Agents on the OpenRouter network consumed close to five times as many tokens as human users, and their consumption has multiplied roughly fourteenfold since February, according to figures published on August 21, 2026 in the a16z newsletter Charts of the Week. Over the same period, site traffic to the three automation platforms that dominated workflow tooling before large language models arrived fell by double digits.

The newsletter, written by Moses Sternstein and dated August 21, 2026, assembles data from OpenAI, OpenRouter, Similarweb, Citadel, Wells Fargo, Indeed, ADP, Gridwise Analytics, Bank of America and the Federal Reserve Bank of Dallas. It carries no product announcement and no vendor claim. What it does carry is a set of measurements on how autonomous software consumes computing capacity, which categories of work are moving fastest, and which incumbent tools are losing ground. Each of those bears on how marketing organisations account for artificial intelligence spending.

Agent token consumption pulls away from human usage

The central measurement comes from OpenRouter, a routing layer that sits between applications and model providers, with charts credited to Peter Walker. According to those figures, agents are using nearly five times as many tokens as human users, and agent usage has grown roughly fourteenfold since February 2026.

The composition of that consumption matters more than the volume. More than 85% of agentic token burn originates in the cached prompt, according to the same OpenRouter data, and cached tokens account for nearly all of the relative growth in token usage.

Sternstein attributes the pattern to a structural difference in how the two classes of user behave. A person tends to work in a prompt-and-response exchange. An agent is built to iterate towards a goal. The opening prompt carries a token-intensive pre-fill containing the core context around that goal, whether policies and procedures, code guidelines or comparable material. From there the agent reads and writes incrementally, adding progressively to the cache as it advances towards a final result.

Cached tokens cost far less than the pre-fill, which improves the unit economics of running agents. They are also, by definition, memory-consumptive, which the newsletter identifies as the reason high bandwidth memory sits in such demand. Agents, in Sternstein's description, "use tokens very differently" from people.

That distinction has a direct bearing on how agency and in-house teams model the cost of autonomous campaign tooling. PPC Land reported in early August 2026 that holding companies had begun reselling model capacity to clients with a markup while transacted agentic media spend remained thin. A cost base weighted towards cheap cached tokens rather than expensive pre-fill changes the arithmetic of that arrangement, and it changes the arithmetic in the client's favour only if the saving is passed through.

Cost discipline has already produced visible engineering responses. One media buying architecture collapsed twelve separate Model Context Protocol calls into a single buyer agent specifically to contain token consumption. KPMG research published in August 2026 found that 49% of surveyed leaders had scaled back agent rollouts when running costs outran the value delivered, with only 35% reporting full visibility into what their AI systems cost to operate.

Automation incumbents lose ground to an agent-native rival

The newsletter's second-order observation concerns what agents may be displacing. According to Similarweb data cited in the piece, traffic to legacy automation and workflow tooling sites has been in serial decline.

n8n, Zapier and Make all predate large language models. Between them they dominate the automation segment measured by site visits. Each is recording double-digit declines on a trailing twelve-week basis, according to the Similarweb figures. Gumloop, launched in 2023 and described in the newsletter as an AI native agent builder, is the only automation platform in the set gaining traction.

Sternstein stops short of declaring the incumbents finished, noting that they have artificial intelligence at their disposal as well, and that agents have barely arrived. The measurement nonetheless carries weight for marketing operations teams, because those three platforms sit inside a large number of production martech stacks. Zapier appears among the eleven launch partners for Google's Data Manager API, and Google's own documentation routes offline conversion imports through it. Similarweb's own Model Context Protocol server, launched in September 2025, integrates with Zapier and n8n among other automation tools. Freelance marketing automation work has clustered around the same three names, with project listings surveyed in October 2025 citing n8n, Zapier and Make as the standard toolset alongside hourly rates from $60 to more than $200.

Site traffic is an imperfect proxy for platform health. It measures visits to a web property rather than workflow executions, seat counts or revenue, and a mature product with stable installed usage can shed visits without shedding customers. The direction, sustained across three separate incumbents over twelve weeks, is what the newsletter treats as the signal.

The adoption gap between firms keeps widening

A separate section addresses distribution rather than aggregate growth. Output tokens have roughly doubled across the typical firm, according to data attributed to OpenAI, but the top decile is pulling further away.

Across all industries the gap in token output between the typical enterprise and the top-decile enterprise runs to roughly eightfold. Top-decile firms have increased output by more than seventeenfold since April 2025. Within Information, the sector covering technology companies, the gap widens to almost twelvefold, with the leading decile outputting 32.5 times as many tokens as it did a little over a year earlier.

