Sam Altman: AI Needs an Entirely New Hardware Form Factor, Not to Be Parasitic on Phones

In a recent interview, OpenAI CEO Sam Altman made a rather striking assertion: AI shouldn't continue to be parasitic on devices designed for humans—it needs an entirely new consumer hardware form factor.

Existing Devices Are Designed for Humans, Not for AI

Altman wasn't stingy with his praise for laptops and iPhones—he called them "one of the greatest inventions in the history of human technology." But then he pivoted: "I feel like AI is constrained by using something designed for humans, not for AI."

Behind that statement lies a structural contradiction that's easy to overlook. Current AI capabilities—especially agent capabilities—are almost entirely unlocked by simulating human actions: clicking buttons, reading screens, filling out forms. AI is borrowing human interaction interfaces to work, and that borrowing itself implies enormous efficiency losses.

Agents (智能体) refer to AI systems capable of perceiving their environment, making autonomous decisions, and executing actions to accomplish specific goals. Unlike chat models that only generate text, agents need to interact with the external world—reading information, calling tools, operating software. Current mainstream implementations include browser automation (e.g., Playwright), desktop control (simulating mouse and keyboard), and screen understanding (OCR, visual localization). The advantage of this "anthropomorphic" approach is its generality—no need for target software to provide interfaces; the cost is low efficiency, error-proneness, and dependence on interface stability. This is precisely the source of what Altman calls the "structural contradiction": an agent's capability ceiling is defined by the design logic of human interfaces, rather than by the task itself.

A Technical Reality That's "A Bit Silly"

Altman used a very direct analogy to describe the absurdity of the current situation: watching AI skillfully operate a computer is indeed impressive, but on closer thought it's "a bit silly"—a computer expends massive computing power to render an interface for human eyes to view, and then AI works hard to "read" that interface back, rather than directly accessing the underlying data.

This observation points to a fundamental bottleneck in the current AI application layer. Whether it's browser automation or desktop agents, they are essentially reverse-engineering on top of the human interface surface. Every screen render, every OCR recognition, every coordinate click is an intermediate step that could have been skipped. Traditional GUIs are optimized for the bandwidth of human eyes and the cognitive habits of the human brain—not for the token throughput of models.

GUI (Graphical User Interface) has been the dominant paradigm of human-computer interaction since the 1980s, with its core being visual hierarchies, icons, windows, and pointer operations that match human perception and cognitive habits. Every frame of interface rendering involves layout calculation, font rasterization, compositing, and refresh—overhead that serves "human viewing." When AI "understands" an interface through screenshot plus OCR or vision models, it's essentially downgrading already-structured data back into pixels, and then having the model reverse-infer the semantics—this back-and-forth causes a double waste of compute and latency. More efficient AI-native interaction should directly consume structured data (such as JSON, API responses, DOM trees), or directly dispatch tasks through intent protocols, thereby bypassing the render-recognize intermediate layer.

Why This Points to Hardware, Not Just Software

If the problem were merely "AI shouldn't read screens," then directly connecting to data via APIs would seem to suffice. But Altman's argument goes a step further: this is a technology that "wants a new form factor."

The deeper meaning of this judgment is that AI's interaction patterns are fundamentally different from humans'. Humans rely on visual hierarchies, touch feedback, and window management; AI is better at structured data streams, parallel task scheduling, and headless intent execution. When the core user of a technology shifts from humans to models, the physical container that hosts it should change accordingly. Historically, GUIs gave rise to mice and windows, touchscreens gave rise to smartphones—every shift in interaction paradigm has been accompanied by a reconstruction of hardware form factors.

Interestingly, Altman didn't describe a specific product form in the interview, but the phrase "consumer hardware" hints at ambitions targeting the mass market, rather than enterprise-grade or developer tools. Combined with OpenAI's previous布局 (layout) rumors in the hardware direction, the direction of this roadmap is already relatively clear.

Interaction paradigm shifts and hardware reconstruction have recurred throughout computing history: command-line interfaces gave rise to keyboards and terminals, GUIs gave rise to mice and window systems, touchscreens and the mobile internet gave rise to smartphones, and voice assistants drove new form factors like smart speakers. The essence of each shift is a change in "who is using computing resources"—from professional operators to ordinary consumers, and now to today's AI models. When the core user becomes a model, the interaction primitives also change: humans need visual feedback and touch confirmation, while models need low-latency data pipelines, parallel task queues, and programmable intent interfaces. Therefore, hardware "designed for AI" isn't necessarily about removing the screen, but about redefining how data is directly consumed and how tasks are natively expressed—this could give rise to entirely new sensor combinations, input/output methods, and device form factors.

What This Means for the Industry

In the short term, agents will still operate in the cracks of human interfaces, with screen reading and simulated operations remaining the mainstream approach. But Altman's statement sends a signal: the competitive window for AI-native hardware is opening. Whoever can define the interaction primitives "designed for AI"—how data is directly consumed, how tasks are natively expressed—may capture the entry point to the next computing platform.

This also poses a sharp question to existing hardware manufacturers: when AI is no longer an app on a phone, but an independent entity with its own dedicated form factor, will today's device dominance remain secure?

Altman didn't give an answer, but he put the question on the table.

One-line Summary

Altman believes AI shouldn't be parasitic on human devices and needs a new consumer hardware form factor.

Paragraph Summary

In an interview, Sam Altman proposed that AI shouldn't continue to be constrained within devices designed for humans, but needs an entirely new consumer hardware form factor. He pointed out that current agents work by simulating clicks and reading screens—essentially reverse-engineering on top of the human GUI surface, causing a structural waste of compute and efficiency: computers render human-eye interfaces, and AI then laboriously reads them back. Altman sees AI as a technology that "wants a new form factor," with interaction patterns fundamentally different from humans, and notes that interaction paradigm shifts have historically been accompanied by hardware reconstruction. His use of the term "consumer hardware" hints at mass-market ambitions. In the short term, agents will still operate in the cracks of human interfaces, but the competitive window for AI-native hardware is opening—whoever defines the interaction primitives "designed for AI" may capture the entry point to the next computing platform.