$800 Billion: The World's Most Famous Wrong Number · Issue #181

Opening

Reader, big tech earnings season wrapped up last week. Alphabet raised its 2026 capex guidance to $195–205 billion, Amazon put out a $220 billion figure, and Meta lifted the low end of its range. And every time one of these announcements lands, the same tally shows up in every article: “The five hyperscalers¹ will spend $800 billion this year.” Nearly every debate about AI investment right now is being fought on top of this single number.

Then, on August 2, Goldman Sachs published a note that challenges that number head-on. The bottom line: $800 billion undercounts global AI investment by roughly $200 billion, while simultaneously overcounting US domestic investment by roughly $200 billion. One number is wrong in both directions at once. Today I want to look at how this wrong number got built, how the verdict changes once you swap the ruler, and where in the world you can check the real answer first.

Why Everyone Cites the Same Number

Let’s start with what the $800 billion actually is. The precise figure is $794 billion, the consensus forecast for 2026 capex from five publicly traded US hyperscalers—Google, Amazon, Microsoft, Meta, and Oracle (based on FactSet aggregation). Looking at the trajectory, it’s easy to see why everyone is fixated on this number. Combined capex for the five companies was $71 billion in 2019, climbed to $154 billion by 2023, hit $412 billion in 2025, and is now forecast to reach $794 billion this year. That’s a fivefold increase in three years.

The reason this number became the standard is simple: it’s easy to count. You only have to track five companies, each figure gets verified quarterly through earnings reports, and analyst estimates are densely layered on top. When you overlay it with a tally TrendForce released on August 3, the reliability of this number actually holds up quite well. TrendForce estimated 2026 capex for the nine largest global cloud providers at $886.7 billion (roughly 90% growth year-over-year), and put the North American five’s share at about 90% of that—which works out to roughly $798 billion. That’s essentially the same as Goldman’s $794 billion consensus. Count the same five companies, and whoever does the counting arrives at the same number.

The problem starts after that. At some point, this number began being used as the answer to two entirely different questions: “What is total global AI investment?” and “What is US AI investment?” But it’s the answer to neither.

How One Number Ends Up Wrong in Both Directions

Let’s follow Goldman’s dissection. First, the overcounting side—the error that appears when you read this number as “US AI investment.”

First, this $800 billion contains spending that has nothing to do with AI. Even in 2022, before the AI boom took off in earnest, capex from the five companies already stood at $158 billion—money that was already being spent on cloud servers and logistics centers. To see the actual “increment” AI created, you have to subtract this baseline first. Second, not all of this money lands on US soil. Tracking the locations of announced projects shows that only about 70% of US hyperscaler capex is actually deployed within the United States, followed by 15% in Asia and 9% in Europe.

Next is the undercounting side—the error when you read this as “global AI investment.” According to Goldman’s credit team analysis, hyperscalers directly account for only about 40% of 2026 AI-related supply. That means the remaining 60% sits entirely outside this tally. Private companies like OpenAI (roughly $102 billion, Goldman’s estimate), publicly traded non-hyperscalers like CoreWeave (roughly $34 billion) and SpaceX (roughly $31 billion), and non-US firms like Samsung Electronics (roughly $52 billion), SK Hynix (roughly $32 billion), Tencent, and Alibaba are all missing entirely. This is also where a definitional question surfaces—whether memory fab expansions should even count as AI investment—and Goldman decided they should.

Once you adjust for all of this, here’s what you get: roughly $1,019 billion in global AI investment for 2026, of which $581 billion is domestic US investment. Compared against the commonly used $794 billion ruler, the global figure falls short by about $200 billion while the US figure runs about $200 billion over. For scale, that $200 billion is roughly twice South Korea’s entire July export total ($98.9 billion).

Of course, this adjusted figure is itself an estimate. So Goldman cross-checks it with two entirely different methods. One works backward from how much AI-exposed public companies’ 2026 gross-profit forecasts have been revised upward since Q3 2022, which yields $1,060 billion. There’s a reason for using gross profit rather than revenue: money paid to memory and foundry makers gets captured again in chip designers’ revenue, so summing revenue figures double-counts the same dollars. The other method traces US national accounts’ commodity-flow statistics to see how much equipment was supplied domestically, which produces a global figure of $1,002 billion. All three methods converge on roughly the same answer: global investment near $1 trillion, with the US figure sitting slightly under $600 billion. Now that the wrong answer is confirmed as wrong, the order of magnitude of the right one looks fairly solid too.

