Andreessen Horowitz’s latest Charts of the Week argues that capital and labor are starting to move away from “bits” and back toward “atoms.” Written by Moses Sternstein and translated by TechFlow, the piece links four developments: ETF theme rotation, the economic weight of data-center construction, higher ride-hailing prices and changing gig work, and the rapid rise of AI agents.
ETF themes have rotated fast
The report says ETFs have become increasingly prominent in public markets, especially for retail investors. Beyond thematic products, the market now spans active and passive ETFs, index-tracking ETFs, and credit ETFs, with a wide range of leverage profiles. The piece attributes the category’s continued expansion to low fees, low barriers to entry, easier distribution, strong marketing, and broader retail participation.
According to Citadel data cited in the article, ETF net inflows are heading toward their strongest year on record, with July setting an all-time monthly high.
The article says ETFs are particularly good at attaching themselves to whatever theme is hottest in the market at a given moment. What stands out more, though, is how far those themes have shifted in only a few years.
In 2020, the top five themes still included clean energy, emerging-market tech, and healthcare. By 2026, that list had been rewritten around AI, nuclear, space, defense, and infrastructure. The piece frames this as a broad ETF push into capital-intensive “atoms over bits” trades. It does not try to predict the outcome, but it does note that the 2026 setup looks very different from what investors were chasing a few years ago.
Data centers are becoming a major local economic force
The second section focuses on data centers. The article notes that while investors may like owning data-center exposure through ETFs, they are often less enthusiastic about seeing those facilities built nearby. Still, whatever the political debate looks like, the economic footprint is already large in some states and, in some cases, may be among the most important forces in local economies.
Measured against non-residential construction spending, the impact can be striking. New Mexico and Wyoming each have less than 3 gigawatts under construction, but because overall construction activity is limited in those states, data centers account for roughly 60% of private non-residential spending. In Pennsylvania, around 3 gigawatts of data-center capacity comes close to 30% of non-residential spending. Texas has much more capacity under construction, but the share is about 10%.
The piece cites Wells Fargo, which looked at the “significant economic benefits” associated with operating and in-construction data centers. Since 2024, counties with operating data centers have shown more housing, higher home prices, lower unemployment, and faster job growth. Counties with data centers under construction also show stronger employment, though they have seen a sharper pullback in new-home construction and less pronounced home-price gains.
The article adds an important caveat: causality is not clean. A large share of existing and new data centers sit in Loudoun County, Virginia, one of the wealthiest counties in the U.S. Many new projects are also in Texas, where housing construction and home-price growth had already surged before 2024, making any later cooling come off a much higher base.
Even so, the core labor point is direct. Building data centers is described as almost certainly positive for employment, and not only because of the number of hard-hat jobs involved. Pay is also materially higher than at comparable employers.
According to Indeed data cited in the article, wage premiums tied to data-center jobs can reach 64% for facilities managers. Even electrical engineers, the lowest premium in the set referenced, still see a 10% premium.
The report also points to a recent Dallas Fed comment from a heavy industrial building contractor: 「We have been paying what we thought were very competitive wages for skilled concrete workers, $28 to $32 per hour. Data centers are paying $45 an hour, plus a $150 per day stipend.」 By the article’s framing, that works out to roughly a 50% wage premium for concrete workers on data-center projects.
ADP data points in the same direction. Using the pay growth of job switchers relative to workers who stay put as a proxy for where demand is hottest, construction, manufacturing, and natural resources/mining show switcher wage growth that is 6 to 9.5 percentage points above that of stayers. The article says that premium is higher than in any other sector it tracks.
The argument here is not a full defense of data centers. It is narrower: saying no to data centers likely also means saying no to higher blue-collar wages and to one of the more powerful nearby investment pulses now forming.
Uber fares are up about 20% since 2024
The third section turns to ride-hailing. If Uber feels more expensive lately, the article says the data supports that impression.
Gridwise Analytics data cited in the piece shows Uber’s average and median fares have both climbed about 20% since 2024, and the trend still appears to be moving higher.
