Andreessen Horowitz, or a16z, said in its latest “Charts of the Week” that capital and labor in the U.S. economy are going through a structural shift from “bits” to “atoms.” The firm pointed to three moves at once: ETF themes have rotated from clean energy and healthcare toward AI, nuclear and space; data centers are offering blue-collar workers a meaningful wage premium; and AI agents are consuming far more tokens than humans while starting to eat into legacy automation tools.
Taken together, the charts frame a broader reordering of where investment dollars are going and where labor demand is building.
ETF themes have shifted sharply in six years
a16z said ETFs continue to expand their reach in public markets, especially among retail investors. The lineup now spans thematic ETFs, active and passive products, index trackers and credit ETFs, with different levels of leverage across the market. The firm said low fees, low barriers to entry, easier distribution and strong marketing have all supported growth, along with a wider rise in retail participation.
Citing Citadel data, a16z said ETF net inflows are on track for the strongest year on record, and that July marked an all-time high.
The firm’s larger point was not just that ETFs chase whatever is popular. It was that the underlying themes have turned over quickly. In 2020, the top five themes still included clean energy, emerging-market tech and healthcare. By 2026, the list had been rewritten around AI, nuclear power, space, defense and infrastructure.
a16z described that as a strong sign that markets are paying more attention to capital-intensive, physical industries. The charts do not try to predict the ending, but they do show that the theme mix in 2026 looks very different from what it did a few years earlier.
Data centers are taking a larger share of local construction spending
The report said investors are comfortable putting data centers into ETF narratives, even as communities are often less enthusiastic about having them nearby. Still, a16z argued that data centers have become economically significant enough that, in some places, they rank among the most important local developments.
For states actively building them, data center projects now account for a sizable share of private nonresidential construction spending.
New Mexico and Wyoming each have less than 3 gigawatts of capacity under construction, but because overall building activity in those states is relatively limited, data centers make up about 60% of private nonresidential construction spending there. In Pennsylvania, about 3 gigawatts of data center capacity comes close to 30% of nonresidential spending. Texas has much more capacity under construction, but because the state’s broader building base is larger, data centers account for about 10% of total spending.
a16z’s conclusion was simple: whatever view one takes of data centers, they are now one of the strongest pulses in the economy.
Wells Fargo and Indeed data point to better jobs and higher pay
According to figures compiled by Wells Fargo, counties with operating data centers have seen stronger outcomes since 2024 across several measures, including more housing, higher home prices, lower unemployment and faster job growth. Counties where data centers are still under construction have also seen better labor-market conditions, though new home construction has cooled more noticeably and home-price gains have been less pronounced.
a16z also noted that causality is not clear. A large share of existing and new data centers sits in Loudoun County, Virginia, already 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 been unusually strong before 2024, making any later slowdown a move from a much higher base.
Even with that caveat, the firm argued that data center construction almost certainly supports employment. Beyond the large number of hard-hat jobs created on site, data centers are paying technical workers more than comparable employers.
Using Indeed data, a16z said wage premiums tied to data center jobs can reach 64% for facilities managers. Even the smallest premium in the cited set, for electrical engineers, was 10%.
Blue-collar workers are seeing wage premiums of roughly 50%
a16z highlighted a recent comment cited in a Dallas Fed report from a heavy industrial building contractor: 「We’ve been paying what we thought were very competitive wages for skilled concrete workers, $28 to $32 an hour. Data centers are paying $45 an hour plus a $150 per day stipend.」
For concrete workers on data center projects, that works out to roughly a 50% pay premium, according to the report.
ADP data told a similar story. Looking at wage growth for job switchers relative to stayers, which a16z used as a proxy for where labor demand is running hot, data-center-linked categories stood out. In construction, manufacturing, and natural resources and mining, pay growth for switchers ran 6 to 9.5 percentage points above that of workers who stayed put. The report said that switching premium was higher than in any other industry.
a16z did not present this as a final defense of data centers. It said the point was narrower: saying no to data centers often means saying no to higher blue-collar pay and to one of the more significant local investment impulses now in play.
Uber fares are up about 20% since 2024
The charts also turned to consumer platforms. If Uber feels more expensive, a16z said Gridwise Analytics data suggests that perception is correct.
Since 2024, both average and median Uber fares have risen by about 20%, and the trend appears to still be moving higher.
Lyft has not followed the same path. Its median and average fares were slightly cheaper than they were at the start of 2024, and overall remained about 24% below Uber, though they too have been rising recently.
