Altman ducks, Nvidia dives

and why it may not matter in the longer term

Here’s a quick take on why I think the launch of DeepSeek, a Chinese hedge-fund development that has evolved into a model that can out-reason OpenAI’s o1, has so rapidly undermined the stock value of Nvidia, what this reveals about the current AI narrative, and why in the end I think Nvidia (and US AI) will recover. That is, why the build-out of hyperscale data centres, alongside the concentration of all kinds of social and cultural value that can be rendered as data, will continue unchecked.

I’m sorry to say it again, but AI was never going to live up to the hype. This was obvious to anyone with even a passing cynicism about venture capital and the tech hype industry that shares its blood supply. It was widely discussed by tech industry insiders and by experts in (yes) artificial intelligence and cognitive science. It’s almost embarrassing to have spent so much time arguing for ideas that can now be found on fridge-magnets: here’s tech rag The Byte, just days before the current share-dive:

‘This [economic] forecast brings into question the logic underlying the incredible economic, environmental, and human costs of AI development… When, for all their advancements, all these half-baked AI rollouts can offer are "leaked photos of heaven", it needs to be asked: who are these advances for? If AI developers are incapable of producing socially compelling use cases for their models, then burning through trillions of dollars in a never ending AI race isn't anymore practical than a one-way trip to Mars.’

Put it on a t-shirt. The story of hype is not a new one, of course. People who have spent the last two years demanding or proclaiming or actively colluding in the reorganisation of public life around this new technology are now nodding wisely and renewing the Gartner Hype Cycle cartoon on their office door. So that’s OK. We all move on. From the carbon that can’t be unburned, the jobs that can’t be made secure again, the knowledge economies that can’t be detoxed.

But there is something puzzling about this hype cycle. Even Gartner is puzzled:

So while GenerativeAI promised virtual persons, spending on it has been mostly on brute force hardware. On cloud credits by business customers, and on hyperscale data centres by AI providers themselves. And this spend is going up despite the ‘disillusionment’ and the lack of ‘functionality’. How to explain?

In my own feed I’m hearing two takes that make sense. One is Brian Merchant’s. He says that the DeepSeek launch has undermined a story about scale. The story is that:

‘Generative AI is the future, it’s improving rapidly, and thanks to that we’re on the road to Artificial General Intelligence. AGI, in turn, will enable enterprise software to “outperform humans at most economically valuable work”… but the models need to be scaled first. Those models need ever more data, they need more compute power, and they need more energy to run.

DeepSeek undermines this story because it was built with a fraction of the dollars and training compute required for the equivalent OpenAI model (it claims ‘only’ 10,000 Nvidia chips were needed to train R1, though this is disputed). Leaving the geopolitics aside for a moment, Merchant argues that the US innovation sector will be running to miniaturise its own offer, and that it may no longer need all the chips and data centres, the demand for which has hyperscaled Nvidia to the largest and most profitable company in the world.

Yes, and no.

I think the story about scale was already on the way out. Altman and Hasabis have both gone on the record in the past year to say that scale ‘is not enough’ to provide the next breakthrough. (Their hands may have been forced, since nothing in the last 14 months has shown anything like the step-up in capability that GPT4 showed from GPT3, while scale has continued unabated). By November of last year the talk among developers was no longer of improving outputs in any significant way, but of matching existing capabilities with models that are smaller and cheaper. It’s a long way down from ‘liberating humanity’ to ‘making the product cheaper’, but venture capital was always going to demand commercially viable product for its noble investment in ‘humanity’ at some point. In those terms, DeepSeek has jumped ahead and US companies will have to respond, because they can’t command as much for their own product as they could before DeepSeek strolled past ChatGPT on the Apple App store.

But the US companies are so powerful, their reserves of capital so immense, their funding from the US state so secure, they can easily weather what amounts to a slightly longer lead-in before their products become profitable. In Altman’s justificatory statement on DeepSeek, he isn’t stepping down at all from the need for ‘more compute’, even though ‘compute’ is (a significant part of) what makes his product so expensive.

One thing Altman may be relying on here is what is called Jevons’ paradox, where making a resource more efficient only increases demand for it. When he says ‘a LOT of AI’ he is assuming that his company in particular, his narrative and the reorganisation of business platforms around it, has made ‘AI’ such an essential commodity (despite its lack of ‘functionality’) that making it cheaper will only accelerate demand. In terms of sustainability he may well be right: even before the current chat-fuelled boom, emissions from ICT outpaced the general growth in global emissions every year, despite efficiencies in production. People just can’t get enough of those chips.

Projected increases in ICT emissions from Freitag et al 2022, produced before the current GenAI emissions boom

We should not forget that along with nVidia, stocks in major energy companies have also taken a hit (this headline in UK investing explicitly tells us that ‘efficiencies’ like the ones apparently available with DeepSeek are a source of ‘concern’ to energy investors, because the more carbon gets burned, regardless of whether it needed to be burned or not, the richer they get).

This is where I think we need another story, perhaps a supplement to the first one about the US losing out on miniaturisation (not itself a novel one). What if building out vast amounts of compute and using it to concentrate, organise and govern vast amounts of data, actually is the project?

What if it’s not about AGI at all? Not even about generative AI and whatever commercially-viable use cases it spits out on its way up and down the hype curve? (Obviously the investors and CEOs and business customers will lap up whatever they can get in the way of cognitive automation, and thank you kindly for a lifetime of service down the use-case mine.) What if it is all about data and compute, and who controls them?

Controlling data, controlling the systems and architectures it flows through, digitising the world’s many cultures and cultural products and privatising them, monetising access to them. This is the promise of generative or cultural AI. Owning data about people and their behaviour, predicting and managing ‘deviance’ from the norm, turning what’s left of the social state (in the wealthy world) into private data-driven services, and the rest of the world into a data farm. This is the promise of predictive AI. There may or may not be profit in these today, but they are surely the route to squeezing the last juice out of capitalism in the years ahead.

I have more on the ways that generative and predictive AI are, in fact, convergent in my next post.

So-called ‘AI’ or data intensive military surveillance is also being used to acquire people-as-data and territory-as-data for targeting and rapid decision making at war. Some of the same companies involved in generative AI have been active in this space also, and the alliance between Silicon Valley and the US military state is one we are already seeing strengthened under the current US regime.

From this perspective, so-called ‘AI’ is just the circus, more effective than crypto at inflating speculative bubbles of private capital, more effective at capturing cultural and social data, more effective at mobilising build-outs of data and energy infrastucture - and more effective in allying with the militarised state. (China, of course, has taken a different path to social governance through data, but one that increasingly relies on indistinguishable technologies and privacy violations.) Yes, in the short term it is a blow to US AI capital that genuinely open models can be built for a fraction of the cost and compute. But that is only the short term. In the medium term, it’s all about the concentration of power and data by vast corporations in alliance with the most powerful military states.

That is why AI has ceased to be a story about the commercial viability of some overhyped product, and become a story about geopolitics, and the battle (literally) for alternative futures.