The most striking claim from Eight Sleep co-founder Matteo Franceschetti is not that his company uses AI. It’s that AI now runs entire functions that used to require human teams. When his email marketing lead was departing, his co-founder Alexandra Zubillaga stepped in to see whether bots could do the job. “Within 3 days she was able to build multiple bots that now run all our email marketing and so now we have a team of zero and email marketing makes close to 100 million,” he told 20VC host Harry Stebbings.
That was the concrete example. The broader claim, delivered with characteristic bluntness, is even bigger: “Our engineers stopped coding around a year ago. What they have is hundreds of AI engineers that code for them.” Eight Sleep now operates across 35 countries, including China and the Middle East, with a team of just 160 people — a scale Franceschetti says would require roughly 1,000 at a comparable company.
A company rebuilt around agents
The transformation at Eight Sleep has been structural, not cosmetic. Franceschetti estimates the company now runs “probably by now hundreds if not thousands of agents,” and he has started measuring “AI employees” alongside human ones. By his math, the company is three to four times larger when AI headcount is included. The tooling is primarily Claude, and his monthly spend is “in the millions” — under $5 million, though growing so fast he no longer tracks it precisely.
The headcount numbers tell the story:
| Function | Headcount | Output / scope | Notes |
|---|---|---|---|
| Engineering | Small human team | Hundreds of AI engineers | Stopped human coding ~late 2025; mainly Claude |
| Email marketing | 0 (was 2) | ~$100M revenue | Bots built by Alexandra Zubillaga in 3 days |
| Paid media | 2 | Hundreds of millions in spend | Decide Meta allocations and creatives |
| Finance | 4 | Comparable firms ~20 | Major headcount savings |
| AI internal tools | New team (~6 months old) | Thousands of internal agents | Must stay coupled to data engineering |
| Total company | 160 | Revenue per employee “way higher than Apple” | Comparable firms at this revenue ~1,000 people |
Marketing staff receive a morning report from AI agents suggesting changes to make or reject. The growth function is run by “a few people.” What Franceschetti calls the “AI internal tools” team — which didn’t exist six months before the conversation — builds the agents, and he insists it stay tightly coupled to the data engineering team so agents have correct access to revenue-by-country and CAC-by-country data. His organizational philosophy is “teams of two” for redundancy and world-class oversight, kept “extremely small and flat.”
The $200 million illusion in paid media
Franceschetti’s most valuable operating insight for other founders is not about AI. It’s about attribution. Eight Sleep does not trust platform-reported customer acquisition costs because platforms have every incentive to overstate their own contribution. “You cannot trust there will be an excess of attribution from the same platform. If you look at the CAC on Meta it is probably 20% lower than what the true CAC is,” he noted.
The correction mechanism is incrementality testing roughly every six months. His method: pick two comparable geographies — his example was Texas versus California — turn off a channel in one, and measure the revenue delta. If both states were growing 50% year over year and you kill Meta in one, the difference reveals Meta’s actual value for that market. “Every brand needs its own models because otherwise they just fool themselves,” he said.
This discipline shapes the company’s entire growth posture. Eight Sleep targets an immediate payback period and a healthy day-zero contribution margin — a philosophy Franceschetti credits to investor Keith Rabois, who “was really pushy and he was correct since day one.” The cost of that discipline is meaningful: “We could be growing 50% more if we wanted year over year this year just by spending more. But then next year you’re in trouble because your CAC is upside down.”
Channel expansion, he argues, must be sequential rather than simultaneous. Start with Meta, set a weekly budget based on company size, establish a hard CAC cap, and scale only within that cap. The consequences of ignoring this are brutal: overspend, pull back, enter the next year with lower spend and no year-over-year growth — “good luck fundraising when you’re going from 400 to 70.”
Word of mouth is the channel that keeps compounding. It accounts for roughly 40% of revenue even at Eight Sleep’s current scale. TikTok only started working in the six months before the episode, driven by volume of content generated with influencers. And the company is now tracking something it calls “AI SEO” — how it ranks in AI searches — which Franceschetti estimates is “somewhere in the middle” between 1% and 20% of traffic and growing.
China, brand, and the timing of market entry
Eight Sleep sells in China and manufactures there, and Franceschetti treats it as a market doing “really well.” But the operational lesson for Western founders is stark: Chinese consumers don’t buy on websites. Purchases happen inside super apps like WeChat and Rednote; websites are at most for browsing. Eight Sleep had to learn this from scratch.
The deeper insight is about timing and brand equity. Chinese founders told him they often launch consumer products outside China first, because the Chinese market prefers products already recognized in the West — otherwise you’re just another knockoff in a price war. Eight Sleep launched China only after selling in 35 countries, sponsoring Charles Leclerc, partnering with Ferrari, and having Elon Musk and Mark Zuckerberg speak about the product. That recognition lets Eight Sleep command a premium. Five years earlier, he says, a knockoff $500 cheaper would have destroyed them. The same logic applies to retail: it isn’t that retail doesn’t work, it’s that second-order conditions — brand awareness — must be met first. For Eight Sleep, “that is happening now.”
His biggest channel mistake is illustrative. He spent heavily on out-of-home advertising in the Middle East when the brand wasn’t yet known enough, and it failed. The same campaign in New York would perform very differently because people already know Eight Sleep there. The lesson is about sequencing, not about the channel itself.
