Community Trust ScoreVerified
What happened
The promise of artificial intelligence has often been equated with seamless efficiencies and rapid returns. But the numbers tell a messier story.
A survey of 2,444 companies by Entelligence AI found that up to 82% of enterprise AI spending gets swallowed before anything useful ships. Bug fixes. Code rewrites. Review delays. For every dollar a company puts into AI, nearly half goes straight to patching errors — not building product. Another big chunk goes to rewriting code that AI generated in the first place. So the machine meant to save engineering hours is, in many cases, creating more of them.
Not cheap.
Lightrun’s 2026 State of AI-Powered Engineering Report adds more weight to that picture. Nearly half of AI-generated code still needs manual debugging even after clearing quality checks. That’s not a draft problem. That’s a production problem — code that passed review, got deployed, and still broke. The gap between what AI tools promise and what they actually deliver in live environments is, based on these numbers, pretty significant.
Meanwhile Oracle has been making aggressive bets on AI infrastructure, and the debt load it’s carrying is staggering — $108 billion in total debt, with additional capital raised through a mix of debt and equity. The company has a substantial portion of its backlog tied to OpenAI as a client. OpenAI, which has itself recorded heavy financial losses, is now a meaningful piece of Oracle’s revenue picture. That’s a lot of exposure to a company that isn’t profitable yet.
And OKX, the crypto exchange, is reworking how it evaluates talent. AI proficiency is now a factor in employee assessments. The logic: if the market is moving toward AI-first workflows, the workforce needs to move with it.
The historical context
The dot-com bubble is the obvious comparison, and it’s probably the right one. In the late ’90s, capital flooded into internet infrastructure and unproven business models at a pace that outran any reasonable analysis of returns. When the correction came, it was brutal. Companies that had raised hundreds of millions on the promise of future traffic and future revenue found out fast that promises don’t service debt.
The AI moment feels similar. Oracle’s financial posture — massive debt, aggressive infrastructure buildout, reliance on a loss-making anchor client — mirrors the kind of expansion-without-clear-profitability-pathway thinking that defined the worst of the dot-com era. Bold strategy, maybe. Fragile, definitely.
The code quality problem has its own historical echo. Early commercial software was famously unreliable. The rush to ship, to capture market share, to be first — it consistently outpaced the ability to deliver something stable. AI-generated code seems to be running into the same wall. The tools are faster. The output is still messy.
Why it matters
Companies are figuring out, sometimes the hard way, that dropping AI into an engineering workflow isn’t a plug-and-play upgrade. It’s a strategic overhaul. Debugging pipelines need rethinking. Code review processes need expanding. Quality benchmarks need resetting. The firms that treat AI adoption as a cost-cutting shortcut are the ones racking up the hidden expenses that Entelligence’s survey captured.
Winners here are probably the companies that went in with realistic expectations — that built error-correction capacity alongside AI deployment, rather than assuming the tools would handle it. Losers are the ones that overcommitted, cut human oversight too fast, and are now quietly eating the rework costs.
For Oracle, the stakes are higher than most. With $108 billion in debt and a backlog heavily weighted toward OpenAI, the company’s upcoming earnings report on June 16, 2026 is basically a referendum on whether its AI-driven strategy can hold. A miss wouldn’t just be a bad quarter. It could raise serious questions about the sustainability of the whole model.
The Entelligence AI data also points to something cultural, not just technical. Organizations aren’t just struggling with bad code. They’re struggling with a gap between what leadership expects AI to do and what engineering teams are actually experiencing. That disconnect — between the boardroom pitch and the production reality — is where a lot of the hidden cost lives.
What to watch
Oracle’s earnings on June 16, 2026 are the clearest near-term signal. If the company misses targets, it won’t be easy to separate AI-related pressure from broader macro issues, but the debt exposure to OpenAI will be the first thing analysts start pulling on. A strong beat buys the strategy more runway. A miss starts a different conversation.
The Lightrun figure worth watching: if the share of AI-generated code requiring manual debugging stays above 40% post-production, that’s not a temporary growing pain. That’s a structural problem with how AI tools are being integrated into development cycles. It would mean the efficiency gains everyone is projecting are, at minimum, being significantly offset.
OKX’s experiment with AI-proficiency-based hiring is harder to track from the outside. But retention rates and productivity metrics over the next few quarters will say something real about whether tying employee evaluations to AI capabilities actually builds a better workforce — or just stresses out people who were already doing their jobs fine.
The broader tension isn’t going away. AI spending is still rising. The infrastructure bets are already placed. But the data from Entelligence, Lightrun, and Oracle’s own balance sheet all point to the same uncomfortable reality: the cost of getting AI wrong is much higher than the industry’s marketing materials tend to mention.
OKX’s headcount strategy is a small data point in a much larger shift — one where AI fluency is becoming table stakes for employment at tech-adjacent firms, whether those firms are exchanges, cloud providers, or somewhere in between.
Oracle carries $108 billion in total debt heading into its June 16 earnings.
