The AI capital flow problem: chips eat, SaaS starves
The same AI wave giving chip workers $100K bonuses is laying off the people who write software. The money flows up, not out.
SK Hynix employees received roughly $100K in profit-sharing last year. Q1 2026 operating profit was up 405%, and Macquarie is projecting something like $900K per capita by 2027. Some cloud companies are apparently funding dedicated production lines for them now, paying upfront to build factories that only produce chips for them.
GitLab announced layoffs this week and also retired their CREDIT values framework (Collaboration, Results for Customers, Efficiency, Diversity, Inclusion and Belonging, Iteration, Transparency, the thing they have been running on since 2015). The new values are “Speed with Quality, Ownership Mindset, Customer Outcomes.” Their stock is down about 50% in the past year. Worth noting that the old framework is still in their public handbook (hooray for version control) if you want to compare.
An employee on HN: “more like ‘get screwed, work more, for less, in worse environments.’”
Both of these are being called “AI investment.”
Where the money actually goes
Every company is “investing in AI.” In practice this usually means moving budget from people (OpEx) to hardware (CapEx).
A senior engineer costs around $300K a year, closer to $500K with benefits. An NVIDIA H200 server costs around $500K. When the market narrative says “AI will replace engineers,” the CEO buys the GPU, not because it actually replaces the engineer, but because investors believe it does, and buying GPUs moves the stock price.
The supply chain tells you who captures the money. ASML makes the EUV machines that make the chips (over $300M each, backlogged). TSMC does the manufacturing ($35.9B revenue in Q1, +41%). NVIDIA and SK Hynix design the chips ($44.1B revenue for NVIDIA, +69%). Cloud companies buy those chips and build data centers, spending a lot of money but mostly passing it upstream. SaaS companies like GitLab try to sell software to the people using those data centers, and watch their stock get cut in half.
The closer to hardware, the more profit grows. The closer to applications, the more pressure there is. By the time you get to the bottom of the stack it is mostly mist.
The three things nobody wants to say out loud
Buying GPUs is easier to explain to investors than keeping engineers
Multiple CEOs have used the phrase “compute crowding out headcount” in earnings calls. It is not a metaphor, it is budget math. Buying GPUs sends the stock up. Keeping engineers does not.
IBM’s Arvind Krishna is about the only major tech CEO who has publicly said “there is no way spending on AI data centers will pay off.” That is probably because IBM both sells AI (Granite, WatsonX) and consumes GPUs, so they see both sides of the bill at the same time. When you are paying and collecting, the math looks different.
Nobody can actually define what they are replacing people with
GM just laid off 500-600 IT workers to hire people with “stronger AI skills.” Worth knowing that GM spent over $10 billion on Cruise between 2016 and 2024, lost $3.48 billion in 2023 alone, and shut down Cruise’s independent operations in December 2024. The same company, the same mouth, now says it is betting on AI again.
One HN commenter asked what “AI-native workflow” actually means: “I don’t understand these words. Does ‘AI-native workflow’ mean vibe coding? I am now seeing a lot of roles asking for ‘AI-enabled engineers’. And I am not sure what that means either.” Someone else answered: “Cheaper younger people who don’t think vibe coding is bad.”
The people being laid off know things that were never written down
When a company fires its senior IT staff, it does not lose “IT headcount.” It loses the person who knows why that one database field is structured the way it is (there was a migration in 2018 that went badly), and the person who knows that a certain API has an undocumented behavior you should not touch, and the person who can tell you which systems will fall over every March during tax season.
None of that is in Confluence. None of it is in Jira. AI cannot replace it, because it was never recorded in the first place.
SaaS tools are getting commoditized, and software engineers are losing negotiating leverage, not because AI actually replaces them (Klarna tried replacing customer service with AI, admitted it did not work after a year, and started rehiring), but because the narrative itself does the damage. The stronger the “AI replaces programmers” story gets, the weaker engineers’ bargaining position becomes, regardless of whether it is true.
One commenter on the GM thread put it simply: “Firing people with institutional knowledge? So what? It’s going to improve profits short-term.” That is the actual thesis. The AI framing is how you say it to the press.
The structural problem
SK Hynix’s 405% profit growth comes from AI chip demand. Those chips are mostly running large model training and inference. Are those models commercially profitable? No, OpenAI is projected to lose $14B in 2026.
So the upstream boom is funded by downstream demand that has not proven its economics yet. If the AI application layer never generates the revenue that justifies all this compute spending, the chip cycle will eventually correct, similar to how Cisco crashed after the first dot-com wave ended.
Short term, probably another two or three years, the spending continues. The hyperscalers have enough cash and credit to keep buying GPUs even without clear returns. HBM already consumes about 20% of global DRAM capacity and is projected to hit 35% by 2027. When AI demand eventually contracts, that specialized capacity is going to be very painful to repurpose.
What to watch
SK Hynix is paying $100K bonuses. GitLab is deleting its values from version control. GM burned $10 billion on self-driving cars and is now calling the same thing by a different name. These are not separate stories.
When a company wraps layoffs inside a market-expansion story, it is worth asking whether they would make the same argument if the thesis were false. In most of these cases, probably yes. The “agentic era” framing is an explanation layered on top of cuts that were going to happen anyway.
Watch what they spend on, not what they write about.