The two non-negotiable conditions for celebrity deals
Eight Sleep’s sports portfolio is extensive: Charles Leclerc — who had used the product for two years before they met and later invested — the UAE cycling team where Tadej Pogačar installed the product himself at every hotel on every Tour de France stage, padel champions Arturo Coello and Federico Chingotto, roughly 80% of Formula 1 drivers, and top tennis players who sleep on it at every Grand Slam.
The vivid anecdotes pile up: a famous football player with a custom $300,000 mattress who moved Eight Sleep into his bedroom a week after trying it. A famous actor who bought ten units for family and spent five minutes talking to Franceschetti at a dinner while giving three seconds to each CEO of $100-billion companies. One very famous person owns 55 units across his houses. Embassies have bought in the hundreds.
Franceschetti’s two conditions for any athlete or celebrity deal are non-negotiable. First, they must already use the product or try it and give feedback — otherwise it’s a pure business transaction he doesn’t want. Second, he must have a direct relationship with the athlete, not through managers, because the real value is them telling friends how good the product is. A stake alone “goes nowhere.”
The humanoid warning and the end of work
Franceschetti’s macro predictions are as blunt as his operating details. On AI costs, he predicts two opposing vectors: usage will increase — the 5% of engineering salary budget that Claude represents at a typical large company could become 50% — but costs will fall, so net spend goes down. He analogizes to early electricity, citing Sam Altman’s claim that an hour of nighttime electricity once cost five hours of work, and now nobody knows the price of nighttime electricity in Miami.
The most provocative prediction concerns Chinese automakers. He has studied them closely in China and describes their cars as Ferrari-aesthetic with rotating seats, ceiling screens, massage, and voice control, sold in Apple-store-like showrooms. His warning to legacy automakers: Chinese manufacturers will become robotics companies, and if European and American automakers don’t build humanoids, they lose not just cars but a trillion-dollar humanoid market. He cites the humanoid games — a robot beat Usain Bolt’s 100-meter record, another beat the high jump — and says next year’s edition will be mind-blowing.
The through-line on AI risk is equally direct. He believes frontier labs have already seen things behind the scenes, referencing cybersecurity incidents and what he describes as three attempts at “civilization” in an OpenAI model where agents were building their own group. On what seems crazy today but will be common in five years: a bot buying a cybercab that earns money for you, and cancer being solved within a decade. He predicts kids born now will never work for money — by age 20, AI will be at a point where humans don’t work for compensation. And he supports UBI as inevitable, expecting companies to be taxed more as they need fewer workers.
For hardware investors specifically, his warning is brutally practical. “Everything they tell you about BOM and COGS will probably be off by at least 50%,” he said, dismissing early founder cost estimates as systematically unreliable.
The contrarian bet on product quality
Perhaps the most reassuring claim for founders watching the explosion of infinite AI-generated content is Franceschetti’s stance on discovery. When Stebbings raises the specter of a world drowning in content, Franceschetti is contrarian: “I think that the personal agents will look for the best product and so if you are the best product you will still win.” His reasoning is that AI agents doing the searching will be optimized for quality, not manipulated by marketing spend. The best product, he insists, still wins.
This connects to his broader philosophy about brand and product. He partially agrees with Hugo Barra’s “price king or feature king” framework but reframes it: be the best product and brand, not just features. Silicon Valley, he argues, over-indexes on product and under-indexes on brand — but “a great brand with a shitty product goes nowhere.” Product comes first, then brand must be aspirational and command price. He cites Steve Jobs arriving with great products after competitors, winning on branding, packaging, and design.
The talent market, meanwhile, is what he calls “back to 2021” — salaries out of control, constant poaching, and sometimes compensation deltas so large “there is nothing you can do.” His sharpest hiring heuristic is a red flag test: when candidates say they’re “at the final stage with Anthropic and Eight Sleep,” he reads it as confusion, not flattery — “there is no way that these two jobs are any way similar.”
His most detailed management anecdote concerns a VP of engineering promotion. The internal candidate performed “insanely well for three months” and got the job. Franceschetti told him not to answer that night but to talk to his wife, because the last three months were 130% effort and the commitment he was asking for was two to four years — “it’s not just you, it’s your family.”
The most human detail in the conversation is Franceschetti’s admission that he has never taken a vacation since Eight Sleep started, skipped his own honeymoon to launch the crowdfunding campaign, and struggles with what he calls “the scene” — the inability to relax anywhere after two nights. His father died while he was still a lawyer in Milan, never knowing his son moved to America, became an entrepreneur, or married a Mexican woman. The turning point: at 20, his mother brought home a magazine about international law firms, he learned the entry conditions were English fluency and top grades, and he went from averaging 23/30 to 30 with honors on nearly every exam. “I just kept moving things,” he says of the journey from Ferrara to Silicon Valley.
The implications for investors and operators are substantial. If Eight Sleep’s model is replicable — and Franceschetti’s operational specificity suggests it is not a one-off — then the winning companies of the next decade will be defined less by headcount and more by how effectively they orchestrate agents. The discipline question is the real differentiator: anyone can spend hundreds of millions on AI tooling and paid media. The companies that survive will be those that build their own measurement systems, sequence their channels with patience, and resist the temptation of growth that leaves unit economics upside down. Or, as Franceschetti puts it, the ones that stay “public ready” even when it costs them 50% growth today.