The numbers are staggering enough to make you question whether you read them right. Analyst estimates for 2026 put combined capital expenditure from the four largest cloud providers — Amazon, Google, Microsoft, and Meta — somewhere north of $600 billion, with some projections reaching $725 billion. Most of that money is aimed at AI infrastructure: data centers, GPUs, custom silicon, networking. At capital intensity ratios of 45–57% of revenue, these companies are committing to AI at a scale that has no clear historical precedent outside wartime industrial mobilization.
Meanwhile, in the actual day-to-day of software development, a different story is forming. A term coined as internet slang — "AI slop" — went mainstream enough in 2025 that Merriam-Webster named it a word of the year. Mentions of the phrase increased roughly ninefold across the web between 2024 and 2025, with negative sentiment peaking near 54% in October. GitHub reportedly floated a kill-switch concept for pull requests overwhelmed by unreviewed AI-generated submissions. Academic researchers have framed the phenomenon as a tragedy of the commons: individual developers externalize the cost of their productivity gains onto reviewers, maintainers, and the broader open-source community.
So you have two simultaneous realities. Capital is betting on AI at an almost incomprehensible scale. And the humans who are supposed to be using the resulting tools are developing measurable trust deficits toward the output those tools produce. It's worth asking whether these two curves are on a collision course — or whether the tension resolves in a way the current discourse isn't fully capturing.
The Bull Case: Infrastructure Leads Demand, Always Has
The most credible version of the optimist argument isn't that AI is already generating broad returns — it's that infrastructure investment necessarily precedes the applications that justify it. Railroads were overbuilt relative to the demand of their era. Fiber was laid far ahead of the traffic that would eventually fill it. The surplus capacity of one cycle becomes the cheap substrate for the next.
GitHub Copilot has reached somewhere around 20 million cumulative users, with roughly 4.7 million paid subscribers as of early 2026. Internal GitHub research has cited productivity improvements on the order of 55% for specific coding tasks, and data showing pull request cycle times dropping substantially for teams that adopt the tooling. These are not nothing. The tool writes 46% of code in repositories where it's enabled — a figure that, if accurate, represents a genuine shift in how software gets produced.
The adoption curve in enterprise is also real, even if uneven. Around 90% of Fortune 100 companies report using AI coding tools in some capacity. The highest-performing deployments tend to share three properties: domain specificity, deep integration into existing workflows, and a deliberate decision to build or buy rather than experiment broadly. That pattern — concentrated wins in narrow, well-defined contexts — is consistent with how most transformative technologies actually diffuse through organizations.
The bull case, stated plainly: we're in the early period where the infrastructure is being built and the uses are being discovered. The ROI looks thin now because organizations haven't yet restructured around the technology. That restructuring, when it happens, unlocks the compounding returns. The spending today is the price of not being stranded later.
The Bear Case: Low-Quality Output Is Eroding the Trust Infrastructure Depends On
The problem with that argument is that trust is not a neutral backdrop — it's an input. And the data on developer trust is genuinely alarming if you take it seriously.
Surveys from 2025 found that 84% of developers use or plan to use AI tools in their development process. In the same surveys, only 29% said they trust the accuracy of AI-generated code — down 11 percentage points from the prior year. That's a widening gap: adoption is rising while trust is falling. Adoption without trust produces a specific failure mode: developers submit code they don't fully understand, reviewers can't assume the author comprehends what they've written, and the review burden increases rather than decreases. Independent code analyses found AI-coauthored pull requests averaged roughly 1.75 times more correctness issues and 1.57 times more security issues than human-written ones.
The trust problem compounds in ways the productivity statistics don't capture. When nearly half of developers admit to committing AI-generated code without reviewing it — partly because review takes longer than generating the code in the first place — the productivity gain is real for the individual developer but negative for the team. Review queues lengthen. Bugs ship. Maintainer burnout accelerates. The cost is just distributed differently.
