← The stand
Cover of The Collection, Volume 1, Number 57: The Leaderboard. Saturday 24 October 2026, Melbourne. Two ranked lists pinned side by side above a wooden desk, one by visits and one by spend, with oxblood threads crossing between them, a turnstile clicker on one side and a stack of coins on the other. Skip cover

Vol. 1  ·  No. 57  ·  Saturday 24 October 2026  ·  Melbourne


The Collection

The Leaderboard

Collected and edited by Newsletter World for AK

Contents

A new week, and the week’s security fire is left on Friday’s shelf. Saturday’s letters are about lists. a16z ranked consumer AI by what people pay for and found half the winners missing from the traffic charts. Its own chart pack shows attention near a ceiling and old GPUs renting above their book value. A sales writer says the easiest agent to sell is the one that keeps score. Two compute forecasts sit twenty-five gigawatts apart. And a fact can be true and still come from the wrong page. They ranked AI by the crowd. The money kept its own list.

  1. iiiEditor’s LetterThey ranked AI by the crowd. The money kept its own list.03
  2. ivThe SpendHalf of America uses it. One in twenty-two pays.04
  3. vThe CeilingAttention is capped. The chips are not depreciating on schedule.05
  4. viThe YardstickSell the agent that keeps the score.06
  5. viiThe ForecastFifty gigawatts or twenty-five, and a CEO living on 2024’s order.07
  6. viiiThe SourceTrue somewhere is not true here.08
  7. ixStanding OrdersFour rules for this issue.09

03  ·  Editor’s Letter

They ranked AI by the crowd. The money kept its own list.

For six editions, a16z’s list of the top consumer AI apps was a list of where people went. The seventh, published this week, also asks where people paid, using card-spend data. The two lists barely agree. Of the 50 products ranked by spend, 29 appear on neither traffic chart. Replit is 48th on the web by visits and sixth by revenue. Midjourney has fallen off the traffic list entirely and sits eleventh by money.

That is the Saturday lesson, and it travels. A leaderboard is an argument about what counts, made before anyone looks at the results. Count visits and you get a popularity contest. Count spend and you find a small, intense group paying a median user’s bill thirty-six times over. Count successful pilots and you find most projects never had a number to beat. Count signed data-centre contracts and you get one forecast; count the electricians and you get half of it.

None of the letters below says the crowd is wrong. They say the crowd is one list, and the decisions that matter, what to build, what to price, what to buy, keep getting made off the wrong one. Pick the list before you read the ranking.

The Collection · The Leaderboard03

04  ·  The Spend

Half of America uses it. One in twenty-two pays.

The first consumer AI ranking by revenue finds a market carried by a heavy few, and a set of winners nobody saw on the traffic charts.

About half of Americans now say they use AI. About 4.5 percent pay for it, double the share a year ago. The new a16z edition, built on YipitData card spend alongside the usual traffic, shows how lopsided the paying side is. The median paying user spends $25 a month. The top 1 percent of payers spend $903. The top tenth of payers bring in more than half of all revenue, the top 1 percent about a fifth, and the bottom half just 16 percent.

The rankings move once money is the measure. Suno is seventh by revenue and 19th by traffic. ElevenLabs is tenth and 25th. Claude has passed Gemini in US paid subscribers, and about 7.5 percent of Anthropic’s subscribers are on the $100-plus Max plan, against roughly 1 percent at ChatGPT and Gemini. ChatGPT still owns the crowd, with 1.2 billion weekly users, and OpenAI was running at about $1 billion a year in ad revenue as of August.

29 of the 50 top earners were on neither traffic list.

The business models are plain. About 85 percent of the listed products charge subscriptions; 13 percent run ads. The letter’s best example of why people pay is a voice clone: one that sounds 99.9 percent like you takes ten seconds to edit, while one at 85 percent takes ten minutes or more. The last few points of quality are where the bill gets paid.

The gaps are as telling as the winners. By the letter’s count, nine of fifteen consumer categories still have no AI product at all. Newer apps are finding payers fast: Instinct users spent more than $1,000 on average in their first month. And the letter relays a line from Eugenia of Wabi that reads like a rebuke to every productivity pitch: most people are not looking to save time, they are looking for ways to spend it.

The Collection · The Leaderboard04

05  ·  The Ceiling

Attention is capped. The chips are not depreciating on schedule.

A chart pack with three lists that disagree with their own assumptions: hours in the week, the book value of a GPU, and what the market pays for security.

The title is the first chart. Netflix growth has slipped to just under 12 percent, its lowest since 2021 and 2022, and BCG puts US media consumption at a ceiling of about 26 hours a week. If attention is near full, a traffic ranking is a ranking of who took hours from whom. It cannot grow the pie. That is one more reason the spend list in the previous piece is the more useful one.

The second chart is about the accounting list. Rental rates for GPUs are up across every model a16z tracks, A100s included, and they sit between 9 and 143 percent above where straight-line six-year depreciation says those chips should be priced by now. The schedule on the books said old chips would fade. The market is paying for them anyway, and token demand on OpenRouter keeps turning upward.

The book said the old chips would fade. The renters did not read it.

