Author: a16z New Media

Translated by: 深潮 TechFlow

Introduction by 深潮: This week, a16z's charts cover four topics: Super platforms recorded unusually high "Other Income" through private equity investment in Q1; AI-generated eBooks are flooding the market but the total amount of quality content is also increasing; Employment in Philippine call centers is rising against the trend, and the cost of voice AI still can't catch up with human labor; AI mobile downloads, revenue, and usage duration have all doubled, with Codex's daily installations surpassing Claude Code. Four charts, four counterintuitive signals.

"Other Income": The Venture Capital Business of Tech Giants

The profits growth in the public market is already exaggerated, and Wall Street expects even higher this year.

However, beneath the profit figures lies an uncommon detail: The revenue of super platforms does not all come from core business. In Q1, "Other Income" accounted for an outrageous proportion of net profit.

Caption: The proportion of "Other Income" in net profit of super platforms in Q1 exceeded one-third, while the historical norm is 5%-10%

In Q1, "Other Income" accounted for more than one-third of net profit, while historically this number has been about 5% to 10%.

Where does this money come from? Mainly from private equity investments of Amazon and Google, totaling approximately $53 billion. Alphabet's CFO mentioned on the earnings call, "Other Income and Expenses amounted to $37.7 billion, primarily from unrealized gains in the private equity portfolio"; Amazon disclosed a net gain of $15.6 billion from its investment in Anthropic in its 10-Q.

In short: Super platforms are quite good at venture capital.

However, tech investment is no longer just a game for the giants. Estimates from KKR show that technology-related capital expenditures are currently the only category driving GDP growth:

Caption: In the 2% increase of US GDP in Q1, technology capital expenditures contributed 1.9%, almost the entirety

In Q1, US GDP grew by 2%, with technology capital expenditures contributing 1.9%. This means that without technology investment, GDP would hardly change.

Widening the scope, according to the statistics from the Bureau of Economic Analysis (BEA) on total capital expenditures by enterprises (including R&D and software), technology now accounts for 55% of all business investments in the US:

Caption: The share of technology in US total corporate capital expenditures continues to rise, currently reaching 55%

This ratio has been rising steadily for a long time, and AI may accelerate this trend. Yardeni Research proposed an interesting framework: there are three factors of production in economics textbooks—land, labor, and capital. Now a fourth should be added: data. AI makes data more useful; the more useful the data, the greater the demand for investment in data and tools for processing data.

Amazon and Google doing VC well is one thing. The bigger fact is: Everyone is now a technology investor.

AI-generated junk books are rampant, but quality content has also increased

Good news: There are far more eBooks on Amazon than before. Bad news: The additional portion is mostly AI-generated junk.

Caption: The monthly publication volume of eBooks on Amazon has tripled since the release of ChatGPT, exceeding 300,000 copies per month by the end of 2025

After the release of ChatGPT, the monthly publication volume of eBooks on Amazon increased from around 100,000 to over 300,000.

This chart can be interpreted in two ways.

The first interpretation is straightforward: AI has arrived, a tsunami of garbage content has flooded in, and Amazon is drowning in low-quality, machine-generated books.

The second interpretation is more worthy of consideration: There is indeed more junk, but there are also more "decent" books than before. A paper by professors from Cornell and Minnesota quantified this—using a nested Logit demand model, it estimated that the eBook selection in 2025 provides about 7% more consumer surplus than a counterfactual baseline of purely human creation. Readers in 2023 received almost no benefits, but by 2025, the benefits became perceptible.

Another interesting finding: AI helps older authors (those who have been publishing before the emergence of LLMs) the most.

Caption: After 2023, the output of "older authors" (authors published before LLM) significantly increased, as AI enhanced their productivity

AI hasn't just created a batch of robot authors; it has also increased the output efficiency of human authors.

