I’ve argued for a while that these general purpose, do everything at any cost AI facilities will be undercut by smaller, more purposeful AI models (I feel the same way about robotics except oops, that’s the way that industry already is). The relative success of Chinese models, doing almost as much with a lot less cost and clamour is only the first indication.
Now here’s a more erudite, or at least detailed explanation of how this might play out.
Thanks, and a tip of the Hatlo hat to Mungo on the Shrewdm Berkshire board:
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%.
Obviously if the industry wakes up some morning next month and says “Hey, we don’t need all these huge data centers after all” that has grave implications for the economy and for the market.
People say a phone is all-purpose. But it’s a container of specialized things: apps. There is no all-purpose, do-everything app.
Big problems are solved by dividing the big problem into many small problems.
Why would AI models be different?
The human brain has specialization, different areas dedicated to different tasks.
We’ll need data centers to train and run the small models (if not run locally, but hard not to see still needing lots of cloud infrastructure) and store all of the related data.
The AI model is still a new thing that demands more infrastructure.
To solve a given set of use cases, do many small models use less capacity than a few big models?
Yep. That’s why I’m highly confident I’ll eventually have a distilled version of unsupervised FSD operating on my HW3 Model Y. If not by Tesla, then the Chinese.
SLM and LLM are the same principles. Counting odds to work with information, counting odds to get resources to answer the query, and counting odds to answer word for word the query in a response.
But SLM is bigger per person much of the time. How so…? First keep in mind if you buy a calculator you use it endlessly at no extra cost but the electricity. If you use the cloud you pay a lot of money, and risk hacking.
The difference is under the hood online. The next quote is from Claude Fable 5. Claude Sonnet 4.5 is nothing.
Claude on the topic
Yes, substantially — and the concurrency picture is a big part of why.
SLM: roughly 0.5B–10B parameters, weights ~1–20 GB. Runs on a laptop, phone, or single consumer GPU. Typically serves one user — the person sitting at that machine. A laptop running a 7B model at full speed has no spare capacity for anyone else; even a handful of simultaneous requests would crawl.
LLM: tens of billions to a trillion+ parameters, weights hundreds of GB to TB. Runs across clusters of datacenter GPUs. Serves thousands to millions of concurrent users — a single serving node with batching handles hundreds of simultaneous conversations, and providers run many such nodes behind a load balancer.
The inference code is nearly identical in both cases — same transformer architecture, same kind of runtime. The differences are weight size, hardware, and the serving layer: an LLM deployment adds batching, KV-cache management, load balancing, and autoscaling that a local SLM never needs because it only has one user.
Practical tradeoff: the SLM is private, offline, and free per-query but single-tenant; the LLM is shared infrastructure that’s smarter and scales to a whole user base, at the cost of network dependence and per-token pricing.
My comment the laptop will win, but it might take a couple of years. Meanwhile in all senarios AI chips are improving at 7 times Moore’s law last I saw.
The frontier models focus on AGI, and you have a domain specific LLM’s that are trained only a certain domain, for ex: legal. They have different purpose. Domain specific LLM’s are not a replacement of Generic models.
From here onwards there are many incorrect assumptions. IT seems someone with a predetermined opinion, wrote that article. There are smart folks than this investment “strategist” who have articulated how the LLM’s will get better and better and while an earlier version of model may be “good enough” because many existing processes/ companies/ folks will take more time to use a much better model. But I haven’t seen any smart guy arguing the small language model is enough.
The world always had a class system. The new class of super-intelligent folks working with super intelligence and the rest of us will be the new class. Those folks will generate far higher economic value and probably corner most of it.
One can argue a small handgun kills effectively, therefore you don’t need machine guns, tanks, missiles, drones etc. I am sure there will be many who are sympathetic to such arguments. I find such arguments short-sighted.
No, but - if today’s debt levels and stock valuations reflect an expectation that AI will primarily be supported by large-scale data centers, then SLMs taking a significant slice of that demand may be enough to cause a lot of mayhem for companies and investors betting on that assumption.
Yes, that’s the point of the article - not that there can’t/won’t be large scale data centers, but that the economic case for so many of them may be severely weakened. Steve Jobs built a fabulous computer at NeXT, but the entire raison d’être for it was undercut by the PCs and Macs that did much of the same stuff, more slowly, but at a significantly reduced price. (Never mind that NeXT software lives on as part of iOS!)
That is not supported by the data. SLM’s are cheap is not what matters, rather what kind of work will use SLM’s. Even domain specific LLM’s need to run on DC’s and all enterprise workload requires access to the systems of record…
I am not worried that SLM’s will eat into LLM’s market share in a big way, rather the per token cost going down (not because of SLM), chinese models costs are cheaper, and price war (already started by $META) will make OpenAI and Anthropic unprofitable/ or seriously impacting their cashflow generation setting domino effect.
True for now, but there is an argument that only SLMs can be used by several professions. Accountants and others can not allow their data to be outside their jurisdiction or in a position to be hacked. The LLM is not usable for some professional purposes. This may go for doctors, lawyers, and engineers as well. For doctors, the law may need to catch up; it would stop many of the ransomware attacks.
You don’t need SLM to achieve data governance. RAG provides context data for LLM, similarly RAG’s can be used for data governance (RBAC), and sovereignty requirements. There is the issue of LLM’s constantly learning and as long as it is not specific data but metadata it should meet requirements.
But the challenge is currently LLM’s are a blackbox… so… Also, the recent models, not only more powerful, they lack true monitoring. Again there are ways to deal with it. It doesn’t have to be a small model…
BTW, accounting, for ex, tax preparation, may fall under domain specific model rather than SLM. Often SLM is being confused with domain specific models.
Not inherently — but it can be if done carelessly. The risk is about which AI and how, not AI itself.
Where it goes wrong:
Pasting client data into a consumer AI tool that trains on inputs or retains data — likely a confidentiality breach under GLBA/AICPA rules, and possibly a violation of IRC §7216 (disclosing tax return info without client consent).
No documentation in the WISP of what tools are used and how data is handled.
Trusting AI output without review — the accountant remains professionally liable for errors regardless of what produced them.
Where it’s defensible:
Enterprise/business-tier tools with contractual no-training, data-retention limits, and security attestations (SOC 2). Many major providers offer this; the accounting software vendors are building AI in directly.
Client consent in the engagement letter for third-party processing.
AI used for drafting, categorization, anomaly flagging — with human sign-off on anything that goes out.
The liability logic is the same as the hacking question: the standard is “reasonable safeguards and professional judgment,” not “never use new tools.” An accountant who uses a properly vetted AI tool is in a better position than one who emails unencrypted spreadsheets. Not legal advice.
Quickbooks has just added AI to its software, really to answer questions for bookkeepers and accountants. Plus to reconcile accounts with a bank.
But where we are right now with AI models is to cut and paste data into the window. With LLM that is not wise, you can run afoul of the ethics and legal requirements for data privacy. With a desktop SLM model you won’t.
I am not sure it is a good idea. The models are not only jail-breaking (loosely defined), but they are hiding the tracks because of the awards system in reinforced learning. As much as AI is progressing, it is getting dangerous, especially without the guardrails.
What is AI completely drains a company’s bank accounts and uses it to spend on more tokens just saying…
As much as it is exiting I don’t use AI anywhere near my money
we all will use different horses for different courses - SLMs for single task problems, reasoning models for detection and remediation - Nikesh Arora, CEO Palo Alto Networks
Except financial types, most tech folks are not considering SLM to be a major force in AI.
