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

Why does it matter? Because the unit of production is no longer a human hour — it's a token.

Speechify's CEO is about to spend more on AI tokens than on all employee salaries combined, and he expects the same to be true across serious tech companies within three years. That's not a forecast — it's a present-tense description of where one 50-million-user company already is. Cliff Weitzman has personally led every department of Speechify except AI, interviewed the CEOs of the top 100 consumer subscription businesses worldwide, and applies what he learned with a relentlessness that makes most operational thinking look polite.

• Token spend will eclipse salary spend at elite AI companies within three years • 40% of billionaires have dyslexia — adversity quotient beats raw intelligence every time • QA is now the scarcest skill in software, and almost everyone is under-investing in it • People don't quit companies — they quit unsatisfied life goals, and retention is a proactive life-design exercise

Speechify will spend more on AI tokens than on salaries next year — and Weitzman expects every serious tech company to get there within three

"We're getting to the point where soon we're going to spend more in tokens than we spend on actual salaries." Weitzman says this casually, then sharpens it: "Next year I expect we'll spend more in tokens than we'll spend on salaries. It's atypical right now, but I don't think it'll be atypical in the long term."

The implication is structural. When the marginal unit of production shifts from human hours to AI inference, the P&L changes shape. Token spend stops being a discretionary software line and becomes a core input — something you budget the way you budget headcount, and hold accountable for measurable output.

He's so committed to this direction that he tracks individual employee credit usage and calls out anyone falling short. "If you don't spend a thousand credits a day, I'm disappointed in you." Enforcement is personal: Zoom screen-shares where Weitzman and senior engineers demo their Claude Code workflows for the whole team, mandatory Loom videos proving daily AI output, and a direct warning to anyone who can't explain their non-usage. He describes aiming for the extreme end of adoption on purpose — "I need to move people from all the way over here on the right to all the way over here on the left, and I'm okay if they meet me in the middle."

The current token spend is concentrated in engineering. The growth and creative teams aren't there yet — more testing infrastructure has to be built before the economics make sense. But the direction is one-way. He's atypical now. He won't be in three years.

40% of billionaires have dyslexia — the same trait that fills prisons builds empires, depending entirely on what failure does to you

The data point lands clean: 40% of incarcerated people have dyslexia, even though it affects only 17% of the population. So do 40% of billionaires. The trait is identical. What splits the outcomes is how a person processes repeated failure.

Weitzman argues that IQ and EQ receive too much weight in hiring and investing conversations. The real predictor is AQ — adversity quotient. "How good are you when things get really tough?" His vehicle metaphor: he doesn't want the fastest jeep. He wants the one that won't get a flat tire in the savannah.

The mechanism is legible. A nine-year-old who fails at the one thing school requires — reading — faces two exits. They conclude the world is stacked against them and follow a dark path. Or they develop a tolerance for failure by failing constantly and getting back up, which gradually becomes a belief that even impossible-seeming problems yield to enough repetition. That second loop is what produces both dyslexic billionaires and founders who survive four and a half years without product-market fit.

His proxy for AQ in hiring is concrete: projects actually shipped to production. Not prototypes — things that required navigating the annoying friction of Apple developer accounts, Cloudflare configs, keeping a side project alive three years after launch. "You don't need to be a genius. You just need to be willing to do the thing." His favorite diagnostic question, borrowed from YC: "What is a non-computer science system you've hacked to your advantage?"

"A little bit of slope makes up for a lot of y intercept. It doesn't matter where you start. It matters how fast you grow — and the growth rate has to do directly with how much pain you're willing to bear."

LLMs let ordinary people overrule expert institutional advice — and Weitzman found his father's cancer months early because he had one open during the appointment

His father's PSA had risen again. The oncologist's plan: sit and wait. The cancer was too small to locate. Weitzman, with LLMs running alongside the conversation, pushed back hard enough to demand a PSMA scan — then went further on his own. He discovered UC Davis operates a U-TOTO Explorer machine capable of imaging at a 2mm voxel resolution versus the standard 4mm China Basin scanner. He tracked down the lab director, got his father scanned in December. They found the cancer in the left seminal vesicle. By January 7th, it had been surgically excised — before the second scan the original doctor had scheduled even arrived.

"I would have never had the confidence to push on the doctor so hard to get the scan if I didn't have LLMs open during the conversation with the doctor."

Separately, for his brother's Lyme-related neural inflammation and psychosis: full proteomics, genome mapped, data loaded into Supabase, an MCP server written on top of it, then cross-referenced against every published paper on genomic anomalies. Four anomalies in gene expression found.

His read on the system is precise: "The problem is not the doctors, the problem is the system where doctors have to treat statistics and they need to make sure they don't get sued." People willing to educate themselves and absorb personal risk can access individualized care the system won't provide. "You are the quarterback of your own medical experience."

The pattern extends beyond medicine. High-stakes domain, asymmetric information, professional gatekeeping — LLMs collapse the asymmetry for anyone willing to use them aggressively.

QA is the last bottleneck AI can't clear — and companies treating it as a second-tier skill are about to feel it

Engineering is being commoditized. Design is being commoditized. What remains irreducible is the judgment to find what actually breaks on a real device, with a real user, in an unexpected situation.

"In a world where software engineering is commoditized and design commoditized, if you try to build stuff with Claude Code, even if you give it all the tools in the world, it will not succeed in QAing itself to perfection." Weitzman finds more live bugs in production than his own QA team does. A friend who works directly with Elon Musk describes the same phenomenon — Musk always surfaces the bugs, always makes it embarrassing.

