Om Malik Reframes AI Models' iPhone Moment Analogy

In a June 2 column on his blog On my Om, veteran technology writer and investor Om Malik argues that comparing foundational AI models to successive iPhone releases is the wrong analogy. He proposes what he calls a "more boring, and more accurate" one. Across past technology cycles, Malik writes, innovations move from "shock and awe" to "ho-hum" and eventually become invisible.

Malik draws a careful distinction. In the PC clock-speed wars, he notes the Pentium 4 climbed from 1.3 GHz at its 2000 launch to 3.8 GHz before the industry pivoted to performance per watt, multi-core designs, and Apple's M1 in late 2020. Smartphones followed a similar arc, with annual upgrades shrinking from essential to marginal. But in both cases, he argues, the technology actually plateaued, which is why interest faded. AI, by contrast, is not plateauing: he writes that capability is still accelerating, citing the jump from GPT-4 to GPT-5 and the spread of open-weight models.

Malik's central claim is that AI resembles optical networking. He recounts how Dense Wavelength Division Multiplexing (DWDM) silently multiplied the capacity of a single fiber for two decades with no launch events or reviews, enabling YouTube, Netflix, and pandemic-era Zoom calls. AI, he contends, will similarly migrate "from topic to infrastructure," staying invisible while its capability keeps compounding, much as inference costs have fallen by roughly an order of magnitude every year or two.

Malik notes this dynamic is unfavorable for the valuations of foundation-model labs and Nvidia, which he says want new metrics to sustain attention. Invoking the optical era, he observes that the companies that built DWDM, such as Nortel and Lucent, were largely forgotten while the layer above them, including Google, Amazon, and Netflix, captured the value, summarizing that "Infrastructure enables; it does not determine who wins."

This is opinion and analysis rather than reported news. Malik's framing is a prediction: he estimates that by 2028 the model will, in his words, have "gone underground," and advises watching not benchmark scores but the moment benchmarks disappear from the conversation.