Change the Ruler, and the Verdict Changes

The reason this correction isn’t just bookkeeping housekeeping is that both sides of the AI bubble debate are using this same number. Optimists say, “They’re spending $800 billion, and demand is holding up.” Pessimists say, “They’re spending $800 billion, and there’s no sign of payback.” Both sides were grading different answers off the same wrong answer key.

Change the ruler, and the verdict on scale changes first. When Goldman weighs its adjusted figure against GDP, US AI investment comes to 1.8% of GDP in 2026, and extrapolating public-company consensus, it climbs to 2.5% in 2027 and 2.8% in 2028. That looks large, but Goldman’s comparison point is that peak investment during the rollout of past general-purpose technologies² like railroads and electricity ranged from 2% to 5% of GDP. So the scale itself falls within a historically normal range. The intuition that “they’re spending too much” doesn’t hold up well yet—at least not when GDP is the denominator.

Instead, a warning light is flashing somewhere else. According to official US statistics, 8% of this year’s nominal investment growth is not real investment at all—it’s price inflation. As component prices, memory chief among them, have risen, the amount of computing you get per dollar has started to shrink. The signal of a bubble tends to leak out first not from “spending too much money” but from “the money spent converting into less real output.”

There’s also an answer buried in this note to a question from earlier issues: the puzzle covered in Why Does the New Fed Chair Talk Like a Startup Founder?—“Why doesn’t $1 trillion in spending show up in the economic data?” Goldman’s explanation is accounting-based. US national accounts don’t classify corporate semiconductor purchases as investment goods, and because AI hardware carries a heavy import component, it gets offset out of GDP calculations. So this $1 trillion is passing through an accounting system that never fully registers it in growth figures. It’s less that the spending is invisible, and more that it’s being counted in a way that makes it invisible.

The Thermometer Sits at Korean Customs

There’s one question left: when does this trillion-dollar cycle turn, and where can you find out first?

Goldman’s answer is interesting. Its judgment tool is a dashboard, and the leading indicators loaded onto it are Taiwan and South Korea’s semiconductor manufacturing equipment imports, electronics-sector new orders and backlogs, export prices, memory prices, and the hourly rental rate for Nvidia GPUs. It’s East Asian trade data, not US data, that fills most of the dashboard. The reason is physical: chips and equipment cross borders before data centers get built, and Taiwan’s and Korea’s trade statistics come out one to two months ahead of official US figures.

ruler The dashboard’s latest reading came in last Saturday, August 1: the Ministry of Trade, Industry and Energy’s July trade figures. Semiconductor exports hit $41 billion, up 178.8% year-over-year, topping $40 billion for a second straight month and accounting for 41.5% of total exports. Exports to the US also rose 68.7% on the back of AI data-center demand, and semiconductor equipment imports climbed in tandem. In the language of Goldman’s dashboard, these indicators are sitting at the top of their range since 2022.

But within the same data, Goldman’s nowcast³ picks up a subtler thread: a signal that AI-related investment in June and July, while still moving at a very solid pace, is showing some deceleration. A red-hot headline growth rate and a subtly cooling pace are two different layers of the story, and right now, the export figures also carry a price effect from rising DRAM fixed prices. The same structure behind the earlier warning—“8% of nominal growth is just price”—turns out to be operating inside Korea’s export statistics too. To be clear, this is a directional signal, not a final figure. Confirmation comes from US statistics one or two months later; the signal itself gets published first, on the first of every month, by Korea.

What this means for Korean readers is clear. The thermometer the world uses to measure the temperature of the trillion-dollar cycle is cargo passing through our own customs. In A World of Paper Wealth, I wrote that Korea holds the real-world bottleneck of memory chips. This note adds one more thing to that: Korea doesn’t just hold the physical bottleneck—it holds the signal too.

Oz’s Lens

When I was doing GTM strategy consulting, I often had to estimate market sizes for new businesses, and one scene kept repeating. A client would lay three research reports side by side—same market, but the numbers differed by two or three times. The math wasn’t wrong. What differed was the definition: what got included, what got left out. That’s how I picked up the habit of reading the footnotes before reading the number itself. Whoever controls the definition controls the conclusion.

