Newsletter · 2026-10-03 · Ashutosh Agarwal

AI Model Spend Becomes a Second Payroll for Lean Startups - How They Build - Week of October 3, 2026

How They Build for the week of October 3, 2026. Podcast synthesis on how AI-native startups and operators are treating model spend as a second payroll: Higgsfield's over $10,000-per-person monthly token bill and $1 billion run rate on close to 400 people, Zuora and SAP setting per-person token budgets, Samsara's 260-person marketing team running 113 agents, a property manager lifting margins from under 20% to 32%, and the model-switching and spending caps turning up at Tesla, UBS and HP.

How They Build

Week of October 3, 2026: AI Model Spend Becomes a Second Payroll for Lean Startups

A billion-dollar AI company spends more on models per person than many startups pay in salary, a CFO gives every employee a "token budget" next to the people budget, and Samsara's marketing team plans to shrink while its agent count triples.

The most revealing line of the week came from Alex Mashrabov, founder of the AI video company Higgsfield, on 20VC. Asked what his team spends on AI models, he did the math out loud: more than $4 million a month, across close to 400 people. "Yes, it's definitely over $10,000 per person."

That one figure captures where startup-building is right now. AI is no longer a cheap tool you hand everyone. For the most aggressive companies it is becoming a second payroll. And founders are starting to manage it like one.

That idea runs through this week's podcasts:

A software CFO now sets a people budget and a token budget side by side, and says he does not care which one a manager spends, as long as the work gets done. (A "token" is the unit AI companies bill by, roughly a word or part of a word.)
Samsara's CMO runs a 260-person marketing team with 113 AI agents, and expects about 200 people and 400 agents within two years.
A property manager with 800 units cut her admin team in half and lifted her profit margin from under 20% to 32%.
And SAP's CEO admitted he finally had to put limits on how much AI his own staff can use.

There is pushback too. The same Higgsfield founder who spends $10,000 a head on AI says his biggest operating mistake was not hiring enough humans in support and legal.

Here's the week.

The Number: $2.5 million in revenue per person

That is roughly what Higgsfield brings in per employee. The company says it just crossed $1 billion in annualized revenue with close to 400 people.

For context, a typical well-run software company is happy with $200,000 to $300,000 of revenue per employee. Higgsfield is running at about ten times that.

The speed is the other half of the story. Mashrabov says it took 18 months to go from $1 million to $1 billion in annualized revenue. He points out that Cursor, the AI coding tool often called the fastest-growing startup ever, took 24 months.

Host Harry Stebbings stressed how unusual the company is: "This company is built with 300 people out of Kazakhstan. It is a complete anomaly."

How they count matters, so Mashrabov spelled it out. Higgsfield takes revenue from the last 28 days and multiplies it by 13. It spreads annual contracts across the year and only counts the slice for that month. "We are not taking like three year enterprise deals and baking into like one billion figure. No, we don't do that." That is the same method most AI labs use. But keep in mind it is a snapshot of recent momentum, not a full year of booked revenue.

Some other details from the conversation:

They nearly died first. Higgsfield burned more than $10 million of a $16 million seed round searching for a product. With under $5 million left, Mashrabov says he stopped chasing "what's hype today, what's the right narrative" and went back to talking to customers. Eight creative directors all told him the same thing: AI video had no camera control. That became the product.
No paid ads. "We don't do paid." Growth came through product-led sign-ups and creators.
One customer went from $99 a month to a $6 million-a-year deal in six months. Much of that demand, he says, comes from Asian e-commerce brands making "hundreds of ads, if not thousands a week," and from the short-form drama industry, where "most of new shows there are made with AI end-to-end."

The lean number is impressive on its own. What makes Higgsfield the story of the week is what it spends to get there. More on that in the Cost Corner below.

