AI ROI Myth: Big Tech's Growth Gambit

Big Tech’s AI bills keep rising even as they promise to cut costs. That’s the contradiction RBC Wealth Management’s piece stakes out: after a multi‑year sprint of investment, firms are finally pointing at scalable returns. Frankly, the claim deserves close scrutiny — “scalable” is a business outcome, not a marketing cue, and the move from impressive models to dependable margins is where the real proof lives.

The article’s core thesis is reasonable: the AI push is entering a phase where spending is meant to produce repeatable revenue. Yes, Big Tech is past the science‑project stage. But the piece glosses over how uneven that journey is from lab to line item. The middle is messy, and that’s where a lot of hoped‑for “scale” quietly dies.

Start with the most obvious gap: research-grade models and dazzling demos don’t equal customer-ready products. Integration headaches, reliability, latency, security, and UX all matter; customers buy outcomes, not benchmarks. You can have a model that looks great on a slide and still fail to clear a CIO’s procurement checklist because it doesn’t plug neatly into existing systems or governance frameworks.

Monetization isn’t a single on-ramp, either. Advertising, subscriptions, cloud services, vertical enterprise offerings — each has its own pricing psychology and sales cycle. An AI feature embedded in productivity software will be monetized very differently from a usage-based API or a niche industry tool. The original piece nods to scalability but underplays how differently these paths behave when budgets tighten or compliance teams push back.

Then there’s the human layer. Sales and support for B2B AI are anything but lightweight. Enterprise buyers expect customization, SLAs, uptime guarantees, audits, and clear accountability when the model misfires. That pushes the cost structure closer to a services business wearing an AI hat than a pure software margin story. The dream of “build once, sell endlessly” clashes with the reality of bespoke deployments and hand‑holding.

From my Goldman days, I learned to separate narrative from runway: investors tolerate big stories, but they pay for recurring receipts. The RBC piece is right that firms are talking more about scale, yet it underplays the operational drag — billing systems, data pipelines, model retraining, monitoring, and customer success. The math doesn’t lie: model accuracy is table stakes; the grind is in everything around it.

Where the article treats infrastructure as background, it should feel like a co-author of the balance sheet. AI at scale is an infrastructure-native business: data centers, specialized chips, networking, storage, cooling, and the software to orchestrate it all. The companies that control more of this stack get better unit economics and better negotiating power. Everyone else gets dependency risk — both technical and commercial.

You can already see the outlines. Cloud incumbents and chip suppliers shape pricing, capacity, and performance ceilings. Software firms might showcase the user experience, but a huge part of the profit pool is determined by who controls the compute economics underneath. Let’s be real: user growth means less if someone else taxes every incremental inference.

History isn’t subtle here. During the early internet buildout, telecoms and backbone providers quietly dictated terms while web companies fought over ad dollars. In smartphones, operating system owners and app store operators captured outsized economics, while many app developers competed for scraps. AI is tracking a similar pattern: control the rails, and you shape the returns.

RBC’s piece does mention risk, including regulation, but treats it more like a background worry than a design constraint. That’s generous. Privacy rules, data residency, security mandates, and antitrust scrutiny don’t just add compliance cost; they shape what products can exist, where they can run, and how data can be combined. For AI, those constraints hit right at the source material.

Regulation also interacts badly with the dream of uniform global scale. Data residency can force regional instances. Sector rules can demand separate environments for healthcare, finance, or public-sector workloads. Each carve‑out chips away at the neat story of one AI platform serving everyone with identical economics.

A common pushback — and the RBC piece hints at this — is that Big Tech’s sheer scale will steamroll these issues. They have distribution, capital, and data. They’ll out‑execute and amortize the costs. True, up to a point. Distribution and data are real moats.

But scale cuts both ways. Large installed bases create expectations for backward compatibility and slow product turnover. You can’t just rip out old workflows without breaking trust or contracts. Those same incumbents are also the ones regulators watch most closely, and that attention translates into friction, delay, and constraints on how aggressively they can move.

There’s a piece missing in the framing: this isn’t just a company-level story, it’s an industrial one. AI will reshuffle who earns what across chipmakers, data‑center operators, cloud hosts, enterprise integrators, and consumer platforms. Gross margin is going to be reallocated, not magically expanded for everyone claiming an “AI strategy.”

A practical example: when a software vendor adds AI features, they may see higher prices per seat, but their cloud bill and support load rise too. Meanwhile, the cloud provider benefits from more compute demand, and the chip supplier benefits from higher‑value hardware sales. The same “AI feature” drives three P&Ls in different ways — only one of them may actually see attractive incremental margin.

So here’s the blunt point the RBC article circles but doesn’t quite land: AI scaling is as much a capital-allocation problem as it is a model-improvement exercise. Big Tech will be rewarded not for building the flashiest models, but for owning the right layers of the stack where AI spend quietly hardens into recurring, defensible cash flow.