(Preview) OpenAI Astride the World, Infrastructure Buildouts and Boundless Ambitions, More on Sora and Creation
Summary
- OpenAI has not earned “Mag Eight” status merely by signing roughly $1 trillion of infrastructure deals. Ben calls founder adulation based on funding rounds “on steroids”: the commitments must be “substantiated and actually paid for” before anyone writes the history books.
- Ben thinks the AI boom is “probably bubbly,” yet its momentum leaves participants little choice but to follow it. His warning sign was universal insistence that “it’s different this time”; he now wonders whether people in markets or senior leadership during the dot-com bubble also recognized that it could end badly but accepted being “swept along with the tides of history.”
- Ben says his insistence that OpenAI make choices was an error. AI is still in “astronomical, massive growth,” comparable to writing software in the 1980s or 1990s on the assumption that processors would improve: “Go for it all and trust the capacity’s going to come.”
- The apparent sprawl is a coherent attempt to make ChatGPT the agentic layer for literally everything. OpenAI’s goal is as many as 8 billion personal assistants, operating 24/7 across devices, email, calendars and even sleep monitoring; accumulated context would create “total lock-in” and require astronomical infrastructure.
- Compute scarcity is already a product and subscription risk at 800 million weekly users. The GPT-5 launch reduced thinking queries for Plus-level subscribers and directed them toward regular GPT-5, with a router deciding when it should think, making paying customers feel something had been taken away. Unlike an ad-supported service offering “a refund of zero dollars,” subscriptions carry an implicit promise.
- OpenAI’s financing and infrastructure push may become a competitive weapon as well as a funding source. Ben says it could draw marginal investment away from Anthropic, Meta or NVIDIA by “making themselves the table where the chips all go in.” NVIDIA’s $100 billion commitment followed by an AMD deal two weeks later—sending AMD’s stock up 40%—is framed by Andrew as an effort to preserve OpenAI’s leverage and reduce dependence on NVIDIA.
Deep dive
1. OpenAI is still a startup-style bet, not a Mag Eight incumbent
Andrew relays a listener’s case for placing OpenAI beside the largest technology companies and Sam Altman beside historic business leaders. Ben’s immediate response: “Roshan needs to relax.”
Signing roughly $1 trillion of infrastructure deals in a month is not the same as delivering them. Ben wants the commitments “substantiated and actually paid for” before the accolades begin.
Ben’s bubble test was the moment everyone denied a bubble and said, “It’s different this time.” He now feels both that this is “probably bubbly” and that participants have no choice but to continue, perhaps knowingly being “swept along with the tides of history.” He was only a teenager and then in college during the dot-com era, but now feels its ecosystem dynamics more directly and wonders whether market participants and senior leaders then had the same acceptance that it would end badly.
2. Compute shortages are colliding with subscription promises
Ben’s former prescription was to deemphasize OpenAI’s API because Microsoft, its partner, offers “the exact same API in Azure” and already knows how to serve enterprises.
The GPT-5 launch sharpened that criticism: OpenAI reduced thinking queries for Plus users and relied more heavily on regular GPT-5 with router technology deciding when to think. Ben would ignore many complaints, but not paying customers believing they bought something that was later withdrawn—“That’s a big problem.”
Andrew points to 800 million weekly users; Ben’s answer is that every broad rollout now demands a step-change in capacity. OpenAI increasingly has Apple’s scale problem: technology may exist before it can be made and shipped to everyone at scale.
3. Ben’s demand for focus was applying mature-company logic too early
Ben’s Microsoft analogy begins with a company whose Windows position was eclipsed by smartphones and that had to choose a new direction. His earlier criticism of Steve Ballmer was that Microsoft’s devices push made little sense; the company needed to focus on services, serve iOS and Android, and prioritize the cloud. Satya Nadella split the Windows team apart to curb its internal power; reunion became safe only once “no one at Microsoft thinks Windows is the future of the company.”
His correction: OpenAI and AI are not yet at that mature stage. Just as developers once assumed processors would become faster instead of optimizing around current limits, OpenAI should trust that capacity will arrive rather than constrain the vision.
“Go for it all and trust the capacity’s going to come” is therefore both permission and rationale for the infrastructure deals. OpenAI is not operating like a Mag Seven incumbent selling financial results, but like a startup selling “a dream and a vision of the future”—thrilling, though it could “very well” end in disaster.
4. ChatGPT is being built as the layer above every platform
Andrew traces ambitions that have broadened every three or four months over the last 1.5 years, from challenging Google to hardware, competition with Meta, entertainment and apps inside ChatGPT—effectively an OS layer—and asks whether OpenAI is attempting too much.
Ben rejects the “spaghetti against the wall” interpretation. With a wide enough aperture, the strategy is “remarkably cogent”: become “the agentic AI layer for literally everything,” or the “Windows of everything.”
WeChat may be the stronger analogy because its mini apps, payments and government services let it sit above the phone itself, diminishing the iPhone’s platform power. ChatGPT similarly aims to travel across every device and context.
The endpoint is a goal of 8 billion personal assistants continuously gaining contextual knowledge about their users through email, calendars and other connectors. The assistant could operate 24/7—even monitoring sleep—and its unique contextual memory would produce “total lock-in”; serving that ambition explains the extraordinary infrastructure demand.
5. Capital dependence and chip diversification reinforce the strategy
Ben sees the aggressive infrastructure and investment push through a game-theory lens: it may starve competitors by diverting marginal investment toward OpenAI rather than Anthropic, Meta or NVIDIA. By becoming “the table where all the chips go in,” OpenAI could make itself the central destination for capital; Altman is good at sketching how the pieces link and is willing to “risk it all.”
NVIDIA’s Jensen Huang committed $100 billion to OpenAI, then OpenAI announced a deal with NVIDIA’s direct competitor AMD two weeks later, sending AMD shares up 40%. Huang said he was surprised AMD was so excited about its next chip and that it would “give away 10% of their company for someone to buy it,” adding, “it’s very clever.” Ben says Altman gave him the impression that NVIDIA had not known about the deal.
The underlying mechanism is Clayton Christensen’s “law of conservation of attractive profits”: one differentiated bottleneck tends to dominate a value chain while surrounding suppliers and customers lose pricing power. Andrew frames the AMD move as an attempt to preserve OpenAI’s leverage and avoid being too reliant on NVIDIA.