Marc Andreessen made a prediction a few years ago on David Perell's podcast: Writing has become too easy, and a flood of poor content will occur; but at the same time, with tools this powerful, quality content should also experience explosive growth. Junk is real, but residual value is also real. Those who write well are now writing more.

The Call Center is not dead, but Voice AI is still too expensive

David George recently wrote an article arguing that AI replacing jobs is a fallacy. He distinguished between "replacement" and "augmentation"—customer service is the most typical candidate for replacement, as AI can answer all questions and has infinite patience.

The logic is sound. But the data does not support it.

Caption: Employment in the Philippines' IT and business outsourcing industry has grown from 1.15 million in 2016 to 1.9 million in 2025, across every significant upgrade in AI capabilities

The Philippines is the global capital of call centers. Data from Apollo shows that employment in the IT and business outsourcing sector has grown from 1.15 million in 2016 to 1.9 million in 2025—spanning across every major upgrade in AI. The industry association predicts that an additional 70,000 jobs will be created in 2026, a year-on-year growth of 3.7%.

The situation in the US is quite similar. Data from Indeed shows that customer service job postings have not decreased, but rather have outperformed the broader market:

Caption: Indeed data shows that the year-on-year growth rate of customer service positions is approximately 10 percentage points higher than overall recruitment, with a turnaround that occurred in August 2025

The year-on-year growth rate of customer service hiring surpasses that of the overall job market by approximately 10 percentage points. Moreover, this turnaround only recently occurred, in August 2025.

Does this indicate that AI is actually beneficial for the customer service industry? Probably not.

The core reason is cost. The output of text LLMs is very cheap, but voice AI is still very expensive. Goldman Sachs conducted an internal test comparing the total costs of AI vs. human customer service:

Caption: Goldman Sachs estimates that the total cost of AI customer service is about $92 per day, while for humans it is about $90 per day, nearly equal

The total cost of AI customer service is about $92 per day, while human costs are about $90 per day. Essentially the same. In comparison to programming agents—pure text outputs, costs are several orders of magnitude lower than human labor. The difference between code and customer service is that the potential demand for code is much greater than for customer service, so the leverage from falling costs is entirely different.

The story of Klarna serves as the best footnote. In early 2024, Klarna announced that it had replaced 700 customer service representatives with AI; the CEO stated that AI had done everyone's job. This became a benchmark case of "AI replacing humans." By May 2025, the CEO changed his tune and began hiring again—service quality declined, with users receiving identical responses.

This situation won't last forever. API costs are rapidly declining, and companies like Decagon are growing quickly; the cost comparison in 18 months may be completely different.

AI's Good Products Spread Extremely Quickly

The penetration speed of AI on mobile is astonishing:

Caption: Q1 data on mobile downloads, revenue, and usage duration of AI applications

Caption: Monetization and usage duration of AI applications in Q1 nearly doubled year-on-year

Downloads, monetization, and usage duration all turned upwards in Q1, with monetization and duration almost doubling year-on-year.

Perhaps people are spending less time on social media because they are using AI vibe coding on their phones? That’s not necessarily a bad thing.

Speaking of vibe coding, a new player has arrived:

Caption: Daily installations of Codex skyrocketed in May, surpassing Claude Code in a single day—the latter had been the king of coding tools for the past year

Daily installations of Codex surged in May, exceeding Claude Code in a single day. Of course, this is just daily data, with a lower base, but it indicates one thing: good products spread very quickly.

Jeff Bezos said in 2012: In the past, you could sell a mediocre product through marketing; now it is becoming increasingly difficult. Good products get users to spread the word.

In the AI field, this logic is taken to the extreme. Signals spread quickly, and user switching intentions are strong; no one will remain loyal to one platform or model.

The same is true on the B2B level:

Caption: YipitData shows the proportion of enterprises using 2-5 and 6-9 AI vendors continues to rise, with those using only one being less than 20%

The proportion of enterprises using multiple AI vendors continues to rise, with those using only one now below 20%. The B2B AI market currently has no winner-takes-all situation.