Newsletter · · Ashutosh Agarwal
Here's Your Salary and Here's Your Token Budget - How They Build - Week of July 4–11, 2026
How They Build for the week of July 4–11, 2026. Tiny teams posted extreme revenue-per-employee numbers (Lovable's $500M ARR with 149 people) while companies from Tesla to Coinbase moved to cap or re-engineer their AI token spend, and one VC predicted a per-employee token budget sitting next to salary.
How They Build
Week of July 4–11, 2026: Here's Your Salary and Here's Your Token Budget
How They Build, week of July 4–11, 2026. This week tiny teams posted enormous revenue numbers, and companies got their first real invoice for the machines doing the rest of the work.
Two stories ran side by side in the podcasts this week, and they are really the same story. On one side, founders showed just how few humans it now takes to build a genuine, revenue-generating company. On the other, those same companies started getting a bill for the AI doing the work, and it is big enough that some are now handing engineers a token budget the way they hand out a salary. ("Tokens" are the units of text an AI model reads and writes; they're what you pay the AI company for. Heavy use adds up fast.)
Here's the week.
The Number: $500 million in revenue. 149 people.
That is the headline figure of the week, and it comes from Clay Bavor, co-founder of the enterprise-AI company Sierra, on 20VC. He was talking about a rival, the AI coding startup Lovable, which had just announced hitting $500 million in ARR with 149 people. (ARR is "annual recurring revenue," the yearly run-rate of a subscription business.) That works out to roughly $3.4 million of revenue per employee, a ratio that used to be reserved for a handful of the most efficient public companies on earth, now posted by a startup most people hadn't heard of two years ago.
"You had a Lovable announce yesterday hitting, I think, $500 million ARR with 149 people... I think the general direction of travel clearly is towards smaller, higher-leverage teams."
The Twenty Minute VC (20VC), "Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need... with Clay Bavor, Co-Founder of Sierra" (2026-07-04)
And this is no longer an anecdote you can wave away. The clearest data of the week came from a roundup of new Stripe and academic research. According to Stripe, solo founders now account for 63% of all C-corporations formed in the second quarter of 2026, an all-time high. The number of one-person businesses ("solopreneurs") earning over $1 million more than doubled between 2023 and 2025. And a joint Harvard Business School / INSEAD study found that AI-native startups are "25% smaller, flatter, and more engineer-heavy, yet equally valued" as their older peers. The host's framing was the memorable part: solopreneurs are "the extreme tail of the efficiency gains possible from AI", they speed-run the experiments that eventually reach everyone else.
"The number of solopreneurs earning a million dollars more than doubled between '23 and '25... This is not a wave of vibe-coded apps hitting a million dollars in ARR, but instead AI helping to fill the gaps that once made hiring mandatory. For small businesses, AI services are now stepping in as the technical co-founder or the sales and marketing first hire."
The AI Daily Brief, "AI Is Making One-Person Million-Dollar Companies More Common" (2026-07-06)
If you want the mental model in one sentence, a founder named Geoff McQueen gave it on Spark of Ages. His stated goal is to build a company doing $10 million in ARR with fewer than 10 full-time employees, against what he called the old rule of thumb of 75 to 100 people for the same revenue.
"A $10 million company with, call it, 75 to 100 employees was probably a pretty standard growth playbook... What AI has done is increased productivity by a factor of five to ten. So it'll be one of the best marketers I've ever worked with, but they'll be doing what would have been five-to-ten-person throughput, doing it with agents and tools to go a hell of a lot faster."
Spark of Ages, "The Career Move That Survives AI / Geoff McQueen, Vibecoding, Revealed Preference, Coaches" (2026-07-10)
What Founders Changed
Two engineers now cover eight sports leagues. Blitz is an AI-native sports-data company that its founders bootstrapped with their own sports-betting winnings. When asked how they think about team-building, co-founder Devin was blunt: keep it tiny on purpose. The company supports eight leagues (MLB, NBA and six others) with a full-time engineering team of exactly two. Their hiring bar is unusual, they want people who understand both AI and sports deeply, rather than a bigger team that only has one of the two.
