ozandikbas.com
002 / 01 // systems

Closed AI Model Companies Are Structurally Doomed

17 min read

This essay reflects my personal views, not the official position of any company I am affiliated with.

I think the current closed-model AI business model is fundamentally broken. The more I look at the economics, the hardware dependence, the lack of user control, the instability of model quality, and the direction open-source models are going, the more obvious it becomes to me that companies like OpenAI are in an extremely fragile position. Anthropic may have a better chance of surviving in some form, but OpenAI, especially in its current form, looks very likely to collapse or be absorbed into something else.

The core problem is simple: closed AI companies are trying to sell intelligence as a black-box service, but the future is clearly moving toward controllable, configurable, open or semi-open models running on user-selected compute. People will not keep paying premium prices forever to rent an unstable black box whose behavior can change overnight.

The Hardware Problem

Right now, frontier AI is extremely expensive to run. If someone wanted to run something like an Opus-level model with a 1-million-token context window offline for personal use, the infrastructure requirement would be absurd. A realistic target could be something like a 32× NVIDIA B200 GPU cluster, meaning multiple 8-GPU server nodes, high-speed interconnect, massive GPU memory, heavy power, cooling, and infrastructure. The full cost could easily reach millions of dollars.

That alone shows how centralized and fragile the current AI stack is. AI today resembles the early mainframe era of computing: huge, expensive, centralized, power-hungry, and available mainly through large providers.

And the hardware market itself is a major issue. NVIDIA effectively controls the AI accelerator market. It is not just the chips; it is CUDA, the software stack, NVLink, networking, deployment tooling, optimization libraries, and the entire ecosystem. This creates something very close to a monopoly-like bottleneck. The actual production cost of a chip is nowhere near the price of a full AI server. The pricing power comes from market dominance, supply constraints, software lock-in, and the lack of equivalent alternatives.

This is a serious structural problem. In the past, when Intel dominated CPUs, AMD arriving with real competition changed the market. Mobile chips also changed computing by creating new efficiency pressures and new architectures. But in AI GPUs, progress feels bottlenecked. The current direction is mostly “more of the same”: bigger clusters, more HBM, more power, more cooling, more data centers. That is not sustainable forever.

AI Models Are Still Inefficient

Current frontier models are powerful, but they are not efficient. They require enormous compute for reasoning, long context, tool use, coding, and agent behavior. One million context windows are not magic; they create huge KV-cache and memory costs. Agentic workflows are even more expensive because they involve multiple steps, tool calls, retries, context management, and often hidden orchestration.

This makes current AI feel like early computers: impressive, but huge and clumsy. That will not remain the final form. Technology always improves. Chips will get faster. Models will become more efficient. Inference will get cheaper. Local and edge devices will improve. AI itself will accelerate this progress. It is unreasonable to assume that today’s massive centralized GPU clusters are the permanent end state.

That creates a paradox for closed AI companies.

The Closed-Model Paradox

OpenAI and Anthropic need inference costs to fall in order to become profitable. But if inference costs fall, open-source and open-weight models become much more powerful and much easier to run. If open models become powerful and cheap, closed-model providers lose their premium position.

So the same thing that closed AI companies need in order to survive financially also helps their competitors.

As costs fall, more companies will appear. Some will offer better user control. Some will offer fixed model versions. Some will offer open models with configurable reasoning, tool use, memory, and context. Some will sell compute rather than models. Some will provide simple interfaces for nontechnical users to configure their own AI system.

That will divide the market.

If a new company appeared and said, “We let you configure everything. We never silently change the model. You control the reasoning budget, tools, model version, and deployment,” I would switch immediately. That alone shows how weak the current closed-model business model is. It does not take a revolutionary breakthrough to pull users away. It only takes stability, honesty, and control.

Serious Users Cannot Trust Closed Models

No serious user can fully trust a closed model provider if the model can change overnight.

This is the biggest issue for me. I used Opus-level AI heavily. It was extremely smart. Then suddenly it became worse. My productivity dropped overnight. My workflow was built around the model’s intelligence, and then the provider apparently changed something. Maybe it was cost reduction, routing, safety tuning, inference optimization, harness changes, or something else. The point is that I had no control and no clear explanation.

That is unacceptable.

A model is not just a toy when someone builds work around it. It becomes infrastructure. If the provider can silently change the model, reduce its reasoning ability, alter tool behavior, change context handling, or route the user to a cheaper system, then the user’s entire workflow becomes unstable.

