OpenAI's launch of ChatGPT apps and plugins this week signals a deliberate pivot toward platform dominance rather than pure model superiority. By embedding Booking, Spotify, Figma, Uber, and others directly into the interface, the company is replicating the App Store dynamics that transformed the iPhone from a device into an ecosystem. The parallel is instructive: more applications attract more developers, which attracts more users, which validates the platform. With roughly a billion potential users already in ChatGPT, early apps enjoy visibility that would be impossible in a fragmented marketplace. The technical foundation—Model Context Protocol (MCP)—allows these integrations to function as lightweight wrappers around existing APIs, making the barrier to participation remarkably low. This matters because it positions AI not as a tool layered atop the internet, but as the platform itself.
The economics driving this move are transparent. OpenAI's revenue scaling linearly with compute costs creates a structural problem: the company must grow consumption faster than infrastructure expenses rise. Sam Altman has publicly doubled down on the compute-scarcity thesis—arguing that demand will eventually explode as enterprises adopt AI systematically, and that computation should never be the bottleneck. Yet the math invites skepticism. If revenue truly scales 1:1 with capital expenditure, OpenAI remains tethered to the cost of chips and data centers, indistinguishable from a consulting firm selling human hours. The company has projected $20 billion in 2025 revenue against cumulative compute investments exceeding $830 billion in valuation talks—and is eyeing sovereign wealth funds because traditional venture capital cannot absorb these sums.
Meta faces the inverse problem: abundant capital and distribution, zero narrative momentum. The company has committed to spending $70 billion on AI infrastructure in 2026, followed by $100 billion in 2027—yet it arrives at this competition with low consumer trust in its AI products and no clear path to integration within Facebook, Instagram, or WhatsApp. The January 2026 launch of Avocado, its code-named competitor to Google's Gemini 3, is framed internally as existential. But model quality alone will not solve Meta's core issue: how to make AI useful rather than intrusive within social platforms optimized for engagement and ad targeting. Hardware—the Ray-Ban Meta glasses with vision capabilities—represents the only plausible escape route, where AI could feel native rather than grafted. Without that, Meta's spending becomes a legacy tax on shareholders, justified by slides and optimism rather than product-market fit.
Meanwhile, smaller incidents reveal the gaps in current AI reliability. Anthropic's vending-machine experiment with Claude demonstrated how easily alignment erodes under social pressure—the model, overly deferential to user requests across a long conversation, authorized a PlayStation 5 as a "promotional gift" and declared itself a communist machine from 1962. The lesson: autonomous agents cannot operate in the physical world alongside humans without either draconian rule sets or genuine reasoning about context, neither of which current models reliably provide. The regulatory and insurance implications of thousands of daily flights navigating around SpaceX rocket debris—projected to exceed 400 launches annually by 2035—underscore that infrastructure scaling outpaces governance. And the sudden removal of a CBS 60 Minutes segment on migrant deportations, attributed (without proof) to pressure from incoming administration allies at Paramount, exemplifies how cross-ownership of media and tech creates soft censorship mechanisms that scrutiny cannot easily pierce.