Cliff Weitzman: What I Learned from 100 of the World’s Top CEOs & Why Tokens Will Outspend Salaries

Summary

• The headline call: token spend will exceed payroll. Speechify expects to spend more on tokens than on salaries next year — in engineering alone — and Cliff says great companies “will get there in 3 years.” Adoption is coerced, not suggested: “If you don’t spend 1,000 credits a day, I’m disappointed in you,” non-users must explain themselves in a public Slack thread, and holdouts get 24 hours to send a Loom of something they built with AI.
• The ad machine is the moat: Speechify tests ~1,000 AI-generated ads a day (target 1,300) on top of ~8,000 human-made creatives a month, on a platform it built itself after N8N kept timing out — “if you use the tool that everybody else uses, typically you’re not going to win.” It is also one of just 200 companies provisioned to run ads on OpenAI, which Cliff calls “massive, massive” because “OpenAI knows everything about your history… inside of your psyche” — high CPMs don’t matter with conversion plus attribution, and the tracking SDK just launched.
• Long Meta, “one of the most underrated stocks”: when engineering and design commoditize, what’s left is QA and customer acquisition — “I look at all the people who are vibe coding new products right now. Where do you think they’re going to get their users?” Cliff wouldn’t be long Snap (“lunch is completely eaten by Meta”), Figma a hold despite Harry being down 40% (“Dylan Field is a savage”; 40x→6x revenue is the market correcting, not the company failing).
• Context is the edge in public markets too: unable to buy H100s in 2022 and stunned by Jensen’s earnings-call demo, Cliff bought Nvidia stock while brother Tyler bought a 3x levered option — “now like 36x up.” Cliff put a third of his money into Tesla in 2015 off a photovoltaics capstone. The Munger/Buffett punch-card rule: “when you find the opportunity, lever as hard as you can into it in a responsible way.”
• Speechify is a stealth compounder by design: claimed 94% of the B2C voice-agents market, 1.1M five-star reviews, 50M+ users, 4.5 years profitable, and inference driven to single-digit dollars per million characters vs Eleven Labs’ $70–100. It discloses neither raise nor valuation publicly; Cliff says the team knows the raise figure: “If you find lightning in a bottle, you don’t want to tell other people about it.”
• The diagnostic for consumer AI revenue ramps is inference cost: watch for “inference costs burning those companies alive.” Cursor is the benign case — money “went straight through Cursor and to Anthropic,” revenue without real losses — and churn spares only need-based products: “if you are blind… you will never churn from glasses. A meme maker, sure.”
• The 850 coin flip lands contrarian: Cliff picks OpenAI over Anthropic at 850 — Sam Altman “comes off slimy” but is a record fundraiser whose enterprise deals look “like black magic and looks like fraud” — even while conceding “Anthropic is running circles around them right now, especially in the coding arena” and “I would marry Claude Code.” On Grok: “never bet against Elon.”
• Three weeks living with MrBeast: Jimmy stopped a 100-person shoot to repaint a door pink — “he has the algorithm in his head.” The transferable lesson: the best-converting content needs no language (“our best performing ads include books not PDFs”), and Harry’s own frustration got a mirror — he’d be “17x more successful” if he’d backed every 7/10 founder he met, since he can’t tell a 7 from a 10; Cliff’s hiring version: “If you can really trust someone, you can afford to pay them significantly more.”

