Newsletter · 2026-08-01 · Ashutosh Agarwal

Tiny Teams Hit Fifty Million in Revenue as AI Token Bills Rival Payroll - How They Build - Week of August 1, 2026

Startups and venture newsletter for the week of August 1, 2026. Founders on the week's podcasts competed over how few people they employ, from a bag brand doing 50 million dollars with six staff to a property manager running a twelve agent org chart, while a second set of operators described AI bills that have grown to the size of the payroll they replaced.

How They Build

Week of August 1, 2026: Headcount Is Now an Insult

This week founders stopped bragging about how many people they employ, and started bragging about how few. Plus: the AI bill has quietly grown as large as the payroll it was supposed to replace.

For most of the last decade, a founder telling you "we have 30 people" was telling you they'd made it. Headcount was the scoreboard. This week, on podcast after podcast, that scoreboard got flipped upside down. The people building companies right now are competing to see who can run the biggest business with the smallest team, and the ones winning that race are describing numbers that would have sounded made up two years ago.

Here's the twist that runs underneath all of it, and the reason this isn't a clean victory lap: the same week founders celebrated firing the org chart, a very different set of operators went on the record about the bill that arrived afterward. The tokens that replaced the payroll now cost, at some companies, about as much as the payroll did.

The Number: $50 million. Six people.

The single most striking efficiency stat of the week came from a D2C bag brand called Hulken, whose two founders host the podcast Unfinished Business. The headline is blunt: they did $50 million in revenue last year with six full-time employees. They've since added one person, so they're at seven now, and their stated goal is to reach $100 million with fewer than ten people.

That works out to somewhere around $7–8 million of revenue per employee, a figure that puts a tiny consumer-products team in the same efficiency league people usually reserve for the most extreme AI-native software companies.

The founders are almost aggressive about it. In their words:

"If you are a brand doing less than 50 mil a year and you have more than 10 full-time employees, you're doing it wrong. I used to see headcount as being a signal of strength… but now when I hear that, it literally signals to me this company is inefficient."

The cleanest way to see how far the mindset has shifted is their payroll math. In consumer brands, the long-standing rule of thumb is that payroll should run about 10% of revenue, and hitting that was considered good. Hulken runs at 4%. As one founder put it: "The gold standard has always been… 10%. Now I hear 8%, I hear 7%. We're at 4%."

How they actually pull it off is the interesting part, and it isn't "we fired everyone and let ChatGPT run the company." It's a deliberate three-layer structure they call the "minimum viable company": a small core of senior in-house leaders, a network of outside agencies for specialist work, and, the layer they're most excited about, an AI project-management layer where every single team member gets their own AI assistant that handles the work you'd normally hand to a junior hire or an intern. One founder has an AI chief-of-staff she named Benny, wired not to the open internet but to the company's own Notion workspace (their "brain"), so it only ever answers using what it knows about Hulken. The team's mental formula: "Minimum viable company equals maximum revenue per employee."

Unfinished Business, "The Minimum Viable Company: How We Run a $50M Brand With Just 7 People" (July 28, 2026).

What Founders Changed

She cut her company from 12 people to 2, then rebuilt the missing 10 in code. On The UpFlip Podcast, Andrea Palacio told the story of buying a landscaping company in 2023 that promptly fell apart: the seller stole from them, the manager quit and told the staff the business wouldn't survive, revenue dropped from $70,000 a month toward $40,000, and the team went "from 12 employees down to 2 employees." With no technical background, she taught herself to build the missing roles as software: an AI voice receptionist that answers every call and books appointments, automated past-due-invoice callers that escalate through three increasingly firm scripts, and automatic after-service photo-and-review messages. Today the business does over $1 million a year, run mostly by her husband, and she does all the building herself: "Right now I could show you a picture of my 1, 2, 3, 4 screens… half of them are Claude Code." When someone offered to buy the company, she said no. The UpFlip Podcast, "249. She Runs a $1Million Home Service Business Using Just 1 AI Tool" (July 27, 2026).