The Twenty Minute VC (20VC), "20VC: $1BN ARR in 18 Months; The Untold Story of Higgsfield | Spending $4M Per Month on Models | Why Moats in AI are BS | Scaling a Content Team to 150 People with Alex Mashrabov" (Sep 28, 2026)

What Founders Changed

Zuora now gives every manager two budgets: one for people, one for tokens. On Run the Numbers, Todd McElhatton, the "COFO" (chief operating and financial officer) of billing software company Zuora, described a real change in how the company plans spending. When someone leaves, managers no longer automatically refill the seat.

"Maybe I had some attrition to organization, and I don't want to back those. I'd much rather have it in token costs. And that's one of the things we move to in the upcoming budget cycle is you'll have a people budget and you'll have a token budget. I'm indifferent as long as you get the deliverable on how you spend it... if a leader tells me they can deliver better with more tokens than people, we'll do more tokens."

Every Zuora employee now has a personal token budget, sized to their role. They can see how much they are using, and they can ask for more if it pays off. (Details in the Cost Corner.)

McElhatton also shared the most striking lean-company stat of the week, about one of Zuora's customers. He describes it as an AI-native company that has hit $40 billion in annual recurring revenue with 2,500 people, and "less than 100 people in their finance function. I mean, it's maybe 50 or 60." He did not name the company.

His view on what this means for org charts is blunt. The layers of people whose job is to gather other people's work and package it for bosses are going away. "I think the layers are absolutely going to fade away." He contrasted that with how startups used to brag:

"Four or five years ago... you hear is we raised X amount of dollars and we added a hundred people this quarter. And that was the best. And now it's like, hey, I've raised some money, but I've got a handful of teams... the output that's coming out is exponentially higher."

In his own engineering team, work that used to come back in "three, four weeks" now comes back overnight: "An engineer is like, hey, I've got a handful of agents that are working on it. I'll have the answer for you in the morning."

Run the Numbers, "Zuora's COFO On Going Private and Pricing AI | Todd McElhatton" (Oct 1, 2026)

Samsara's CMO: 260 marketers, 113 agents, and you can't get a top rating unless you build agents. On The GTMnow Podcast, Meagan, CMO of fleet-software company Samsara, gave one of the most detailed looks yet at an AI-run marketing department inside a large public company.

The basic setup: 260 people and 113 AI agents, with headcount held flat. The host called that ratio out, and asked where it goes next. Her answer: "I think in two years, your orgs are smaller, right? I think you're 200 people, maybe. And you're 400 agents."

What changed in practice:

Performance reviews now require it. "You can't get a four or five, the highest rankings, unless you're actively working with AI building agents." She set that expectation a year ago.
Building agents is how you get promoted. "If someone came to me and said, I don't need to hire two or three people. I built these agents and they're doing the work of two or three people or 10 people... I am definitely going to promote them."
Content output jumped. A content writer who used to produce "a piece every week or every two weeks" now does "10 a day."
Custom sales landing pages went from 2–3 weeks to under 30 minutes. These are pages built for one target company, a tactic called account-based marketing. The old process meant working with an agency and a sales rep. Now an agent researches the company, pulls account data from Salesforce, finds matching case studies, applies brand rules, and builds the page in the website system. "You could do unlimited, right?"
A rebrand rolled out in three and a half weeks. Designing the new brand still took a year with an agency. But rolling it out across every slide deck, case study and ad was done in 3.5 weeks, using a "branding agent" that sales reps call from Slack: give it an old deck, get back a rebranded one.
Most agents live in Slack. They are built mostly with a tool called Gumloop.

Her advice to other CMOs: "Don't take six months to build an AI strategy." If your team lacks the skills, "go hire engineers" into marketing.

She also named the trade-off out loud. Time spent building agents has to pay back: "We either need to deliver more or do it in less time or both."

The GTMnow Podcast, "Lessons from building 113 AI agents and reinventing a GTM function with AI | Meagan (CMO, Samsara)" (Sep 29, 2026)

An 800-unit property manager cut admin staff in half and lifted margins from under 20% to 32%. This one is not a tech startup, and that is the point. On The Property Management Show, Ping Xu of Spotted Properties, which manages about 800 rental units in Ontario, explained how she rebuilt her back office around AI. She built the tools herself.

