GitLab's layoff cost less than four hours of NVIDIA buybacks
Four divisions you can do on your phone, every input from the companies' own releases. Who gets paid first in the AI boom.
On May 11, 2026, GitLab told its staff that about 14% of them, 350 people, would be leaving, and that the company would pull out of 22 countries. The quarter had closed eleven days earlier. When the numbers for it came out on June 2, they showed $264.2 million in revenue, up 23%, an 86% GAAP gross margin and $149.2 million in operating cash flow. The same release priced the plan at $30 million to $35 million in severance, termination benefits and retention payments, between $86,000 and $100,000 for every role removed.
Below are four divisions. Every input comes from a company’s own release, linked where it appears. I will say what I got each time. Do them yourself and check me.
1. The buyback clock
NVIDIA’s quarter ended on April 26, four days before GitLab’s. In those thirteen weeks the company returned $20.0 billion to shareholders through buybacks and dividends. Thirteen weeks is 2,184 hours. Divide $20.0 billion by 2,184, then divide GitLab’s $35 million by what you got.
I get $9.2 million an hour and 3.8 hours. GitLab’s whole layoff cost less than four hours of NVIDIA’s buybacks. For the same quarter SK hynix reported a 72% operating margin and 35 trillion won of net cash, and TSMC a 66.2% gross margin on $35.9 billion of revenue. GitLab’s 86% gross margin is the highest number in this paragraph, and GitLab is the company that cut.
2. Two points of gross margin
A year earlier GitLab’s GAAP gross margin was 88%. Revenue grew 23% over that year. The cost line under its subscription revenue grew from $19.3 million to $30.6 million, which is 59%. GitLab does not say what is in that line. Its own strategy post says “Enterprise AI bills are climbing as fast as adoption.”
Take 2% of $264.2 million. Multiply by four. Then divide the $35 million severance by the quarterly figure.
That comes to $5.3 million a quarter, $21 million a year, and six to seven quarters before the recurring line has cost more than the one-time charge. The severance is paid once, and the two points come back every quarter for as long as that bill keeps climbing.
3. The token bill grows with the square of the conversation
An agent resends its whole history on every turn. If each turn adds d tokens of context, turn k sends k × d tokens, and a run of N turns sends d × N(N+1)/2 input tokens in total.
Take d = 2,000 and N = 40, then price it off a public list. Google’s page lists Gemini 3.1 Pro Preview input at $2.00 per million tokens for prompts up to 200,000 tokens and $4.00 above that, and Gemini 3.7 Flash at $0.75 through the end of 2026 and $1.50 from January.
My phone says 1,640,000 input tokens, $3.28 a run at the Pro price and $1.23 at the Flash price, output tokens not included. Now set N = 80. The count goes to 6,480,000 and the bill to $12.96, just under four times the money for twice the turns, and the eightieth prompt is 160,000 tokens, close to the line where the per-token price doubles. Context caching takes some of that back, at $0.20 per million cached tokens plus $4.50 per million per hour of storage on the same page, so run the sum again with your own cache hit rate.
Last step. GitLab’s letter says the developer platform market “used to be measured in tens of dollars per user per month, this year it is hundreds/user/month and headed to thousands.” Take a $30 seat and divide it by $3.28. I get nine agent runs per user per month before a dollar of salary, rent or sales commission. The same post keeps subscriptions and adds “consumption pricing for the work agents do”, and I read that line as the answer to this division.
4. Payback
Sales and marketing was $119.4 million in the quarter and $107.6 million a year earlier. Over that year, quarterly revenue grew by $49.7 million, and at an 86% gross margin each new revenue dollar brings 86 cents of gross profit.
Crude version, with the knobs in the open. Assume four quarters of selling at somewhere between those two rates produced the $49.7 million, and divide by the new gross profit per quarter.
I land on ten to eleven quarters, before R&D, G&A or the AI bill. Some of that spend renews existing customers rather than finding new ones, and some of the growth came from existing customers expanding at a 117% net retention rate, so move the inputs and see how far the answer moves. In every version I tried, the chip vendor in division 1 was paid when the parts shipped.
The thread
The HN thread on the announcement was submitted by an account named AnonGitLabEmpl and reached 703 points. Its first comment read “Oh and it won’t be done until June 1st, so the employees can have some anxiety until then. As a treat.” Simon Willison called the strategy section “quite an aggressively optimistic vision for the future of the software industry to tuck into a ‘workforce reduction’ announcement” and noted that the cuts did not appear to target engineers. Six replies deep, one commenter wrote that “Gitlab’s decision is being taken in spite of LLMs, not because of them.”
The announcement lists “rewiring internal processes with AI agents, automating the reviews, approvals, and handoffs” as the fourth of four operational changes, and says two paragraphs later that the changes and the strategy “are related but independent.” The letter to employees under it says “this is not an AI optimization or cost cutting exercise.” The restructuring charge is itemized as severance, termination benefits and retention. No line in either document claims that an agent has taken over a role. I read it the way that commenter did. The cost decision came first, and the agent paragraph gave it a name.
The part the charge does not price
The $30 million covers the exits. Off the books is whatever leaves with them, like the person on the platform team who knows which job falls over every March and why, or the one who remembers that a database column is shaped the way it is because of a migration in 2018 that went sideways. Little of that is written down. One commenter on the thread pointed out that voluntary separation windows tend to take the high performers first, the ones who already have offers elsewhere. GitLab’s window closed on May 18, two weeks before anyone knew the final shape of the company.
My bet
For the next two years I expect the bottleneck vendors to keep winning and generic AI wrappers to get crushed. I expect established SaaS companies to hold their gross margins by charging separately for agent work and to defend operating margins with fewer people. Divisions 3 and 2 are the arithmetic under those two moves. GitLab’s own letter promises to reinvest “the majority of savings”, and it reaffirmed its guidance in the same post that announced the cuts.
I also expect most AI-era headcount cuts to survive when the productivity numbers disappoint. While the payroll saving has a hard effective date, the promised AI output remains trapped in a slide deck. When the memory cycle turns, the HBM orders get cut, and I do not expect the 350 roles to come back with the next upcycle.