The AI tsunami
Are you still sipping a cocktail on the beach?
Keeping up
Ah. I remember the good old days. When there were three years between AI model releases (like GPT-3 & GPT-4). When there was actually time to understand what a model could do before it was obsoleted.
These days things are different. We’re lucky if there are three days between model releases. Since Deepseek R1 (Jan 20th) we’ve had another three models from China - Qwen 2.5, Doubao 1.5, and Kimi 1.5. Are they any good? I don’t know. And probably never will - I don’t have the time to invest exploring each model. Nor do benchmarks help - each model has found a subset of benchmarks which it can use to claim to be the best…
We are now beyond the point where it is possible to keep up - and it’s only going to get harder. So how do you cope in this world?
Crowd mentality
It’s simple - we follow the crowd. Crowd following is a survival mechanism that's served us well for millennia. Why risk your life testing every berry in the forest when you can copy what others are eating? We're seeing this play out with the DeepSeek-R1-berry, sorry, AI models.
Over the past week R1 has gone viral. Somehow it’s become a winner - and now everyone wants it. Not because it’s the best (sure it’s good) but because everyone was talking about it. And in a world where the news is almost overwhelmingly gloomy a new AI model makes a pleasant change.
This pattern will only intensify. Periodically a model will go viral and capture attention - and users - through a combination of timing, marketing, and crowd psychology. Others, potentially equally capable, will miss out and be destined to niche usage or, worse, fade into obscurity.
We saw this playing out last week - here is data for DeepSeek vs Claude from Google Trends:
Note how DeepSeek came from nowhere. You’ve got to feel a little sorry for Anthropic.
Add in ChatGPT and things are a little different. It’s still the most popular:
And poor old Gemini (green) and xAI (invisible purple):
If you think I’m being unfair on xAI, then rest assured that no-one is searching for x-AI and Grok either.
And the geographical split is illuminating:
The story these charts tell is clear: viral adoption is alive and well. A new model can explode from nowhere, capturing mindshare through timing and momentum.
The casualties of speed
This is a brutal race. There will be casualties. The companies that can’t keep up. Overlooked technologies. Countries and regions. Europe. Australia. India. Russia.
Take Europe. Aside from Mistral (who are not, currently, thriving), Europe isn’t in the race. And the UK? The country that helped kickstart AI research. Which produced Geoffrey Hinton - the so-called godfather of AI. Which produced DeepMind, the model which arguably kickstarted the current AI race. We too are watching from the sidelines. It's a familiar story arc. We saw it with the jet engine. We saw it with the computer. And now we're seeing it with AI.
The strategic game
This is more than just a technology race. Open source is currently driving innovation, but it won't stay that way. The models have strategic importance - both commercial and national. We've already seen the drawbridges rising:
OpenAI's shift from open research to closed development (ironically they are now one of the most closed companies despite their name).
The emerging divide between US/China and the rest.
The concentration of compute power among a few players.
Access to the best models will become restricted. For those outside China, US and their allies this is not good news.
Europe has just one trump card. ASML. They are the sole supplier of the lithography machines used by TSMC, in Taiwan, to make the chips used to train and run the AI models. And TSMC are the sole company making these chips. Conventional wisdom is that ASML have a 5+ year lead; TSMC 2-4 years. It’s an interesting position to be in.
Looking forward
Three trends are emerging that will shape what happens next.
First up, infrastructure evolution:
If you want to use DeepSeek - and don’t want to send data to China (or run it locally) - then Groq is a good option. Groq is a company that builds their own hardware to run AI models. They host DeepSeek R1, Llama and others. And they are blindingly fast.
Part of DeepSeek’s special sauce was a smart shortcut - they bypassed NVIDIA's CUDA software (think of CUDA as the translator between AI models and graphics cards) to talk directly to the hardware. It appears that got them 10x better performance, albeit with a maintenance cost as their code is now optimized for a specific hardware generation.
Second, we're starting to really see Jevons' paradox. Back in 1865, an English economist noticed something counterintuitive - making coal burning more efficient led to increased coal usage, not less as it opened up markets that hadn’t previously existed. The same is happening with AI:
As models become more efficient, usage explodes.
Even Anthropic, with their massive compute resources, are struggling with capacity - apparently overwhelmed by the sheer volume of code generation requests via Cursor & Github Copilot.
If you thought DeepSeek was bad news for Nvidia you are wrong - it’s great news as we’re going to need more and more compute.
Third, progress is accelerating:
Last December marked a turning point, and the pace keeps increasing.
We're seeing increased specialization - groups focusing on ever-smaller areas of AI.
o3 is coming in the next few months, likely putting the ball firmly back in the US court.
And if US companies can replicate DeepSeek's tricks, they'll pull even further ahead.
Countries are starting to see AI models as national assets, not just commercial products. As this mindset takes hold, even the most powerful models may be kept behind closed doors for strategic advantage, regardless of how much money could be made by selling them.
In time, the cost of failing to compete for those outside the US and China will become clear. But for now many remain in a bubble, sipping their cocktails on the beach as the tsunami races towards the shore…