AI App Crisis, OpenAI Does Math, Big Nvidia Deal

Revenue Cat, a subscription infrastructure provider serving more than 75,000 developers, published its 2026 State of Subscription Apps report, drawing on analysis of more than 1 billion in-app subscription transactions representing approximately $11 billion in annual developer revenue. The dataset spans iOS, Android, and web applications, making it one of the more comprehensive looks at where subscription money is actually moving. The core finding is stark: AI-powered apps churn subscribers at meaningfully higher rates than non-AI apps across nearly every subscription interval. At the 12-month mark, AI apps retain only 21% of subscribers acquired on day one, compared to 30.7% for non-AI apps—a gap of nearly 10 percentage points. On a monthly basis, AI apps retain 6.1% of subscribers versus 9.5% for non-AI apps. Refund rates follow a similar pattern, with AI apps seeing median refund rates of 5.3% compared to 4.2% for non-AI apps, and outlier AI apps reaching refund rates as high as 15.6%. Despite weaker retention, AI apps outperform on monetization speed and realized revenue per user. Median download monetization for AI apps sits at 2.4% versus 2.0% for non-AI apps. Monthly realized lifetime value (LTV) is approximately $19 per user for AI apps compared to roughly $13.50 for non-AI apps; on an annual subscription basis, the gap widens to $30 versus $20. The pattern suggests AI apps extract more revenue per acquired user but struggle to sustain engagement long enough to compound that value. The discussion attributes the retention problem primarily to overpromising and underdelivering—a dynamic accelerated by the pace of AI hype cycles. When a product fails to meet inflated expectations on first use, users rarely return, even if the underlying technology improves. The report also highlights that only 27% of apps in the dataset are categorized as AI-powered, meaning the majority of the subscription app market has not yet integrated AI features. Category adoption is uneven: 61% of photo and video apps incorporate AI, while gaming sits at just 6.2%, travel at 12.3%, and business apps at 19.1%.

OpenAI has launched dynamic visual explanations within ChatGPT, a feature designed to make mathematical and scientific concepts manipulable rather than static. Instead of receiving a text description or a fixed diagram, users can adjust variables directly—for example, modifying the sides of a triangle to observe real-time changes to the hypotenuse—and watch equations update accordingly. The feature currently covers more than 70 math and science concepts, including compound interest, exponential decay, Coulomb's law, Ohm's law, kinetic energy, and Hooke's law. The strategic rationale is clear given the scale of existing usage: OpenAI reports that more than 140 million people use ChatGPT weekly for math and science assistance, out of a broader base of approximately 900 million weekly users. The feature represents a shift from answer delivery toward conceptual exploration, with potential implications for tutoring access and STEM education at scale. Google's Gemini introduced comparable interactive diagram functionality in late 2024 as part of its own push into education, signaling that interactive visual reasoning is becoming a competitive battleground among frontier AI platforms.

Thinking Machine Labs—the AI startup founded by former OpenAI CTO Mira Murati—has announced a multi-year strategic partnership with Nvidia to deploy large-scale computing infrastructure beginning in 2027. The agreement includes deploying at least 1 gigawatt of Nvidia's Vera Rubin AI systems, one of Nvidia's newest compute architectures. Nvidia is also making a direct strategic investment in the company. Thinking Machine Labs has raised more than $2 billion since its founding last year and is valued above $2 billion. The company's stated focus is building AI models that produce more replicable and reliable outputs. Its first commercial product, an API called Tinker, launched last year. The 2027 start date for the compute deployment suggests the company's most significant product scaling is still ahead. The deal fits a broader pattern of aggressive infrastructure commitments across the AI industry. Nvidia CEO Jensen Huang has projected that the industry will spend $3 trillion to $4 trillion on AI infrastructure by the end of the decade. OpenAI's $300 billion compute partnership with Oracle, announced last year, represents the high-water mark of such agreements, but mid-stage AI companies are increasingly following suit as access to compute becomes a primary competitive constraint.

- AI apps monetize downloads faster and generate higher per-user LTV than non-AI apps, but 12-month subscriber retention (21%) lags non-AI apps (30.7%) by nearly 10 percentage points—a gap the data attributes largely to overpromising and underdelivering on product capabilities. - Only 27% of subscription apps in Revenue Cat's dataset are AI-powered, indicating that the majority of the app market has not yet integrated AI features; photo and video lead adoption at 61%, while gaming (6.2%) and travel (12.3%) lag significantly. - ChatGPT's new interactive visual explanation feature—covering 70+ math and science concepts with real-time variable manipulation—signals a strategic move by OpenAI to deepen utility and reduce churn among its 140 million weekly math and science users. - Thinking Machine Labs' 1-gigawatt Nvidia compute deal, starting in 2027, reflects an industry-wide pattern in which AI companies are securing infrastructure capacity years in advance, with Nvidia's Jensen Huang projecting $3–4 trillion in AI infrastructure spending by decade's end. - The convergence of retention data, product feature expansion, and infrastructure investment points to a maturing phase in AI commercialization where durable utility—not novelty—will determine long-term winners.