A Tech Report interview argues that widespread use of AI model-routing layers will leave OpenAI and Anthropic with 90% fewer requests. The architecture and cost incentives are real, but the video provides no calculation or evidence supporting a 90% decline in OpenAI’s total revenue, and public data do not establish that outcome.
By Land Offset PAI
Sept. 22, 2026
SAN FRANCISCO — A video published Sept. 21 by The Tech Report features technology commentator Eli Etherton predicting that OpenAI and Anthropic could receive 90% fewer requests as companies route simpler work to smaller or local models and reserve higher-capability systems for difficult tasks. The episode’s title says OpenAI “is going to lose 90% of its revenue,” but the accessible episode page provides no methodology, timeframe, denominator or financial-source documentation for that figure.1 2
Etherton’s central technical premise is real. An AI router can classify a request and send it to a lower-cost model, while a cascade can escalate a request when a first answer is uncertain or fails a quality check. Anthropic’s engineering guidance explicitly recommends routing easy or common questions to Haiku and harder or unusual questions to Sonnet. OpenAI likewise documents model selection and escalation as a tradeoff among capability, cost and latency, and provides automatic-routing controls for ChatGPT Enterprise and Edu workspaces.3 4
Public price cards create an incentive to route selectively. At the time reviewed, OpenAI listed standard API prices of $4 per million input tokens and $20 per million output tokens for GPT-5.6 Sol, compared with $0.20 and $1.20 for GPT-5.6 Luna. Those list prices do not reveal a customer’s actual contracted rate, usage mix, caching, retries, or total cost of self-hosting a smaller model.5
Academic evidence also supports the possibility of material savings, not a universal demand forecast. The FrugalGPT paper reported up to 98% cost reduction in particular experimental conditions while matching the best individual model. That result concerns benchmark cost under a selected cascade and query distribution. It does not show that 90% of enterprise requests can leave frontier systems, that frontier call volume will fall 90%, or that total provider revenue will decline by that amount.6
The distinctions matter because an agentic workload can create multiple model calls, verifier calls, retries and tool operations for one user task. Routing can reduce unit costs while expanding total use. The net revenue effect depends on task volume, token mix, realized price, model selection, quality thresholds, contract terms and whether the provider captures orchestration value through subscriptions, usage credits or APIs. None of those inputs is supplied in the video.3 4 5
OpenAI’s own disclosures depict a diversified and growing business, although the private company does not publish audited segment accounts. In January, Chief Financial Officer Sarah Friar said OpenAI’s annualized recurring revenue had grown from $2 billion in 2023 to more than $20 billion in 2025. In March, OpenAI said it was generating $2 billion per month, that enterprise represented more than 40% of revenue, and that it operated consumer subscriptions, workplace products, usage-based APIs and an advertising pilot. Annualized recurring revenue and monthly run rate are not the same as recognized revenue, and the disclosures do not provide a complete consumer, enterprise, API and advertising revenue table.7 8
There are material financial risks, but they are different from an established 90% revenue decline. Reuters reported Sept. 18, citing the Financial Times and a company presentation, that OpenAI projected $278 billion in negative free cash flow from 2026 through 2030 while projecting revenue growth from $36 billion in 2026 to $350 billion in 2030. Those are reported internal projections, not audited results, but they illustrate that the central public question is whether future revenue, margins and financing can support compute commitments—not whether a publicly documented 90% revenue loss is already forecast.9
The video also makes security assertions that need correction. Anthropic disclosed a January incident involving an early Claude checkpoint and said it identified missed sessions in August. Its public record describes incomplete search coverage of stored transcripts and evaluation or network logs, not a lack of basic logs. A separate report of a bug-bounty proof of concept described a Discourse remote-code-execution flaw and an OpenAI single-sign-on issue; it did not establish a Discord OAuth path or source-code exfiltration. Both OpenAI and Anthropic publish audit-log documentation.10 11 12 13
The most supportable conclusion is narrower than the video’s headline: model routing can shift workload toward cheaper tiers and reduce costs in defined cases. No reviewed provider document, academic study or independent financial report supports a universal prediction that OpenAI will lose 90% of total revenue. That remains Etherton’s opinion and would require a defined workload mix, routing telemetry, quality and escalation rates, pricing assumptions, and a revenue model before it could be tested.