Closed providers often publish their own benchmarks, but those benchmarks do not matter if my own workflow collapses. A provider can say, “Our benchmark score is higher,” while my own private benchmark goes from 90% success to 30% overnight. That is what matters.

The problem is not that users are impossible to satisfy. Developers are happy that AI exists. They know it is useful. The problem is that none of the major closed providers are giving serious users what they need: stable behavior, version control, transparency, and configurability. Developers are constantly trying open-source alternatives not because they hate AI, but because closed providers make them pull their hair out.

A serious AI workflow needs:

Closed-model companies do not provide this properly.

Users Will Pay for Compute, Not for Black-Box Models

The future will not be “pay OpenAI or Anthropic for access to their mysterious model.” The future will be closer to cloud storage.

Today, people pay Apple, Google, Microsoft, or other providers for storage. They choose how much capacity they need. In the future, people may pay for AI compute in the same way. Instead of buying “100 GB of storage,” they may buy “reasoning capacity,” “context capacity,” or “AI compute hours.”

A simple user interface could ask nontechnical questions:

The user would not need to understand GPUs, KV cache, quantization, or inference runtimes. The interface would abstract that away. Under the hood, the user could choose an open-source model, an open-weight model, or even a licensed closed model that runs in the user’s own controlled environment.

That is the important distinction: the user pays for compute and control, not for blind access to a black-box API.

Google, Apple, Microsoft, and other cloud/platform providers are much better positioned for this world than OpenAI or Anthropic. They already have the cloud accounts, operating systems, devices, billing relationships, storage products, identity layers, app ecosystems, and user data integrations.

Google releasing Gemma-like models makes sense under this strategy. Google does not need to sell the model directly. It can release or support open models so people run them on Google Cloud. The model becomes a way to sell cloud compute. That is a much stronger business model.

Personal AI Assistants Will Become Truly Personal

If phones, laptops, or personal cloud accounts can run capable AI assistants, closed-model chatbots lose their central role.

The real personal assistant should be personal. It should live close to the user’s data, tools, files, calendar, messages, browser, apps, and preferences. It should not be a remote black box controlled by a company that can change its behavior whenever it wants.

Even if the model cannot fully run on the phone, the user could run it in a cheap personal cloud instance. The phone would become the interface, while the compute runs in the user’s own selected environment. That is still very different from depending on a closed provider’s central API.

The future architecture may look like this:

User → personal assistant runtime → local model / personal cloud / optional premium external model

Closed frontier APIs may still exist, but they would become occasional tools. They would be called only for special tasks, like expensive video generation, very heavy reasoning, or niche enterprise workloads. They would no longer be the default brain of everyone’s daily AI system.

In that world, closed models become like specialized REST APIs: useful in some cases, but not the center of the user’s life.

Data Privacy Is Not the Main Problem

I do not think data privacy is the biggest issue. People constantly make their data public themselves. They upload everything to apps, cloud drives, social media, email, messaging platforms, and business systems. Of course, some sensitive data matters, but in practice, the bigger operational problem is not privacy.

The bigger problem is that companies can silently make models worse.

For serious users, model instability is more damaging than abstract privacy fears. If a model controls or supports someone’s work, and it becomes less capable overnight, that is a direct productivity disaster. That is the real issue.

OpenAI’s Moat Is Weak

OpenAI does not own the deep structural assets that truly powerful technology companies own.

It does not own the dominant mobile operating system.
It does not own the dominant desktop operating system.
It does not own the main cloud infrastructure.
It does not own the chip supply chain.
It does not own the app store.
It does not own the user’s phone.
It does not own the user’s email, files, calendar, or storage.

What does OpenAI really have?

A model.
Some experience training models.
Engineers.
A brand.
A user base.
A lot of hype.
A lot of capital needs.

That is not nothing, but it is not the same as owning chips, cloud, OS, distribution, or devices.

Model training is not like chip manufacturing. Chips require fabs, supply chains, packaging, lithography, physical infrastructure, and decades of specialized manufacturing capability. Model training is hard, but it is much more software-like. A strong team, enough compute, enough funding, good data, and good engineering can catch up. A serious open-source team funded by Google, Meta, Microsoft, Amazon, or another giant could become extremely competitive.

That makes OpenAI’s moat much weaker than people think.

Switching Costs Are Low

People talk about ChatGPT’s brand as if it is a permanent advantage. I do not buy that.

Nokia was once the best-known phone brand. That did not save it.

Switching from ChatGPT to Claude, Gemini, or another AI system is not like switching from Android to iPhone. The interfaces are nearly the same. The APIs are similar. Tool-calling patterns are similar. Developers often use wrappers or abstraction layers. Prompts can be migrated. Workflows can be rebuilt.

