Newsletter · 2026-09-12 · Ashutosh Agarwal

Eight Sleep Runs a 100 Million Dollar Email Channel With a Team of Zero - How They Build - Week of September 12, 2026

How They Build for the week of September 12, 2026: Eight Sleep's co-founder described an email marketing channel doing close to 100 million dollars now run by nobody after a co-founder built the bots in three days, Cloudflare's CEO explained why his 1,100-person layoff hit the measurers rather than the coders, Uber began cutting micro-teams, and one AI evaluation firm burned 1.5 million dollars of tokens in a month, ten times its salary bill.

How They Build

Week of September 12, 2026: Eight Sleep Runs a 100 Million Dollar Email Channel With a Team of Zero

This week a founder said the quiet part out loud: his email marketing team used to be two people, now it's nobody, and it still drives close to $100 million. One co-founder built the bots that replaced the team in three days. Meanwhile, the layoffs stopped hitting the coders and started hitting the middle of the org chart, the managers, the "measurers," the people whose whole job was coordination.

The Number: 2 to 0 people, running a roughly $100 million channel

The most striking efficiency claim of the week came from Matteo Franceschetti, co-founder of Eight Sleep (the smart-mattress company), talking to Harry Stebbings on 20VC. Here is the story in his words:

"Before we had an email marketing team, now we have zero. Everything is done through AI and email marketing makes a very large amount for us... we had a team of two. Then the person leading the team was leaving. And at that point, my co-founder, Alexandra, jumped in and said, okay, let me see how I can use AI to really make this work. And within three days, she was able to build multiple bots that now run all our email marketing. So now we have a team of zero and email marketing makes close to 100 million."

Read that again: a marketing channel doing close to $100 million is now run by no one, and the bots were built by a single co-founder over a long weekend. The bots decide which email to send, when, and with what copy, based on the company's own historical data.

And email is just the loudest example. The same thing has happened to paid advertising and finance:

• Paid media: 2 people, "hundreds of millions." "Right now, our paid media team makes hundreds of millions, and there are two people." Every morning the team gets a report from AI agents suggesting what to change; the humans just approve or reject.
• Finance: 4 people instead of roughly 20. "Probably a company like ours you have, I don't know, a team of 20 people in finance. Our team is four people... I just really like to keep teams small and try to see if I can 5X the output."
• The engineers stopped writing code a year ago. "Our engineers stopped coding around a year ago. Since then, they didn't code. What they have is hundreds of AI engineers that code for them. And that is how we can achieve what we are achieving at scale in 35 countries." He now literally counts two workforces: "the human employees, but then there is also the AI employees. And... we are probably three, four X bigger if you start including all our AI employees."

The punchline stat: Eight Sleep is 160 people, and Franceschetti says a company at their revenue would normally be around 1,000. "Our revenue per employees is way higher than Apple... Our output is 5x what is average in Silicon Valley." His three-year plan is "250 people making a billion."

Why this is the number of the week: we have covered plenty of "solo founder, $1M ARR" stories, but this is different, a real, scaled consumer brand quietly deleting entire functions, not jobs. The lesson for founders isn't "fire everyone." It's the mental model underneath it: Franceschetti says he now designs the company "in terms of teams of two," two world-class people overseeing an area that AI makes enormous. The Twenty Minute VC (20VC), "7 Predictions for How AI Changes the World... with Matteo Franceschetti, Co-Founder @ Eight Sleep" (Sep 12, 2026).

What Founders Changed

$35 million in revenue. 32 people. Bootstrapped. The cleanest lean-SaaS confession of the week came from Adam Robinson, founder of Retention.com, on the ProfitLed podcast. A former Wall Street derivatives trader who spent seven brutal years on an earlier company (including two in-person mass layoffs, "27 out of 35 of our people" the first time), he now runs a business that clears roughly $1.1 million of revenue per employee with no outside investors:

"We're at 35 million ARR and we're like 32 people. I mean, by any measure, that is a significant win, right? Especially as a bootstrapper."

