# The AI Bull Case Quietly Assumes Intelligence Gets Cheap > Source: > Published: 2026-07-06 18:15:38+00:00 The AI buildout is the largest infrastructure bet in history. If you’re a technologist, the most important thing to understand about it is where the money won’t end up. There is a number the industry keeps repeating without quite absorbing it. In 2026, the four largest hyperscalers—Alphabet, Amazon, Meta, and Microsoft—are on track to [spend a combined ~$725 billion in capital expenditures](https://valueaddvc.com/blog/ai-hyperscaler-capex-compared-why-microsoft-google-meta-and-amazon-are-all-spending-at-once), up roughly 77% from a 2025 figure that was itself a record. McKinsey puts [cumulative data-center investment at nearly $6.7 trillion by 2030](https://www.mckinsey.com/featured-insights/future-of-asia/countries-and-regions/southeast-asia/southeast-asia-perspectives/beyond-the-spillover-asia-pacific-the-next-engine-of-data-center-demand). The IEA expects [data-center electricity demand to roughly double, to around 945 terawatt-hours](https://www.datacenterdynamics.com/en/news/iea-data-center-energy-consumption-set-to-double-by-2030-to-945twh/)—more than Japan’s entire annual consumption—by the same year. Those aren’t technology numbers. They’re infrastructure numbers, the kind the economy produces only a few times a century: the railroads, the telephone and internet networks, the interstate highways, the electrification of America. And there’s a pattern those buildouts share that almost nobody bakes into the current AI narrative. When a foundational technology arrives, the durable fortunes rarely go to the companies that own the core infrastructure. They go to the people who build on top of it once it becomes cheap, boring, and ubiquitous. If that’s right, it changes what a technologist should be building right now. That argument—the full version, with the history and the receipts—is the subject of a new Techstrong white paper, *The Dynamo and the Token*. This is the short version, and the one claim inside it worth arguing about over coffee. ### The Optimist Has a Chart, and It Might Be Right Daniel Newman, CEO of The Futurum Group, recently posted the cleanest possible statement of the bull case: a chart of the combined free cash flow of the four giants from 2006 out to 2030 (S&P Capital IQ historicals, Bloomberg consensus estimates). It climbs for years, then gets crushed in 2026–27 as the capex wave hits the income statement—Amazon’s line actually dips negative—and then, in the estimate years, goes vertical. Roughly $210 billion in 2028. More than $630 billion by 2030. Newman’s read: this is simply how investment cycles work. Underinvestment is a bigger risk than overinvestment. The returns go parabolic. “Strap in,” he writes. He may be right. The easy move is to wave the chart away as hype, and it isn’t hype—it’s a serious claim backed by serious money, and his own numbers actually run hotter than consensus (more than $10 trillion in AI infrastructure by 2030). Any argument worth making has to take the parabola seriously. But a graph of cash flows captured *during* a buildout tells you nothing about returns retained *after* the thing it produces becomes a commodity. Those are two different questions. The entire argument sits in the gap between them. ### Why Intelligence Commoditizes—And Why It Isn’t Optional Here’s the part that should interest anyone who actually ships with these models. The frontier race looks permanently unwinnable by anyone: each release tops the last, the gaps feel meaningful, surely whoever owns the smartest model owns the future. That’s a misread, and the reason isn’t a bet on model architectures. It’s an economic constraint the whole business case quietly depends on. AI displaces human labor only where it’s cheaper than the human doing the task. That’s the entire proposition—the reason a CFO signs the check. And right now, for a large and growing share of real workloads, the math is uncomfortably tight: once you count the tokens burned on reasoning, retries, context, and orchestration, running the work through a model can cost *more* than paying a person. The whole industry conversation about [“tokenmaxing”—wringing more useful output from fewer tokens](https://techstrong.ai/videos/backstory-tackles-enterprise-ai-token-economics/)—is evidence that the cost ceiling is real and already shaping how systems get built. The implication: if token costs don’t fall substantially, adoption stalls at exactly the margin where the value was supposed to come from. So there’s relentless, structural downward pressure on the price of intelligence—not because the labs are generous, but because [their customers’ ROI case collapses without it](https://techstrong.ai/articles/corporate-finance-chiefs-grapple-with-unpredictable-token-bills-as-ai-costs-skyrocket-survey/). Price compression isn’t a risk that *might* hit the model providers. It’s a precondition the whole edifice needs in order to pay off. Now look back at the parabola. That vertical line assumes mass adoption. Mass adoption assumes intelligence gets cheap enough to clear the human-cost bar. The bull case and the commoditization case aren’t opposed—the parabola *depends* on the very [price collapse that strips pricing power away from raw intelligence](https://techstrong.ai/articles/openai-considers-aggressive-price-cuts-to-rival-anthropic-ahead-of-historic-ipos-report/). The optimist’s own math carries you to the commoditizer’s conclusion. There’s a second force, one every engineer routing between providers already feels: substitutability. Different brands of gasoline make elaborate claims about additives, and most drivers just pull into whatever station is cheap and close. Basic intelligence is heading the same way. When the cost of switching providers approaches zero and the quality gap on routine work is marginal, the product is a commodity no matter how miraculous it looked three years ago—and increasingly the switching isn’t even a human decision. The emerging default architecture [routes each request to the cheapest model that passes eval](https://therouter.ai/news/datadog-state-ai-engineering-2026-multi-model-routing-production/). When