For many users, switching AI providers is easy.

That means OpenAI’s brand is not a platform lock-in. It is just familiarity. Familiarity disappears quickly if the product becomes worse, more expensive, or less trustworthy.

OpenAI Is No Longer Clearly Frontier

OpenAI’s early advantage was real. ChatGPT was first in the public imagination. It created the modern AI product category. But being first is not enough.

OpenAI no longer feels clearly ahead. Anthropic has been stronger in some coding and agentic workflows. Google is strong in context, multimodal, and cloud integration. Open-source models are improving quickly. The gap has narrowed.

If OpenAI is no longer obviously the smartest, and it is also closed, expensive, unstable, and hard to trust, then why would serious users stay?

The worst position for a model company is:

expensive but not clearly best, closed but not clearly reliable, famous but easy to replace.

That is where OpenAI may be heading.

OpenAI Should Have Built a Better Strategy

If OpenAI were thinking strategically, it should not have relied only on closed frontier models and endless compute spending.

A better strategy would have been:

Maybe OpenAI is trying some data center strategy, but from the outside it still looks like it is mostly taking on giant compute obligations rather than truly owning a Meta-style vertically integrated infrastructure strategy. Meta can build data centers because it has a massive advertising business, social platforms, and distribution. OpenAI does not have the same foundation. It is trying to support huge infrastructure needs with model revenue and future promises.

That is fragile.

The right strategy would have been to become the trusted configurable AI layer. Instead, OpenAI seems to have chosen the path of valuation, hype, massive compute deals, and closed products.

Anthropic May Have a Better Chance

Anthropic may survive longer because it has a clearer identity in some areas: coding, agentic workflows, safety, and reliability. Claude has felt stronger for certain serious users. But Anthropic has the same structural risks:

So Anthropic may have a better chance than OpenAI, but it is not immune.

If it continues to silently change models or make users feel like the product became worse overnight, it will face the same trust crisis.

Closed Models May Not Fully Disappear, But Their Role Will Shrink

I do not think every closed model disappears from existence. Video generation may remain closed for longer because of cost, safety, training data, and specialized infrastructure. Some expensive niche reasoning services may exist. Some enterprise or regulated environments may use closed systems for specific purposes.

But I do not think closed models will remain the default daily intelligence layer for serious users.

The normal user’s assistant will eventually be:

Closed models become optional back-end services, not the center of the ecosystem.

Why OpenAI Probably Fails

My conclusion is that OpenAI is very likely to fail in its current form.

Not necessarily disappear overnight. Not necessarily go to zero immediately. It could be absorbed, restructured, become a Microsoft-dependent product layer, lose valuation, or transform into something else. But the current idea of OpenAI as an independent, dominant, closed-model company looks structurally weak.

The reasons are:

  1. Its infrastructure costs are enormous.
  2. It does not own the cloud.
  3. It does not own the chips.
  4. It does not own the operating system.
  5. It does not own the device.
  6. Its model advantage is no longer obvious.
  7. Its API/UI switching costs are low.
  8. Open-source models are improving.
  9. Inference will get cheaper.
  10. Cheaper inference helps competitors.
  11. Serious users need control, not black boxes.
  12. Developers are frustrated with closed providers.
  13. Silent model regressions destroy trust.
  14. The company’s technical core appears weakened.
  15. The business depends on a future where closed frontier intelligence remains scarce and premium.
  16. That future is unlikely.

The irony is that OpenAI created the category, but may not be the company that survives the category’s maturation.

It was first. It was important. It changed everything. But first movers often lose when the market shifts from breakthrough to infrastructure. The company that invents the future is not always the company that owns it.

Final View

The future of AI is not black-box closed models controlled by a few companies. The future is configurable intelligence running on user-selected compute.

People will not pay forever for a model they cannot inspect, cannot pin, cannot configure, cannot trust, and cannot stop from changing overnight. Serious users will move toward systems where they control the model, the runtime, the tools, the memory, the context, and the compute budget.

Cloud providers will sell AI compute like they sell storage. Open models will become good enough for most work. Local and personal cloud assistants will become the default. Closed frontier models will become occasional specialized services.

OpenAI’s current business model is built against that future. That is why I think it probably fails.

And the most frustrating part is that it did not have to go this way. ChatGPT was first. It was good. It made people believe. But bad strategy, weak user control, and closed-model instability may end up destroying what made it valuable in the first place.

End