What makes it a "How They Build" story is the target he set himself years ago and then beat: "25 million ARR, 35 people max. Like this is insane. And it's like getting better by the day." The screenshot-worthy part isn't a productivity multiplier, it's that a two-time founder who once had to fire a room full of people now treats a 30-person, $35M company as the goal, not a way station on the road to headcount. Lean stopped being a constraint and became the win condition. ProfitLed Podcast, "Living the Dream at $35M ARR, 32 people | Adam Robinson, S3E10" (Sep 9, 2026).

Zero to $5 million ARR in one month, with a team of 12. Eric Simons, CEO of Bolt (the AI app-builder, formerly StackBlitz), told the Product Podcast one of the wildest revenue-ramp stories in software. His company spent seven years and got nowhere: "before that, we had spent seven years getting to 500K." A board meeting to wind the company down was already on the calendar. Then they launched Bolt in October 2024:

"In the first month, we went from zero to five million of ARR... And then the month after we went from 5.5 to 20.5 or something... And our team was like 10 or 12 people at that time. And we overnight, we just woke up with tens of thousands or hundreds of thousands of paying customers."

The building lesson he keeps repeating is "the best advice is don't die," stay in the game long enough to catch the wave. Normally, he notes, "to grow the team of a company that's at 20 million of ARR, usually you have minimum two years to prepare for that. And we, that happened in two months for us. And so there were no playbooks." A dozen people, a product that went 40x in eight weeks, and no operating manual for that speed. The Product Podcast, "Bolt CEO on Turning AI Prototypes Into Production Code Engineers Trust... a 5M-to-20M ARR Ramp in Two Months | Eric Simons | E311" (Sep 9, 2026).

Build the AI salesforce three years early, and pay your humans for what the bots close. Carles Reina was employee number four and the first go-to-market hire at ElevenLabs, which he helped scale from $0 to over $600 million in revenue in under four years ($0 to $100M in 20 months, then plus $100M in 10, plus $130M in 5, plus $270M in 6). On the SaaStr podcast he described pitching an AI sales team back when the founders thought it was crazy, 2.5 years ago, at a 30-person offsite:

"I pitched them this idea of... I want to build an AI go-to-market person. I want to build an AI account exec, an AI SDR, an AI customer success manager... and everyone was like, no, the technology is just not there."

He kept pushing, got one developer dedicated to it, and built agents that answer inbound within 30 minutes and run upsells on the long tail of SMB accounts. The clever org move was on compensation: to stop his human reps from fearing or fighting the agents, ElevenLabs pays the human account owner commission on revenue the AI generates. "You're double compensating... but compensate the process so that you remove any friction."

The other screenshot number: Reina's controversial "20x quota" rule, a rep on a $100K base is expected to bring in $2 million a year. It drew accusations online that he was "slaving people," but he says it worked because the AI tooling made it achievable, and the team now runs at 167% average quarterly quota attainment, with some reps at 300% to 600%. "I would rather have a smaller team getting very well compensated than a bloated team that is not reaching quota." The Official SaaStr Podcast, "SaaStr 877 CRO Confidential: 0 to $600M in Under 4 Years. The ElevenLabs GTM Playbook with Carles Reina" (Sep 9, 2026).

The layoffs moved to the middle of the org chart. The biggest shift this week wasn't founders bragging about lean teams, it was a public-company CEO explaining, in unusual detail, which jobs AI actually erased. Matthew Prince, CEO of Cloudflare, laid off about 1,100 people (roughly 20% of the company) in early May, right after a record first quarter. His framework, from a widely-shared Wall Street Journal op-ed and this Empire interview, splits a company into three groups:

"There are builders, there are sellers, and there are measurers... the AI can now do all the measuring. The AI does the audits, the finance, the legal, the compliance, the middle management."

