The Maps Have Stopped Tracking the Territory
Two narratives dominate coverage of AI right now, and both are wrong in ways that are easy to verify. The first says we’re in a bubble — that the data center buildout is a tulip craze with better PR, that Microsoft canceling a lease or two portends the whole thing collapsing, that GPUs are worthless scrap metal in 24 months. The second says that rigorous academic work shows AI doesn’t really work — that studies demonstrate it slows developers down, that serious economists have calculated its impact will be negligible, that the hype is overblown. Both narratives carry institutional weight. Both dissolve on contact with the facts.
Part One: The Buildout
Historical analogies
The “bubble” framing reaches instinctively for tulips, and it’s the wrong grab every time. Tulip bulbs produced no cashflow and left behind no infrastructure. The comparison is rhetorical, not analytical.
The right analogy is the 19th-century railroad mania. Capex peaked at 5–9% of US GDP in the 1880s. Multiple crashes, in 1873 and 1893. Many operators went bankrupt. The tracks stayed laid. The physical network still forms the spine of US freight transport 150 years later. That’s the shape of bubbles in durable infrastructure: the equity gets crushed, the operators get reshuffled, and the concrete remains.
The more recent rhyme is the dot-com cycle, which makes the same point louder. The 2000–2002 crash destroyed trillions in market value and bankrupted hundreds of pure-play internet companies. It did not unplug the internet. Fiber stayed in the ground. Bandwidth kept growing. Amazon, Google, eBay, and Salesforce all survived the crash. Tesla, Facebook, Gmail, YouTube, Twitter, Airbnb, and Uber were all founded during or immediately after the supposed recession. “Bubble” is a statement about near-term equity prices. It is not a statement about whether the technology or the infrastructure persists. The distinction matters because the bear case, even when correct, doesn’t imply what people think it implies.
The capex reality
The scale is the part that should end the conversation. In 2026 alone, the Big Five hyperscalers are on track to spend over $600 billion on infrastructure. The 14 largest public data-center operators are tracking to $750 billion. Goldman projects $1.15 trillion in cumulative hyperscaler capex from 2025 through 2027. McKinsey’s ten-year number is $5–7 trillion. The Marshall Plan, adjusted for inflation, was ~$170 billion over four years, 5.2% of 1948 US GDP, publicly funded via grants. The AI buildout is running at similar GDP share — annually — privately funded. 2026’s single-year capex exceeds the full-program inflation-adjusted totals of the Manhattan Project, Apollo, and the ISS combined. This is straightforwardly the largest private-capital megaproject in recorded history.
US data center construction spending ran at roughly $9–10 billion per month from 2020 through early 2022, per the Census Bureau. By December 2025 it hit $45.1 billion per month — an 85% two-year jump. Over 23 gigawatts of capacity was under construction globally at the end of Q3 2025.
The “worthless GPU” claim dissolves on the same kind of examination. Hyperscalers depreciate servers over 5–6 years; Azure K80s ran nine years in production; Google’s 7–8 year old TPUs still hit 100% utilization; CoreWeave’s A100s from 2020 are fully booked in 2026. H100 rental contracts are being renewed at the same rates they were signed at two to three years ago, some going out through 2028. SemiAnalysis reports that capacity coming online through August–September 2026 is already fully booked. Meanwhile the tax treatment — Section 179 expensing, bonus depreciation, state-level property tax abatements stretching 20–30 years — means the hardware is being written off aggressively against income while generating revenue at 30–40% gross margins, then cascading down to inference workloads (projected to be ~80% of AI compute by 2030) for the back half of its operational life.
Even the cancellation narrative, correctly framed, supports the opposite conclusion. If a meaningful fraction of the announced US data-center pipeline is being delayed or canceled, that is evidence of a pipeline so massive that, even after historic attrition, what remains still constitutes the largest capital-formation wave ever. Cancellations are the pipeline hitting power, permitting, and supply-chain constraints. They are not the pipeline failing to exist.
Part Two: The Academic Failure
Structural problems
Peer review takes 9–18 months across most disciplines, longer in economics. Data collection usually predates submission by another year. What this means in practice: the “authoritative” academic literature on AI in 2026 is measuring models from 2023.
Daron Acemoglu’s “The Simple Macroeconomics of AI” is the clearest case. Published as an NBER working paper in May 2024, then in Economic Policy in January 2025, it concluded AI would produce at most a 0.53–0.66% TFP gain over a decade. Its three central empirical inputs were published in early-to-mid 2023 and based on GPT-4 and ChatGPT-3.5. The exposure measure — literally GPT-4, rating whether GPT-4-class systems could do various jobs — was frozen to March 2023 capabilities. No reasoning models. No tool use. No agents. No computer use. No persistent context.
By the time the paper was shaping public discourse, OpenAI’s o1 had introduced reasoning, Claude had shipped computer use, and Claude Code had launched the agentic coding paradigm. OpenAI’s own GDPval benchmark, released September 2025, showed Claude Opus 4.1 matching or beating human industry experts on 47.6% of real economic deliverables — up from GPT-4o’s 13.7% fifteen months earlier. The paper everyone cites as the sober economist’s take is a fossil.
The METR “19% slowdown” study is the coding analog. Sixteen experienced open-source developers, working on mature codebases they already knew cold, using Cursor with Claude 3.5/3.7 Sonnet — a tool most of them weren’t fluent in. It is precisely the scenario where AI help is weakest, and the authors said so. Their February 2026 follow-up noted the developers with biggest AI speedups were self-selecting out of the study because they refused to work without AI. The paper itself is honest. The citation chain that turned it into “AI makes developers slower” is not.
Brand maintenance
The structural problems — publication lag, cheap models, normalcy bias, arm’s-length study design — would be bad enough on their own. What makes this actively corrosive is the amplifier class. Cal Newport is a Georgetown CS professor who could install Claude Code in ten minutes and still writes about AI coding tools as “helper utilities” for “tedious boilerplate.” Gary Marcus has spent seven straight years predicting scaling would plateau, and every year moves the goalposts when it doesn’t. Acemoglu publicly dismisses ChatGPT passing the bar as having “nothing to do with reality.”
These people are not failing at rigor. They are succeeding at something else: maintaining a brand. Newport’s entire intellectual franchise — deep work, slow productivity, the pre-digital contemplative life — would be materially damaged if AI turned out to be a large productivity multiplier for knowledge workers. So he consistently describes the technology as it existed around ChatGPT’s launch and cites whichever study confirms the brand. Marcus has made himself the go-to quote for “AI skepticism” in legacy media; updating would destroy the positioning. The skepticism is the product.
Why this combines badly
The two failures compound. The structural lag guarantees that peer-reviewed findings are always describing a previous capability regime. The amplifier class then launders those findings through platforms that strip the methodological caveats and elevate the flawed premises into universal claims. A narrow result from a 2023 GPT-4 snapshot becomes “economists have calculated that AI will barely affect the economy.” A 16-person study of experts on familiar codebases becomes “rigorous research shows AI slows developers.” The credentialing system that’s supposed to filter signal from noise is, on this topic, amplifying noise and calling it signal.
The result is an epistemic landscape that bears little resemblance to the territory it claims to map. Meanwhile, the concrete is getting poured. The GPUs are getting installed. The power substations are coming online. The capability curve is tripling in fifteen months on real economic work. The map and the territory have divorced, and the people still drawing maps aren’t visiting the territory. That’s not honest disagreement about an uncertain future. That’s misinformation produced by the prestige class, with all the trappings of rigor, and it does a disservice to everyone who trusts institutions to keep their maps current.
