The conventional read on today’s AI investment surge is simple: corporations and governments are pouring trillions into a technology that isn’t delivering commensurate returns, making it the greatest capital misallocation in modern history. But that framing conflates a real problem — undisciplined AI spending — with a false conclusion: that AI itself is the con.
The Thesis
I believe the “AI is a money pit” narrative is one of the most strategically dangerous half-truths circulating in boardrooms right now. Not because it’s wrong about the waste — it isn’t — but because it’s wrong about the cause. The waste isn’t a property of AI. It’s a property of how most organizations have chosen to deploy it: without clear outcome metrics, without integration into core workflows, and with a reflexive assumption that buying the most expensive model is the same as building a competitive advantage.
That distinction matters enormously. Executives who internalize “AI wastes money” will pull back investment across the board. Executives who internalize “undisciplined AI spending wastes money” will get precise — and precision is exactly what this moment demands. The difference between those two responses, compounded over the next three to five years, is the difference between leading a category and watching someone else define it.
Supporting Argument: The Spending Numbers Are Real, and So Is the Chaos
Let’s not minimize the critique. Enterprise AI spending has ballooned at a pace that has genuinely outrun measurable output. Major cloud providers are reporting AI-related capital expenditure in the hundreds of billions annually, and yet productivity surveys across industries routinely show that most employees either don’t use AI tools provided to them, or use them in ways that don’t touch revenue-generating work. Pilots proliferate; production deployments stall. That’s not a technology problem — it’s a governance problem dressed up as a technology problem.
The pattern is familiar. Early internet adoption saw companies spend fortunes on websites that did nothing and intranets nobody used. The spending looked idiotic in 2001. It looks rational in retrospect because the organisations that spent purposefully — not abundantly — survived to compound the advantage. The companies that cut all digital investment after the dot-com bust spent a decade catching up.
The same dynamic is playing out with AI infrastructure today. The data center build-out driving a blue-collar jobs wave isn’t speculative excess — it’s load-bearing infrastructure for a computing paradigm shift. The question isn’t whether to build it; it’s whether the applications running on top of it are generating value. Most aren’t yet. That gap is real. But closing it requires discipline, not retreat.
Supporting Argument: The Overlooked Angle — Waste Is Concentrated, Not Universal
Here’s what the “AI is a waste” argument consistently misses: the waste is not evenly distributed. It is almost entirely concentrated in two categories — vanity deployments (chatbots and copilots that don’t connect to actual business logic) and infrastructure over-provisioning (buying GPU clusters to run experiments that could run on smaller hardware).
Meanwhile, a quieter cohort of organizations is generating measurable, compounding returns. They tend to share three traits: they started with a specific, measurable problem rather than a mandate to “use AI”; they integrated AI outputs into decision loops rather than treating AI as a standalone product; and they invested in data infrastructure before they invested in model access. These organizations aren’t making headlines. Their ROI doesn’t fit the “AI hype cycle” narrative that gets clicks.
There’s a structural irony here worth naming: the organizations generating the loudest AI announcements are often the same ones generating the least verifiable return, while the organizations quietly compounding AI advantages are too busy shipping to issue press releases. This creates a selection bias in the public debate — the visible AI spend is disproportionately the wasteful kind, leading analysts to generalize a governance failure into a technology indictment. It’s like judging the ROI of cloud computing exclusively from the failed SaaS startups of 2011.
This also explains why some of the sharpest AI capability research — like the architectural choices behind Anthropic’s Claude Opus 4 and its bet on honesty and coding utility — gets less boardroom attention than it deserves. Executives chasing the flashiest model are missing the point. The question isn’t which frontier model you’re subscribed to; it’s whether the model’s output is wired into a workflow that generates a decision or removes a cost.
Supporting Argument: The Infrastructure Bet Has Historical Precedent
Critics of AI spending often point to the mismatch between investment and near-term revenue. They’re right that the gap is wide. What they underweight is that transformative infrastructure almost always looks like waste before it looks like value.
The electrification of American industry in the early twentieth century took roughly three decades to show up in aggregate productivity statistics. Factories that wired up early didn’t see immediate gains; they had to redesign workflows around the new capability before the numbers moved. Economists who measured electricity investment against 1905 output would have called it a bubble. By 1930, it had restructured the entire industrial economy.
