Tunguz uses SaaS-pricing intuition from a decade at Redpoint to argue that per-engineer AI spend is compounding fast enough to breach fully-loaded salary costs (~$300K–$400K) by around 2029. His evidence is the derivative, not the current level: power-user spend has moved from $50/month to $500–$2000/month in eighteen months, a cadence no finance team has modeled.
The editorial argues no CFO will approve a $350K/year Cursor bill for a single developer, so the actual outcome is quiet rationing of agent access rather than a clean salary-crossover. The 'AI replaces engineers' framing distracts from the more boring reality: AI vendors become the second-largest engineering COGS.
The editorial reframes Tunguz's projection by pointing out that agentic workflows where a single Claude Code session can burn $20 in tokens don't behave like fixed per-seat SaaS. The cost curve looks like early AWS spend before FinOps teams existed, meaning the governance problem is usage-based token accounting rather than seat licensing.
The editorial draws a direct parallel to Datadog, which started as a cheap $15/host tool and grew into a line item requiring VP-level sign-off on log retention. AI coding vendors are on the same trajectory — starting cheap per seat, then compounding into a top-tier COGS item that forces engineering leadership to police individual usage.
Tomasz Tunguz — the Theory Ventures partner who spent a decade at Redpoint watching SaaS pricing curves — published a post arguing that AI tooling spend per engineer is compounding fast enough to cross fully-loaded salary costs by around 2029. The piece hit Hacker News at #1 with 97 points and a comment thread that read less like a debate and more like a group of engineering managers checking their own invoices.
The numbers he lays out aren't wild. A senior engineer in the US costs a company roughly $300K–$400K fully loaded. Today's per-seat AI stack — Cursor at $20–$40, GitHub Copilot at $19–$39, Claude Code and API usage on top, plus whatever RAG-over-your-codebase vendor sneaks in — lands somewhere between $1K and $5K per engineer per year at most shops. That's noise. The problem is the derivative. Token-heavy agentic workflows are pushing power-user spend from $50/month to $500–$2000/month in eighteen months, and nobody in finance has a mental model for a line item that 10x's on that cadence.
Tunguz's straight-line projection is the click-baity part. The interesting part is buried under it: if you extrapolate current agent adoption — where a single Claude Code session can burn $20 in tokens on a nontrivial refactor — the per-engineer AI cost curve doesn't look like SaaS. It looks like cloud compute in 2015, before anyone had a FinOps team.
The headline framing — "AI will cost more than the engineer" — is the wrong fight. No CFO is going to approve a $350K/year Cursor bill for one dev. What actually happens is more boring and more painful: the seat gets rationed.
We've seen this movie. Datadog was the last vendor to pull this trick on engineering budgets. It started as a $15/host monitoring tool and ended as the line item that made VPs of Engineering personally sign off on log retention windows. The pattern isn't "AI replaces engineers" — it's "AI vendors quietly become the second-largest COGS in your engineering org, and then someone has to justify the spend head-by-head." Tunguz is early on this, but the trajectory is legible.
The HN thread caught something the post didn't quite say out loud: almost nobody is measuring AI-assisted output rigorously. GitHub's own studies claim 55% faster task completion for Copilot users. Uplevel's replication came back with essentially no measurable throughput gain and *more* bugs. METR ran a controlled study on experienced open-source devs and found AI tools made them 19% slower — while the same devs reported feeling 20% faster. That gap between perceived and actual velocity is the number that should terrify anyone underwriting a five-figure per-seat AI budget.
There's a second-order problem the post glosses over. Anthropic and OpenAI are burning cash to keep inference prices below cost. Claude Sonnet dropped in price twice in 2024 and once in 2025, and API margins on the frontier models are, by most external estimates, still negative for long-context agentic workloads. The moment the frontier labs need to actually make money — and the equity story shifts from growth to unit economics — the per-token price will move against you, and your "productivity" math will re-underwrite overnight.
The comment section on HN split cleanly along a line worth naming. Staff+ engineers running large agentic workflows reported real gains — refactors that would've taken a sprint compressed to an afternoon. Mid-level engineers reported a wash: the AI writes plausible code, they spend the saved time reviewing it, net zero. Junior engineers who leaned hardest on the tools reported the biggest gap between output and understanding. If your budget model assumes uniform multiplier across seniority, you're wrong in a way that costs money and, eventually, institutional knowledge.
Three concrete things to do before your next planning cycle.
Put AI spend in the same review process as cloud. If you have a FinOps practice for AWS, you need one for Anthropic and OpenAI. Per-engineer usage varies by 20-50x in most teams I've seen data on. The engineer burning $1800/month in Claude tokens might be your best investment or your worst — you cannot tell without instrumenting. Log every API key by team, tag every agent invocation by workflow, and put the top 10% of spenders in a monthly review. Not to punish them — to figure out what they're doing that everyone else isn't.
Stop paying per-seat for tools you can't measure. If you can't articulate what shipped this quarter that wouldn't have shipped without Cursor, you're not buying productivity, you're buying developer satisfaction — which is fine, but it's an HR line item, not an R&D one. That distinction matters when the budget gets cut. Set up a lightweight measurement: PR throughput, time-to-first-commit on new tickets, revert rate. Compare heavy AI users to light users on comparable work. If there's no delta after a quarter, you have a signal.
Assume prices go up, not down. Every FinOps war story from 2015-2020 starts with "we architected for the promo price." Frontier model pricing is currently a customer-acquisition subsidy. Your 2027 budget should model 2-3x the current per-token cost, especially for anything using extended thinking or large context windows. If your projected ROI doesn't survive that stress test, you're building on quicksand.
The 2029 breakeven is almost certainly wrong in its specifics — straight-line projections of exponential curves usually are. But the direction is right, and Tunguz is naming something the industry has been too polite to say: AI tooling is on track to become the most-scrutinized line item in engineering budgets, and the shops that can't tie spend to shipped work are going to lose the argument. The winners of the next two years won't be the teams with the biggest Anthropic invoice. They'll be the ones who can point at that invoice and say, with a straight face, exactly what it bought.
Analogous statement:Evian use 1.25 million litres of water per employee per year. When can we expect other non-bottled-water corporations to rise to this level of water usage?
A.I suffers from the last-mile problem. It can do 90% of the work in 20 minutes but then the remaining 10% ends up taking 20 million hours to actually finish. It frustrating to the point that I sometimes want to throw the whole thing out and start from scratch.
Open-weight models are going to completely shatter these forecasts. It takes a little more effort – right now, probably won’t be true in three months – but you can achieve the same at 1/10th of the cost.
Mr. Mark Zuckerberg is particularly not happy about these stats. He was promised something else and he has already fired like half of the company.It is really crazy people didn't think this through.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Garbage. You can't include training by the companies that develop an llm in the comparison against companies that merely use the same llm. Apples and potatoes.