"Tejas and I wanted to keep the team small for as long as possible. It is faster to build things than it was ever before, and quite frankly, we know what we're building better than anyone else... We'd rather have just a few people who can build a lot by themselves than a much larger team who maybe only has one of those two skills. In terms of full-time engineers, it's just Tejas and I right now."
The Betting Startups Podcast, "Ep. 218: AI-first sports intelligence infrastructure w/ Devon & Tejas from Blitz" (2026-07-07)
The design department stopped using Figma. This was the most vivid "before and after" of the week. On Empire, a founder from the crypto research firm Blockworks described what happened when his team piped its design system and brand assets into Anthropic's new "Claude Design" tool. Work that used to bounce across a chain of people (head of product, designer, copywriter, engineer) now collapses into a day, and sales decks assemble in minutes. He said the team's use of Figma, the industry-standard design tool, has dropped to about 5% of what it was.
"We need a new landing page for our API, for this monitoring product, for investor relations, for our 'contact us' form. That would have previously... [gone] to the product designer, who mocked it up, sent it back, 'ah, we need some work', then off to someone good at copy. I'm not kidding: this took 24 hours. The whole thing... And for every sales deal we have a Slack channel, you upload the notes, combine it with the HubSpot record and the Slack, say 'I'm pitching [customer], it's a $275K deal'... it will make you a deck in two minutes that's better than we ever could have had. I couldn't be doing what I'm doing now without AI. I would need a much, much larger team."
Empire, "Crypto's Value Capture Problem & Why Robinhood Built Its Own Blockchain" (2026-07-10)
Three times the revenue, same 120 people. Abbas Mohammed built his first business, a real-estate operation, to $1.7 million in annual revenue over four years by hiring 25 virtual assistants. His current company, Remote Leverage, is now running at roughly $1.5 million per month. His claim: AI let him roughly triple revenue while holding internal headcount flat at about 120 people, by automating the low-level work so his specialists could aim at higher-value tasks. His hiring philosophy is a rebuke to the "cheapest labor" instinct, he argues a slightly more expensive expert delivers "literally 10x better results."
"It took me four years to get to $1.7 million in revenue in my first business. Now we're at $1.5 million per month, and we did that in less than two years... That extra 20%, 30%, 40% in payroll you pay for a really good expert will give you literally 10x better results than someone who's slightly cheaper."
Grit Daily Startup Show, "Abbas Mohammed on Why Better Talent Beats Doing Everything Yourself" (2026-07-07)
$1 million in revenue, two weeks after launch. Raj Singh, a serial founder whose last company reached 100 million people before being acquired, built a new AI life-coaching app called Purpose with best-selling author Mark Manson. His team built the whole product in about six months, and it crossed $1 million in ARR within two weeks of launching. The interesting part is the product insight, not just the speed: he thinks the mainstream chatbots are engineered to please, and he set out to build the opposite.
"All the foundation models, ChatGPT, Claude, they're not just too much sugar. They're candy. It's all sugar, because they are maxing for engagement... The product challenge was: how do you give a sycophantic LLM a backbone? Over 40% of Purpose users now say they've gotten, quote, 'life-changing value.'"
Subversive, "How Purpose Hit $1M in ARR Two Weeks After Launch" (2026-07-09)
An 11-person media company out-earning a much bigger one. On his own show, Scott Galloway interviewed the founders of TBPN, a daily tech program, and laid their numbers on the table with visible envy. Eleven employees, $5 million in ad revenue in 2025, tracking past $30 million in 2026, bootstrapped, profitable, zero outside capital. That is roughly $2.7 million of revenue per employee this year, on a path toward well over $4 million. Galloway noted his own ProfG media business does about $20 million with more downloads and video views, and yet TBPN pulls in 50% more revenue. Their trick was commercial discipline, not headcount: annual, fixed-rate sponsorships (pitched "like sponsoring a Formula 1 team") sold to a deliberately narrow audience of maybe 200,000 operators and investors.
