In my regular research (behind a paywall), I have been saying for a while that I think the future of AI is not large language models (LLM), but small language models (SLM) run on local desktop computers or even mobile phones. In May, a team from Stanford University published research that compared these SLMs with the performance of LLMs run in data centres. If their results are true, then we will hardly need any data centres in the future, and the hyperscalers are wasting hundreds of billions of dollars in investments.
Seriously, if you are an investor trying to figure out where to invest in the AI hype, you need to read this paper in full. But to get you started, let me give you some highlights.
First, they ran a series of SLMs (QWEN 3, GEMMA 3, GPT-OSS, GRANITE 4.0) that can be downloaded on a local PC and compared their performance with cloud-based state-of-the-art LLMs (ChatGPT 5, Claude Sonnet 4.5, Gemini 2.5 Pro).
They ran these SLMs on local PCs powered either by an Nvidia chip or an Apple M4 chip, as they are readily available in current high-end desktop computers (the entire study was done before Nvidia presented its AI chip for PCs, which will only accelerate the move away from datacentres to models run on desktops).
Then they traced the performance of these SLMs vs LLM between 2023 and October 2025 on both chat tasks and reasoning tasks.
The chart below shows the Win/Tie-ratio for SLMs vs LLMs in chat requests, which still make up the vast majority of requests today. As you can see, in every domain, the best SLM is able to find the same or better answers than an LLM in 90% or more of the cases, with an average across all domains of 98.6%.
Win/Tie-ratio of SLM vs. LLM in chat requests
Source: Saad-Falson et al. (2026)
When it comes to reasoning tasks, which are obviously more demanding, SLMs are catching up fast. On average, they provide a better or at least as good an answer as LLMs in 62.5% of the cases.
Win/Tie-ratio of SLM vs. LLM in reasoning tasks
Source: Saad-Falson et al. (2026)
However, in real life, the tasks for SLMs and LLMs are typically a mix of chat requests and reasoning tasks, so the third chart shows the weighted average of chat request performance and reasoning performance based on the frequency of tasks in each domain. As you can see, on average, SLMs are as good if not better than LLMs in 81.2% of the cases, with the LLMs having a significant advantage only in areas like engineering, life sciences, transportation and computer sciences.
Win/Tie-ratio of SLM vs. LLM in chat and reasoning tasks
Source: Saad-Falson et al. (2026)
But it’s not just accuracy. SLMs achieve this performance at energy and compute costs that are between 50% and 85% lower than for an LLM, depending on the SLM and hardware used in the computer.
What is more, SLMs are catching up rapidly in reasoning tasks. The final chart shows the performance of SLMs as a function of difficulty level and model generation for reasoning tasks alone.
In 2023, the success rate of SLMs in reasoning tasks was typically 50% or so across all five difficulty levels. By October 2025, the SLMs achieved 99% success for the easiest reasoning tasks in levels 1 and 2, 85% to 92% success in harder tasks (levels 3 and 4) and only lagged LLMs in the hardest tasks of level 5 (51.5% success rate).
Success rates of SLMs in reasoning tasks
Source: Saad-Falson et al. (2026)
This already means that one can replace data centres and their expensive cutting-edge semiconductor infrastructure in four out of five use cases. The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
There clearly are areas where LLMs are still way ahead, particularly in agentic AI applications, where SLMs currently only achieve accuracy and success rates of less than 50%. Similarly, it is difficult to run these SLMs on smartphones so far. The models that can be run on an iPhone are significantly worse than the models that can be run on a desktop PC.
But – and this is important – the models run on a desktop PC, and even more so, the ones run on a smartphone are much more energy efficient than the ones run in the cloud. The inference per Watt of these SLMs is typically seven times larger than that of LLMs. And that means that when you encounter a task that can be solved on a desktop or even a mobile device, it is cheaper to do so locally than send it to a data centre.
This has important implications for investors, in my view:
We need many fewer data centres than we think. If we can already replace 70% to 80% of the tasks that are expected to run on LLMs with SLMs, the hyperscalers have simply no revenue growth in the future that is nearly enough to justify the capex. In fact, if this research is true and this trend continues, data centres may be the worst investment in the AI space one can make right now.
While we continue to need enormous investments in semiconductors of all sorts, we do not need to invest in the most advanced Nvidia chips. The cheaper ones that run on desktop PCs will be enough. The best case for Nvidia is that it can replace its high-end data centre GPUs with its new chips for desktop PCs. What will that do to Nvidia’s margins and revenue growth going forward?
We still need LLMs for the most advanced tasks, and companies like OpenAI, Anthropic and others will be able to ‘dumb down’ their models to an SLM and sell them instead of LLMs. But given the already fierce competition from Chinese providers like QWEN or IBM’s GRANITE, the profit margins for these models will be much smaller than for LLMs. So what does that mean for the valuation of these companies in their planned IPOs and their growth trajectory?
While agentic AI is still better on LLMs, this may only be a temporary advantage, similar to what we have seen in reasoning tasks and single chat requests. If that is the case, the true winners of the AI boom will not be the providers of advanced hardware and data centres but the boring manufacturers of desktop computers like Dell and Apple.
Watching this race unfold is going to be fun, and I am increasingly convinced that many people will be badly burned because they invest in the wrong technology.
Yeah, a massive over-investment boom in tech-related stuff which turns out to be reasonably useful but not a money-printer - we've already been there and got the t-shirt (or the top-hat for those remembering the 1840s railway mania).
Great post and interesting research by Stanford too!
If the use of SLMs and open source (and open weight) models continue to proliferate, which I think they will after people en masse figure out how to implement them in organisations, then Google may be in an interesting position as they're Nvidia's only real competitor especially as ASICs could take a bit more market share from Nvidia GPUs. It'll likely be a thing where both will coexist and the buyers of chips pick either or depending on use cases and their requirements.
This is all in spite of Google's culture clash and being at war with itself (Demis's vision is different to what Google wants wrt LLMs - Demis is more of a believer in 'world models'). By all accounts, Google should be leading in the 'AI race' but bureaucracy and 'timidity'(?) seems to be getting the better of them. Bear in mind Google had an LLM 1 year before OpenAI released ChatGPT and chose to not because they thought it'd cannibalise their search business. Deepmind had understood the limitations of LLM's and therefore did not pursue that research aggressively, hence him believing in 'world models' which remiains to be seen. If Google had not publicly released the transformer paper and kept it as a 'trade secret', there would probably be no ChatGPT.
Apple with its new CEO John Ternus - who previously led the hardware division - will likely revamp Apple's hardware, especially the Mac lineup including the Mac Mini, so they can capture value.
The future could well be specialized edge models that know only ONE thing - and know it well. Such models don’t require the energy and compute necessary to be “the everything” model. They will live comfortably as edge models; likely in reduced format in either computers with ring-fenced propietary data (for enterprise) or in our pockets on our smart phones/PCs (for consumers).