The Cognitive Revolution

Written By Vas Bakopoulos

Welcome back, embedders,

This week we start with the big picture and debate the future of cognition, responding to Sequoia's essay from last week. Its main point is that we have made this trip before. The Industrial Revolution externalized muscle: machines took the physical work, the jobs built on that work disappeared, and after a long and painful stretch the world came out better. The question is whether it is different this time, because the machines that took muscle never came for judgment.

- Vas

The big picture

The Cognitive Revolution

Sequoia's Konstantine Buhler published an essay last week with a provocative premise: the Industrial Revolution moved work from human muscle to machines. AI is now beginning to do the same with the human mind. Buhler argues that this should ultimately make us more productive, not obsolete. As intelligence gets cheaper, we will consume more of it, and humans will move toward higher-value forms of thinking.

But what if AI keeps moving up the same ladder? If AI progresses from calculation to analysis to judgment, and increasingly helps us decide which questions to ask, where exactly do humans move next?

And there is a second, perhaps harder question: if AI does more of our thinking, does it free us to think at a higher level, or slowly erode the cognitive muscles we need to get there? That question becomes particularly uncomfortable when you apply it to education. What should we still require children to do themselves, even when AI can do it better.

I spent a long walk arguing these questions with ChatGPT. It became the first installment of a new series I'm calling Retriever.

Read: The Cognitive Revolution: AI Is Raising the Ceiling and Removing the Ladder

AI, the big picture

AI infrastructure could become a $31.6 trillion build

PwC estimates global data-center investment will reach $31.6 trillion through 2050, with annual spending rising from about $800 billion today to $1.8 trillion.The interesting part isn't just the scale. Unlike traditional infrastructure, the spending may never really "finish." Servers and chips have to be replaced continually. PwC expects ICT equipment to grow from 70% of data-center investment today to 93% by 2050, turning the AI buildout into a recurring investment cycle rather than a one-time construction boom.Source: PwC Global Data Centre Outlook

­

Only 22% of companies have scaled AI. 85% are spending more anyway

Gartner finds that only 22% of organizations have scaled AI across multiple business units or become AI-first. Yet 85% of functional leaders plan to increase AI spending this year.The gap is worth watching: investment is accelerating much faster than organizational maturity. Among Gartner's high performers, 81% of AI initiatives generated positive returns. Low performers couldn't even identify the ROI of 29% of their initiatives.Source: Gartner C-Suite AI Survey

­

Which AI model is "best"? It depends on what you ask it to do

Two benchmarks published two days apart produced different winners.Artificial Analysis ranked Claude Fable 5.1 first on its broad intelligence index, followed by GPT-6 Astra. Neon tested 42 models on 100 support tickets and found GPT-5.3 Codex the most accurate.That isn't necessarily a contradiction. It's a useful reminder that model rankings increasingly depend on the task and the benchmark. Neon itself cautions that its results apply to one workload, not to models in general.Sources: Artificial Analysis · Neon

­

a16z: AI may strengthen systems of record rather than kill them

One popular AI thesis is that agents will weaken traditional enterprise software. a16z's Seema Amble argues the opposite may happen.Companies that already own critical enterprise data can move their AI from simply retrieving that data toward making judgments and taking action on it.The opening for AI-native companies: become better at doing the work itself, even when an incumbent owns the underlying data.

AI and marketing

ChatGPT ads hit a $1 billion run rate, and OpenAI is building the missing ad stack

OpenAI says ChatGPT advertising has reached a $1 billion annualized run rate, excluding ad credits.

But advertisers have been asking for something more basic: better measurement, targeting and campaign-management tools.

Now OpenAI is beginning to deliver them, including an Ads Manager plugin for creating and managing campaigns through natural language, improved custom audiences and conversion matching, carousel reporting and total-budget pacing.

The interesting story is no longer whether ChatGPT can become an advertising platform. It's how quickly OpenAI can build the infrastructure required to become a serious one.

­

Google brings exact and phrase match into AI Mode

Google has begun allowing exact and phrase match keywords to trigger text ads inside AI Mode, not just broad match and automated campaign types.

Google describes it as a small experiment and says ads appear only when the user's intent is explicit and direct. But strategically, it matters: some of the familiar controls of search advertising are beginning to follow users into AI search experiences.

Source: Search Engine Roundtable

Sources

Vas Bakopoulos
Next

Only 6% are getting real value from AI, barely more than a year ago