"11 employees, $5 million in ad revenue in 2025. You're tracking past $30 million in 2026. Bootstrapped, profitable, zero outside capital... And yet you guys have 50% more revenue [than ProfG]."
The Prof G Pod with Scott Galloway, "Why OpenAI Bought a Podcast, with TBPN's John Coogan and Jordi Hays" (2026-07-09)
The counter-current: efficiency doesn't automatically mean fewer people
Two of the week's most senior voices pushed back on the "AI equals layoffs" reflex, and one company got publicly burned for believing it.
Ken Griffin, founder of the hedge fund Citadel and the market-maker Citadel Securities, described an internal system that does genuinely PhD-level work. His team built an agentic system (meaning AI that carries out multi-step tasks on its own) to reproduce academic finance papers, the kind of grind that normally takes a team of masters and PhD holders weeks.
"It takes roughly six to eight weeks to reproduce a paper... My colleague built an agentic AI system that would read a paper, reproduce it, verify the results, produce the results out of sample, and do all this work in about, on average, two to three hours. And here's the key point: there's no reduction in headcount at Citadel on the back of this breakthrough. I have incredibly talented people... I'll take every single productivity gain I can get, because with the talented people we have, we just have more to go after."
Exchanges, "Ken Griffin on US-China Tensions and AI" (2026-07-09)
And a cautionary tale: Ford reportedly rehired more than 300 veteran quality inspectors and engineers after AI-driven quality checks failed to match human experience, the company having wrongly assumed that feeding design requirements into a model would produce high-quality output without the seasoned engineers. It's a useful reminder that the replacement stories don't always stick. Valuetainment, "'We're Team Human' - Ford Hired Back 350 Engineers They Fired Due to AI" (2026-07-07)
Geoff McQueen made the same point from the builder's chair, and it's worth heeding before you fire your engineering team: a slick AI demo is not a product. He compared a one-shot "vibe-coded" app to a soapbox derby car with a nice paint job.
"As soon as you get to a hill in the rain, when you've got your kids in the back, that's when bad things really happen. There's a big difference between what you can vibe-code as a demo and what it takes to actually have a real product."
Spark of Ages, "The Career Move That Survives AI / Geoff McQueen" (2026-07-10)
The quiet version: hold the line on headcount, let revenue climb
For companies not chasing a solo-founder dream, the practical move was subtler, keep the team the same size and let AI absorb the growth. On AI to ROI, Evan Schwartz of AMCS Group (which makes software for waste and recycling fleets) walked through the math that actually moves a board-level number, not just "I saved three hours."
"I can now add two to three times the number of touches to my customers than I could before, because AI does the low-value but necessary work... We were able to reduce our churn from 6% down to 3%. We manage 700,000-plus trucks globally, and route optimization saves 17 gallons of diesel per truck per month."
AI to ROI, "AI Governance, Ethics, and the Stewardship Framework: A Conversation with Evan Schwartz, Chief Innovation Officer at AMCS Group" (2026-07-07)
His discipline is worth stealing: every AI project has to tie to a board metric (EBITDA, free cash flow, SG&A) before it gets funded. He gave a beautifully plain example of why "time saved" is a trap: shaving 40 minutes off a truck's route is worthless on its own, "I still have to pay the guy a full day." The money only appears when you re-org the routes and park three trucks, at roughly $1 million each in all-in annual cost. That's $3 million of real free cash flow, not a fuzzy productivity claim.
The Cost Corner
Last week the theme was shock, companies discovering their AI bills. This week the podcasts moved on to the mechanics: how companies are actually getting the bill under control, and the uncomfortable new idea that a token budget is becoming a line on the payroll next to salary.
