Why Does OpenAI Need a 'Stargate' Supercomputer? Ft. Perplexity CEO Aravind Srinivas

OpenAI’s planned “Stargate” supercomputer is framed as a compute arms race and an AGI accelerant: Microsoft’s willingness to fund a massive new cluster hinges on OpenAI delivering meaningful capability jumps, and those jumps are expected to track with large, near-term increases in training and inference power. The centerpiece claim is that Stargate would deliver at least a 100x jump in compute—“orders of magnitude” beyond what Microsoft currently supplies—while landing around 2028, with earlier stages coming online sooner. More compute, in this view, translates directly into stronger “frontier” models, which then become the substrate for whatever comes next in the 1–4 year AI timeline.

The argument starts with a conditional: Stargate moves forward only if OpenAI can improve its models enough to justify the investment. That improvement is tied to expected model milestones—GPT 4.5 “in the spring,” GPT 5 “at the end of this year or the beginning of next,” and then even later generations. A key supporting thread is a claim that such a project is “absolutely required” for artificial general intelligence, defined less as a single benchmark and more as the kind of system people would feel comfortable hiring for most jobs.

Skepticism about AGI timelines is met with a practical counterpoint: if AGI were truly imminent, hiring at the current pace would be unnecessary. Perplexity CEO Aravind Srinivas is used to press that question—why keep scaling teams and operations if a near-term AGI breakthrough is already within reach. The transcript also emphasizes the non-glamorous reality of running frontier systems: clusters must be maintained, GPUs selected, failures handled, and production code debugged—tasks that still require human operators, even if models become more capable.

The compute story is paired with an energy story. A “mathematical discrepancy” is flagged: Stargate’s compute gains are described as enormous, yet the power draw is said to be comparable to running several large data centers today. The response leans on a semiconductor trend from TSMC: energy-efficient performance is projected to improve roughly 3x every two years, implying chips could be nearly 10x more energy efficient by 2028. That efficiency curve is presented as the bridge between “100x compute” and “manageable watts.”

Beyond raw competition with Google, the transcript lays out multiple reasons for Stargate. One is capacity parity: Google is portrayed as having more near-term compute and more AI server chips than OpenAI, with Microsoft leadership describing the strategic advantage as compute, data, and IP rather than personnel alone. Another reason is to train larger future model families (the transcript name-checks GPT 7, GPT 7.5, and GPT 8 as targets for later training cycles). A third reason is “long inference”—letting models think longer via chain-of-thought search, framed as a way to boost reliability and unlock capabilities that show up in demos where responses arrive after sustained reasoning.

Finally, Stargate is positioned as a multimodal platform. The transcript points to OpenAI’s voice system (described as able to imitate a voice from about 15 seconds of audio) and to text-to-video generation exemplified by Sora, arguing that more compute supports richer audio/video/robotics capabilities—along with the risks of high-fidelity impersonation and deepfake-like content. The overall takeaway is that Stargate is less about a single product like Sora or a voice feature and more about manufacturing intelligence at a scale that could reshape what AI can do across tasks, timelines, and modalities.