AI is earlier in that curve than most commentary acknowledges. The compute build-out — which critics cite as evidence of excess — is, in my view, analogous to building the electrical grid. It’s a prerequisite, not a product. The concentration of that infrastructure in companies like Nvidia creates real competitive concerns, but it does not make the infrastructure itself wasteful. Grids are not bubbles. The applications running on them can be, and some are.
What should also give pause to blanket pessimists is the direction of hardware innovation. Research into light-powered valleytronics chips and other post-silicon architectures suggests the compute efficiency curve is far from flattened. If inference costs drop by another order of magnitude over the next five years — a reasonable expectation given the current R&D trajectory — today’s “wasteful” infrastructure investment looks dramatically cheaper in hindsight.
The Strongest Counterargument
The most intellectually honest version of the AI-as-waste argument doesn’t claim the technology is useless. It claims that the pace of spending has decoupled from the pace of verifiable value creation in a way that structurally resembles prior asset bubbles. The argument, made seriously by economists and technology skeptics alike, goes like this: when capital allocation is driven by fear of missing out rather than modelled return, you get mispriced assets. Firms are spending on AI not because they’ve calculated a positive NPV, but because their board demands an AI strategy and their competitors are announcing one. That competitive anxiety, not commercial logic, is the primary driver of current spend.
This is a genuinely strong objection, and I don’t want to dismiss it. FOMO-driven capital allocation is real. The majority of enterprise AI pilots that never reach production are evidence that many of these decisions aren’t rigorous. If you’re a CFO signing off on AI spend with no defined success metric, the skeptics are talking to you.
But the counterargument proves too much. Fear of missing out also drove early enterprise internet adoption, early cloud migration, and early mobile investment. In each case, the FOMO was partially irrational and the spending was partially wasteful — and in each case the underlying technology restructured the competitive landscape anyway. The companies that used FOMO as a reason to sit out those transitions didn’t avoid the waste; they just deferred it and paid a larger catch-up premium later. Calling the investment a bubble because the decision process is messy is a category error. Bubbles burst and leave nothing behind. Infrastructure transitions are messy and leave transformed industries behind.
There is, however, one area where I think the skeptics land a cleaner hit: the reliability and trust gap in AI outputs is a genuine brake on deployment velocity, and no amount of capital spending closes it. Until enterprises can trust AI outputs in high-stakes decisions without expensive human review layers, the ROI ceiling is real. That’s a research problem, not a spending problem — and it’s one that more money alone cannot solve.
Why It Still Holds
Even accepting that the process is messy and some spending is pure FOMO, the core thesis holds: AI is not a money-wasting scheme. It is a capability platform that most organizations are currently using badly. The distinction is actionable. “AI is a waste” leads to divestment. “We’re deploying AI badly” leads to a governance overhaul, tighter criteria for AI projects, investment in data quality, and a focus on workflow integration over model selection.
Organizations that internalize the right lesson will emerge from the current hype trough with real capability advantages. Those that internalize the wrong lesson will re-enter the market in three years as laggards — paying higher prices for talent, compute, and competitive catch-up. The AI disruption already crushing pre-ChatGPT startups is a preview of what happens to incumbents who treat skepticism as a strategy.
I also think there’s a subtle safety dimension here that gets lost in the ROI debate. The push for AI safety mechanisms and governance frameworks isn’t anti-progress — it’s what makes progress trustworthy enough to actually deploy at scale. Treating safety investment as part of the “waste” column is a false economy. It’s the cost of making AI usable in consequential domains.
The Prediction
Within 24 months, I expect a visible split to emerge between two classes of enterprise: those that imposed rigorous ROI gates on AI projects in 2024–2025 and are now compounding targeted advantages, and those that either spent indiscriminately or cut entirely and are scrambling to catch up. The companies in the first group will not be the ones with the biggest AI budgets — they’ll be the ones with the smallest ratio of AI pilots to AI deployments. If that split fails to materialize and AI adoption remains broadly diffuse with no clear winners by late 2026, I’ll concede the skeptics had the more accurate model. But I don’t think it will.
