DeepSeek spills Big AI's open secret: Bright people with good ideas can beat billion dollar binges

You can’t make monopoly money without a monopoly, but you sure can lose it

Opinion It would take a heart of stone not to explode with joy at the massive infusion of schadenfreude provided in recent days by the DeepSeek AIpocalypse.

Trillion-dollar markdown in tech stocks, slack-jawed panic at tech companies previously "too big to care" suddenly caring a whole lot, and the mainstream media unable to talk about anything else. All through a single app from an unknown Chinese company with the on-the-nose logo of a whale in its death throes. 

If that's not enough, the sound of the comically misnamed OpenAI grunting in indignation over abuse of intellectual property in training data by an actually open AI is funnier than any three Monty Python sketches combined. "What's yours is mine, what's mine is mine too" makes a great state motto for Silicon Valley, but "information wants to be free" might be even better. 

Some have argued that this is a Sputnik moment, when a disdained rival suddenly leapfrogs a complacent establishment figure. That certainly fits the New Cold War narrative, but there's a better analogy closer to home – the rise of the PC architecture at the expense of the mainframe. IBM was the establishment, and it created the PC to keep things that way. It didn't bother to burden the design with too much intellectual property in either OS or hardware, reckoning that since it owned so much of the IT market anyway, it couldn't lose. Then Compaq overwhelmed the tiny legal defense of the copyright BIOS and the barbarians poured in through the gates. (As we explained here, back in the early '80s, when Big Blue started building the first IBM PC, the only part of the computer it had copyright control of was the BIOS ROM chip, and before long Compaq had figured a way to reverse-engineer it so that it could sell IBM-compatible systems for less than Big Blue was charging.)

DeepSeek has pulled off a similar trick, using a variety of existing AI ideas to not only replicate the magic beans that the giants have built their beanstalk business models on, but to start giving it away.

Oh, and it (apparently) costs much less in chips and watts to train. While DeepSeek has spent many millions on thousands of its own GPUs and research and development costs to build its open source V3 and R1 LLMs – one analysis reckons the Chinese lab has spent $1.6 billion on hardware – the lab claims training its V3 in the cloud would cost less than $6 million (2.78 million Nvidia H800 GPU hours at $2 an hour.) The argument goes that DeepSeek is able to match the West without demanding or requiring the zillions in investment that Silicon Valley has talked up.

Of the three underlying assumptions that have passed for genuine rationales in the AI bubble – LLMs will drive innovation, they will create untold billions in added value, and only the very largest tech companies can play – all but the first has just been disproved. 

The one that's left, the assumption that LLMs will change the world in wonderful ways, is going to be much easier to test if DeepSeek's claims about training costs are true. AI works best when it is reined into specific tasks instead of generalities, and many more companies, institutions, and researchers will be able to experiment here. That's badly needed, with Apple, Google, and Microsoft desperately trying to force-feed the bad stuff down our necks. 

It's also bad news for the parody of an industrial policy recently adopted by the US and UK governments, that simply pouring billions into giant ocean-boiler datacenters will compensate for lack of thought or investment in infrastructure that will stay relevant for more than five years, which is good news for those of us left in the reality zone. But then, if you hadn't twigged what it means that the initial investors in Stargate are Oracle and Softbank, that may not be a zone you have a visa for. 

There is a notion at large that none of this matters, as taking the ideas, indeed the code, of DeepSeek and decanting that into giant ocean-boiler datacenters will allow Big Tech to regain the advantage. Microsoft seems to have reached this conclusion remarkably quickly. Yet this presupposes that throwing money instead of smarts at LLM development will always work. When you have more money than smarts, that's a compelling argument, even though it's just been shown not to be true. 

All this is saying nothing about DeepSeek's nature as a product of China. The usual caveats apply: don't run the app unless you trust the Chinese state to behave itself, do get stuck into running it locally if you're interested, and don't be surprised if restricting what high tech a country can buy results in uncomfortable consequences over time. Best jam a plan B in there. 

It is in the nature of technology to constantly evolve, just like living systems, and just as hard to predict what small furry mammal is going to eat your dinosaur lunch. And, again like biological evolution, each new burst of novelty acts as scaffolding for the next. In this case, we've filled the planet with networked pocket computers of incredible power. On top of that, a massive gene pool of open source components can spawn novel organisms at tremendous speed. 

This can give a new idea universal impact virtually overnight. It took 24 years for IBM to be driven out of the PC market it created. 24 hours may soon be closer to the mark. ®