Nvidia’s Jensen Huang says he would be “deeply alarmed” if a $500,000 engineer burned less than $250,000 in AI tokens a year. On the latest Bricks & Bytes, four operators picked that idea apart. Their actual spend ranges from £36 a month to a projected $200,000 a year on Claude alone, and they flat out disagree on whether token usage means anything at all. The split matters because “use more tokens” is quietly becoming a management metric, and most of the people pushing it are the people selling the tokens. Here is the debate, and here is who we think has it right.
There is a new productivity number making the rounds, and it is a strange one. Instead of measuring what an engineer ships, some leaders now want to measure how much AI compute they consume. Spend more, the logic goes, and you must be getting more done.
The most extreme version came from Nvidia’s Jensen Huang, who said on a recent podcast that he would “go ape” if a top engineer on a $500,000 salary spent only $5,000 on tokens in a year. His target? At least half their salary, burned in compute. When asked if Nvidia was spending $2 billion a year on tokens for its engineering team, his answer was simply, “We’re trying to.”
That framing landed in the middle of a Bricks & Bytes panel where Owen Drury, Dustin DeVan, Patric, and Martin compared how they actually run AI inside their own companies. The numbers were all over the map. So was the philosophy. What follows is the disagreement laid out plainly, because if you run a team, this argument is heading your way whether you want it or not.
One camp treats token spend as a sign of ambition and pulls forward deliverables by burning more. The other treats it as a vanity metric with no link to outcomes. The honest answer sits closer to the second camp, with one important exception for cash-burning startups.
The case for spending big
Why pulling deliverables forward can pay for itself
Dustin DeVan, who runs a venture-backed ConTech company, made the strongest version of the spend-more argument, and it is not the lazy one. His point is about cash flow, not bragging rights. Most tech startups are burning capital every month. If you can spend $5,000 in tokens to accelerate a deliverable by a month, you have effectively pulled that revenue or that milestone forward by a month. For a company losing money each month, that math often works.
So at his company the engineering token cap exists, but it gets switched off regularly. He expects to spend somewhere around $200,000 on Claude this year across roughly 25 engineers. Crucially, he does not see that as waste. He sees it as buying speed in a window where speed compounds, since the longer a feature is in market, the longer you get to sell it.
This is the kernel of truth inside Huang’s pitch. If an AI tool genuinely makes an expensive engineer dramatically more productive, then refusing to spend on it is like a chip designer insisting on pencil and paper. Under-use can be a real problem. Nobody serious is arguing you should starve your best people of good tools.
- You are cash-flow negative and racing a clock. Pulling a deliverable forward a month is worth real money when every month burns capital.
- The work is genuinely token-heavy. Coding, large-scale analysis, and agentic workflows can justify serious compute in a way email drafting never will.
- You trust the people spending it. DeVan’s model only works because his teams have equity and are treated as adults making business decisions.
The case against the metric
Token spend has nothing to do with outcomes
Here the panel turned sharp. DeVan himself, despite spending the most, was the most scathing about Huang-style mandates. His argument: token consumption has nothing to do with an outcome. The best engineers now use AI to write code while they think hard about architecture, and hard problems take time to think about, not tokens to brute-force. Pushing someone to keep an AI running just to burn budget is, in his words, completely contrary to the culture you would want to build.
Patric, who invests in and advises startups, made the cynical-but-accurate observation. In a company that mandates token spend, the smartest way to hit your quota is to do the bare minimum for your salary and burn a pile of compute building your own side project on company time. Mandate consumption and you do not get productivity. You get theatre, and possibly a competitor incubating on your dime.
Then there is Patric’s German idiom, which should be printed on a poster in every boardroom debating this: if you want to drain a lake, do not ask the frogs. The loudest voices telling you to use more AI are very often the ones who profit most from you using more AI. That does not make them wrong. It does mean their incentive is not your incentive.
If you want to dry up a lake, don’t ask the frogs. Don’t listen to the chip maker about how much AI you should be using.Patric, on the Bricks & Bytes panel
Cut through the AI noise
We track what’s actually working in construction tech, not what the vendors want you to think is working. One sharp email a week.
Join 2500+ ReadersThe Martin question
What if your business just doesn’t need to spend much?
The most useful voice on the panel might have been the quietest. Martin runs a small design team of about six people. His total AI bill is roughly £36 a month, two subscriptions, and he uses it to summarize emails, help with writing, and do a bit of creative thinking. His honest question to the room: does every business actually need to spend a lot on this? For what his team does, he genuinely could not see where the extra money would go, except maybe waste.
