Algorithmic Iron Curtain III: As Models Grow Larger, Does Knowledge Shrink? The Crisis of AI Industry Concentration

# Algorithmic Iron Curtain III: As Models Grow Larger, Does Knowledge Shrink? The Crisis of AI Industry Concentration

When five companies control global compute, does knowledge democracy still exist?

Author: 阿克斯星門 / Assistant Professor, Center for General Education, Chung Yuan Christian University

Editorial Preface

Generative AI appears as a chat tool available to everyone, but underneath lies a highly concentrated compute economy. When model training depends on expensive chips, cloud data centers, and vast amounts of electricity, knowledge production is no longer only about who has ideas, but who has GPUs, who has power, and who has capital. This article continues the 'Algorithmic Iron Curtain' series, examining power concentration in the era of large models, and whether public knowledge can retain democratic character under compute monopoly.

Model Democratization Often Only Reaches the Login Page

The most enchanting myth about generative AI is: everyone can use models, so knowledge has finally democratized.

This statement is only half right. Yes, users can ask questions on their phones, call AI to write summaries, revise copy, or create presentations. But from 'using a model' to 'creating a model', there lies a very real trench: chips, electricity, data centers, cloud contracts, and a series of bills that make accountants suddenly miss the abacus.

In the past, intellectuals debated who controlled publishing machines; later we debated who controls search engines; today we must ask: who controls compute?

Compute Becomes New Land

In the AI industry, compute is no longer just a technical resource, but new-era land. Without land, farmers cannot cultivate; without compute, researchers with ideas can only write elegant models on whiteboards that won't run.

Large models require expensive GPUs, high-speed networks, cooling systems, stable power, and engineering teams. These conditions are inherently uneven. A few enterprises and cloud platforms can secure the latest chips, sign long-term supply contracts, and build data centers whose electricity consumption approaches a city's total. Universities, public research institutions, small media outlets, and local communities are often left in the position of 'you may use APIs, but don't ask where the models come from'.

This is the quietest side of the algorithmic iron curtain: it does not need to block you. It only needs to make you unable to afford rent.

Do Larger Models Necessarily Mean More Knowledge?

The AI industry often equates 'big' with progress: larger models, more parameters, longer context windows, and bigger data centers. But knowledge does not necessarily become freer as models grow. On the contrary, larger models narrow entry points.

When model training costs are highly concentrated, public knowledge may be repackaged into paid interfaces. Human-accumulated text, images, audio, code, and educational resources are absorbed by platforms and returned to society via subscription. This resembles a giant restaurant sweeping up all produce from the public market, cooking it into set meals, then telling you: congratulations, this is innovation.

Of course, corporate investment in R&D is not a sin. The problem lies in whether education, research, media, and public services increasingly depend on a few models; then model errors, policy changes, price adjustments, or service outages become risks to the public knowledge infrastructure, not merely commercial issues.

Open-Source Models Are Not Panaceas, But They Matter

Open-source or open-weight models provide a counterforce. They allow more researchers to inspect model behavior, enable local languages, educational contexts, and small professional domains to build their own fine-tuning systems, and free communities from perpetual dependence on single platforms.

Yet we cannot be naive. Open-source models without compute, data governance, maintenance resources, and safety assessments may only be beautiful showpieces. True public AI infrastructure is not merely downloading a model; it requires someone who can train, deploy, maintain, evaluate, update, and subject data sources and usage rules to public oversight.

Taiwan Should Ask Not 'Does It Have AI' but 'Who Has AI'

Taiwan has advantages in the semiconductor supply chain, yet still needs clearer institutional design for public compute, educational models, cultural data governance, and local-language AI. Especially if Indigenous languages, local knowledge, educational content, and public media data enter the AI era, they cannot rely solely on platform goodwill or one-off tenders.

We need a public compute pool, a local-model test bed, language data licensing regimes, community co-governance, and foundational services usable jointly by universities, local governments, Indigenous communities, and media. Otherwise, the most ironic future scene may be: Taiwan produces the world's most important chips, yet its public knowledge continues to rent someone else's model brain.

The Next Front of Knowledge Democracy

The AI problem is not just whether models will answer, but who can ask questions, who can train, who can correct, and who can refuse to be used for training.

True knowledge democracy means every society can participate in designing the knowledge infrastructure. Otherwise, so-called intelligent eras may simply become new versions of cloud feudalism: lords reside in data centers, commoners pray within monthly subscription plans that prices never rise.