Platform fees appear to be driving much of the increase
a16z said one major source of the higher Uber price is a rise in platform fees. Those charges have kept climbing over the past year and a half, with median fees posting a visible jump in October. Lyft’s platform fees, by contrast, fell sharply for a period before starting to move up again.
The firm said that helps Uber and Lyft alike. Drivers are also benefiting. Average gross driver earnings per trip have continued to rise since 2024 and recently reached a record high.
a16z said it remains unclear whether higher ride-hailing prices will weigh on demand over the long run. For now, the picture is more straightforward: riders are paying more, and the gains are showing up on both the platform and driver side.
The report added that while the overall hiring market remains soft, more people are not only starting their own ventures but also moving into gig work, at least in part because earnings have improved.
Social commerce is the fastest-growing gig category in the charts
Ride-hailing may be getting more common, but a16z said it is not the fastest-growing type of gig work. That title, in the latest charts, goes to social commerce, which the firm described as the QVC of the social media era.
Based on Bank of America client account data, most gig categories outside vacation rentals have grown, but social commerce rose by more than 30%, easily outpacing the rest even from a smaller base.
a16z did not settle on one explanation. It listed several possibilities, including the shift in media consumption from television to social platforms, broader strength in e-commerce, and the chance that AI has cut the cost of running other parts of a social-media-based selling operation. The report also suggested that Instagram ad targeting may be part of the story, or that all of these factors are contributing at once.
The gap between typical AI users and heavy adopters is widening
On AI, a16z returned to a trend it has flagged before: demand is rising, but growth is far from evenly distributed. There is already a large usage gap between median users and heavy users, and recent OpenAI data suggests that split is widening.
Token output for a typical company has roughly doubled, but the top 10% of companies are pulling away much faster.
Across industries, the gap in token output between a typical company and the top 10% is about 8x, according to the figures cited by a16z. Since April 2025, output from the top decile has increased more than 17x. In some sectors, the spread is wider. In information, meaning the tech sector, the gap is close to 12x, and token output from the top 10% of companies is 32.5x what it was more than a year earlier.
a16z said the reason heavy users are generating more tokens is not casual conversation with AI. Many of them have moved past chat and into more advanced tooling.
Among the top 10% of companies, adoption of plugins is about 2x that of a typical company, while Skills adoption is about 6x. Even so, the firm said heavy users still have a long way to go before matching OpenAI’s own internal level of adoption.
Legal professionals stand out in Codex adoption
The report also said the fastest growth in more powerful AI tooling is not necessarily coming from tech. If Codex adoption is used as a measure of maturity, knowledge workers are improving across the board, but legal professionals stand out the most.
Since February 2026, Codex adoption among legal professionals has surged 108x. a16z said it is open to debate how much of that reflects broader Codex distribution, but legal was the clearest outlier in the data it cited.
AI agents now use nearly five times as many tokens as humans
As AI adoption moves from chat toward more complex execution, the composition of token growth is changing too. a16z said full agent deployments still represent only a small share of adopters, yet agents already account for a large share of token demand.
Citing OpenRouter and chart author Peter Walker, the firm said agents use nearly 5x as many tokens as humans, and agent usage has increased about 14x since February.
The way those tokens are used also looks very different. More than 85% of agent token consumption comes from cached prompts, according to OpenRouter, and cached tokens make up nearly all of the relative increase in usage.
a16z said the logic is straightforward. Humans tend to use AI in a question-and-answer format. Agents are built to iterate toward a goal. Their initial prompts include token-heavy prefills containing the core context around a task, whether that is policy workflow, code standards or something else. From there, the agent keeps reading and writing incrementally as it pushes toward the result, adding more material to cache along the way.
The report said cached tokens are much cheaper than prefills, which improves the economics of agent use, but they also consume far more memory. That, in a16z’s telling, helps explain why high-bandwidth memory has become so sought after. A growing pool of busy agents needs memory to keep working without restarting from scratch each cycle.
Legacy automation tools are seeing double-digit traffic declines
a16z said one possible knock-on effect of agent growth is weaker demand for traditional automation and workflow tools.
According to Similarweb data, traffic to automation-tool websites has been falling across the board except for Gumloop. N8N, Zapier and Make all predate the LLM era, and while they still dominate the automation category by visits, each has posted double-digit declines on a rolling 12-week basis.
Gumloop, launched in 2023 as an “AI-native agent builder,” is the only automation platform in the cited Similarweb data that is still gaining momentum.
a16z stopped short of saying tools like Zapier will be displaced. Those companies have AI strategies of their own. Still, the report’s final point was clear: agents are early, but they are already changing the shape of the market in a measurable way.