Future Markets: There Will be Blood
The AI bubble is bursting. Few people want to acknowledge it.
AWS was frequently down in Q1 (less frequently in Q2) as a result of firing a bunch of experienced engineers to conserve capital, as GPU clusters are not cheap and the economy is weak. 80% of the remaining engineers, many of them inexperienced, were forced to complete their weekly coding tasks using Amazon’s substandard in-house IDE - Kiro. Predictably, code broke and continues to break. The most senior engineers are now spending their time cleaning up the mess. My guess is that AWS CEO Matt Garman will be forced out this year.
If you don’t write code you don’t know how bad forced usage mandates are. There are plenty of stories of Amazon’s Kiro unnecessarily re-writing entire applications, wrongly deleting code bases and the like. No engineer wants to use subpar tools. When forced to do so, bad things happen.
At some point Amazon (AMZN) will have to pull back on CapEx. You simply can’t have AI shipping code to the point where it breaks the entire AWS business. The fact that a few weeks ago Matt Garmin had to mandate that all AI produced code be reviewed before shipping raised an eyebrow or two. You mean to say that unreviewed code was flying out the door? If you ship AI written code without reviewing it you will get bitten.
At some point AWS will quietly hire engineers back if it has not already done so. The pace of shipped code will slow - my guess is this has already happened, which is why AWS code breaks are down, not because AWS is producing better code.
AWS will never admit that they are screwing up, but when reported numbers slow, both the economy and poor AI scaling will be to blame.
AWS was never an “AI” company to begin with. Sure, AWS was early to leverage machine learning in its cloud business, but AWS doesn’t exactly have an AI lab that rivals OpenAI, Anthropic or DeepMind.
AWS isn’t alone. Microsoft (MSFT) has failed miserably with Copilot despite forcing employees to use it.
Salesforce (CRM) cut engineering heads too fast, too early a few months ago.
Goodness knows what’s going on at Oracle (ORCL), they don’t have an AI lab, yet founder and CTO Larry Ellison says AI is writing 80% of Oracle’s code! That OpenAI Oracle deal never happened. I knew it wouldn’t.
As I have written over the past few days:
1.) companies are not hiring young people - exactly the cohort who would build AI agents and drive the next leg of AI growth.
2.) Anthropic and OpenAI will start to slam into the systems integration wall as they sell into the enterprise beyond the Software Engineering function as knowledge work is famously siloed and big AI projects can’t be executed across siloed data.
3.) Software applications need to be rewritten to CLI to optimize for agentic AI, as agents struggle to engage with traditional SaaS applications.
The LLM builders can keep buying NVIDIA (NVDA) chips and building models, but growth is about to slow, and the problem can’t be solved with capital or technology alone. Time will be the gating factor. It takes time to knock down data silos. I don’t believe many companies will be taking on AI projects with the 10YR approaching 5% (and going much higher in my view) and the price of oil north of $100. Companies will push to a degree, they won’t want to admit they were wrong to the Board, but the truth about AI will come out (it’s more expensive to deploy than we initially thought). Growth will slow and many CEOs will be fired. There will be blood. All avoidable, had only the CEOs themselves done more personal due diligence around LLMs they would have been better positioned to push back on their Boards when Boards pressured them to “look into AI, my friend’s company is doing it.”