Models Become Commoditized, Gateways Become Expensive

Core Thesis

The future of AI frontier models is not a question of 'smarter chatbots' but a question of cognitive infrastructure ownership. If techno-feudalism becomes reality, its core will not be models dominating humans, but rather a structure where rent extraction solidifies through a handful of platforms controlling compute resources, data, distribution channels, identity/payment/workflow, and regulatory certification.

My assessment is this: while models themselves are likely to commoditize rapidly, the control plane surrounding them faces a high risk of feudalization. In other words, more important than 'model monopoly' are compute monopoly, workflow monopoly, user state/memory monopoly, and agent platform monopoly.

1. Frontier Models Have Not Yet Slowed

'Scaling is over' claims remain unpersuasive. According to Stanford HAI's 2026 AI Index, over 90% of notable frontier models emerged in 2025 alone, with some surpassing or matching human baselines in PhD-level scientific questions, multimodal reasoning, and math olympiad challenges. Coding benchmarks like SWE-bench Verified also went from the 60% range to nearly 100% within one year.

Epoch AI's estimates point in the same direction. Since 2020, training compute for frontier language models has increased fivefold annually—doubling approximately every 5.2 months—while pre-training compute efficiency has improved roughly threefold per year. Some estimates suggest LLM inference costs, holding performance constant, have halved approximately every two months.

Thus, 'the technological future of frontier models' is not simply about increasing parameter counts. The frontlines ahead divide roughly into five areas.

First, test-time compute. Rather than a single larger model, what matters is spending more time thinking through problems, exploring multiple paths, and attaching verifiers.

Second, agentification. Models are moving beyond answer generators to become execution layers manipulating browsers, IDEs, databases, payments, internal systems, robots, and laboratory equipment.

Third, multimodal and world model evolution. The trajectory moves from text-centricity toward integrating images, audio, video, 3D, sensor data, and behavioral logs.

Fourth, domain-closed data. In medicine, law, semiconductors, pharmaceuticals, defense, finance, and manufacturing, internal enterprise data and experimental data become more important than open web data.

Fifth, verifiability. More than benchmark scores, 'will this agent cause incidents in actual enterprise processes,' 'is it auditable,' and 'is accountability clear' become economic adoption bottlenecks.

However, performance gaps do not guarantee permanent monopoly. The AI Index assesses that the model performance gap between the U.S. and China has effectively disappeared, citing cases where Chinese models have traded the lead with top U.S. models. As the frontier advances rapidly, catch-up accelerates as well. The core of the techno-feudalism debate lies precisely in this duality.

2. What is Techno-Feudalism?

Techno-feudalism goes one step beyond saying 'Big Tech makes a lot of money.' Capitalist profit and feudal rent differ in nature. Capitalist firms earn profits by selling goods in competitive markets. Feudal lords hold land or gateways and extract tolls, tenant fees, and licensing charges from anyone seeking to do something on or through them.

This is precisely what Yanis Varoufakis's techno-feudalism thesis addresses. He views platforms like Apple, Amazon, and Meta as having become a kind of 'cloud fiefdom,' with their owners extracting wealth through means closer to rent than profit. App store commissions, platform-locked merchants, and structures where user activity enhances platform asset value are examples. His book introduction also advances the claim that Big Tech owners are rewriting global power and economic rules like modern feudal lords.

This argument also connects with Shoshana Zuboff's 'surveillance capitalism.' Zuboff defines surveillance capitalism as an economic logic that converts private human experience into behavioral data, packaging it as predictive products for sale. In the AI era, this data extraction intensifies further. The more users interact with models, the more model providers can learn their intentions, preferences, work patterns, internal organizational knowledge, and decision habits.

What AI adds is not simple platform rent. It is cognitive rent. Past platforms controlled search, social, shopping, and app distribution. AI platforms become the interface for thinking, writing, coding, decision-making, work automation, customer service, research, and education. This creates much deeper dependency.

3. The Actual Structure of AI Techno-Feudalism

Even if AI feudalism arrives, it won't look like 'a single superintelligence on a throne.' A more realistic picture looks like this.

