John Coogan misreads OpenAI features as a new economic model
By Klaus Berger · Reporting from Frankfurt ·
OpenAI shifts developer liabilities onto consumer subscription accounts, creating a textbook moral hazard that John Coogan mistakes for a grand economic breakthrough.
On the tech podcast TBPN, co-host John Coogan discussed OpenAI's new "Sign-In with ChatGPT" feature. This mechanism allows users to run third-party software using their own subscription tokens. Coogan remarked that it feels "potentially like the beginning of a new economic structure." His co-host Jordi Hays explained the distress that prompted this change. He recalled how early developers scaled quickly but got "stuck with like a $30,000 API bill." What Silicon Valley hails as a revolution is just a belated accounting adjustment. It merely shifts variable liabilities from fragile start-ups to centralized balance sheets.
Shifting the Marginal Cost of Computation
For three years, software start-ups operated under a dangerous financial illusion. Eager to capture market share, developers integrated token-heavy artificial intelligence features. They absorbed the variable inference costs on their own corporate accounts. They systematically underpriced operational risk. They burned venture capital to subsidize raw compute. Under the new arrangement, developers shift the marginal cost of execution. They place it directly onto the end user’s monthly subscription allowance. The industry calls this Bring Your Own Compute.
Advocates argue that offloading compute liabilities democratizes software creation. They believe lean independent teams can now ship ambitious, inference-hungry products. These include specialized language learning tools or document editors. Teams can do this without facing sudden insolvency from unexpected user spikes. For the consumer, consolidating expenses under one primary subscription tab appears cleaner. It seems better than managing fragmented API keys and credit card relationships.
Yet this arrangement creates a textbook moral hazard. When developers do not pay for compute, they face zero economic discipline to optimize their software. If a developer bears no liability for token consumption, the incentives point five years out. They will build inefficient, token-hungry workflows that drain central server capacity. Meanwhile, central infrastructure providers like OpenAI assume the unhedged risk of unmetered usage spikes. Historical patterns in utility markets show that obscured marginal costs inflate consumption. Eventually, the provider must ration access or raise baseline fees. This penalizes disciplined users to subsidize wasteful code.
Theological Delusions and Balance Sheet Realities
This appetite for magical thinking extended into the discussion of Anthropic’s overtures to the Vatican. Coogan detailed reports that Anthropic co-founder Chris Olah lobbied advisers to Pope Leo XIV. He urged the Holy See to take the possibility of model consciousness seriously. This campaign prompted tech investor Neil Khosla to warn publicly against an encroaching "AI psychosis."
The case for model consciousness rests on the notion that scaling neural networks might mimic subjective experience. This would supposedly require proactive moral guidelines. But Pope Leo XIV was entirely sound in his refusal to grant spiritual or moral standing to software. As a matter of institutional rigor, an algorithm remains a matrix of statistical probabilities. It calculates these over ingested human work. It possesses neither legal agency, operational intentionality, nor balance-sheet liability. Attempting to clothe statistical calculation in theological solemnity is a transparent effort to fabricate moral weight. It merely distracts from mundane commercial liabilities.
Software enterprises cannot evade microeconomic realities through structural abstraction or spiritual performance. Tech firms must eventually confront the balance sheet. They cannot avoid this by shifting compute liabilities or seeking papal blessings. Durable enterprise stability is built on clear rules and explicit pricing of risk. It requires boring competence, not subsidized compute and grand gestures.