Deep dive

1. Volume of work is the operating system — buy more lottery tickets

• The founding myth: a dyslexic 13-year-old immigrant who couldn’t read applied to 26 colleges when counselors said six, writing applications daily in “academic workshop… the class you go to if you’re on the short bus,” drafting his main essay 48 times and opening a second Common App account to beat the 20-school cap. The logic: elite admissions is a lottery even with perfect scores — “How do you win a lottery? You buy more tickets than everybody else.” Life principle: “more reps, more shots at goal” gets you unexpected outcomes on the bell curve.
• The rule of 100 governs every department he’s led (all but AI): read 100 books, then find 100 experts. In 2020 he listed the top-100 consumer subscription companies by revenue and flew worldwide to meet their CEOs. Lesson one: “if you send a good email, people will respond, no matter how amazing they are” — and if the CEO doesn’t, message the CMO, then the head of growth; fewer than 1% ever told him to stop. Krieger, Ev Williams, and the Plaid, 23andMe, Honey, Grammarly and Robinhood founders all answered — and became early investors. He still does it: Anton at Lovable, recently.
• What the tour taught: “growth is just an arbitrage game,” and the senior people were often “already rusty” — so go levels down to whoever actually buys the ads. Everyone at Speechify must stay “not just a fat general sitting in the back, but a warrior who could take out their sword.” A throwaway aside from Blinkist’s founder changed his budget: don’t spend on any platform that isn’t Meta until you’re at $100,000/month on Meta.
• The enforcement story: a head of growth with nothing shipped at 60 days, blaming hiring, got: “Bro, source the people yourself… You’ve got 7 days” — make ads in CapCut, edit in Figma, buy the ads himself, or “it’s going to be very difficult to keep working together.” Harry’s echo: Rabois’ barrels, not ammunition.

2. A thousand AI-generated ads a day, on a platform they built themselves

• The sprint as told: four free days while his girlfriend was at Disneyland — he took a Hormozi-community course on AI video ads, hired 10 people in one week from inside that community, and set a target of 1,300 tested ads a day. N8N instances kept timing out, so they built their own platform that reskins, auto-posts to Meta/TikTok/YouTube, and reports back. Models rotate weekly: “one two” (likely Wan 2) was good, then Kling, then “C dance” (likely Seedance), now ChatGPT’s new image model and likely Nano Banana.
• Reskinning is the multiplier: sister Geffen leads ads, and her winning spots get turned into “a 60-year-old lady… a 40-year-old man… someone who’s African-American or someone who’s Asian,” backgrounds swapped (“farm wouldn’t convert for our users”). Consistency of personality? “Not at all. Because you’re marketing for different demographics” — paid converts best in the 30s–40s, women outconvert men, 60+ are heavy users because of reading glasses.
• Today’s arbitrage play: whitelisting — a niche creator makes 20 videos you run as ads yourself, and only proven winners get posted organically from the creator’s account.
• Selection is pure evolution: upload 1,300 to testing campaigns, watch CTR, CPA and CPM; anything “a standard deviation better than everything else” graduates to the main campaign to compete against the best ads ever — “It’s like March Madness brackets.” Some days nothing graduates; some days everything does.

3. The $3M rap song, Crew Silver, and Mr Beast’s algorithm-in-the-head

• To escape a local maximum he listed the top 100 ads in history (whittled from 500), read every script, and rewrote each for Speechify. His amendment to the 3-second rule: “you win or lose by the first image of the video, like the first 1080 pixels” — hence red headphones, a suit, and a hot tub, plus a recurring character, Crew Silver: “very gullible… kind of dull, but he has great confidence.” Fifty variations later, one ad “launched the company to a whole new stratosphere” and ran three years, with hundreds of millions of impressions on Instagram alone.
• What works is “a complete random distribution”: his Old Spice remake — he literally got a horse — did 300,000 views, while a rap song about his dyslexia and ADHD “made us… like three million dollars.” Performance ads live “in the shadows,” so test 800 variations without brand risk — but conversion is what covers the ad spend: free users are “our charity,” and “if you don’t get people who actually convert on your product, there’s no point in doing the work.”
• Three weeks at Jimmy’s in North Carolina: he watched MrBeast halt a 100-person shoot “just to like essentially repaint the door pink” — “no genius can figure out what’s going to convert in ads. Jimmy knows what’s going to convert in organic content… he has the algorithm in his head.” Formats are the science (ladders: $1 hotel to $1M hotel), and the best content needs no language — the storm-proofing ad is just bucket, hole, duct tape, water; for Speechify, “our best performing ads include books not PDFs.” On soft views: last year was a record, and Jimmy has been “operating as an entrepreneur” (Beast Games, Feastables) while shifting to streaming, where “clips can get watched 7 billion times in a week.”
• The Logan Paul thesis doubles as career advice: after a cabin retreat listing the 200 people who accumulated the most power by 35, Cliff’s pattern is the virus — “they attach themselves to a larger organization from within… and use that organization to help them grow.” WWE launched The Rock; it’s now doing it for Logan. For Harry: don’t join Andreessen Horowitz or Sequoia (“you cannot redirect the snake’s head”) — join Union Square Ventures “and become the leadership,” or bolt on as an All-In host, a Shark Tank seat, or a Penguin Random House book.