First, compute fiefdoms. Frontier models require GPUs/ASICs, memory, networks, power, cooling, data centers, and long-term cloud contracts. The Stanford AI Index notes that the U.S. holds 5,427 data centers—more than ten times any other country—and that most leading AI chips are manufactured by a single company, Taiwan's TSMC. Epoch AI also estimates global AI computing capacity has grown approximately 3.3x annually since 2022, with Nvidia AI chips accounting for over 60% of total compute.

Second, power fiefdoms. AI infrastructure looks like a software business but is actually also a power, real estate, cooling, and transmission grid business. Goldman Sachs estimates that AI infrastructure construction from 2026–2031 may require cumulative capital expenditure of approximately $7.6 trillion across compute, data centers, and power. While this forecast rests on strong assumptions, it's clear AI is not pure software but a massive physical infrastructure competition.

Third, data/memory fiefdoms. When users entrust their email, calendars, documents, codebases, health information, financial information, and decision logs to an AI assistant, the model provider becomes not merely a tool provider but an institution storing the user's cognitive continuity. Switching costs here far exceed those of changing search engines.

Fourth, agent gateways. Future AI will be less 'model that gives answers' and more 'agent that handles tasks.' When users accumulate accounts, payments, permissions, workflows, automation scripts, team collaboration, and audit logs within a specific agent ecosystem, that ecosystem becomes a stronger fiefdom than an app store.

Fifth, regulatory certification fiefdoms. The EU's AI Act requires transparency, copyright, and safety/security obligations for general-purpose AI models, with separate obligations for cutting-edge models with systemic risk. The EU's GPAI Code of Practice is a voluntary tool supporting such compliance, organized into transparency, copyright, and safety/security chapters. This doesn't mean regulation is unnecessary. However, as regulatory compliance costs grow, certification barriers can become disproportionately favorable to large operators.

When these five layers combine, 'AI model companies' become less like simple software vendors and more like digital lords. Users are bound not as serfs, but more precisely because they cannot leave with their work state, data, habits, automation routines, and organizational memory.

4. But the Claim 'Techno-Feudalism is Inevitable' is Wrong

There are strong counterarguments. Quite strong ones, in fact.

First, models are commoditizing with remarkable speed. Holding performance constant, inference costs are falling rapidly and training efficiency is improving. Today's frontier becomes tomorrow's basic API feature at a fast pace. Feudal fiefdoms are typically built on long-lasting scarce resources, but model performance spreads faster than expected.

Second, U.S.-China competition and open model competition are forces breaking monopoly. As the AI Index notes, the performance gap between the U.S. and China has already narrowed significantly. Add open-weight models, small specialized models, on-device inference, and model routers, and enterprises become less tied to single vendors.

Third, customers aren't fools. Enterprises are already building multi-model strategies, internal evaluation sets, in-house data layers, and vendor-switchable architectures. The smartest companies don't bet on 'who wins among OpenAI/Anthropic/Google/Mistral.' They treat models as replaceable compilers and keep the real assets—data, evaluation, workflows, permission layers—in-house.

Fourth, techno-feudalism theory often too quickly labels strong variants of capitalism as 'post-capitalism.' Platform commissions, network effects, and monopoly rents are long-standing phenomena within capitalism. Thus, diagnoses that 'capitalism is dead' may be exaggerated. A more accurate formulation is likely that feudal rent extraction within capitalism is intensifying.

My position accepts these counterarguments. Techno-feudalism is a good warning but risks becoming a bad total theory. AI's future is not a singular feudalism but rather a collision between a commoditizing model layer and a feudalizing infrastructure/platform layer.

5. The Most Probable Future: 'Models Become Commoditized, Gateways Become Expensive'

Looking toward around 2030, the most plausible scenario is this.

Model performance continues rising. But the center of differentiation shifts from 'who has the biggest model' to 'who provides the most trustworthy agent execution environment, permission management, auditing, payments, tool ecosystem, and enterprise integration.'

Closed frontier models remain. Especially in national security, advanced science, automated research, large-scale coding, cyber, bio, and robotics, top-tier models will continue to matter. However, in most commercial work, the range where open models and mid-sized models suffice will broaden.