A property manager built a 12-agent "company" and grew profit 168% without adding units. On STR Global Unlocked, Alex Zemianek, CEO of JC Vacation Rentals, described replacing his tech stack with an org chart of AI agents. He started with one agent, scaled up to 20, then settled at 12: a financial agent wired directly into QuickBooks and his property software that "can draft a P&L in under a minute," a marketing agent, operations and listings agents, a "librarian" that keeps shared knowledge clean, a CEO agent that writes his morning brief by 6 a.m., and a security agent with veto power. He says the system lets him work "about a hundred times faster" than before and drove net operating income up 168% without adding meaningful inventory, "maybe just two or three units." The most telling detail: the security agent has overruled him. "I've had times where I tried to veto, and the security agent kept shutting me down." And what did he do with the extra profit? He hired a human, a "head of experience," because "people are still going to be yearning for that soul and hospitality that we can't lose." STR Global Unlocked with Simon Lehmann, "041: This Property Manager Increased Profit by 160% With AI | Alex Zemianek" (July 28, 2026).

A freight brokerage grew revenue 80% in a year without hiring a single person. On The Dynamo Show, the founder of Freight Hero explained how one customer pulled this off. In freight brokerage, a human "hero" typically touches each shipment six to eight times, checking trucks, chasing drivers, handling paperwork. Freight Hero's agent, "Robin," drives that down to "close to one or even under one" touch per load, with the human staying in the loop only for exceptions. The freed-up labor doesn't get cut, it gets redeployed to booking more loads and winning more customers, which is how a brokerage "was able to grow revenue 80% year over year without having to change headcount." His framing of why this matters in a brutal industry (brokers have spent four years in what's called "the great freight recession," many operating at negative margins): "You don't do that work anymore. Give it to us… and then you get your best people and you focus them on revenue-generating roles." The Dynamo Show, "AI Powered Operations for Freight Brokers" (July 29, 2026).

The org chart itself is being redesigned, flatter, more senior, and full of "multi-specialists." On Tech Deciphered, the hosts laid out what an AI-native company actually looks like structurally, and it's less "one person runs everything" and more a specific set of shifts. Middle management shrinks first, because a lot of a middle manager's job is "routing," passing tasks around, aggregating and synthesizing them back up, and "AI and agents in general are very good at that." What rises in its place is what they call the "multi-specialist": the end of hyper-specialization where one person only does front-end, or only design, or only product. "You can combine designing and shipping code, product managing and shipping code, being an analyst and deploying." And the hiring instinct has inverted, from blitzscaling to deliberate restraint: "the good old days of… let me go and hire 300 people to scale my go-to-market" are over, replaced by "the tendency to under-hire rather than over-hire." As investors, they said, the question is no longer "is this startup doing something in AI?" but "is this an AI-native startup? Organizationally, culturally… how is the team working?" Tech Deciphered, "79 – The Cognitive Age" (July 31, 2026).

The "50% rule," and why the authors think it's actually low. On AI First with Adam and Andy, Adam Brotman and Andy Sack (co-authors of the book Agents Inc.) put a number on the labor question: they estimate almost any business, restaurants, retail, law firms, accounting firms, finance, "probably can be run with at least 50% fewer people today." And they think that's conservative: "It might be actually generous. It might be greater than that… I'm seeing evidence of 70 to 80%." When they set out to redesign one real process for a 90% time cut, they landed at 85%. They're now building their own company to prove it: a startup that will be "totally agent-native from day one," where "we may only have two people or 10 people doing what eventually normally would have required 100 people." They were careful to note the optimistic counter-argument too, that AI-proficient people tend to get busier, not idler, and chase new "unlocks" rather than shrink, but their base case is stark. AI First with Adam and Andy, "The 50 Percent Rule: How AI Is Reshaping the Future of Labor" (July 29, 2026).

The Other Side: "You cannot shrink your way to dominance"

The lean-team story got its sharpest pushback this week too, and it's worth taking seriously, because the counter-evidence came with names and numbers.

Companies are quietly re-hiring the people they let go. Elon Musk Podcast (a news-recap show, despite the name) walked through a striking reversal: Alphabet and the railroad giant CSX are "actively expanding their headcounts right now, reversing a long stretch of AI-fueled job cuts," with executives "literally calling back laid-off workers because they'd hit a wall." The reason isn't sentimental, it's cost. "When a company decides to replace a department with an automated system, they aren't just buying software. They are taking on the continuous cost of cloud computing power… the perceived savings just vanish." The episode framed the real choice as two strategies: the "lean model" (cut 30% of staff, keep output flat, pad the margin) versus the "multiplier model" (keep everyone, use AI to multiply output tenfold). Its blunt verdict: "You cannot shrink your way to dominance. The lean company is playing defense. The multiplier company is playing offense." It also quoted Lattice CEO Sarah Franklin's warning that "having coding agents doesn't negate the need for engineers," because models trained on past data are "a terrible way to invent something new." Elon Musk Podcast, "Why AI layoffs are backfiring" (July 28, 2026).