indifference is engineered into the stack as a design principle, commoditization stops being a behavior you observe and becomes a structural fact. One honest qualification, because it matters. A token is not a watt. Electrons are perfectly fungible; model outputs differ in quality, voice, and reliability, and for some high-value work—senior-engineer hours, large-scale legal review—AI is already far cheaper than the human alternative, so the ceiling doesn’t bind there at all. Intelligence won’t commoditize into a *pure* commodity like electricity. It commoditizes into a *differentiated* one—closer to gasoline with additives, or specialty coffee: thin margins on the base product, a brand and switching-cost premium at the edges, and genuinely specialized models for specialized tasks. That’s not a hole in the thesis. It *is* the thesis, stated precisely. The base layer—good-enough general reasoning, the bulk of the volume—commoditizes. The premium frontier holds a thinner and thinner slice. The fortunes follow the volume. ### Electrons, Bits, Tokens: Has This Pattern Happened Before? This isn’t the first time a general-purpose technology has come down to a fight over a single unit of account. Electrification gave the economy the electron, billed as a kilowatt-hour. Computing gave it the bit, distributed as a packet. Generative AI is giving it the token, billed by the million. These aren’t metaphors for each other—they’re a lineage, and each ran the same play: a heroic decade of capital formation, a wave of triumphalist coverage, a quieter decade of price collapse, and then a long, profitable second act lived almost entirely on the layer above. Bits are the example every technologist lived through. The 1990s telecom buildout laid millions of miles of fiber, lit it, and watched the long-haul price of a bit fall through the floor. The companies that built the pipes—Global Crossing, WorldCom, the regional Bells—mostly didn’t survive the decade after the buildout in recognizable form. The trillion dollars of enterprise value created *on top of* cheap bits went to an entirely different set of names: Google, Amazon, Netflix, and the whole tail of SaaS built on the assumption that bandwidth was, for their purposes, free. The white paper traces the electron version in detail—including the wrinkle that the biggest early winner, General Electric, won precisely because it owned the grid layer *and* the appliances on top of it. That detail turns out to matter a lot for reading who wins in AI. ### So Who’s the Utility, and Who’s GE? Put the chart and the history together and the picture gets more useful than “infrastructure owners lose.” None of the four companies in Newman’s parabola is a pure grid operator. Each owns the intelligence infrastructure *and* a dominant application layer—Search and YouTube, Office and Copilot, Instagram’s ad engine, AWS. They’re not the regulated utilities of this story. They’re GE: dynamos and appliances under one roof. If that vertical cash-flow line materializes, it almost certainly arrives *through* the application layer—ad optimization, cloud margin, Copilot seats, [agentic products](https://techstrong.ai/agentic-ai/databricks-launches-genie-ai-agents-cost-control-tools-to-fight-runaway-corporate-tech-bills/)—not through selling raw tokens at a premium. Read that way, the bull chart quietly *supports* the up-the-stack thesis instead of refuting it. So who plays the regulated utility—the essential, indispensable, margin-compressed grid operator? The pure-play infrastructure sellers. The neoclouds. The standalone data-center operators. [The GPU-as-a-service platforms renting out undifferentiated compute](https://techstrong.ai/videos/crusoe-optimizes-ai-inference-beyond-hyperscalers/). They own the wires without owning the appliances, which is exactly why they’re the most exposed to commodity economics. They aren’t in Newman’s chart. That absence is the whole story. ### What Should You Build Now? The practical question for a technologist stops being “which model is winning this month” and becomes “what am I building that still has a moat when the model underneath me is a swappable commodity.” The white paper makes the full case; the short answer is that defensible value migrates to a few specific places—proprietary data and the customer relationship, workflows rebuilt around abundant intelligence rather than a model bolted onto an unchanged process, [orchestration that assumes models are interchangeable](https://platformengineering.com/features/your-newest-platform-user-is-an-ai-agent-build-it-a-golden-path/), and the memory and context layers that persist across whichever model is cheapest this quarter. The load-bearing word there is *robust*. You don’t have to buy any particular timeline for the price collapse. Whether it fully arrives in eighteen months or four years, the same bet pays off: build for the world where intelligence is cheap, and don’t stake your edge on privileged access to a model that won’t stay privileged. The people who get rich in a gold rush are the ones selling picks and shovels—but [that fortune fades once the commodity sets in](https://techstrong.ai/articles/new-data-suggests-ai-investments-are-starting-to-pay-off-though-risks-persist/). A disclosure, since it’s the point. Techstrong runs on the very vendors this thesis says are most exposed to margin compression—the cloud, infrastructure, and AI-tooling companies selling shovels. This isn’t an argument I’d make if I were talking my book. I’m making it because I think it’s true, and because our audience—practitioners, builders—happens to be standing on the layer where the value is actually heading. Most haven’t clocked it yet, because the entire conversation is fixated on the layer below them. The capex spectacle will continue. The parabola may even arrive. Strap in, by all means—just be clear about which seat you’re in. *The full argument — including the electrification history, the four charts, the regulation question, and the case for where durable returns actually land — is in the Techstrong white paper “The Dynamo and the Token.” It’s the first in a series on infrastructure cycles and where value migrates. *