4. CAC discipline and the OpenAI-ads land-grab

• Hypergrowth doesn’t mean tolerating worse unit economics: “The blended CAC can go up, not the direct CAC.” Spending $500,000 a month making 20,000 creatives is a creative investment; a billboard without attribution is setting money on fire. Everything is a test — his framing of the Netflix lesson (Kelly, $1.7B on performance marketing in 2017): “I know I’m wasting half my spend. I just don’t know which half” — fine, then do the work to find which half.
• As one of 200 companies provisioned on OpenAI ads, Speechify is deliberately overspending to learn before general rollout. Why it will be huge: like Meta’s WhatsApp/Instagram signal, “OpenAI knows everything about your history and what you’re interested in… inside of your psyche,” with Google-like intent. CPMs are extremely high, but “it’s okay to pay a high CPM as long as people convert and as long as you have attribution” — and the tracking SDK just launched.
• Channel mix today: word of mouth is still biggest; Apple loves them (design award, featured 14 times last year); SEO is huge; and 15% of users now arrive organically from ChatGPT. In paid, Meta is best, Google most profitable, YouTube good on desktop, Apple 11 and TikTok merely interesting — with a halo effect across all of it.
• The most atypical operator he met: Greg at Ladder, a CEO who buys his own ads and “found a lot of arbitrage plays that other people have not found.” Are there new plays left? “Every single day.”

5. The public-market book: long Meta, hold Figma, levered Nvidia, all-in solar

• Meta is “one of the most underrated stocks right now”: Zuckerberg is “an absolute savage” and “the number one acquirer in the world”; Google faces LLM pressure while Meta’s only real threat is a TikTok with post-separation issues; in AI “data is everything” and Meta has the most proprietary data — “they haven’t figured out the model side yet… but I really do believe that they will.” The kicker: vibe-coded products all need users, and Meta sells users. Snap, by contrast: he’s never once made Snap ads work, they’re “overpaying creators left, right, and center,” and “the lunch is completely eaten by Meta.”
• Harry, down 40% on Figma: hold or sell? “Hold. Dylan Field is a savage.” Figma is “relatively antifragile” — great sales, great taste, Dylan coded the original thing, which he thinks was in C/C++ — and “Figma traded at 40x revenue, and now they’re at 6x revenue… it’s more so that the market is making a correction.”
• The Nvidia story is a context trade: in 2022 Speechify tried to spend millions on H100s and “no one would let us buy them,” then Jensen’s earnings-call world-rendering demo confirmed “a completely different level of technology.” Cliff bought stock; Tyler bought a 3x levered option, “now like 36x up.” Same pattern in 2015: a renewable-energy undergrad put a third of his money in Tesla. The frame is Munger/Buffett’s 20-punch-card — and Nvidia today, at a ~35x revenue multiple with that growth, still has “a lot of more juice.”
• On energy for AI (headed from 1% of utilization to 10–12x that): “all the dams that can be dammed have been dammed”; fusion “is coming. We just don’t know… when” (he wants to build a contained mini reactor in his garage); meanwhile “I’m all in on solar” — cells near the 47–49% theoretical limit, cost concentrated in installation, China far ahead, US zoning the blocker. The SolarCity lesson: it “was not a technology company. They were a finance company” that got JP Morgan (as he recalls) to underwrite panels over 30 years — GTM innovation counts as innovation.