The third is about what investors reward. Cyber stocks sold off this month, after gains of up to about 125 percent, while revenue growth across the group barely moved. The pack’s security note to portfolio chief executives, which this paper covered at length on Friday, comes down to two lines: double a security budget that typically runs near 10 percent, and treat every AI agent as an insider threat.

Two more charts round out the pack. Prices for electrical equipment have surged while domestic production capacity fell and imports outran output. And ADP data shows hours per worker rising faster than headcount, with job-switchers earning a premium of about five percentage points. Firms are buying more hours from the people they already have.

The Collection · The Leaderboard05

06  ·  The Yardstick

Sell the agent that keeps the score

Most AI projects stall because nobody agreed what would count as success. A monitoring agent brings its own number.

The surveys agree on the failure rate. Orgvue asked 1,163 leaders and found 78 percent of AI projects had either failed, 35 percent, or stalled in pilot, 43 percent. Eighty-four percent said they should have set ROI targets first. Gartner asked 782 infrastructure and operations leaders and found only 28 percent of use cases fully succeed, with 20 percent failing outright. IDC found 69 percent see a gap between what they spend and what they can measure.

IDC’s Tony Olvet supplies the test that cuts through: can the customer report the result on an earnings call? Most pilots cannot, because the number was never chosen. The writer’s answer is to sell the agent whose job is watching, a monitor whose output is the measurement. It arrives with its own yardstick, so the buyer does not have to invent one.

The contract is the off-ramp, not the finish line.

He borrows three questions from the sales trainer Mahan Khalsa to ask before any build. How do you measure it? What is it now? What would you like it to be? Then he points at the vendors that already price this way. Sierra moved to outcome pricing in December 2024. Zendesk charges per resolution, and only counts a resolution after a 72-hour quiet period in which the customer does not come back. That quiet period is the whole idea in one clause: the score is kept by what happens after the agent stops talking.

The Collection · The Leaderboard06

07  ·  The Forecast

Fifty gigawatts or twenty-five

An analyst’s count of contracts against an investor’s count of what can be built, and a chief executive who says today’s compute was ordered two years ago.

SemiAnalysis forecasts 50 gigawatts of data-centre capacity coming online in 2027, counted from signed contracts and construction already visible. Altimeter’s Brad Gerstner called that “too aggressive” on the All-In podcast and put the figure nearer 25. SemiAnalysis answered with satellite images of the sites. IREN’s chief executive, Daniel Roberts, added how his company pays for its GPUs: customers prepay about half, and debt covers most of the rest. Newcomer’s summary is fair to both sides. The money is arriving. Whether the buildings and chips are physically running by 2027 is the open question, and Gerstner, it reports, found the latest case more convincing.

Sundar Pichai gave the other half of the answer to Big Technology. “The compute we have today is a function of decisions I made in 2024,” he said. Everyone is short, he added, and “everyone wishes you had two times what you had planned for.” Google Cloud, which the letter says grew 82 percent last quarter, is “forgoing short-term revenues” to keep frontier research supplied.

“The compute we have today is a function of decisions I made in 2024.”

Pichai also pushed back on the loudest list in the industry, the countdown to machines that improve themselves. “We are pretty far from the end state of RSI,” he said, while allowing that agents already help researchers build the next models. Newcomer, meanwhile, carries a ranking of its own: Stanford’s Ilya Strebulaev and Ohio State’s Blake Jackson put a16z at the top of 50 AI-focused venture firms, ahead of Lightspeed, Menlo, Sequoia and ICONIQ, on valuations, growth, dilution and distributions. As Anthropic’s filing nears and OpenAI is reported by CNBC to be raising $30 billion at a $1.25 trillion valuation, every firm on that list is about to be re-ranked by the only measure LPs keep: cash back.

The Collection · The Leaderboard07

08  ·  The Source

True somewhere is not true here

A retrieval answer can pass every source check and still describe the wrong customer.

Kotrotsos opens with a made-up support reply: “Your plan includes a 30-day refund window.” The assistant retrieved a policy with a 30-day window and an account record. Every important phrase in its answer appears somewhere in what it read. Yet the policy may belong to a different plan, or the assistant may imply the account record itself grants the refund.

A source check that pools all the retrieved text will pass that answer. The preview names the research it draws on, ProvenanceGuard, and the failure it targets: cross-source conflation in tool-using agents, where a claim is supported by some evidence on the table but credited to the wrong source.

A fact found in the evidence is not always a fact about this customer.

It is the same problem as the rest of this issue, at the scale of one sentence. The check was real. It was scoring against the wrong list. What the full piece proposes as a fix sits behind Medium’s paywall, and this paper has not read it.

The Collection · The Leaderboard08

09  ·  Standing Orders

Four rules for this issue

  1. I

    Choose the list before the ranking.

    Visits, spend, resolutions and contracts each crown a different winner. Decide which one your decision depends on, then read the chart.

  2. II

    Find the heavy few.

    When the top tenth of payers bring in half the revenue, build and price for them by name, not for the average.

  3. III

    Ask the three questions first.

    How do you measure it, what is it now, what should it be. A project that cannot answer them is a pilot waiting to stall.

  4. IV

    Check where the fact came from.

    An answer is grounded only if each claim traces to the source that actually covers this case, not to whatever happened to be retrieved.