AI giants like OpenAI and Google risk repeating IBM's 1980s mistake, where technical superiority lost to accessible, user-controlled personal computers. Local Small Language Models (SLMs), enabled by local embedded data and improving hardware, could disrupt centralised foundation models by offering privacy, speed, and cost advantages once bundled software eliminates complexity barriers. As someone who was using desktops in 1984 it all seems familiar to me.
Technical Superiority Isn't Enough
In 1981, IBM dismissed PCs as inferior to mainframes, which processed 1,000 transactions per second versus PCs' 10% performance. Yet PCs won by being "good enough" for business needs—VisiCalc spreadsheets, WordStar processing—whilst offering user control over IBM's $10,000-50,000 monthly rental fees. Market dynamics shifted permanently when PCs crossed the functionality threshold around 1984, when I started using them as a consultant at McKinsey.
Foundation Models Repeat Mainframe Mistakes
AI giants operate like 1970s IBM, renting access to massive, centralised systems in their search for godlike AGI. OpenAI charges $20/month for ChatGPT Plus whilst API usage generates additional per-token fees. Their assumed moats - training costs ($100-500 million for frontier models), talent concentration, infrastructure requiring thousands of H100 GPUs - mirror mainframe-era thinking. Industries now depend on these computational providers as rental landlords.
Small Models Outflank
Genuinely open SLMs signal opportunity. Chinese models like Qwen-2.5 use Apache 2.0 licensing allowing unrestricted commercial deployment, unlike Meta's Llama which restricts commercial use for services. Retrieval augmented generation enables SLMs to access external data, offering offline operation and instant responses.
Current limitations are substantial: most local SLMs require 16-24GB RAM (beyond most business laptops), score 65-75% on reasoning benchmarks versus GPT-4's 85-90%, and lack integrated search or multimodal capabilities. However, these represent packaging problems rather than fundamental barriers and their performance is improving. Bundled applications could integrate base models, hosting, embedding systems, search, sound and vision capabilities behind unified interfaces—eliminating complexity without requiring technical expertise, as is currently the case. In addition, network effects, data advantages and organisational syrup may delay the displacement of LLMs by SLMs.
Economics Favour Local Deployment
Hardware costs ($2,000-$4,000 for capable systems) compete favourably against subscription models for businesses spending $200-2,000 monthly on AI APIs. Local inference approaches zero marginal cost after hardware purchase. Foundation model companies face potential stranded assets as billions invested in infrastructure become worthless if inference migrates to edge devices.
Transition Accelerates
Hardware improvements favour distributed systems: Apple's M4 provides 2x the neural processing power of M1, whilst model compression (quantisation) achieves 4-8x size reductions with minimal performance loss. Hybrid architectures - local models handling routine tasks, cloud for complex reasoning - provide a transition pathway for Foundation Model firms whilst preserving some centralised revenue. Expect Google, OpenAI and Anthropic to produce something like this soon.
The shift to SLMs could have significant stock market implications where most recent gains have been tech led. AI's PC moment approaches as bundled software could eliminate complexity barriers. Incumbents must adapt beyond technical superiority or risk repeating IBM's strategic miscalculation.