AI companies' real moat is not model weights but user state. Entrusting email, calendars, documents, code, work history, personal memory, team knowledge graphs, customer data, and action logs to one AI ecosystem binds users to that ecosystem. At that point, 'my AI assistant' is both a convenience tool and a potential feudal fiefdom.

Nations pursue sovereign AI. The U.S., China, and EU are building different regulatory frameworks, data regimes, semiconductors, clouds, and military/intelligence infrastructures. The EU explained that GPAI rules apply from August 2025, with new models subject to enforcement after one year and existing models after two. Such trends are more likely to create a blockized AI order than a single global AI market.

Labor markets polarize. Some ordinary knowledge work sees prices plummet. Conversely, people with problem definition, verification, data ownership, system design, and organizational execution authority gain leverage. 'Individuals who use AI well' matter less than 'people who design environments where AI can work.'

6. Implications for Korea

If Korea takes this issue seriously, it must not stop at 'let's build one Korean-style hyperscale model.' That's far too shallow.

Korea's strategy should be cognitive infrastructure sovereignty rather than model sovereignty. This includes the following.

National and industrial-level shared compute pools, long-term power contracts, data center sites, cooling and transmission infrastructure, semiconductor supply chains, defense and cyber-specialized models, public sector evaluation sets, Korean language/legal/medical/manufacturing data governance, and model-switchable standard APIs are necessary. The Stanford AI Index mentions Korea shows notable innovation density in AI patents per capita. But patents and research capacity alone are insufficient. The core of feudalization is not paper counts but gateway ownership.

Korean enterprises are the same. Entrusting enterprise-wide intelligence entirely to a specific frontier model API is a strategic mistake. A better structure looks like this.

Even if models are purchased externally, keep the data layer internal. Keep evaluation sets internal. Keep agent permission management internal. Manage prompts and workflows in a vendor-neutral way. Deploy model routers. Maintain audit logs. Prepare open model fallbacks. In other words, even if model companies provide the brain, memory, limbs, and nervous system should remain self-owned.

7. Indicators for Testing Techno-Feudalism

This debate should be a testable hypothesis, not ideology. Signals strengthening the techno-feudalism hypothesis include the following.

A handful of clouds preempt most long-term GPU/ASIC contracts. AI assistants become the default interface handling users' email, calendars, payments, documents, code, browsers, and internal systems. When enterprises try to switch vendors, the cost of moving data, memory, and workflows becomes prohibitively high. Regulatory compliance costs work overwhelmingly in favor of large firms over startups. App store-style commissions are replicated in AI agent ecosystems. Users consume free or cheap services, but those interactions continually grow the platform's model, advertising, prediction, and automation assets.

Conversely, signals weakening the techno-feudalism hypothesis also exist.

Open-weight models catch up to closed top-tier models within 3–6 months. High-performance inference becomes sufficient on local devices or cheap commodity cloud. Enterprises adopt model routers and internal evals as standard. Data portability and agent memory portability are mandated by law or market standards. API margins structurally collapse. Users can easily transfer their AI assistant's memory, tools, and permissions.

Ultimately, the core question converges to one.

Is exit possible?

If you can leave, it's a market; if you cannot leave, it's a fiefdom.

Conclusion

AI frontier models will grow stronger. But more important than 'who has the strongest model' is the question 'who are the gatekeepers of the world in which that model operates?'

Techno-feudalism is not a predetermined future but a warning. The conditions under which this warning becomes reality are clear. The moment compute, data, identity, payments, workflows, agent memory, and regulatory certification become tied to a few platforms, AI becomes not a liberating technology but a rent-extraction apparatus. Conversely, if the model layer commoditizes, data and memory become portable, compute access decentralizes, and evaluation and auditing remain open, AI can become a productivity explosion rather than feudalism.

Reduced to its sharpest form, it's this.

The political economy of the AI era is shifting from 'ownership of means of production' to 'ownership of means of inference and means of memory.'

And the criterion for freedom is not how smart the model is, but whether you can discard that model anytime and still take your knowledge, memory, tools, and organization with you.