Lyft looks different over the same period. Its median and average fares are slightly cheaper than they were at the start of 2024 and run about 24% below Uber overall, though they too have risen recently.
The article says one major source of Uber’s price increase appears to be higher platform fees. Those fees have kept rising for more than a year, with median fees showing a clear jump in October. Lyft’s platform fees, by contrast, fell sharply for a period before only recently starting to move up again.
The gains are not limited to the platforms. Drivers are also getting paid more. Average gross driver earnings per trip have been rising since 2024 and have recently reached a record high, according to the article.
Whether higher ride-hailing prices will eventually weigh on demand remains an open question in the piece. For now, though, the direction is simple: fares are climbing, and the benefit is flowing to both platforms and drivers.
Social commerce is the fastest-growing gig category
The article adds that even though the broader hiring market remains soft, more people are either starting businesses or moving into gig work, likely helped in part by stronger earnings. Ride-hailing may be more popular, but it is not the fastest-growing slice of the gig economy.
Using Bank of America customer-account data, the piece says most gig categories are growing aside from vacation rentals. Social commerce stands out the most, posting growth of more than 30%, far ahead of other categories, though from a smaller base.
The article does not settle on a single explanation for that growth. It lists several possibilities: media consumption shifting from television to social platforms, broader e-commerce strength, lower operating costs due to AI, or increasingly effective ad targeting. The piece presents these as possibilities rather than conclusions.
AI agents are still early, but they are already changing usage patterns
The final section looks at AI adoption. The article says overall demand keeps growing, but the distribution is far from normal. Usage among median users and power users is diverging sharply, and recent OpenAI data suggests that gap is still widening.
At the enterprise level, output tokens for a typical company have roughly doubled, while the top 10% of companies have pulled much farther ahead. Across all industries, the gap in token output between a typical company and the top decile is about 8x. Since April 2025, token output for that top 10% has grown by more than 17x. In the information sector, the gap is close to 12x, and token output for the top decile is 32.5x what it was a little over a year earlier.
Those heavier users are not simply chatting more. The article says they have moved further into advanced AI tooling. Adoption rates for plugins and Skills among the top 10% of enterprises are about 2x and 6x those of a typical company, respectively, though even these heavy users still trail OpenAI’s own internal adoption levels.
Using Codex adoption as a marker of maturity, legal professionals stand out the most. The article says Codex adoption in legal has surged 108x since February 2026. It also notes that some of that move may reflect broader Codex promotion, but legal still separates itself from the rest of the field.
That shift toward more complex usage is also changing how tokens are consumed. Data from OpenRouter, cited alongside chart author Peter Walker, shows that full agent deployments still represent only a small share of AI adopters, but agents already account for an outsized share of token demand. Agent token consumption is nearly 5x that of humans and has grown about 14x since February.
OpenRouter data also shows that more than 85% of agent token consumption comes from cached prompts, and cached tokens make up almost all of the relative growth in token usage. The article explains the distinction this way: humans tend to interact in a question-and-answer loop, while agents iterate repeatedly toward a goal. Their initial prompts contain token-heavy prefills, such as policy context or coding rules, and then they read and write incrementally while adding more material to cache as the task progresses.
Cached tokens are much cheaper than prefills, which improves the economics of agents, but they are also memory-intensive by definition. The piece links that directly to strong demand for high-bandwidth memory, arguing that busy agents need memory to keep running rather than restarting from scratch at every cycle.
There is also a second-order effect. Agents may already be weakening demand for traditional automation and workflow tools. Similarweb data cited in the article shows traffic declining across automation-tool websites other than Gumloop. N8N, Zapier, and Make all predate the LLM era and still dominate automation by visits, but each has posted double-digit declines on a rolling 12-week basis. Gumloop, launched in 2023 as an AI-native agent builder, is described as the only automation platform in the Similarweb set still gaining momentum.
The article stops short of declaring the older platforms finished. They are building with AI too. But its conclusion on this point is clear enough: agents are only just arriving, and they are already beginning to change the competitive picture in tangible ways.