The leaders are not producing that volume through chat. Plug-in and Skills adoption among top-decile enterprises runs at roughly twice and six times the level of the typical firm respectively, on the newsletter's reading of the OpenAI figures, and even those power users sit below OpenAI's own internal adoption levels.

Measured by adoption of Codex, the fastest-moving professional category is not technology. Legal workers have increased their Codex adoption by 108 times since February 2026. Sternstein flags the obvious caveat, querying how much of the shift reflects the broader Codex rollout rather than legal-specific demand, but records that the category stands apart from the rest.

Knowledge workers across the board are registering higher adoption of more advanced tooling. For marketing organisations, the relevant implication is that a sector-level adoption average conceals an order-of-magnitude spread between firms in the same industry. Survey evidence has pointed the same way. LayerX research published in May 2026 found only 18.24% of enterprise employees using AI tools weekly, with 47% of enterprise AI conversations occurring outside corporate governance. A concentration of usage in a small group of firms, and within those firms a small group of users, is consistent with both datasets.

Concentration also sits on the supply side. France's competition authority named OpenAI, Google and Anthropic as holders of more than 84% of the global AI agent market as of May 2026, citing Sensor Tower data, and flagged lock-in risk in a July 2026 opinion.

Thematic ETF flows rotate from bits to atoms

The newsletter opens on exchange-traded funds. According to data from Citadel, ETF net inflows are pacing for their strongest year on record, with July 2026 setting an all-time monthly record. Sternstein attributes the trend to a combination of low fees, low barriers to entry, easier distribution, marketing and higher retail participation generally.

The composition of thematic products has turned over almost completely inside six years. In 2020 the top five themes included clean energy, emerging markets technology and healthcare. By 2026 the leaderboard is occupied by artificial intelligence, nuclear, space, defence and infrastructure. The newsletter characterises the shift as a takeover by the capital-intensive world of atoms over bits, and declines to forecast whether the reallocation pays off.

For media planners, thematic fund flows are a proxy for which sectors are raising and deploying capital, and therefore which sectors carry advertising budgets attached to growth stories rather than maintenance.

Data centre construction and the blue collar wage premium

Data centres occupy the newsletter's longest section, and the numbers describe a construction impulse concentrated in a handful of states.

New Mexico and Wyoming each have less than three gigawatts under construction. Because neither state builds much of anything else, data centres account for roughly 60% of all private non-residential construction in both. Pennsylvania, a considerably larger state, has around three gigawatts under way, amounting to nearly 30% of its non-residential spending. Texas is building far more capacity in absolute terms, yet that construction represents about 10% of the state's total.

Wells Fargo tabulated concurrent economic effects. Since 2024, counties with operating data centres show a uniformly better picture across four measures: more housing, higher home values, lower unemployment and stronger job growth. Counties currently building show a clearly better employment picture, but a worse drawdown in new housing and weaker home value appreciation.

Sternstein records the causation problem directly. A large share of existing and new data centres sit in Loudoun County, Virginia, one of the wealthiest counties in the United States. A large share of new builds sit in Texas, which recorded historic residential construction and price appreciation before 2024, so any housing retreat there is measured against a far higher base.

The wage data is less ambiguous. According to Indeed figures cited in the newsletter, the data centre wage premium runs as high as 64% for a facilities manager and as low as 10% for electrical engineers. A comment in a Dallas Fed report, quoted in the newsletter from an unnamed heavy industrial construction contractor, described paying skilled concrete workers $28 to $32 per hour while data centres offered $45 per hour plus a $150 per diem, a premium of roughly 50%.

ADP data on job-switchers points the same way. Wage growth for workers changing jobs in construction, manufacturing and natural resources or mining runs 6 to 9.5 percentage points higher than for those staying put, a switching premium the newsletter describes as far above any other sector. Sternstein's summary is that "saying no to data centers almost certainly involves saying no to much higher wages" for blue collar workers.

The advertising relevance is indirect but not remote. Advertising revenue funds a substantial share of the capital expenditure behind that construction. Meta announced a strategic venture with BlackRock on July 28, 2026 to develop a one-gigawatt data centre in El Paso, Texas, in a quarter when its profit fell 8% to $15.8 billion and capital expenditure guidance sat at $125 billion to $145 billion.

Rideshare fares rise and gig work composition shifts

According to data from Gridwise Analytics, both the average and median Uber fare have increased roughly 20% since 2024 and appear to be rising still. Over the same period, median and average Lyft fares came in slightly cheaper than at the beginning of 2024, and run about 24% cheaper than Uber overall, though Lyft prices have been rising lately.

Platform fees account for much of the increase. Uber's platform fees have been rising for more than a year, with the median stepping up notably in October. Lyft's platform fee began rising only recently, following a period of substantial decline. Average gross driver pay per trip has also risen since 2024 and recently hit an all-time high. Whether higher prices eventually suppress demand remains open in the data.

