Follow the Money: Power Laws and the Agent Wave in Consumer AI

The seventh edition of a list published for three years changes the measurement itself: for the first time, the team places US consumer card spending next to web traffic. Sourced from the YipitData panel, this spend data excludes enterprise and small-business payments and tracks only the individual wallet. According to the a16z team, the new lens makes visible dozens of products that never reach the traffic lists, from desktop apps to agents living inside messaging platforms. The message is blunt: consumer AI is now measured in money, not just visits.

On the traffic side there are signs of maturity: only 11 new products entered the web and mobile lists in this edition, the weakest refresh in the series. Yet the calm surface should not be misread; every product that stayed on the list kept growing visibly. The narrators stress the gap between the buzz in Silicon Valley conversations and the landscape of the wider world. AI is seeping everywhere, but the shop window is slowly settling. Call it traffic maturation .

The spending rank tells a very different story: 29 of the top 50 products by spend sit on neither traffic list. Where the money flows and where the clicks flow have come apart. That split stands out as one of the list's most provocative findings. The segment hiding in the dark looks like the main story of the next edition. The spend lens drags a hidden economy into daylight.

Power law: a paying minority, a spending core

The gap between breadth of use and depth of payment is striking: about half of Americans say they use AI, and a quarter are in contact with it nearly every day. Yet the share paying for a subscription hovers around 4.5 percent, with some measures pointing to the 2 to 2.5 percent band. The rate has doubled in a year, but the picture holds: the crowd tries, a minority pays. The gap shows consumer adoption moving at two speeds.

Inside the payers, a sharp power law operates: the most generous 10 percent of spenders generate more than half of revenue, while the top 1 percent alone carries a fifth of it. The bottom half keeps only 16 percent. A top-1-percent user paying on a personal card spends $93 a month, while the median user stops at $25. Money pools in a small, passionate core.

The core's profile is no surprise: the biggest spenders spend in order to produce. Developer tools, productivity apps, and creative production tools are heavily over-represented in this group. Names such as Granola, Higgsfield, and Manus sit far above average in the baskets of top spenders. As one speaker puts it, people use personal cards to write code they never wrote and produce videos they never produced. The result is a maker economy : payment is made for the joy of creating, not for saving.

The personal agent wave

The hottest topic of recent months is personal agents: OpenClaw staged a surge just outside the last measurement window that would have placed it high on a retrospective ranking, yet it is absent this time after its traffic collapsed. The team moving to OpenAI, with work continuing through an open-source foundation, writes the second act of the story. Grockbot and Town, pro-consumer hybrid agents, plus OpenAI's Dots move, are candidates to fill the gap. The cycle recalls a classic agent pendulum , where pioneers light the fire and successors take the stage.

All eyes on the consumer side are on Meta's Muse app: 500,000 downloads and 250,000 active users in the first 12 days after launch sparked excitement across tech circles. But the scale check sobers things up: live only in the US and Canada, Muse reached 5 million downloads in the same window while Threads hit 16 million in 22 days in the same geography. According to TechCrunch, citing Sensor Tower data, Muse later passed 3.4 million downloads and gained backing on the Meta Connect stage with video chat, computer use on the Mac, and smart-glasses integration. The race to the average user remains far from the mainstream threshold .

The boldest early agent numbers come from Instinct: 100,000 users, 10 percent day-over-day growth, 40 percent connecting a card in the first three weeks, and average first-month spending above $1,000 per user. According to the table compiled by MasterNodeAI, Meta's Muse app is playing for the store crown with 600,000 daily active users while Instinct went out to raise at a $2.5 billion valuation. The figures still reflect niche excitement, but the momentum is undeniable. The picture signals a consumer agent market starting to consolidate.

Distribution strategy is deliberately cautious: serving agents costs so much that companies do not want the product to explode overnight. Unlike Threads, personal tools like Muse need no critical-mass threshold; the value lies in personality, not crowds. The key to staying power is seen in the ecosystem: a marketplace and business network woven from hundreds of partnerships will carry the product from one-time trial to daily habit. The speakers' thesis is that compounding platform effects will operate for agents too.

Platforms and the trust wall

Platform posture runs both ways: Amazon shuts the door on agents like Muse browsing and buying on its site, while hundreds of partnerships on the Shopify side are already live. A deeper gap remains: no assistant product has yet unlocked person-to-person network effects. The more intimately software knows its user than any past product, the less appetite there is to carry that intimacy into rooms with other people. As with the split between work and personal accounts, the privacy-network tension in agents is unresolved.