"That's what separates a product from a great product. It's just QA."

The workflow in practice: build a feature with a well-crafted Claude Code prompt. Spend the next 48 hours finding everything that breaks — different phones, dropped Wi-Fi, users doing the unexpected — until the thing is genuinely solid. Then ship. It's unglamorous and irreplaceable.

He also uses QA as a talent diagnostic. Engineers who struggle to ride features all the way to production sometimes excel at finding failures in code someone else wrote. Before cutting them, put them in QA. If they're good there, that's a real role worth keeping and coaching. If they can't find bugs either, that's the answer. The scarcest engineering skill right now isn't writing code. It's knowing when the code is actually done.

Growth is a volume-and-arbitrage game — and the ad that made $3M was a rap song nobody predicted

The arbitrage advantage disappears the moment everyone uses the same tool. So Weitzman built his own. Four days while his girlfriend was at Disneyland: a crash course in AI-generated video ads, 10 new engineers hired in a week, and a custom platform that auto-reskins creative, posts to Meta, TikTok, and YouTube, and feeds daily Manus reports. Target: 1,300 ads tested per day. Current run rate: roughly 1,000 AI-generated per day, alongside 8,000 human-made organic creatives per month.

"Growth is just an arbitrage game. You are competing with every single other person in the world who wants to get their product in front of users." The two levers are content and distribution. Both require finding edges before they're mainstream.

The clearest rule from studying 100 consumer subscription CEOs: "Don't even bother spending money on any platform that's not Meta until you reach $100,000 a month." He heard it from the Blinkist founder as a casual aside and immediately stopped spending elsewhere.

The $3M ad: a rap song about his experience with dyslexia and ADHD. He'd expected the Old Spice–style production number to dominate. That one got 300,000 views. The rap made $3 million. The lesson isn't about rap — it's that outcomes are distributed randomly enough that the only reliable strategy is volume. More formats, more tests, more shots. Senior growth executives at major companies are often the wrong people to learn from — they've gone rusty. Go two levels down to whoever is actually buying the ads.

Bulking and cutting are opposite organizational states — any company oscillating between them in the same quarter is failing at both

"Any Harvard MBA can cut costs if you're smart — it takes a genius to grow revenue."

Companies cycle like bodybuilders: bulk, then cut. The error is treating them as dials you can dial simultaneously. "You can't oscillate between the two in one week. You have to commit for like six months to doing one." The mindsets are genuinely incompatible. Growth demands pouring your entire concentrated will into a single lever. Cutting demands analytical detachment and margin discipline. Do both at once and you're pressing the gas and brake simultaneously. The company slows. The culture gets confused about what it's optimizing for.

Speechify spent four and a half years finding product-market fit. Then four and a half years running profitably. Now it's in hyper-growth again — a deliberate re-entry, not a drift. The sequencing was the point.

Even within a growth phase, the discipline holds. Blended CAC can rise; direct CAC cannot. An experiment on a new channel — Speechify is one of 200 companies provisioned to test OpenAI ads — counts as a capital investment even when it costs real money, because being ahead on a new channel before it opens broadly has compounding value. Spending on a billboard with no attribution? Never. Every dollar has to teach something.

The same logic maps to creators: Logan Paul milked YouTube, bulked into boxing and WWE, launched Prime, now has a fund. Distinct phases, fully committed to each in sequence.

Employees don't quit managers — they quit unsatisfied life goals, and finding those goals is the actual retention strategy

Most retention conversations happen after someone gives notice. By then you're negotiating. Weitzman runs his intervention earlier and goes much deeper.

A key early engineer wanted to leave because he had no friends in the Bay Area. Weitzman didn't counter with equity. He handed the engineer his own phone, opened a notes file, and asked him to write his ten top life goals. Friendships dominated the list. Assignment: send me a Google Sheet with 50 people in the Bay Area you want to meet, by end of day. The engineer delivered. Weitzman DMed all 50 on Instagram. Thirty-five showed up to dinner on his patio. The engineer made friends. He stayed.

For his future COO, Weitzman read immigration law himself after 15 lawyers refused to help with a visa situation, and secured a green card directly. For a Ukrainian engineer considering leaving just before the war started — Weitzman flew there despite US government warnings, signing a waiver releasing the airline from liability. He ran a three-day hackathon. The engineer stayed.

"People typically leave not for money. They leave for another thing in their life that's not satisfying them. Figure out what's not satisfying and solve that thing."

The diagnostic is the same every time: find the actual unmet goal and solve it specifically. Money is occasionally the real answer — in which case, just solve it. But it's rarely the root.

The firms building aggressive AI judgment now are compounding an advantage that won't be closeable in three years

Everything Weitzman describes converges on the same underlying shift: the limiting resource is no longer engineering capacity or capital — it's cultivated judgment. Who finds the cancer the algorithm missed? Who QAs what Claude Code can't? Who retains the engineer by diagnosing the unsatisfied life goal? These are the same skill: human discernment deployed at leverage points AI can't reach on its own.

The token-spend-exceeds-salaries prediction is the tell. In that world, a company's value lives almost entirely in the people who know where to direct the inference and what to do with what comes back.

The people building that judgment now — aggressively, at a thousand credits a day — will be untouchable. Everyone else will be writing better prompts.

---

Topics: AI, growth marketing, performance marketing, entrepreneurship, voice AI, consumer subscriptions, hiring, leadership, product management, adversity quotient, dyslexia, AI tools, company strategy, retention, LLMs, ad testing