He is emphatically not firing the coders. The opposite: "If I can have a developer [who] is now 10 times as productive, I'm going to hire as many developers as I possibly can." Salespeople went from spending roughly 40% of their time with customers to roughly 90%, because AI does the presentations and CRM busywork. What got cut was the coordination layer, and Prince explained the mechanism precisely. The old Harvard Business School rule was roughly 6 direct reports per manager; AI tools let Cloudflare push that toward 12, and "as you do that, the organization flattens... we just didn't need as many middle managers."

He's blunt that it was awful ("real, dear friends of mine that we let go... I got a bunch of death threats") but frames waiting to cut as "chicken shit leadership": once you know you'll restructure, "every day you wait is actually a disservice to the employees you're eventually going to lay off." And the demand for humans hasn't vanished, Cloudflare is hiring over 1,000 interns this year ("everybody else has stopped hiring interns... we like to zag when everyone else is zigging"), and embedding them in senior teams to teach the veterans new AI tools. A telling anecdote from the host, at Blockworks: he offered a team lead another headcount and got turned down, "I've got eight agents right now. Don't give me another human... I hate human management, but I've never had more fun managing these agents." Empire, "Cloudflare CEO: The Man Deciding The Internet's Next Business Model" (Sep 7, 2026).

The Other Side

The "delete the middle" story has a dangerous failure mode, and this week produced a genuinely sharp warning about it.

When you cut a "micro-team," you might be firing the person holding the department together. A breakdown episode of the Elon Musk Podcast laid out the new corporate purge: Uber is cutting about 3,300 jobs (10% of its workforce), specifically targeting "micro-teams," managers with only one or two direct reports, and slicing that group by nearly half. It's a pattern: "Intel is cutting layers of management from 12 down to 6... Google is trimming their small team managers by 35%. Coinbase is doing the exact same thing." Middle managers went from having the safest seats in the building to being "pure overhead."

The problem is that executives make these cuts from a spreadsheet. The episode's cautionary tale: "Alex Butterworth," a senior legal counsel at Uber, showed just two direct reports in the HR software, a textbook "inefficient micro-team." But the reality: Alex was actually coordinating four offshore contractors plus an indirect report (a seven-person international unit), and personally "drafting and negotiating hundreds of corporate contracts every single year." The software counted the offshore people as vendors, so the dashboard showed a team of two. "The human resources dashboard is functionally lying to the executive... You are basically yanking the engine out of the car to make it lighter and then wondering why it won't drive."

And the "player-coach" model that's supposed to replace those managers? The episode is skeptical: asking one person to be an elite individual contributor and manage a team means "the company is extracting two full-time salaries worth of labor from a single person under the guise of organizational agility," and the coaching quietly disappears, because mentoring doesn't show up on a quarterly dashboard the way shipped code does. Elon Musk Podcast, "Why Big Tech Is Axing Micro-teams" (Sep 6, 2026).

Most companies buying AI still aren't getting a return on it. The soberest data of the week came from a Moody's economist using Ramp's spend data. The uncomfortable finding, in plain terms: "The vast majority of firms that are using AI are not getting ROI from it." That's why, she argues, AI still barely shows up in national productivity statistics, the gains are real but concentrated in a tiny group of intense users. More on the numbers below, but the headline for founders is a useful gut-check: buying the tools is not the same as getting the leverage. Moody's Talks, Inside Economics, "The AI Series: Adoption Ramps Up" (Sep 8, 2026).

Even the AI-native sales leader says the agents aren't good enough at the human part. Carles Reina, for all his AI-salesforce enthusiasm, drew a hard line: "I do still believe that AI agents are not that good today [at] something that is very relationship-driven." Big deals still need a human on a plane. LLMs, he says, are fantastic at analyzing data and working inside the "distribution" of what's normal, but real selling and genuinely creative campaigns (like the ElevenLabs grants program that pulled demand away from competitors) happen "outside the distribution," where humans still win. SaaStr 877 (Sep 9, 2026).