That is not a confession of being behind. It is a reasonable read of his situation. A traditional design studio has a different token appetite than a venture-backed software company shipping code daily. The mistake is assuming one number fits both. The right spend is whatever the work demands, and for some teams that is a rounding error.
This is where Patric’s “mission” model is worth stealing. Rather than open the floodgates, he gives deliberate, project-specific token budgets. Someone has to define the mission, the objective, the milestones, and the expected outcome, then say how much compute the bet is worth. Sometimes that is $10,000 on a single push. The point is the spend is bet-driven, tied to a result, not a quota to hit for its own sake.
So who is right?
Measure the outcome, fund the bet, ignore the mandate
Put the three approaches side by side and a clear answer emerges. The spend-more camp is right that under-use is a real cost and that speed can be worth paying for. The against camp is right that consumption is a terrible proxy for value, and that anyone mandating it should be viewed with suspicion. Martin is right that plenty of businesses simply do not need much.
The synthesis is Patric’s mission model. Tie spend to a defined outcome, give people real budgets for real bets, and judge results by what shipped, not what was consumed. That gives you DeVan’s speed when speed matters, avoids the theatre Patric warned about, and respects Martin’s reality that some teams just do not need to burn much. Construction has an extra reason to be careful here. You can only move as fast as your customers can adopt, so racing a 12-month roadmap nobody is ready for is its own kind of waste.
The token economy is real and it is growing. But “how much did you spend” is the wrong question. “What did it get you” is the right one. Anyone selling you the first question is, gently, a frog at the edge of the lake.
| Approach | Who holds it | Core logic | Best for | Main risk | Token mindset |
|---|---|---|---|---|---|
| Spend big, uncap | DeVan (ConTech CEO) | Pull deliverables forward; speed compounds | Cash-burning startups racing a clock | Building faster than customers can adopt | A tool worth buying |
| Mission budgets | Patric (investor/advisor) | Fund defined bets with set token limits | Teams that want control without paralysis | Overhead of defining every mission | A bet to be sized |
| Spend lean | Martin (design studio) | Only pay for what the work demands | Traditional, non-code-heavy businesses | Genuinely missing a useful tool | A small line item |
| Mandate consumption | Huang-style edict | High spend signals high productivity | Almost nobody, honestly | Theatre, gaming, tech debt, side hustles | A target to hit |
Speaking on the All-In Podcast around Nvidia’s GTC 2026, Huang said he would be “deeply alarmed” if a $500,000 engineer consumed less than $250,000 worth of AI tokens in a year, and that a $5,000 figure would make him “go ape.” He has framed token budgets worth roughly half an engineer’s salary as a Silicon Valley recruiting tool. (Source)
On its own, no. Consumption measures input, not output. The Bricks & Bytes panel argued it has no link to actual outcomes, and the wider debate backs that up. Microsoft’s heavy AI-coding push in 2025 drew criticism that code quantity is not the same as code quality. The better measure is what ships and whether it works. (Source)
Some are pushing hard. Coinbase CEO Brian Armstrong gave engineers one week in 2025 to onboard to AI coding tools like Cursor and GitHub Copilot, and those who did not comply were let go. Coinbase later reported around a third of its new code was AI-generated, with mandatory human review. Note this was an adoption mandate, not a token-spend quota. (Source)
It depends entirely on the work. A code-heavy software team can justify large bills; a traditional design or admin team often needs very little. The panel’s most practical model was Patric’s “mission” approach: fund specific projects with defined token budgets tied to clear outcomes, rather than setting a blanket quota or removing all limits. (Source)
Most early-stage tech companies are cash-flow negative, burning capital every month. If spending more on tokens pulls a deliverable forward, it can pull forward revenue or a funding milestone, which makes the spend a net positive. An established, profitable firm with steady output has far less reason to chase speed at any cost. (Source)
Yes, somewhat. Construction adopts technology slowly, so you can only move as fast as customers and crews are ready for. Burning compute to race a roadmap your market cannot absorb yet is wasted speed. The mission model fits well here: tie AI spend to specific, adoptable outcomes rather than chasing a consumption target. (Source)
Tom’s Hardware: Huang on engineers using tokens worth half their salary
Business Insider via AOL: Huang “deeply alarmed” by low token spend
Fortune: Coinbase CEO fired engineers who wouldn’t adopt AI
WinBuzzer: Huang’s AI token pay proposal explained
Entrepreneur: The wider corporate AI mandate trend
Related Articles
Nemetschek Acquires HCSS: Why the Infrastructure Bet Is the Biggest Signal in ConTech Right Now
The Bricks & Bytes Weekly Newsletter
