Good Enough AI
It’s not a space race, but a race to the bottom.
Summary: DeepSeek’s open source model signaled to the market we can do more with less, turning prior assumptions on its head. Developing frontier models is likely to become a high-cost, low-margin business where the value will be almost entirely captured by consumers.
↓ Go deeper (9 min)
Have you ever asked yourself why airplanes don’t fly faster? The Concorde was once the promise of the aviation industry, the first supersonic airliner to enter service flying passengers across the Atlantic at twice the speed of sound. But anno 2025 the aircraft is no longer in use.
Commercial airplanes fly much slower nowadays, because flying fast is expensive. As it turns out most people just want cheap, reliable flights that get them from A to B without bankrupting them. And so, air travel didn’t keep getting faster, it got more efficient.
Now you can maybe guess where I’m going with this, but given everything that has happened over the past week — I’m of course talking about DeepSeek — it’s becoming clear that AI is likely to be headed for the same fate.
The consumer wins
The tl;dr version is that a relatively small Chinese AI lab built an extremely cost-efficient reasoning model on-par with OpenAI’s o1. As if that isn’t surprising enough, they open-sourced the whole thing and published in detail how it was made, basically giving the model as well as the training recipe away for free.
Many described it as a ‘Sputnik-moment for AI’, a reference to the first artificial satellite launched by the Soviet Union, which kicked off an arms race with the US. A tempting comparison, but ultimately a misguided one.
What we’re witnessing here is not a space race, but a race to the bottom. It was long thought only deep pockets could create highly performative models, but DeepSeek has shown the world you can do more with less. By making its models and training recipes freely available, everyone can now learn from and replicate reasoning-style models that are relatively cheap to run.
On the podcast Prof G Markets, Robert Armstrong, a financial commentator for the Financial Times talked about the implications:
“We had a certain vision on Friday of what the future economic structure of the AI industry would be, and that vision has significantly changed. (…) Now we have a vision of the AI industry where it is much more competitive, and the profits are shared. Much of the value is going to be captured by consumers rather than companies.
He goes on to make the same airplane-analogy as I did at the start, explaining:
“[the aviation industry] is an incredible industry in terms of what it gives us and what it can do for us (…) but is a terrible industry, economically.”
In other words, being in the foundation model business isn’t going make anyone rich, despite being hugely beneficial to people. Presumably, the AI labs know this, which is why we’re seeing companies like Anthropic and OpenAI aggressively expanding their product portfolios as they risk running out of runway much quicker than anticipated.
The best price is free of charge
At the application layer, things are getting heated too. DeepSeek’s chatbot, which is free to use, quickly rose to the number 1 spot for most download app in the App store. Their release seriously undermines OpenAI’s 200 dollar monthly subscription, and in an attempt to do damage control, OpenAI was quick to push out not one but two product launches this week: a new reasoning model o3-mini and a research agent called Deep Research.
AI agents are seen as the next frontier with companies like Salesforce, Microsoft, and Atlassian rolling out products left and right, but this future is not guaranteed to arrive as soon as people think it will.
Knowing all that, let’s entertain a little thought experiment together: what if most people will settle for Good Enough AI? What if the majority of people just want AI with average intelligence that is free or almost free?
What if the billions of dollars that are currently being spent on making AI “smarter” — Stargate-sized projects, announced with much fanfare, which are believed to be required to train the next generation of models — only serve an incredibly small group of power-users who are obsessed with math and coding benchmarks? What if the real disruption is not superintelligence, but mediocre AI that runs on your device?
Surely, the frontier labs would have you believe they are building rockets. ‘A country of geniuses in a data center’, is how Anthropic’s CEO Dario Amodai likes to describe it. Call it a hunch, but I think there’s good chance these endeavors will end up much like the Concorde: a loss leader that will never make up for its investments. Lest we forget, OpenAI, valued at $157 billion, has yet to turn a profit.
The reality is that the average person doesn’t notice the difference in performance between DeepSeek’s R1 and OpenAI’s o1 on any given day. To me, this tells you all you need to know.
Speak soon,
— Jurgen
Not done reading yet?
Check out my guest article on AI Supremacy for a deep dive on DeepSeek and how this company was able to catch Silicon Valley by surprise. AI Supremacy is a flagship newsletter on AI and emerging technology, written and curated by .
Or browse my latest articles you may have missed:
Concorde is an interesting analogy, that we write about the intro to The Future Normal (my book)
My 2c - it was a combination of push & pull, beyond just cost:
There were the sustainability & noise issues.
But also, private jets became more accessible (for those with lots of $$$), while offering greater flexibility than a single route 'mass' option.
Plus, on longer routes, wifi & lie-flat beds made the flying experience far less 'costly' (in terms of lost productivity) for business travellers.
Thanks Jurgen for this analysis. I particularly appreciated it - as all your issues, by the way - because it put the emphasis on the consumer, on the possibilities and on the "victories" that potentially emerge. I think that a consumer-oriented perspective, together with the more technical one, is also very important to understand, analyze and break down the narratives and key factors that drive certain discourses and make them evolve from more technical to "pop".