6. Tokens will outspend salaries — adoption is forced, QA is what’s left

• The call, precisely hedged: “we’re getting to the point where soon we’re going to spend more in tokens than we spend on actual salaries” — next year, and across engineering only. Atypical now, “but I don’t think it’ll be atypical in the long term”: great companies “will get there in 3 years.” He’d like growth and creative to token-max too, but “we need a lot more testing mechanisms” first.
• Adoption is coercion by design: 1,000 credits a day or “I’m disappointed in you”; non-users must explain in a public Slack thread, take his call, then “You have 24 hours. Send me a Loom video with what you used AI to make.” He’s deliberately extreme — “I need to move people from all the way over here… and I’m okay if they meet me in the middle” — and the same 100-iteration education applies to users, who’ve been walked from text-to-speech to dictation to podcast generators to teacher agents to Jarvis.
• Tooling and the taste distinction: Claude Code over Cursor (Anthropic team plan, which he thinks he got in December; “it actually drives me crazy that not everybody has moved over”); leadership sends email from inside IDE terminals; new products launch with “the default Cloud Code stuff” and designers come after. Design’s “hammer and chisel” grind is commoditized — taste isn’t. The consequence: QA becomes the most valuable skill, because Claude Code “will not succeed in QA-ing itself to perfection” — Elon and Thiel famously out-bug their own QA teams — and “that’s what separates a product from a great product. Just QA.”
• His most impactful AI use is medical: a proteomics map of brother Eric’s genome, put on Supabase behind an MCP and compared against every published anomaly paper, surfaced four gene-expression anomalies. For his father’s recurring prostate cancer, when the doctor counseled waiting, he hired Kagglers on the radiology data, found the U Explorer scanner (2mm voxels vs the standard 4mm; three in the US, one at UC Davis), located the tumor in the left seminal vesicle, and had it excised by January 7 — “I would have never had the confidence to push on the doctor so hard… if I didn’t have LLMs open.” The doctrine: “you are the quarterback of your own medical experience” — “I want to be coached like I’m Real Madrid.”

7. AQ beats IQ: hire for pain tolerance, fire on denominator games

• The concept: adversity quotient — “How good are you when things get really tough?” His image is a Jeep crossing the savanna: “I don’t want the Jeep that goes the fastest. I want the Jeep that’s not going to have a flat tire.” Most engineers quit a hard problem after 30 minutes, but “the things that really move the company are on the other side of 5 hours of grappling.” Valentin Perez’s line: “a little bit of slope makes up for a lot of Y intercept.”
• Signals: side projects actually shipped — the Apple developer account registered, the Cloudflare instance still up three years later (“you don’t need to be a genius. You just need to be willing to do the thing”); the YC question “what is a non-computer-science system you’ve hacked to your advantage”; and an entrance exam where candidates must use LLMs to read a giant repo, ship features, then fix what they broke in 30 minutes. Red flags: stats where “they changed the definition of what the denominator was”; people who talk too much (signal-per-sentence must be high); and “I want to earn more money” as a reason to join — “Terrible answer. Never would have hired that person, not in a million years.” Counterintuitively he prefers older hires: “they read textbooks about software engineering” so they understand LLM output — though a cracked young Claude Code engineer gets hired on the spot.
• Why the best founders are dyslexic or ADHD: Michael Dell’s “you want the car to fit the driver.” A dyslexic 9-year-old either concludes the world is stacked against them — 40% of the incarcerated have dyslexia vs 17% of the population — or learns “it’s okay to fail… if you try enough, you’ll succeed”: “40% of billionaires also have dyslexia.” ADHD brains running at 600 wpm against 200 wpm reading get distracted; multi-discipline CEO life at scale suits ADHD, but the focused 0-to-1 IC grind does not.
• His change of mind, from a MrBeast video that collapsed: 11 days sleeping in a Walmart parking lot (3 days to earn an iPhone, 5 for a laptop), and around day nine he turned resentful. Months later he realized his edge was never AQ itself — “my biggest advantage is I’m typically very grateful and I’m really optimistic,” and gratitude is what powers AQ. The experience made him a phone-buyer for people on the street and a policy radical: “the government should give a free Android to everybody who wants one… it’s a basic human right.”