The results she reported:

Admin team down to three people, from about six, over roughly a year (last summer to this summer). The virtual assistant role went first, then the full-time finance manager.
Profit margin from under 20% to about 32%. "Before I had such a big problem breaking 20%... with the AI right now we're at 32." She says that came from cutting payroll.
55% of the tedious work in maintenance and accounting is now automated.
Invoice day shrank from all day Thursday to one or two hours on Friday. The old process had seven or eight steps, with a person double-checking another person at each one.
Legal notices that took an admin hours are now "one click." That includes Ontario's standard rent-increase and eviction forms.

Her method is worth copying. She started with the employee who pushed the most work back to her, and asked a simple question: what do you hate doing? "They give me a list. I'm like, okay, let me handle this." Once she automated that list, the employee had spare hours, and she either found them higher-value work or helped them move on. "It was not a surprise to them."

Getting the team to actually use the tools took about five months, from November 2025 to April 2026. The early problem was that staff went back to their old routines the moment the AI glitched. She now treats her own team as customers: "If you churn, that means that the AI product is not strong enough." The turning point came when staff started asking her for new features.

She also sees a payoff beyond margin. A property management business usually sells for one to 1.5 times earnings. An AI-run one, she argues, "we can look at maybe three or four."

The Property Management Show, "Building an AI-Native Property Management Company" (Sep 30, 2026)

A Fortune 500 healthcare company halved its outsourced call-center staff with one "AI employee" named Clara. On Be Customer Led, Gabe Larsen, chief revenue officer of Autonom, described a customer deployment. Keep in mind this is a vendor describing its own client, so it is a sales story. But the numbers are concrete.

The client is a large healthcare company that relied on an outsourced call center (a "BPO," or business process outsourcing firm). Costs were "going through the roof." The result:

Outsourced reps went from 300 to about 150, heading toward 100.
One AI agent, "Clara," handles about 5,000 inbound calls a day, plus a few hundred chats and a few hundred emails. It does the whole first-level support job: looking up the customer, creating and closing tickets, updating Salesforce, and handing off to a human when needed.
It builds its own knowledge base. The company did not have good documentation. So at the end of each week, Clara reports what it couldn't answer, for example: "I got asked the same question 15 times. We ought to create a knowledge source."

Larsen pushed back on the idea that you can simply swap people out: "You don't fire everybody and throw AI in." The goal, as boards now put it, is to "scale output, not headcount."

Be Customer Led, "Gabe Larsen On How to Staff Your Customer Service Team with Autonomous Cloud Employees" (Oct 1, 2026)

Accounting firms: the bottleneck isn't the AI, it's the way the work is organized. On the Jason On Firms Podcast, accounting-firm advisor Jason Staats gave a progress report that should interest any founder running a services business:

Tax data entry: from 15 minutes to under 90 seconds. Seven months ago he showed Claude pulling "70 plus pages of tax documents" into a "60-some tab Excel workbook" in about 15 minutes. "That same process, less than 90 seconds now with greater accuracy."
Claude as the accounting system. Four months ago he showed Claude managing a 7,000-transaction-a-month bookkeeping file, stored as a plain text file through a free tool called Beancount. One firm owner told him they had since moved "a good number of clients" over.

His main point is about process, not tools. Most firms "stumbled into" workflows that AI can't fit into: numbers keyed straight into tax software, review done by eye against 50 pages of documents. The fix is to put the whole job in one place, like a single standard workbook, so AI can move it from step one to step ten. "We have to understand AI enough right now to be able to build a workflow that AI can be a part of."

And the scale argument: "Anything that you can get AI to do for you can be deployed at any scale... You can do that for a thousand clients simultaneously, 10,000 clients simultaneously."