This dynamic extends beyond code. Open-source projects have reported being overwhelmed by AI-generated bug reports and vulnerability disclosures that consume maintainer time without surfacing valid findings. Stack Overflow's signal quality has degraded. The informational commons that developers depend on — documentation, Q&A forums, tutorials, third-party libraries — is increasingly contaminated with output that looks plausible but isn't verified.
The bear case isn't that AI coding tools don't work. It's that the conditions required for them to work well — careful prompting, rigorous review, genuine understanding of what the model produced — are being skipped at scale. And the infrastructure spending is predicated on continued adoption, not adoption followed by a trust collapse that forces a reset.
On the organizational side, the numbers aren't encouraging either. One widely cited survey found that only 29% of enterprises report significant ROI from generative AI, despite the fact that 97% of executives claim to be benefiting from it. IBM reportedly estimated that 75% of AI initiatives don't deliver expected returns. An analysis of S&P 500 companies found that only about 21% could cite a measurable AI benefit at all. These figures come from surveys and analyst reports with their own methodological limitations, so treat them as directional rather than precise. But the direction they point in is consistent: individual productivity wins are real; organizational transformation is lagging badly.
What a Mismatch Would Actually Mean
Here's the structural uncertainty that neither the bulls nor the bears resolve cleanly: the spending and the demand operate on different timescales with different feedback loops.
Hyperscaler capex commitments run two to four years ahead of the capacity coming online. Data center construction has its own constraints — transformers, switchgear, grid interconnections — that don't respond quickly to changes in demand signals. Estimates suggest that data centers representing something like 7 gigawatts of new capacity announced for 2026 haven't broken ground. The physical infrastructure, once committed, doesn't flex.
Developer trust, by contrast, operates on much shorter cycles. A team that gets burned by AI-generated code in a critical system doesn't wait years to revise its workflow. It puts gates in place, requires authors to explain what they submitted, or bans the tool outright for production paths. These micro-decisions aggregate quickly across organizations.
The most plausible version of a mismatch scenario isn't a dramatic bubble pop — it's something more gradual and harder to see in aggregate numbers. Adoption figures stay high because developers continue using AI tools. But the high-value use cases narrow: the tools work well for drafting boilerplate, writing tests for well-specified functions, translating between languages. They work poorly for complex reasoning about system behavior, security-sensitive code, and anything requiring genuine knowledge of the production environment. If the use cases that justify the infrastructure spending are actually a narrower slice than the current trajectory implies, the ROI math gets uncomfortable.
The honest answer is that nobody can tell you right now whether the infrastructure build is ahead of its time or ahead of demand. That uncertainty is real and should be treated as such.
What You Should Actually Watch
If you're trying to form a grounded view on this rather than defaulting to either narrative, a few signals are more informative than the headline spending numbers.
Watch trust, not just adoption. The gap between "84% of developers use AI tools" and "29% trust AI-generated code" is the most important metric in this space right now. If that gap closes — trust recovering toward adoption rates — the bull case strengthens materially. If trust keeps falling while adoption grows, you're watching a slow-motion quality problem that will eventually force a correction in how the tools are used, regardless of how much infrastructure exists.
Watch organizational restructuring, not individual productivity. The productivity gains at the individual level are reasonably well-documented. What isn't documented is whether organizations are actually restructuring their processes around those gains — fewer senior engineers reviewing more output, different team compositions, genuinely different software architectures. Without that restructuring, the gains stay at the individual level and the ROI gap persists.
Watch the open-source commons. The health of documentation, forums, and open-source repositories is a leading indicator for developer infrastructure trust broadly. If the signal-to-noise ratio in those shared resources keeps degrading, the tools that depend on them — including AI models trained and fine-tuned on that data — degrade too.
The spending is real. The productivity gains, in specific contexts, are real. The trust deficit is real. The honest position is that all three can be true simultaneously, and that which one dominates the next three years depends on choices that individual teams, open-source maintainers, and tool builders are making right now.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
