Goldman Sachs puts a $7.6 trillion figure on the AI buildout — and the number doing the rounds on FX desks
A bank research note quantifying the AI buildout at $7.6 trillion through 2031 circulated on 7 July 2026. Two days earlier, the same desk told clients the yen is heading to 165. The two calls together are doing more work than either alone.
On 7 July 2026, a Goldman Sachs research note circulated across the Telegram channels that startup operators, venture platform staff and product teams actually read — AngelList and Product Hunt first, then the usual forwarding tree. The number on the slide: AI capital spending of $765 billion in 2026, rising to roughly $1.6 trillion a year by 2031, totalling $7.6 trillion across the five-year window. The slide package did not arrive alone. On 5 July 2026 — two days earlier, per an Unusual Whales post — a separate Goldman FX desk forecast told clients the Japanese yen would weaken to 165 against the US dollar.
Read together, the two notes sketch a single posture: the bank is telling its clients that the AI buildout is large enough, and durable enough, that the dollar's financing role remains the spine of the cycle — even at the cost of a weaker yen and a steeper bill for every economy whose debt is denominated in anything but greenbacks.
What the $7.6 trillion actually is
The $7.6 trillion figure is a sum of parts, not a single capex line. Goldman's framework, as it appeared in the Telegram-circulated table, breaks the spending into three buckets: hyperscaler infrastructure (the cloud platforms building the data centres), enterprise IT (chips, networking, on-prem compute), and the power and grid buildout required to keep the first two running. The annual run-rate roughly doubles between 2026 and 2031, which implies the curve does not flatten — a forecast that depends on continued returns from generative AI products large enough to justify the spend.
That last condition is the one the note does not stress but that the wider market keeps stress-testing. If model margins compress, if inference becomes commoditised faster than training scales, or if regulators in any one jurisdiction slow the data-centre pipeline, the second-half numbers soften. The forecast's shape is therefore best read as an upper-bound commitment rather than a baseline — the amount that will be spent if the AI cycle continues to look like the AI cycle has looked so far.
Why the yen call matters more than usual
A weaker-yen forecast from a US bank is not normally a story. It is normally a tactical FX view, the kind that gets a couple of paragraphs on a rates desk and a footnote in a wider dollar-strength piece. The 5 July note is doing more work than that.
The Bank of Japan has spent the better part of two years edging away from yield-curve control and from the era of suppressed yen volatility. The currency's slide toward 165 against the dollar is the flip side of that normalisation: policy divergence is doing what divergence usually does, and a weaker yen is the cost of letting domestic rates rise more slowly than US rates. The Goldman call is, in effect, telling clients not to fight that divergence.
For an AI buildout denominated in dollars and priced in dollar-funded capex, a weaker yen has a specific structural value: it lowers the dollar cost of Japanese capital goods — lithography equipment from ASML's Canon and Nikon suppliers, semiconductor materials, precision components — at exactly the moment the supply chain is being asked to absorb the largest sustained order book in its history. The weaker the yen, the cheaper Japan's contribution to the global compute stack looks on a US-invoice basis. That is not an accident. It is the trade that the FX desk and the capex desk are both implicitly recommending.
The structural read
The $7.6 trillion capex figure and the 165 yen call, placed side by side, describe a specific kind of cycle. It is a cycle in which the AI buildout is funded overwhelmingly in dollars; in which the dollar's strength is itself part of the cost-curve that makes the buildout affordable; and in which the burden of adjustment falls on the currencies and economies that sit downstream of dollar funding.
That is not a new arrangement. It is the arrangement that has governed US tech capex since at least the smartphone era, and arguably since the personal-computer buildout of the 1990s. What is new is the scale, and the fact that a single research desk is now putting a five-year, multi-trillion-dollar figure on it — and tying it, in the same week, to a specific FX target.
What we verified / what we could not
Verified against the source items. The $765 billion 2026 figure, the $1.6 trillion 2031 run-rate, and the $7.6 trillion cumulative total all appear in the Telegram-circulated summary attributed to Goldman Sachs research. The 165 yen forecast is independently confirmed by the Unusual Whales post of 6 July 2026.
Not in the source items. The underlying Goldman research PDFs, the named lead economists on each note, the methodology sections, the FX desk's specific stop-loss and time-horizon assumptions, and any client-tier distribution data. The Telegram and X posts give the headlines and the directional calls; they do not give the working papers. A reader wanting to audit the model would need to reach the notes through a Goldman client login or through the bank's published research portal.
What the sources do not specify. Whether the $7.6 trillion is gross or net of depreciation and refresh cycles; whether the FX call assumes a particular Bank of Japan path; and whether the two notes were coordinated across desks or produced independently. Both notes are consistent with a unified view; the source material does not prove that view was deliberate.
Stakes
If the forecast is roughly right, the next five years of AI capex will be larger than the entire annual GDP of every country on earth except the United States and China. The financing of that capex — the bond issuance, the equity raises, the dollar funding lines that make Japanese and Korean component supply possible — becomes a macroeconomic event in its own right. A weaker yen is one small, visible piece of that.
If the forecast is wrong, the second-half numbers will soften first in the grid buildout, then in enterprise IT, and only last in hyperscaler infrastructure — because hyperscalers can self-finance in ways that smaller buyers cannot. The pattern of any disappointment would therefore be uneven, and would land hardest on the suppliers furthest down the stack.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://t.me/s/AngelList
- https://t.me/s/producthunt
- https://x.com/unusual_whales/status/1941478