Jason On Firms Podcast, "647 AI Just Changed How We Should Run Our Accounting Firms [It's time to acknowledge the truth]" (Sep 29, 2026)

HP's software chief: the "assistant" layer becomes agents, and managers become individual contributors. On The Product Podcast, Faisal Masud, president of HP's digital services business, sketched how he thinks org charts will change. He used retail as the example. Today a buyer, a planner and a merchandiser each have a junior assistant next to them.

"With AI, I think those second levels are agents. And those first level buyers are actually now not managers, they're actually ICs managing agents."

("IC" means individual contributor, someone who does the work rather than managing people.) The host, who is also deploying agents heavily, raised the problem nobody has solved yet: "Who is going to manage this fleet of agents? How good are they? When do we actually need a human instead of an agent?"

Masud also explained how HP proved its own AI support product before selling it. They tested it first on HP's own 80,000 devices. It now promises customers 30–40% fewer IT support tickets, and says some customers have seen fix times fall from "6, 7 days" to "2, 3 days." He thinks pricing will follow: fewer human users means seat-based pricing gives way to fees tied to results.

The Product Podcast, "HP President on Replacing Middle Management and Cutting Tickets 40% with AI Agents | Faisal Masud | E314" (Sep 30, 2026)

The Other Side

Not every founder this week said AI is shrinking their team.

Higgsfield's biggest regret was cutting humans too soon. Mashrabov expected AI to mostly replace legal and customer support. "That's obviously one of the main mistakes operationally, which we have done in the company, that we didn't ramp these teams quickly." Higgsfield now has 10+ people in legal and 40+ in customer success. AI could probably handle "over 60%" of first-line support, he says, "but when it especially comes to B2B, like AI just doesn't work." The reason is surprising: the company ships so fast that the AI support agents can't keep up. "If context and rules change pretty much twice a week, it gets a little difficult." In other words, faster product teams make AI support harder, not easier.
"Workslop" has a price. On Business of Tech, Dr. Gleb Tsipursky cited Stanford and BetterUp research published in Harvard Business Review. It found that workers say 15% of what colleagues send them is poorly made AI output. Fixing each one takes about two hours. The total cost: $186 per employee per month.
AI adopters are hiring more, not less, in some research. Tsipursky also cited findings that companies adopting AI grow headcount 6% faster and revenue 9% faster than those that don't. He argues the gains go to owners, not to smaller teams.

The Twenty Minute VC (20VC), "20VC: $1BN ARR in 18 Months; The Untold Story of Higgsfield..." (Sep 28, 2026); Business of Tech: Daily 10-Minute IT Services Insights, "AI Margins and MSP Growth: Dr. Gleb Tsipursky on Passing Savings vs. Competing Away Profit" (Sep 26, 2026)

The Cost Corner

This was the richest theme of the week. The pattern: companies told everyone to use AI as much as possible. Now the bills are in, and leaders are building budgets, limits and model-switching to control them.

Higgsfield: $4 million a month, and a $30,000 week. Back to Mashrabov. Internal AI model spend is "over 4 million" a month, more than $10,000 per person. That rose when the creative team (not just engineers) started building their own tools with AI, which people call "vibe coding."

"This month, I just caught a guy who spent over $30,000 in a week on Astra model... just went like five nights straight on Astra... Many people spend over $10,000 in a week."

He admits his finance team thinks he is too loose: "Sometimes I feel it goes like, it really goes out of control." And it is going up. He expects top "10x" engineers and creatives to spend $50,000 to $100,000 a month within a year. He also expects them to ask for matching raises.

The margin side is just as important. Higgsfield earns over 80% gross margin when it runs customers on its own or open models, but only 20–30% when it resells closed models from the big labs. So it picks the model itself in over 40% of cases, which it calls "tokenomics." For any AI app company, which model you route traffic to now decides your margin.