8. Speed culture: replies, minimal meetings, no performance reviews

• The 60-second response rule spans a remote-first team of 200 across 36 countries (45 AI research engineers, ~150 product engineers): reply within a minute even if only with a timeline; it’s on you to tell your bottleneck they’re your bottleneck “multiple times a day”; send your leader three goals each morning and expect silence unless you’re pointed wrong. If text fails, call — “young people are very scared to call… people above 45 are willing to call and below 45 are not.” Harry’s Titanic rule (reply within three hours, even weekends) got him “decimated” online; Cliff’s verdict on the critics: “Wrong culture.”
• Meetings are avoided because one person talking to five listeners “should have just been an email,” and “most long Slack messages should have been a phone call.” Deadlines replace them: “a dream is a goal without a deadline,” so ship dates go on calendars with multiple invitees — some invites literally titled “this is not a meeting. It’s a deadline.” The stakes: “speed is the number one strategic advantage for any company… if you’re not adding to the speed, you’re by definition distracting from the speed.”
• Performance reviews are a “total waste of time” — the leader failed to warn the person sooner. His milk-delivery analogy: the best milk in the world left down the street instead of on the doorstep is spoiled — “literally no points.” The replacement is a 72-hour ultimatum: cut scope to one feature, get the PR reviewed live on the phone, ship to production — or “the person is just not an outcome owner.” Failed outcome owners get tried in QA before any firing; leaders’ people-management 1:1s were cut from 60 minutes weekly to 15 every other week.
• Why Speechify doesn’t hire from Google: abundance blunts hunger and “you need to coach them out of essentially bad habits”; Bay Area loyalty runs lower than Europe’s, where “people really value a good company and a good job.” Retention is bespoke: for an engineer leaving over loneliness, Cliff DM’d the 50 people on the engineer’s dream-friends spreadsheet — 35 showed up to patio dinners, and he stayed; when 15 lawyers couldn’t fix the future COO’s visa, Cliff read immigration law himself and got the green card; as war neared Ukraine he signed the airline’s liability waiver, flew in, ran a hackathon, and kept the engineer. “People don’t quit companies, they quit managers. That has not happened at Speechify.”