More than half of employers now regret the cuts. On Download This Show, the panel cited that "more than half of employers are now reported saying they regret cutting jobs for AI," diagnosing it as "poor leadership, poor change management and carelessness" plus a basic misunderstanding: "volume is not equal to service." An AI can handle an outrageous number of calls and still fail the customer, the example given was a call-center startup whose AI loses the thread the moment a caller "ums and ahs." The deeper point: technology only delivers real productivity when you actually redesign the business around it, not when you "just… play with it" and fire people. Download This Show, "Have AI models truly 'gone rogue'?" (July 31, 2026).

The layoff scoreboard, with receipts. Grumpy Old Geeks ran through the week's job cuts with the actual figures: BuzzFeed laid off 180 people, about 33% of its remaining workforce, following what the hosts called "a failed AI strategy," with annual losses reaching $58 million by 2025 and the cuts expected to save up to $32 million a year. Uber cut 10% of its customer-service team and cited AI directly, with a VP memo explaining the org "has become too complex and siloed" and needed to be reshaped "to layer AI on." Patreon's CEO announced it was "flattening the organization." The hosts' skeptical read: a lot of "we must use AI or we'll be dead in three years" is executives keeping their name in the news while dodging the harder question of how any of it becomes profitable. Grumpy Old Geeks, "This Time With Chicks" (July 30, 2026).

And a reminder that more code isn't more value. On the economics podcast Justified Posteriors, two economists dug into a new research paper measuring what AI coding tools actually do to output. The raw numbers are enormous: adopting an asynchronous coding agent produced a 17x increase in lines of code, a sync agent about 10x, and plain autocomplete about 2x. But watch what happens as you move from "typing" toward "shipping": by the time you reach actual software releases, the same agents delivered only a 30% increase. They call it "attenuation," the gains evaporate up the chain. And zooming out to the whole ecosystem, more apps are being produced but "we don't seem to see an increase in downloads." The takeaway, in their words: "If you look around, you see a lot more code, but it's not clear that it's resulted in amazingly better software for us yet." Justified Posteriors, "Is AI Replacing Programmers or Boosting Them?" (July 27, 2026).

The Cost Corner: the bill caught up

This is the theme worth flagging loudest this week. For a year, the story was "AI lets you replace payroll." This week the story became "…and then the AI bill grew to the size of the payroll."

Start with the number everyone is quoting. At NVIDIA's developer conference in March, CEO Jensen Huang laid out a thought experiment that has since taken on a life of its own: take a software engineer you pay $500,000 a year, and at year-end ask what they spent on tokens. If the answer is less than $250,000, Huang said he'd be "deeply alarmed." In other words, a well-equipped engineer should be burning a quarter-million dollars of AI a year. As one skeptic dryly noted, "Jensen Huang has a commercial interest in doing this. This is the guy who runs the pencil factory selling you pencils." But the line is spreading, one speaker described three consecutive presenters at a Las Vegas conference quoting it as if it were settled fact.

What "$250K per engineer" feels like from the bottom. The best reporting on this came from The Catalyst by Softchoice, which followed a financial analyst it called Benjamin through a corporate AI pilot. He was given a token budget, first $150, then $250 a month. Within two days he'd burned $60 of it on Excel work, panicked, and started rationing: "I started using it for high-priority Excel model changes, just trying not to burn through my $250 in the month." He now hoards tokens for an emergency that hasn't happened, doing work "the slow way on purpose," and admitted the tool had "changed the way that I work extensively," until the meter took it back: "I almost lost my self-reliance." Meanwhile, at the other end of the spectrum, the same episode described an engineer at Meta who built an internal dashboard called "Claude Economics" that ranked all 85,000 employees by token consumption, the top 250 earned badges like "Token Legend" and "Cash Wizard." In a single 30-day window, Meta staff burned more than 60 trillion tokens, with the single heaviest user running up roughly 281 billion tokens on his own (Mark Zuckerberg didn't crack the top 250). Meta shut the dashboard down two days after it leaked.

