Two real, well-sourced figures appeared on the same podcast in the same breath — and the compression that made that possible is exactly what makes large AI numbers so hard to reason about.
On the MAD Podcast, host Matt Turck asked Sachin Katti, OpenAI’s head of industrial compute, to confirm two figures in quick succession. The first — that OpenAI was planning to spend roughly $50 billion on computing in 2026 — was sourced as well as a number can be: OpenAI president Greg Brockman stated it as the company’s projected compute expenditure in sworn testimony during the Musk litigation, as reported by Bloomberg. Katti confirmed it with a qualifier: “that sounds pretty accurate.” The second — that “the entire industry” was also going to invest $700 billion in computing expenses — received a single word: “Yes.”
Nothing about that exchange was dishonest. A podcast answer is not a regulatory filing, and neither Katti nor Turck said anything that could fairly be called bad faith. But the two figures are not the same kind of quantity, and placing them adjacently — with the same year on both and dollars on both — makes them look as though they sit on a common axis. They do not. The $50 billion is what OpenAI pays for compute. The second figure traces to a Futurum Group estimate that Microsoft, Alphabet, Amazon, Meta, and Oracle have committed to between $660 billion and $690 billion of capital expenditure in 2026. That is the construction budget of five named sellers, not computing expenses across a whole industry.
The Futurum estimate is a well-behaved number: its components are named and its arithmetic is checkable. Amazon at roughly $200 billion, Alphabet at $175–185 billion, Meta at $115–135 billion, Microsoft tracking toward $120 billion or more (a floor, not a midpoint), Oracle at around $50 billion. Take midpoints where ranges were given and they sum to approximately $675 billion — comfortably inside the stated $660–690 billion band. The problem was never the estimate. It was the label the estimate travelled under once it entered the conversation.
The key insight: Two substitutions happened inside a single sentence. Capital expenditure became “computing expenses.” Five named companies became “the entire industry.” Neither is dishonest — that is exactly what spoken compression does. But the category widens far more easily than it narrows: five firms become an industry in four words, and getting the qualifier back takes a paragraph most readers never reach.
Five companies building, not an industry spending. The distinction is the whole piece.
The Structural Read
Placing a purchase and a construction budget on the same axis produces a number that looks like a market share and is not one. Spend and capital expenditure are not comparable quantities even when both are denominated in dollars and cover the same year — one measures what a buyer pays, the other measures what sellers build. That is a general property of financial comparison, not a complaint about this conversation.
Here the gap is more specific than that. The two figures are not merely different in kind — they are structurally coupled. A large part of OpenAI’s $50 billion flows as revenue to some of the very companies whose capital expenditure constitutes the larger figure. The buyer’s spend and the sellers’ build move together by construction rather than varying independently the way a share and its base would. This publication does not compute any ratio, share, or percentage between the two figures. The tempting division is explicitly refused: the result would describe nothing at all, because the denominator already contains the numerator as one of its inputs.
Map of AI — Structural Observation
“The well-sourced figure was the one confirmed with a qualifier. The looser one received a single word of assent. That is not a criticism of anyone. It is how conversation works — and it is why the label a number travels under outlives the care anyone took with it.”
There is also a smaller point about what a confirmation is worth when the interviewee did not supply the number. “I read somewhere” introduced both figures; an executive’s agreement sent them out into the world carrying the same apparent authority. In this case the $50 billion is about as solid as a number gets — sworn testimony — so the confirmation added nothing and cost nothing. But the general property holds regardless: agreeing with a figure somebody else read somewhere is a different speech act from disclosing it. The substitution is directionally reasonable — those five companies are an enormous share of the AI infrastructure build — which is exactly what makes it durable. A sentence that feels right even though it is imprecise about what it counts is harder to correct than one that is obviously wrong.
Three Implications
CREDIT THE ESTIMATE, AUDIT THE LABEL
The Futurum Group figure is better-behaved than most large AI numbers precisely because its components are named and its parts sum to its whole. Dismissing it because of how it was framed in a podcast would be a poor reading. The right move is to restore the correct label — five named companies’ capex commitments — and use the number for what it actually measures.
CATEGORY WIDENING IS ASYMMETRIC AND FAST
Five firms become an industry in four words. Getting the qualifier back afterwards takes a paragraph that most readers never reach. Every analyst, investor, and operator working from secondary sources on AI infrastructure spending should treat scope as the first question — not the last. What does this number count, exactly? The answer changes the figure’s usefulness entirely.
COUPLED FIGURES RESIST RATIO ANALYSIS
When a buyer’s spend feeds directly into sellers’ revenue — and those same sellers’ capex constitutes the larger figure — the two numbers cannot be placed in a ratio without circular reasoning. The structural relationship between OpenAI’s compute spend and the hyperscalers’ infrastructure build is not a market-share story. It is a dependency story, and it reads differently once the coupling is visible.
Business Engineer Framework
The Map of AI: Where Each Layer Sits in the Stack
The Map of AI tracks 200+ companies across nine layers of the AI stack — from silicon and cloud infrastructure at the foundation through models, tooling, and applications at the surface. This week’s numbers live at layers one and two: the infrastructure capex of hyperscalers and the compute spend of foundation model providers. Understanding which layer a figure belongs to is the prerequisite for knowing what it can and cannot be compared against. The label problem this piece describes is, at its root, a layer-confusion problem.
Two accurate numbers, both worth knowing, travelled through a single spoken exchange and arrived on the other side sharing a label neither was built to carry — and because the compression was directionally reasonable, it stuck. The $50 billion is what OpenAI pays for compute, sourced from sworn testimony. The ~$675 billion midpoint is what five named hyperscalers have committed to build, sourced from a named research firm whose arithmetic checks out. Use both. Compare neither. The moment you place them on a shared axis, the number you produce describes the relationship you assumed, not the one that exists.
The $50 billion is Greg Brockman’s sworn testimony during the Musk litigation, as reported by Bloomberg, and it describes OpenAI’s compute spend — what the company pays for compute.The figure of roughly $700 billion is, as reported, a Futurum Group estimate that five named companies — Microsoft, Alphabet, Amazon, Meta and Oracle — have committed to between $660 billion and $690 billion of capital expenditure in 2026. It is not industry-wide computing expenses. Component figures are midpoints where ranges were given, and the Microsoft figure is a floor rather than a point estimate. No ratio, share or percentage between the two figures is computed or stated anywhere above. They are opposite sides of the same transaction — a customer’s purchase and suppliers’ construction budgets — so dividing one by the other would describe nothing. Nothing above imputes bad faith to Sachin Katti, to Matt Turck or to anyone else, and nothing above alleges misleading, spin or inflation. A podcast answer is not a filing, and the compressions described are the ordinary work of spoken conversation. OpenAI’s revenue, losses, margins and cash position, the share of its compute supplied by any named company, any contract value, any company’s spend to date against its commitment, 2027 figures and whether any commitment will be met are not established and do not appear. Nothing above predicts capital expenditure, a bubble, a correction or whether spending continues, and nothing above is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