The composition of gig work is changing alongside the pricing. Among Bank of America customer accounts, every category of gig work has increased except vacation rentals. Social commerce increased by more than 30%, far ahead of any other category, albeit from a smaller base. Sternstein offers three candidate explanations without selecting one: shifting media consumption from television to social platforms combined with the broader ecommerce expansion, artificial intelligence lowering the cost of the other operational components of running a social selling business, or advertising targeting on Instagram being effective enough to sustain the activity.

That growth rate sits above the pace recorded for the wider creator economy. United States creator economy advertising spend reached $37 billion in 2025 and is projected at $43.9 billion in 2026, a growth rate near 18.6%, according to IAB research published in December 2025. Platform-level social commerce figures have run hotter still: TikTok Shop reached $15.82 billion in United States ecommerce sales in 2025, growing 108% year over year, according to EMARKETER data.

Why the measurements matter for marketing

Three of the newsletter's findings bear directly on decisions being taken inside advertising organisations now.

The first is cost structure. If more than 85% of agent token consumption sits in the cached prompt, then the dominant variable in agentic tooling cost is context length and iteration depth, not model choice alone. That reframes procurement conversations that have largely centred on per-token list prices, and it gives clients a specific question to put to vendors and agencies about what portion of a token bill reflects cached rather than pre-fill volume.

The second is displacement risk inside the martech stack. Zapier, n8n and Make are load-bearing components in a large number of marketing data pipelines, from conversion imports to lead routing. Traffic declines are not the same as failure, and none of the three has announced a change in trajectory. Teams with production dependencies on those platforms nonetheless now have a public datapoint indicating that an agent-native competitor is the only one in the category gaining ground.

The third is the measurement problem that follows machines consuming a growing share of computing capacity and, separately, of web traffic. Automated requests overtook human ones on web-page traffic for the first time on record in mid-2026, reaching 57.5% of requests by June, per Cloudflare figures cited in IAB Australia guidance. The United States accounts for 53.5% of global bot traffic, according to Decodo analysis published on August 11, 2026. HUMAN Security research released on April 9, 2026 found automation growing eight times faster than human traffic. Lunio survey data found that 51.1% of marketers named automated bidding algorithms optimising toward non-human converters as a leading agentic risk to their 2026 strategy, while only 5.3% of the 131 marketers surveyed ran a dedicated invalid traffic platform.

The token data and the traffic data describe the same underlying shift from two angles. Agents are the fastest-growing consumers of inference capacity and the fastest-growing category of web requests. Neither trend has yet produced correspondingly large agentic media budgets. Magnite's chief executive placed the market in a discovery phase in August 2026, describing partner activity as one buyer transacting with one seller rather than the one-to-many pattern that would move meaningful spend.

Sternstein ends the section on agents with a caveat that applies to every figure in it: the technology is very early in its development, and it is already changing the landscape in substantial ways.

Timeline

Summary

Who: Moses Sternstein, writing in the a16z newsletter Charts of the Week, compiling data attributed to OpenRouter with charts by Peter Walker, OpenAI, Similarweb, Citadel, Wells Fargo, Indeed, ADP, Gridwise Analytics, Bank of America and the Federal Reserve Bank of Dallas.

What: A set of measurements covering agentic token consumption at nearly five times human levels with more than 85% of that burn in the cached prompt, roughly fourteenfold growth in agent usage since February 2026, double-digit trailing twelve-week traffic declines at n8n, Zapier and Make against growth at Gumloop, an eightfold token output gap between typical and top-decile enterprises, a 108-fold rise in Codex adoption among legal workers since February 2026, record ETF net inflows with thematic leadership rotating to artificial intelligence, nuclear, space, defence and infrastructure, data centre construction accounting for roughly 60% of private non-residential building in New Mexico and Wyoming, wage premiums of 10% to 64% for data centre roles, roughly 20% higher Uber fares since 2024, and social commerce gig work growth above 30% among Bank of America customer accounts.

When: Published August 21, 2026. Underlying data spans 2020 through July 2026, with several series indexed to April 2025 or February 2026 baselines.

Where: United States data throughout, with state-level construction figures for New Mexico, Wyoming, Pennsylvania, Texas and Virginia, and network-level token figures covering the OpenRouter routing layer globally.

Why: The figures quantify where autonomous software consumption is concentrated, which cost components drive it, and which incumbent tooling categories are losing share. For advertising and marketing organisations, the cached-prompt share reframes how agentic tooling costs are modelled and negotiated, the automation platform declines flag dependency risk inside production martech pipelines, and the widening firm-level adoption gap indicates that sector averages conceal order-of-magnitude differences between competitors in the same industry.