The macro framing from the same episode is the part to internalize: as vendors switch from per-seat to per-token pricing, "some companies have seen the cost go up 7x, 10x, 20x," to the point that "you're literally now seeing some people that are paying as much for their AI tokens as they're paying almost for their engineering employees." Uber's CTO said back in April the company had burned through its entire 2026 AI budget, the whole year, gone in four months. And Gartner's latest forecast puts worldwide AI spending at $2.59 trillion this year, up 47%, while Gartner's own analyst concedes "CIOs are struggling to prove value." The Catalyst by Softchoice, "The Token Burn Episode: What Happens When Your Software Bill Has No Ceiling" (July 29, 2026).

The backlash against "token maxing." On Tech Talks, a SnapLogic leader pushed back hard on the Jensen framing, calling the "spend more to be better" logic "motivated reasoning" and citing Goodhart's law: "the minute the measure becomes the goal, it ceases to be a useful measure." SnapLogic's own approach is telling: every engineer gets their own API key visible in telemetry, with "soft limits" that pause usage if it spikes, followed by a conversation: "Hey, what caused this spike?" The point isn't to cap cost, it's to develop skill. He also flagged a new site, tokenspend.org, where a company called TechWolf is publishing its own per-person AI spend so others can benchmark against it, the beginnings of an actual market rate for "what should this cost per employee." Tech Talks, "The Truth About Token Maxing: Why Your AI ROI Is an Illusion" (July 29, 2026).

Token budgets are becoming a line item like any other. On [Un]Churned, Saumyo Mukherjee of Braze described how the company is formalizing it: every department head gets a token budget sized to their team, and "every individual will have X dollars per month of tokens to use." Braze watches usage for a quarter, then reallocates, an engineering team will justifiably burn far more than an HR team. His philosophy on the tradeoff between cost and capability: "We should not care about the bare-bones dollars, as long as we can justify the ROI… Braze made the right choice, yes, even if it costs more, as long as the cost is justifiable." [Un]Churned, "Vibe Coding Is a Bandaid on a Bandaid ft. Saumyo Mukherjee (Braze)" (July 29, 2026).

The pressure valve: swap in the cheap model. The counter-move to runaway bills is already happening in production, per The Rundown's deep dive on China's new Kimi K3 model. Companies are learning to route work by cost: "DoorDash said they use the Moonshot models for lower-level work, and they use Anthropic's models for their hardest cutting-edge tasks." Airbnb runs its customer-service agents "mostly on Alibaba's Qwen model," CEO Brian Chesky says it's "fast and cheap," and even Microsoft is reportedly testing whether Kimi K3 could power features in Copilot. The pricing gap is the whole story: Kimi K3 runs about $15 per million output tokens, versus roughly $30 for OpenAI's top model and $50 for Anthropic's Fable 5, a 50–70% discount for near-frontier performance. The market noticed: one analyst estimated that the day after Kimi dropped, about $314 billion got shaved off the valuation estimates for OpenAI and Anthropic. The Rundown, "Deep Dive: Did China Break the AI Trade Again?" (July 25, 2026).

And one number that captures the whole shift. On The Accounting Podcast, the hosts cited data from the spend-management company RAMP: since June 2025, RAMP's customers have spent 20.7 times more on token expenses. Part of that is simply the shift from "we all paid $20 a month for unlimited" to usage-based pricing, but as one host put it, "Is there any other comparable business expense that's been that drastic of an increase?" The only thing they could think of was NVIDIA chips. The Accounting Podcast, "IRS Chief Spied on Colleagues & Out of Control AI" (July 31, 2026).

The through-line

The founders building $50M businesses with six people and the CIOs staring down 20x token bills are describing the same phenomenon from opposite ends. AI really is letting a handful of people do what used to take a hundred: Hulken, JC Vacation Rentals, Freight Hero, and Andrea Palacio's landscaping company are not hypotheticals, they're operating businesses with real numbers. But the cost of the labor didn't disappear; it moved. It went from a payroll line you could predict to a token meter you can't, and at the frontier the two numbers are starting to converge. The winners of the next couple of years won't be whoever fires the most people or whoever spends the most tokens. They'll be whoever figures out the exchange rate between the two.