The trust wall stands as the tallest barrier to growth: no mass adoption without trusting software that reaches your email, calendar, and credit card. Fear of rogue actions or sharing private material with the wrong person sits in every early adopter's mind. People intuitively know what to share with whom; now we expect the same manners from software. As a society we are learning these new norms together, and the lesson shows the debate over trust in AI is a human question, not a technical one.

Cost, usage, and the empty box

Unit economics remain foggy: some assistant startups are said to absorb hundreds or even thousands of dollars a month in serving cost per user. Data compiled by the team behind an assistant benchmark, with over 1,500 early users, tracks more than 170 agents, and the number-one use case is coding and technical automation. For products aimed at the mainstream consumer, cost runs at tens of dollars per user instead. The spread puts numbers on the adoption gap between early adopters and average users.

The product handed to a new user often looks like an empty box: a reservation or flight-booking demo means nothing to someone with no such need right now. On social media people see what others try and add riff upon riff, yet in AI this shared discovery loop has not formed. The experience is deeply personal; nobody sees what anyone else asks their agent. Outside coding and productivity, the absence of this social discovery mechanism stands as the quiet barrier to a mainstream leap.

The business-model inversion and the return of ads

The revenue table runs against internet history: 85 percent of products on the web list earn through subscriptions, 62 percent through credits or token payments, and only 13 percent have ads. Yet today's giant consumer companies make their money mostly from advertising and transaction fees. ChatGPT entered the global top 20 consumer subscriptions within three and a half years; most other names there are media and platform companies. The core claim of the segment: as inference costs fall, the subscription mandate will loosen and ads and transaction-fee models will return.

The first proof of the return comes from OpenAI: as Reuters reported, ChatGPT ads raced to a $1 billion annualized revenue run rate in under 200 days. According to OpenAI's official announcement, the platform reached tens of thousands of advertisers, and its Ads Manager opened across India, Europe, the Middle East, and North Africa. According to The Decoder, citing OpenAI's developer-event figures, weekly users approach 1.2 billion, a density that explains why ad inventory filled so fast. Because pushing ads into personal conversation can damage trust, delivery must follow labeling and moments of commercial intent ; vertical cases such as Open Evidence, reaching 50 to 60 percent of physicians, show how a targeted audience makes advertising valuable.

Three giants diverge

The big labs are drifting apart: ChatGPT runs at roughly 6 times Claude and 2 times Gemini in web traffic, with about three times the paid subscribers of both rivals combined in the US. The surprise is Claude overtaking Gemini in paid subscriber count despite a far smaller installed base. Anthropic keeps the ads book shut and leans on subscriptions; the share of subscribers on plans above $100 runs at 7.5 percent, clearly above the 1 percent at rivals. The users of the three products diverge too, turning assistant divergence into one of the report's sharpest structural findings.

Creative tools show split fronts: in audio, specialists such as ElevenLabs and Suno quickly seized the space that labs, unwilling to wrestle with IP headaches, left open. In image and video the picture flipped; Midjourney fell off the traffic list yet returned in the revenue ranking, and according to WinBuzzer the company stays a profitable exception with revenue above $200 million. OpenAI's image model and Google's Nano Banana and VO moves absorb independent makers' consumer traffic, while Chinese firms with broad data access gain an edge in video. The rising wave is agent-friendly tools: models that roam the internet and plug into other services give compatible creative products an extra tailwind.

An opening at the interface layer

The quiet engine of growth is the slide from individual to enterprise: products such as Granola, Whisper Flow, and Superhuman get loved in personal use, pulled into work, then add privacy, security, and team plans to reach enterprise revenue. Where Canva took more than six years for that journey, today's crossing runs far faster and cheaper. Incumbents' inability to cannibalize their own interfaces widens the opening; neither Google Docs and Gmail nor the big labs, outside chat and coding interfaces, have been reinvented. On hardware, according to TechCrunch, Plaud proves the window with over 2 million devices sold and $100 million in annualized revenue. And context playbooks in products like Town, which learn a user's voice and setting, build a hard-to-copy lock-in: the gap between a 99.9 percent on-voice email and an 85 percent one is ten seconds of fixing versus ten minutes of rework.

The white-space map

The white-space map whets the appetite: today's consumer AI clusters almost entirely around search replacement, productivity, and design editing. Network categories such as dating, recruiting, social AI, and shopping sit nearly empty on the list; in gaming and entertainment, models have yet to win the average viewer, with booming AI micro-dramas the exception. The founder of Wabby sums up the segment: people look for ways to spend time, not save it. Everything curious tinkerers try, share, and package into products moves value from the model to the experience layer .