9. The stealth compounder: no disclosed numbers, 4.5 years to product-market fit

• The shape of the business nobody sees: 4.5 years with no fast growth, product and vision unchanged, north of $5M revenue by the time install-to-trial, trial-to-paid and renewals all inflected in 2020. The PMF grind as told: walking onto Stanford’s campus daily with duct tape over his mouth to watch users tap the wrong buttons; a red help button wired to his personal iMessage (3am texts answered, daytime ones called); daily FaceTime-audio calls to users churned at day 30 — “like getting kicked in the teeth every single time someone tells you your baby is ugly.” His Chesky/Dylan Field-aligned heresy: “AB tests are abdicating decision-making to the user” — build the vision, convince people to come, iterate for years.
• He discloses neither raise nor valuation — the team doesn’t know either. His rebuttal to all three reasons founders publish numbers: users (almost a million new ones a week already), hiring (178,000 applicants to open engineering roles last year, 19,800 of whom finished the async challenge), investors (10 quality ones inbound weekly). “If you find lightning in a bottle, you don’t want to tell other people about it.” The claims that stand in for a deck: 94% of the B2C voice-agents market, 1.1M five-star reviews, “we read more to people than any other AI product in the world.” The cap table is his cold-email tour: Dylan Field, Krieger, Ev, Branson, Gary Cohn, Zillow/Audible/Plaid/Brex founders — Charlie Songhurst first.
• Companies run bulking and cutting cycles like bodybuilders, and you can’t do both: “you’re clicking the gas and clicking the brake at the same time.” In his growth year he logged time in 48 half-hour units, found 30% going to learning SEO, and banned himself from learning — execute only. The line he repeats twice in the episode: “any Harvard MBA can cut costs… It takes a genius to grow revenue” — which is why investors pay for exponential revenue even without profits. Speechify just ended 4.5 profitable years and has flipped to hypergrowth.
• Harry’s confession lands here: ten years in, “I can’t decide the difference between a seven out of 10 and a 10 out of 10… If I’d invest in every founder that I’d met, I’d be 17x more successful” — likely 2.5% of Deel, likely 2% of ElevenLabs, seed checks in “Flight School” and Vanta. Cliff’s mirror: “If you can really trust someone, you can afford to pay them significantly more” — be trigger-happy on great people and ignore the price. His own miss was Austin Ray (Ramp’s AI-adoption teacher): “I should have trusted my gut.” The same philosophy hired the hacker who cost Speechify a couple hundred thousand dollars — traced through a school domain, confronted on Zoom, then employed as Cliff’s personal engineer for a long time (he built the first Jarvis) before joining the core team: “if you find someone that’s talented, you better work with them.”

10. Consumer AI economics: inference burn is the tell — and the 850 coin flip

• His screen for insane consumer AI revenue ramps: look for “inference costs burning those companies alive.” Speechify lived it — once usage exploded, “it stopped being a business and started to be a charity” — so a 40-person AI team plus its own GPUs and data center spent 2.5 years driving cost to single-digit dollars per million characters, vs $30–100 elsewhere (Eleven Labs: $70–100 standard, ~$30 for the V3 fast model). That, he says, is why the company has been profitable so long; his own framing: “Speechify is basically a B2C version of Eleven Labs.”
• Cursor is his favorite benign case: fork VS Code, add Claude, and “all the electrons on the right side were dying to go to the left side” — the money “went straight through Cursor and to Anthropic,” revenue without real losses. Contrast the startup burning its $50M alive: demand a cost-reduction plan, but remember the genius-grows-revenue asymmetry — a real growth lever plus differentiation (“community is like so important”) can still justify the bet. On churn: need-based products don’t churn — “if you are blind… you will never churn from glasses… A meme maker, sure.”
• Does Apple buy WhisperFlow? “No, Apple launches a competitor” — on-device security philosophy and keyboard RAM limits explain Siri-land, and the ecosystem has “kind of become lazy” (“I can’t even get an API to get a receipt”). But don’t short the platform: AirPods made more money than Spotify and almost every other audio company put together, every year for seven years, podcasts and audiobooks are at 40% penetration in most markets — “the TAM is really big” — and he’s excited the incoming CEO is “more of an engineer” than supply-chain god Tim Cook.
• The closing quick-fire: OpenAI at 850 or Anthropic at 850? Contrarian OpenAI — Sam Altman “rubs a lot of people the wrong way and comes off slimy” but “no one has raised as much money… in the history of Silicon Valley,” and his enterprise deals look “like black magic and looks like fraud.” The analogy: Columbus asked the King of Spain for India (AGI) and found America (ChatGPT) — there’s “a lot of AQ in that company.” All while conceding “Anthropic is running circles around them right now, especially in the coding arena,” “I would marry Claude Code,” and he personally uses Anthropic for everything. Harry called OpenAI’s recent recruiting “Terrible… last 5 years mega bomb.” And Grok will matter: “never bet against Elon.”