Nintendo's example shows how China has built AI at a fraction of cost of the US: David Richards
Some of my best work came in years with no money to spend. Scarcity does the work most boards pay consultants to do: it settles what matters and pushes you to build not rent.
In the late 1970s an engineer sat on the Shinkansen, Japan’s bullet train, and watched a bored salaryman press the keys of a pocket calculator to pass the time out of Tokyo. He saw a games machine nobody had bothered to design.
The engineer was Gunpei Yokoi. The company was Nintendo, then reported to be seven or eight billion yen in debt after failed arcade ventures and the oil shock.
The timing was luck. Sharp and Casio were locked in a calculator price war, and their suppliers were sitting on cheap, unwanted LCD screens and 4-bit processors.
He had no route to anyone who could say yes. Then the president’s chauffeur called in sick. The engineer drove.
He made his case at the wheel and got a polite hearing, nothing more, with no way of knowing that his passenger would spend the meeting beside the head of Sharp, and a Sharp executive turned up at Nintendo the following week.
Yokoi called it lateral thinking with withered technology. Ignore the frontier. Find a radical use for what your competitors have written off as dead, and buy it cheap. Game & Watch ran on button batteries. Its graphics were fixed segments on a calculator screen, and the clock gave a grown man a reason to carry a toy.
By 1982 Nintendo wanted Donkey Kong in a pocket. The clamshell had to fold flat, which ruled out a joystick, so Yokoi’s team laid a plastic cross over a central pivot. A hinge produced the directional pad. It became the standard control under the left thumb.
Yokoi projected 100,000 units. Nintendo sold 43.4 million across nearly 60 models. The debt cleared. Four billion yen went into the bank. The profits paid for the machine that became the NES, and the clamshell returned as the DS.
China has been handed Yokoi’s problem. Washington barred China from Nvidia’s best chips. Denied brute force, the Chinese labs went round the bottleneck. Rather than dense networks in which every parameter fires for every token, they refined mixture-of-experts designs, so DeepSeek’s R1 carried 671 billion parameters and woke only 37 billion of them per token.
Moonshot’s Kimi K3 landed in July, open-weight and free to download, within touching distance of the American frontier. Airbnb, DoorDash and Coinbase are running Chinese models on local servers, and they are doing it because the arithmetic works. Washington starved a rival of compute. It learned to need less.
The American valuations assume that none of this matters. Anthropic’s May round valued the company at $965bn post-money, and a listing could come as early as October. OpenAI was priced at $852bn post-money in March. Moonshot was valued at $35bn in July. Anyone buying those shares is buying the proposition that a lead costing hundreds of billions cannot be matched for a fraction of the money. The counter-argument is free to download.
Yokoi’s rule outlived the crisis that produced it. The Game Boy shipped in 1989 with a monochrome screen and old hardware, against colour machines from Sega and Atari that ate their batteries in hours. The Game Boy line sold 118.7 million. In the year to March, Nintendo took 2.3 trillion yen in sales and 424 billion in net profit.
Necessity is the mother of invention. Nintendo could not afford the frontier in 1979 and ended up defining it for the next four decades. The American labs valued at close to a trillion dollars apiece were built with the best chips money could buy, and their challengers were forced to build without them, at a fraction of the cost.
That is what should worry the people buying the shares.
David Richards MBE is managing partner of Yorkshire AI Labs