Frontier American models charge tens of dollars per million tokens, while China’s Kimi K3 is priced near the cost of compute
| Photo Credit:
Go Nakamura
In 1884, aluminium was used to cap the Washington Monument, like a jewel, because it was worth more than silver. That expensive, it could decorate a monument but not build an aeroplane. Powered flight wasn’t waiting on an idea; engineers had sketched it for decades. It was waiting on a metal cheap enough to fly, which arrived only after 1895, when hydroelectric power at Niagara Falls crashed the cost of refining it. Within 20 years, the metal that once graced an emperor’s table was wrapped around sandwiches and riveted into airframes. The Wright brothers didn’t invent aviation. Cheap aluminium did.That pattern, expensive novelty first, then a price collapse, then a new economy nobody could picture beforehand, repeats with tedious regularity: electricity, priced like jewellery in 1882, took decades to get cheap enough to produce the assembly line; a transistor cost eight dollars in 1958, and a dollar now buys tens of billions, which is how you get the smartphone economy rather than a cheaper mainframe. Artificial intelligence is living through that story, and its second act arrived earlier than expected.The three points that matteredOn July 16, Moonshot AI, a lab based in Beijing, released a model called Kimi K3. On the Frontend Code Arena, where human evaluators blind compare two AI outputs and pick the better one, Kimi K3 scored three points higher than Anthropic’s newest model. It’s a narrow result on one benchmark, not a sweep; on broader indices the American model still leads, and Moonshot’s own notes concede the point. But the benchmark it won is hardest to game and most tied to daily enterprise use: writing and reviewing code. The price gap dwarfs the performance gap: frontier American models charge tens of dollars per million tokens, while Kimi K3 is priced near the cost of compute and ships free from July 27.This isn’t the first time a Chinese model has drawn level with the American frontier; DeepSeek did something similar in late 2024, and the gap reopened within months. What’s different now is the breadth of the field behind the leader: Moonshot, DeepSeek, Zhipu, Alibaba, and ByteDance are all fielding frontier or near-frontier models on a rolling basis, several of which are freely downloadable, while the credible American frontier rests on two companies. Days after Kimi K’s release, China’s president framed open-source AI as a deliberate national strategy, announcing training placements for engineers from developing countries and joint AI centres across the Global South. That’s not charity; it’s industrial policy aimed at the customer base India has assumed was naturally its own.Washington’s instinctThe American policy response has moved fast, revealing something uncomfortable about how this fight will actually be fought. Reporting this month indicates the Commerce Department has weighed adding Chinese AI labs to its Entity List, cutting off American access without a licence, alongside an executive order requiring US companies hosting Chinese models to guarantee security and accept liability for breaches. A more candid version skips the ban entirely: simply direct agencies to issue advisories that create enough uncertainty, regardless of substantiation, that regulated enterprises quietly back away on their own. That approach has drawn sharp criticism, on the grounds that regulation built on manufactured doubt corrodes the rule of law.There is a genuine security case underneath the manoeuvring. A freely downloadable model isn’t the same as an open one; Moonshot has handed over a finished product, not the ingredients, and no outsider can verify what it was trained on. Running a Chinese model instead of an American one swaps one black box for another. But a policy built on making enterprises nervous, rather than demonstrated harm, looks closer to protecting two incumbent labs’ pricing power than genuine security policy. American and Chinese labs are already training on each other’s models, blurring whose is really whose, and cutting off access mainly pushes users towards providers with no compliance regime at all.Where India actually sitsIndia is not currently a producer of frontier models. It is a renter of intelligence, and the terms of that rental are being set by other people’s export controls and subsidies unless India builds enough capacity to negotiate rather than simply accept.The temptation is to treat freely downloadable Chinese models as a shortcut, building India’s AI stack on them and skipping the capital cost of training from scratch. There’s a real case for that; cost matters, and Indian enterprises need cheap intelligence now. But it carries the same dependency problem as leaning on American models, just with a different government’s opaque pipeline underneath, one explicit about treating AI exports as a tool of influence over markets India competes for.The more durable answer isn’t picking a side but reducing how much either side can determine India’s trajectory: treating sovereign compute and the ability to train models domestically as core infrastructure, and backing India’s own freely available models seriously enough that tuning a foreign base model isn’t the ceiling of ambition. Worth remembering, too, are the eighteen days this June when one major American model went dark worldwide after Washington’s export rules briefly changed. Diversifying which black box you depend on isn’t sovereignty, but it’s safer than depending on one.The argument that actually mattersThe benchmark score that started this debate will likely look dated within a quarter; these leads swap back and forth on cycles now measured in weeks. What won’t reverse as easily is the economics: a technology moving from luxury to commodity in a handful of cycles, the same collapse that took aluminium two decades and electricity half a century.Governments can slow that at the margins; reasonable people can disagree about whether they should try. No government has stopped a commodity price from finding its floor in 140 years of this pattern repeating. The useful question for policymakers in Washington, Beijing and Delhi isn’t who wins the next leader-board, but who builds the infrastructure to use cheap intelligence once the price falls.Kaushal is a tech and social entrepreneur and Programme Director (Eastern India) at WHEELS Global Foundation. He writes on AI, economy and geopoliticsPublished on July 21, 2026
