AI is no longer just a software story: 17 charts on how capital, work, and compute are shifting tracks

Capital turns toward heavier, slower, and costlier themes

The first set of charts lays out the shift in capital allocation.

Net inflows into ETFs are on track to break records in 2026. According to Citadel Securities data, as of August 10, the pace for the year is roughly $600 billion ahead of last year's record. July alone saw net inflows of about $346 billion, marking a new monthly high.

The overall numbers are striking, but the destination is even more telling. In 2020, the leading themes were clean energy, emerging market tech, and healthcare. By 2026, AI, nuclear, space, defense, and infrastructure have climbed the rankings. These sectors all depend more heavily on capital expenditure, physical infrastructure, energy, and long construction cycles.

1. Annual totals: 2026 ETF net inflows are on pace to set a record. Source: Citadel Securities / Global Market Intelligence.

2. Monthly peak: July 2026 hit roughly $346 billion, a series high. Source: Citadel Securities / Global Market Intelligence.

3. Structural shift: leading themes move toward AI, nuclear, defense, and infrastructure. Source: Bloomberg / J.P. Morgan Asset Management.

This marks a narrative shift from bits to atoms. Previously, investors favored rapidly scalable software and consumer internet companies. Now, capital is flowing toward electricity, factories, satellites, defense manufacturing capacity, and compute infrastructure.

Yet capital flows only demonstrate a change in investor preference. They don't prove that these projects will ultimately be profitable, nor do they guarantee strong returns after heavy inflows. Thematic ETFs are particularly good at compressing a complex industry into a tradable label—and particularly prone to conflating 'the direction may be right' with 'the price is still reasonable.'

Compute construction is repricing skilled labor

If the first set of charts is about expectations, the data center charts show the real-world frictions that emerge once those expectations begin to materialize.

In New Mexico and Wyoming, data centers account for roughly 60% of private nonresidential construction spending. That number is large, but it shouldn't be read in isolation: both states have less than 3GW of data center capacity under construction, and the high share reflects relatively little other construction activity. Pennsylvania, with a similar 3GW pipeline, accounts for close to 30% of the state's nonresidential spending. Texas, meanwhile, has a much larger construction pipeline, but its share is around 10%.

That's what makes data centers unusual. They are simultaneously technology infrastructure and heavy industrial projects involving land, power, HVAC, concrete, electrical, and network engineering. In some states, they are beginning to dominate local construction cycles.

County-level data also paints a broadly positive picture. Counties with existing large data center operations have outperformed the national average in housing construction, home prices, unemployment rates, and job growth since 2024. Counties where data centers are still under construction show clearer employment gains, though housing and price trends are more mixed.

This shouldn't be mistaken for causation. Data centers tend to locate where power, land, network, and tax conditions are already favorable. Virginia's Loudoun County was already wealthy, and Texas experienced unusually strong residential construction before 2024. The charts show an association between data centers and better economic outcomes, not the results of a controlled experiment.

Wage data offers a closer look at genuine supply and demand shocks. Comparing job postings for similar roles on Indeed, median wage premiums for data center positions range from about 10% for electrical engineers to roughly 64% for facilities managers. Construction managers see a premium of about 29%, while site supervisors earn about 26% more. Dallas Fed interviews with local firms are even more direct: skilled concrete workers in the region typically earn $28–$32 per hour, but data center projects have pushed offers to $45 per hour, along with a $150 daily stipend.

1. Construction share: a high share requires looking at both gigawatts under construction and the local construction base. Source: Goldman Sachs Global Investment Research.

2. County-level outcomes: operating counties look better overall, but site selection and existing economic conditions create selection effects. Source: Wells Fargo / FracTracker / US Census.

3. Same-role wages: facilities managers see about a 64% premium, electrical engineers about 10%. Source: Indeed, January–June 2026 job postings.

4. Switching premium: construction, manufacturing, and natural resources/mining lead in job-switcher wage gains. Source: ADP Research.

ADP job-switching data provides another point of verification. In construction, manufacturing, and natural resources/mining, workers who change jobs see wage growth that is about 6 to 9.5 percentage points higher than those who stay. The job-switching premium isn't unique to data centers, but it does indicate rising demand for 'hard hat' roles.

Local debates about data centers shouldn't focus only on grid load, water usage, taxes, noise, or long-term operational jobs. Rejecting a project could mean forgoing a wave of construction spending and wage premiums for skilled workers. Accepting one doesn't mean all county-level prosperity stems from the data center itself. The real trade-off is a full set of costs and benefits, not a simple yes-or-no question about AI.

Ride-hailing costs more, but platform and driver accounts don't merge

Uber is getting more expensive, and it's not just perception. Using Gridwise data, a16z's charts show that between early 2024 and the second quarter of 2026, both Uber's average and median fares rose by about 20%. Lyft remains cheaper overall.

Gridwise's Q2 2026 page offers a more easily verifiable snapshot: Uber's average fare is $24.62 per trip, while Lyft's is $19.83. The medians are $18.01 and $14.99, respectively. Uber's average fare has risen 14.2% over two years, partly because premium vehicle types make up a larger share of trips rather than because all comparable rides got uniformly more expensive.

The extra revenue isn't flowing in just one direction. Average platform fees have reached $5.30 per trip for Uber and $3.11 for Lyft, up 14.0% and 26.4% year over year respectively—both outpacing fare growth. Meanwhile, average gross earnings per trip for drivers rose to $15.78, up 7.6% year over year and exceeding the 2021 peak.

1. Fares: Uber's rise is more pronounced; Lyft stays relatively cheaper. Source: Gridwise Analytics.

2. Platform fees: both are climbing, with Uber higher in absolute terms. Source: Gridwise Analytics.

3. Driver gross earnings: $15.78 per trip, above the 2021 peak; this is gross pay. Source: Gridwise Analytics.

4. Gig mix: social commerce grows fastest in the BofA sample, but methodology isn't fully disclosed. Source: Bank of America internal data.

