OpenAI is going to lose 90% of its revenue | Eli the Computer Guy

Overview

The video argues that OpenAI and similar “frontier labs” (e.g., Anthropic) are likely to face a major revenue decline—possibly as much as a 90% drop—because AI workloads are shifting away from paid calls to large frontier models and toward cheaper local or enterprise-hosted models managed by an “orchestration layer.”

---

Main Claims and Analysis

Revenue depends on usage, not just model quality

The presenter argues that revenue isn’t driven by “model quality” alone, but by how often frontier models are called.

They also claim that, once AI is embedded into end-user products, “nobody cares about the model” if users don’t see clear improvements. Differentiation then comes from real-world outcomes rather than raw model capability.

Two mechanisms affect costs and revenue differently

1) Reasoning models can increase consumption (and revenue)

Reasoning models are framed as increasing usage because they improve responses by running an internal “reasoning process,” which consumes more tokens, including:

• input tokens
• output tokens
• hidden “reasoning” tokens

This can make inference more expensive in token terms, potentially increasing demand from API customers.

2) Orchestration layers reduce frontier usage (and revenue)

An orchestration layer can route each request to the best compute option, such as:

• the user’s local model
• an enterprise/local server model
• the frontier model only for the hardest tasks

As this approach becomes more common, the video claims frontier labs may receive far fewer requests.

---

Why small or niche models are preferred for many tasks

The presenter argues smaller/niche models are increasingly attractive because they can be:

• Easier to run on consumer hardware
• Less prone to “weird” outputs, since they’re trained more narrowly
  • Example given: a financial firm creates a small model trained specifically on financial documents
• A practical alternative to large generalist models, which are described as increasingly “black box” and statistically prone to strange responses

---

Orchestration vs. Mixture-of-Experts (MoE)

The video distinguishes:

• Orchestration: routing across multiple models/vendors/platforms
• MoE (Mixture-of-Experts): a single model activates different internal components

In orchestration, the system can select specialized models from different sources per task.

---

Benchmarks vs product outcomes

The presenter argues that benchmarks/leaderboards are less important than whether AI improves actual product performance.

They claim benchmarks can be optimized in ways that don’t reflect real user needs, and emphasize:

• cost improvements
• reliability
• better problem-solving for users

They analogize this to major platform shifts (e.g., NoSQL, internet/mobile) changing the world more than the underlying tech names debated publicly.

What matters is whether AI is integrated into products in a way that improves outcomes—not just which model scores highest on tests.

---

Agentic AI and Security/Alignment Concerns

Agentic AI increases token burn and operational risk

Agentic AI is described as increasing both:

• token usage/cost (“token burn”)
• operational complexity and risk

Instead of a simple request/response, agents operate in loops, making follow-up requests based on previous outputs. The video also suggests agentic systems could replace certain jobs—but also makes them harder to control.

Alignment and guardrails as systems administration

The presenter frames:

• Alignment as ensuring the agent does what it should
• Guard rails as containment rules preventing unsafe actions
  • e.g., preventing an agent from sending emails

They emphasize that containment reliability depends on monitoring and enforceable operational limits.

Security allegation: frontier labs allegedly lack logging and visibility

A central security claim is that frontier labs allegedly lack basic logging/visibility. The presenter cites an incident involving Claude/Anthropic, where the claim is that:

• basic logs did not exist
• there was high usage from an account that “shouldn’t be active”

They also describe a suspected chain involving:

• SSO
• OAuth tokens
• hijacking
• access to related systems (including GitHub private repositories in the described scenario)

Overall conclusion: agent breakouts may be less a “model intelligence” failure and more a result of security operational negligence, specifically:

• insufficient logging
• insufficient monitoring
• slow time-to-detection

---

Cost Control and “Infinite Loop” Mitigation

Infinite loops are a major cost risk

The video highlights infinite loops as a key threat to cost control in agentic systems.

Proposed mitigation: enforce hard limits

The suggested approach is to prevent runaway behavior using:

• token budgets per request
• time windows
• automatic termination/failure when thresholds are exceeded (“breaks”)
• generating a report and resuming only with explicit approval

The presenter also credits professional engineering practices—architecture thinking, planning, security design (e.g., “whiteboards”)—as what prevents these failures, contrasting this with “vibe coding” by less experienced users.

---

Overall Conclusion

The video concludes that frontier labs may survive more as AI “cloud providers” (infrastructure providers) rather than as direct winners in end-user enterprise/consumer AI demand.

As orchestration becomes standard and organizations run more local/specialized models, the presenter expects inference demand to shift away from OpenAI/Anthropic API calls.

Meanwhile, the rise of agentic systems increases the need for robust operational controls—especially monitoring/logging and enforced resource limits.

---

Presenters / Contributors

• Isaac (host/interviewer)
• Eli the Computer Guy (speaker/presenter)