Microsoft is repositioning around reasoning-as-a-service and orchestration ^[8][9]. Each is a retreat from the layer being commoditized to the layer above it — from selling the model to selling the integration. The pressure behind the retreat is arithmetic. OpenAI's inference costs hit $8.65 billion in nine months, more than its revenue ^[10]. Its ad revenue projection is off by 90%, against $75 billion in Oracle GPU obligations ^[11]. DeepSeek trained R1 for $294,000; Sam Altman puts the cost of a frontier model above $100 million ^[12]. You cannot win a price war against a competitor whose training bill is three hundred times smaller than yours. The honest complications don't rescue the US position; they sharpen it. Independent benchmarkers still put Chinese models three to nine months behind on frontier capability ^[13]. Anthropic accuses Chinese labs of distilling its models at industrial scale — meaning part of that price gap is IP extraction, not pure efficiency ^[14]. If the cheap tier is partly built on borrowed capability, the US labs' position is worse, not better: they are being undercut by their own work, priced at a fraction of what it cost to produce. Which leaves the question the retreat cannot answer. Microsoft alone carries $175 billion in capital spending and $329 billion in long-duration lease commitments — half a trillion dollars on one balance sheet, priced against a future where the model layer earned it back ^[9]. The services layer is now being asked to recover that money. Whether consulting, integration, and orchestration can earn back half a trillion dollars when the models beneath them are being priced toward zero is the open question the whole industry is now betting on. The export controls were meant to deny China the tools to compete. They manufactured the chips that set the price floor the US labs are now fleeing upward from.