The blunt instrument: caps. The backdrop is now well established, Uber famously blew through its annual AI budget in four months and imposed a $1,500-per-month cap for non-coding roles; Walmart moved from unlimited internal-tool use to hard token budgets. This week added a fresh name to the list: Tesla has applied token-spending caps evenly across the entire company as a first, blunt response. One host framed the whole shift as the move from the "AI subsidy era" to the "token scarcity era." The AI Daily Brief, "The Big Ways AI Just Changed" (2026-07-04) and The AI Daily Brief, "AI Costs Are Surging and the Cheap Model Fix Might Not Last" (2026-07-08)
The smarter instrument: swap the model. Coinbase's approach got the most attention among engineers. Rather than cap usage, it changed which models do the work, swapping its default models away from the priciest frontier models toward cheaper open-weight options (GLM-52 and Kimi-2.7), then adding caching and "difficulty-based routing" (send easy jobs to cheap models, hard jobs to expensive ones). The result:
"They were able to cut spend to nearly half of what it was at its peak, yet they increased their actual token usage. And in terms of output, it didn't fall off."
Everyday AI Podcast, "Ep 813: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem" (2026-07-07)
The moat-building instrument: fine-tune your own. The most striking economics of the week came from fine-tuning, retraining a smaller model on your own proprietary data so it beats a giant general model at your specific task, for a fraction of the cost. The example everyone cited was a Bridgewater study run with Mira Murati's Thinking Machines Lab. On the AI Daily Brief's telling, big general models (from GPT 5.2 up to Claude Opus 4.8) scored 74–78% accuracy on the task at a cost of $20 to $90, while Bridgewater's fine-tuned specialist hit about 85% for single-digit dollars. Microsoft is productizing the same idea with "Frontier Tuning," claiming its tuned models match GPT-class quality at up to 10x lower cost.
"People sometimes ask, why fine-tune when general-purpose models keep getting better? Bridgewater's work is a good reminder that with the right data, here, expert judgments, you can beat prompting-only approaches by a lot."
The AI Daily Brief, "AI Costs Are Surging and the Cheap Model Fix Might Not Last" (2026-07-08)
The unsettling new idea: tokens as a comp line. Back on 20VC, Clay Bavor put words to where this is heading. He's seen top engineers who lean hard on coding agents burn more than $100,000 a year in tokens, "a meaningful fraction of an engineering salary." His prediction is that companies will start budgeting tokens per employee, right alongside pay.
"I think the direction we're headed is some amount of token budgeting on a per-employee basis. For CFOs in the future, capital allocation will look more like: here's your salary, here's your token budget, have at it."
The Twenty Minute VC (20VC), "Open Models vs Frontier Models... The $100,000 Token Budget Every Engineer Will Need... with Clay Bavor, Co-Founder of Sierra" (2026-07-04)
He also offered a number worth arguing about at your next partner meeting. Salesforce's Marc Benioff has said he spends about $300 million a year with Anthropic on his dev teams, which sounds enormous but works out to roughly 3.8% of developer salaries. Bavor thinks that number is "wildly off" from where it settles, and bets it converges closer to 20%. If he's right, the implication is stark: a lot of AI application companies are, in his words, "grossly overvalued" at today's assumed spend, and undervalued if spend really quadruples.
The middle path: tiered access. Not everyone is capping or swapping. PwC's chief people officer described giving all 80,000-plus staff a democratized base layer of tools (Copilot, an internal "ChatPwC"), while the more expensive frontier tools require a justification, with the whole budget "very closely monitored, not just by our leaders, but by our board." Her framing is the one to watch: she refuses to treat AI spend as optional, comparing it to the ROI question people once asked about laptops. But she also flagged the wild card underneath all of this, that today's prices are "so highly subsidized... we don't know the full cost", and that if the true cost surfaces, offshore or entry-level humans could again become cheaper than the machine for repetitive work. Future Ready Leadership With Jacob Morgan, "PwC's Chief People Officer on Training 80,000 People for the AI Era With Human Skills at the Center" (2026-07-06)
The tension that closes the week: a founder at Blockworks who is thrilled with Anthropic's new Fable 5 model in one breath admitted in the next that it is "ridiculously expensive", it "chews through tokens like no model I've ever seen", and that his firm is fielding more "I ran out of credits" complaints than ever. The productivity is real. So is the bill. Empire, "Crypto's Value Capture Problem & Why Robinhood Built Its Own Blockchain" (2026-07-10)