The Cost Corner

The theme underneath everything this week: the AI bill is starting to rival payroll, and the numbers are getting absurd.

"Almost an employee's worth of salary in tokens." On the a16z podcast, Rayan Krishnan of Vals (a firm that evaluates AI models) shared a Fortune 10 company's setup: each engineer gets roughly $100 a day of Claude Code budget, recently raised to $300 a day per employee. His framing: "almost an employee's worth of salary in tokens for them to use." The daily limits are so binding that they've reshaped the workday, the rate limit resets at 4 p.m., "so the most productive hours of work are actually now 4 to 6 p.m.," with a dead zone in the early afternoon where "people go on walks or get a coffee because they just don't have the rate limits." Krishnan's blunt prediction: "token spend may start to eclipse salary spend," which means "you actually have to justify the ROI much more keenly than you've seen over the last six months."

His own cautionary experiment is the screenshot stat: Vals gave its team unlimited access to coding tools for a month. "We had a lot of engineers spending between one to two billion tokens a day. Peak day was one engineer spending six billion." The damage: "We spent roughly $1.5 million worth of tokens... it was 10x more we were spending in tokens than employee salary for that month." The fix wasn't a blanket cap, it was measurement: they built a tool to figure out which model to use for which task, found that some agents (they cite Cognition's Devin) are far more "token efficient," and now auto-recommend a starting model per ticket. A counterintuitive nugget for anyone tuning their own bill: Anthropic's cheaper Sonnet model can end up more expensive than the premium Opus, "because it is so token hungry." The a16z Show, "Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan" (Sep 9, 2026).

The spend is wildly lopsided, and the price of intelligence is now falling. Moody's, using Ramp's itemized receipt data, put hard distribution numbers on the market:

• The top 1% of businesses spend about $7,400 per employee per month on AI. The top 10% spend about $650. The median spends about $12, basically one chat subscription per person.
• 90% of all AI spend comes from less than 10% of customers. This is a power-law market, not a broad one.
• The price lever the model companies were counting on is eroding: the effective price per million tokens "has fallen like in a month, like 30%, 40%." OpenAI just cut its frontier model's price by another 20%, and Anthropic hadn't yet responded.
• Cheap models still aren't winning, though: only about 6% of businesses use open-source models, and even those keep spending more on OpenAI and Anthropic. Anthropic's most expensive frontier model has only cracked roughly 5% to 10% of business usage, partly on price, partly because government-mandated safety checks force 30-day data retention, "a non-starter for many businesses."

The strategic read for founders: demand ("quantity") keeps climbing, but "there's no price lever anymore... it's all going to be a volume play." If you're modeling your own AI costs into 2027, assume the per-token price keeps dropping, and assume your usage rises fast enough to eat the savings anyway. Moody's Talks, Inside Economics, "The AI Series: Adoption Ramps Up" (Sep 8, 2026).

The five-to-seven-million-dollar surprise. Reinforcing the trend from the enterprise side, an episode with Dell's Allen Clingerman described cloud AI bills of $5 million to $7 million a month that "shocked" customers, even as the price of tokens fell, and made the case for routing each query to the lowest-cost model and location (edge, core, or cloud) rather than reflexively reaching for the priciest frontier model. Same lesson Eight Sleep is living on the other end of the size spectrum: Franceschetti won't even quote his monthly Claude bill precisely because "it keeps growing so fast that I don't know," and he framed the investor question of the moment. Salesforce reportedly spends $300M a year on Claude, about 5% of its engineering budget; does that 5% stay flat, climb to 50%, or collapse to 1% as models get cheaper? His bet: usage rises and price falls at the same time, and "net-net, I think it will go down," the way nobody worries about the cost of electricity anymore. Between Fires and Futures, "The Cost of Intelligence: Why Your AI Bill Exploded While Tokens Got Cheap, with Dell's Allen Clingerman" (Sep 7, 2026); 20VC, Eight Sleep (Sep 12, 2026).