Seen through that experience, the real problem with $800 billion isn’t that it’s wrong—it’s that it’s easy. It’s a number you get just by adding up five companies’ disclosures, so everyone grabbed it, and the number that was easy to count crowded out the number that actually needed counting. Measurability beat importance. This correction won’t immediately change the conclusion of the bubble debate. But the unit of the debate needs to change—away from the earnings story of five companies, toward the story of an entire supply chain that stretches across memory fabs, power grids, and private labs. And the moment you swap the ruler to an industry-wide unit, Korea stops being a spectator in this debate and becomes part of what’s being counted. Samsung Electronics’ and SK Hynix’s investments are already listed on the global AI capex ledger.

One last note: this piece isn’t a basis for making investment decisions about any specific asset or stock. It’s about interrogating how numbers get defined, so please make any investment decisions after checking the primary sources and disclosures yourself.

Closing

Here’s the summary.

First, the hyperscaler $800 billion figure is a wrong answer in both directions—$200 billion too low for the global figure, $200 billion too high for the US figure. The corrected numbers are $1,019 billion globally and $581 billion for the US.

Second, changing the ruler changes the verdict. Scale relative to GDP falls within the normal range seen during past general-purpose-technology buildouts, but the new warning light is that 8% of nominal growth is just price inflation.

Third, the data most likely to reveal the inflection point first is Taiwan’s and Korea’s trade statistics. The semiconductor and equipment line items in the trade figures released on the first of every month make up the world’s fastest dashboard, so I’d encourage you to check them yourself on the first of next month.

Does your industry have a number like this too, Reader? A “standard figure” that everyone cites, but that looks strange the moment you dig into its source and definition? Tell me in the comments what the number is and where it falls apart. If enough examples come in, I’ll dedicate a follow-up issue to dissecting industry-standard numbers.

💬 Tell me about the “famous wrong number” in your industry, Reader—I’ll factor it into a follow-up issue. 📨 If you have a colleague who follows AI investment news closely, please pass this along.

📎 References & Further Reading

Primary sources

Joseph Briggs & Sarah Dong, “Assessing the Current Pace of AI Investment,” Goldman Sachs Global Investment Research, August 2, 2026. ··· This research note is the backbone of today’s issue. It’s an institutional subscription resource, so I’m leaving the citation instead of a link.
TrendForce, “CSP资本支出预计增长90%，2026年AI服务器出货量增幅上调至近31%” [“CSP Capex Expected to Grow 90%, 2026 AI Server Shipment Growth Raised to Nearly 31%”], August 3, 2026. ··· This is the source of the $886.7 billion tally for the nine major CSPs. I used it to cross-check against Goldman’s consensus figure.
Ministry of Trade, Industry and Energy, “July 2026 Trade Trends,” August 1, 2026. / Financial News, “July Exports Hit $98.9 Billion, Second-Highest Ever… Semiconductors Top $40 Billion for Second Straight Month,” August 1, 2026. ··· This is the source for the $41 billion semiconductor export figure and the rise in equipment imports cited in the “thermometer” section.
Digital Daily, “The Q2 US Big Tech ‘AI Report Card’ Is In… The Numbers Global Markets Are Watching,” July 31, 2026. ··· This is where the Alphabet, Amazon, and Meta capex-guidance increases mentioned in the Opening are laid out.

Background

Newspim, “Big Tech Q2 Earnings Kick Off… The Challenge of Justifying AI Capex,” July 20, 2026. ··· Shows how this number evolved over time, including the four companies’ $725 billion guidance and the 2027 consensus of $900 billion.

Related past issues

Why Does the New Fed Chair Talk Like a Startup Founder? ··· Covers “why doesn’t it show up in the data” from a narrative angle—it pairs well with today’s accounting-based answer.
A World of Paper Wealth, and San Francisco’s Barter Economy ··· The story of Korea’s real-world bottleneck. Pairs well with today’s “Korea holds the signal” argument.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Hyperscaler: A cloud provider like Google, Amazon, or Microsoft that builds and operates massive data centers directly. Because they operate at such enormous scale, their capex is often used as a proxy indicator for the entire AI infrastructure market. ↩
General Purpose Technology (GPT): A technology like electricity or railroads that reshapes the way an entire economy produces things, not just one industry. These technologies share a pattern of decades-long, large-scale infrastructure investment during their rollout phase. ↩
Nowcast: A technique for estimating the current state of something in real time using faster, alternative data before official statistics are released. It’s less a forecast than a live weather report of right now. ↩