The Twenty Minute VC (20VC), "20VC: $1BN ARR in 18 Months..." (Sep 28, 2026)

Zuora: per-person token budgets, and the $400 bottle of wine. McElhatton says Zuora moved almost all its AI tools from per-seat pricing to per-token pricing this year. The hard part is that employees can't see what they are spending. His analogy:

"It's almost like going to a restaurant... someone giving you a wine list and saying, these are all really good wines. And you just order one. And then, like, I didn't realize that wasn't a $40 bottle of wine. That was a $400 bottle of wine."

So Zuora is building live dashboards and giving every employee a budget based on their role. The goal is for people to learn trade-offs like "if I use this model here, it's a third of the cost. And I pretty much get the same output." He was also clear that heavy use alone proves nothing: "Because you're using the most tokens doesn't mean you're the smartest engineer around."

Run the Numbers, "Zuora's COFO On Going Private and Pricing AI | Todd McElhatton" (Oct 1, 2026)

SAP's CEO: "my CFO says, hey, where is this going?" On Big Technology Podcast, SAP CEO Christian Klein admitted he resisted limits at first, because they clashed with his own message to staff to use AI. But about three to four months ago, SAP introduced token budgets by job profile: for developers, product managers, designers, finance and HR. Managers can approve more for a specific project. The top 1% of users, the people improving SAP's AI agents, have no cap. "Let them run."

Klein says the bigger lever was switching models, which "can cut your cost by... a factor 10." SAP's product managers must now test every agent on several models, and many agents are already on "the fifth model." His core warning is one every founder should hear:

"It doesn't help you if some of your employees is getting 20% more productive, if at the same time, the cost is going 30% more up."

Host Alex Kantrowitz cited Ramp data showing the top 1% of users cut monthly AI spend from $7,976 to $7,205, a 9.7% drop in one month.

Big Technology Podcast, "SAP CEO: AI Won't Kill Software, But It Will Change Your Job, With Christian Klein" (Sep 30, 2026)

The end of the $200 all-you-can-eat plan. On Everyday AI, host Jordan Wilson showed how big the gap is between subscription plans and real usage. He says he uses about 2 billion tokens a week on a $200-a-month OpenAI plan. If he ran that same work through Anthropic's Fable 5 at its new pay-per-use price ($10 per million input tokens, $50 per million output), "that is $200,000 a month." His roundup of companies pulling back:

Tesla has capped employee AI tool spend at $200 a week.
UBS found 60% of enterprises it surveyed are already throttling AI spend.
A friend at a big tech company has a $100 a month limit.

His contrarian point: if the giants are rationing AI, smaller companies that keep spending may gain an edge.

Everyday AI Podcast, "Ep 872: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem (Start Here Series Vol 31)" (Sep 29, 2026)

Cheaper tools for the same job. Two more data points on cutting cost without cutting usage:

$40,000 down to $6,000 a month for a 20-engineer team. On The VentureFizz Podcast, Jack O'Brien, CEO of inference startup Subconscious, described a beta customer that moved its coding agents onto open models through his platform. Engineers reported "tokens are being generated faster" with no drop in quality. (This is a vendor describing its own customer.) O'Brien's view: "Companies care about costs for the first time," now that open models are good enough for most work. His own startup spends "roughly the same amount on really talented, you know, expensive people as we are on... GPUs."
2 cents versus $4.70 per document. On the Elon Musk Podcast, the hosts covered TypeSafe AI's model JEV, which picks from preset answers instead of writing text. In one test on Texas government contracts, it cost $0.02 per document versus $4.70 for Claude Opus. At 100,000 documents, that is about $2,000 versus nearly half a million dollars. Early users reported it was 89% faster and 39% cheaper (Polylane), and up to 400x cheaper on some sorting tasks. The lesson: "Reading is cheap, but writing is what you pay for." If a job only needs a yes or no, don't pay a model to write a paragraph.

The VentureFizz Podcast, "Episode 446: Jack O'Brien - CEO & Co-Founder, Subconscious" (Sep 28, 2026); Elon Musk Podcast, "Why Jev refuses to write sentences" (Sep 27, 2026)