These charts are most easily misread as a win-win for platforms and drivers. But earnings per trip are gross pay, not net income. Fuel, vehicle depreciation, insurance, deadheading, wait times, and the number of trips completed per hour can all change what drivers actually take home. Higher platform fees directly improve unit economics for the platform, but higher earnings per trip for drivers don't necessarily translate into better hourly wages or net profit.

The gig structure is shifting too. In Bank of America's internal account data, ride-hailing, delivery, and content creation are all growing, but since the 2024 baseline, social commerce has risen more than 30%—the fastest pace of any category. This could reflect attention moving from TV to social platforms, maturing e-commerce infrastructure, or AI lowering the costs of scripting, creative assets, customer support, and ad placement.

'Could' is the key word here. BofA's charts don't disclose the full sample or methodology, and social commerce's low base amplifies the growth rate. It's useful as an early signal, not as proof of causation.

Agent usage is reshaping token consumption and toolchains

The final six charts push the perspective from 'how many people are using AI' to 'how a small group of companies is already using it.'

OpenAI's Enterprise Signals defines frontier enterprises as those in the top 10% by output tokens per active user. As of June 2026, these companies generate 8.3 times more output tokens per user than typical enterprises, up from just 2.6 times in January. The widening gap isn't just about more chat interactions; it's about connecting company context, tools, and repetitive workflows to agents.

Adoption rates reveal this stratification. Among frontier enterprises, 21% of active weekly users engage with Plugins and 19% use Skills. For typical enterprises, those figures are just 9% and 3%. OpenAI's internal Plugins usage rate of 95% serves more as an upper-bound reference for what deep adoption could look like, rather than a benchmark typical companies should currently meet.

Role diffusion extends well beyond engineering teams. Since February 2026, weekly active users of enterprise Codex have grown 108-fold in legal roles, 41-fold in sales and recruiting, 26-fold in marketing, and 5-fold in engineering. The legal figure is striking, but low baselines, product access, and rollout timing all amplify these multiples. The data shows tools spreading across functions, not that legal productivity has increased 108-fold.

1. Enterprise stratification: the per-user output token gap between frontier and typical firms keeps widening. Source: OpenAI Enterprise Signals.

2. Tool adoption: frontier firms more often wire context and capabilities into agents. Source: OpenAI Enterprise Signals.

3. Role diffusion: legal grew 108x, but low baselines and product launches inflate multiples. Source: OpenAI Enterprise Signals.

4. Token scale: agent-classified traffic is about 7.3 trillion tokens, roughly 14x since the start of the year. Source: OpenRouter / Peter Walker.

5. Cache structure: over 85% of agent token consumption comes from cached prompts. Source: OpenRouter.

6. Toolchain pressure: legacy platforms see traffic fall while Gumloop rises; visits don't equal actual workflows or revenue. Source: Similarweb.

Traffic patterns on OpenRouter illustrate why agents so quickly amplify compute demand. a16z cites Peter Walker's categorization of OpenRouter traffic: agents' 7-day average consumption is roughly 7.3 trillion tokens, up about 14-fold since the start of the year and nearly 5 times the volume of human traffic. More than 85% of agent token consumption comes from cached prompts.

This aligns with how agents work. Humans typically ask one question and receive one answer. An agent, by contrast, continuously reads, calls tools, writes results, checks for errors, and proceeds to the next step toward a single goal. The initial large context gets reused repeatedly, with new content entering the cache at each turn. Cached tokens are cheaper, making long tasks economically feasible. At the same time, this shifts the system bottleneck from 'how long is the generation' to 'how many tasks are continuously running and how long must context be retained.'

However, a high volume of cached tokens alone doesn't tell us how much high-bandwidth memory the entire industry needs. It's evidence of a shift in inference workload structure, not a complete calculation for HBM demand. A more rigorous assessment would also require data on concurrency, context length, cache hit rates, model architecture, and hardware utilization.

Older automation tools are already feeling the pressure. Similarweb website traffic data shows N8N, Zapier, and Make all experienced double-digit declines over the last 12 weeks, while the AI-native agent builder Gumloop saw gains. The direction is notable, but it's premature to declare the end of the Zapier era: website visits don't equal actual workflow executions, paid seats, API calls, or revenue, and established platforms are integrating AI into their own products.

Three ledgers will determine how far this cycle can go

Whether this shift is sustainable ultimately comes down to three ledgers.

First, whether capital expenditure translates into sustained output. ETF flows, data center construction, and HBM demand can all rise quickly, but what determines the quality of the cycle is utilization, unit economics, and downstream willingness to pay.

Second, how scarce factors get repriced. Data centers are already pushing up wages for concrete workers, electrical engineers, and facilities managers in certain regions. Agents are raising the value of high-quality context, tool integrations, caching, and governance capabilities. AI's bottleneck increasingly looks less like 'is there a model available' and more like 'can the real-world resources be organized.'

Third, whether gains remain concentrated at the top. The token gap between frontier and typical enterprises has widened from 2.6 times to 8.3 times. Even when companies have access to the same models, the depth at which they actually deploy them is diverging rapidly. If this gap continues to widen, the first visible consequence may be further divergence in organizational execution capacity, rather than a uniform productivity lift across all companies.

The next phase of AI will appear simultaneously on construction payrolls, in enterprise tool permissions, in cloud cache bills, and in how long capital markets are willing to wait. The final outcome will be decided by a very practical checklist: whether there's enough electricity, whether the talent exists, whether workflows are connected, and whether unit costs can come down. Themes only become productivity after passing through those four gates.