Goldman Puts AI Capex Near 4% of US Output by 2027. The Buyers Are Funding It With $302 Billion of New Debt as the Fed Hikes insight cover
Markets / EventNVDA · MSFT · AMZN•27 min read

Goldman Puts AI Capex Near 4% of US Output by 2027. The Buyers Are Funding It With $302 Billion of New Debt as the Fed Hikes

Goldman Sachs Research's November 12 forecast of $1.4 trillion of AI-related hardware and intellectual-property spending — nearly 4% of US output by end-2027 — turns the AI buildout into a macro variable, and the Fed's own September minutes already name surging AI investment as an inflation pressure. The filings show where the stress actually lands: the five largest US hyperscalers spent $691 billion of cash capex in the trailing twelve months against $724 billion of operating cash flow, added roughly $302 billion of net new debt and $57 billion of net new equity, and converted all of it into just $33 billion of combined free cash flow. A 25 basis point hike costs those five about $1.4 billion a year against $607 billion of combined net income, so the hike is not their problem — the marginal project and the credit layer funding 35% of 2027 capex are. Own the asset-light toll booths and the power bottleneck; be careful with the balance sheets funding the last mile.

Published Nov 12, 2026Updated Nov 12, 2026

AI hardware and IP spending, annualized end-2027

$1.4T

Goldman Sachs Research forecast, Nov 12, 2026 — nearly 4% of US output

Net new debt raised by the five largest US hyper

$302B

Aggregated from company cash flow statements, trailing twelve months to Sep 30, 2026

Added growth in 2027 US GDP from AI investment

+0.6pp

Goldman Sachs Research, Nov 12, 2026; +0.2pp excluding semiconductor spending

Fed funds target after the December 9, 2026 meet

4.00–4.25%

Reuters poll, Nov 10, 2026 — all 101 economists expect a 25 basis point hike

AI hardware and IP spending, annualized end-2027

$1.4T

Goldman Sachs Research forecast, Nov 12, 2026 — nearly 4% of US output

Net new debt raised by the five largest US hyperscalers

$302B

Aggregated from company cash flow statements, trailing twelve months to Sep 30, 2026

Added growth in 2027 US GDP from AI investment

+0.6pp

Goldman Sachs Research, Nov 12, 2026; +0.2pp excluding semiconductor spending

Fed funds target after the December 9, 2026 meeting

4.00–4.25%

Reuters poll, Nov 10, 2026 — all 101 economists expect a 25 basis point hike

The forecast

Four Percent of Output, Six Tenths of a Point of Growth

Goldman Sachs Research put AI-related hardware and intellectual-property spending at $1.4 trillion on an annualized basis by the end of 2027 — close to 4% of the US economy — and above $900 billion by the end of 2026. The same forecast adds 5.3 percentage points to 2027 business capital spending growth including semiconductors, or 4.3 points excluding them, and lifts 2027 US GDP growth by 0.6 percentage points, or 0.2 points without chips. Total US business investment growth is forecast at 8% on a quarterly basis in 2027, close to double the 4.2% consensus, with technology accounting for 64% of the gain.

Goldman Sachs Research's 2027 forecast, Nov 12, 2026
MeasureGoldman forecastRead-through
AI hardware + IP spending, end-2026More than $900B annualizedAlready well past 2% of US output before the buildout peaks
AI hardware + IP spending, end-2027$1.4T annualizedClose to 4% of the US economy
2027 capex growth contribution+5.3pp including semis; +4.3pp excludingSemiconductors carry more than half the impulse
2027 GDP growth contribution+0.6pp including semis; +0.2pp excludingA 4%-of-output boom moves growth far less
US business investment growth, 20278% quarterly versus 4.2% consensusNearly double what the market expects
Technology share of the capex gain64%One sector decides the investment cycle

Those two numbers look contradictory and are not. Strip out semiconductors and Goldman's contribution to 2027 growth falls from 0.6 points to 0.2, while the Fed's own median for 2027 growth is 2.4%. The mechanism is simple: the buildout squeezes power, chips and construction capacity, yet it is too small a share of domestic supply to move the aggregate much. Goldman itself frames the risk as financing rather than demand — higher borrowing costs and tougher scrutiny of returns.

The headline is a spending number, not a growth number. Goldman puts AI at nearly 4% of US output by end-2027 and 0.6 points of 2027 GDP growth, against a Fed median of 2.4%. Model the economy off the 4% figure and you will over-forecast growth and under-forecast rates.

Stress test

The Buyers Already Run at a $922 Billion Annualized Pace

Goldman's number is an exit rate, not a calendar-year total, so the cleanest way to test it is against reported spending. In the September 2026 quarter alone, Microsoft, Amazon, Alphabet, Meta and Oracle spent $230.4 billion on property and equipment — an annualized pace of $921.7 billion. Goldman's separate September work put the top five hyperscalers' AI infrastructure capex at $1.2 trillion for 2027 and $1.4 trillion for 2028. Bridging the two requires the five to lift their total capex run-rate roughly 52% from where they exited the third quarter, with the rest of the AI chain adding on top.

Annualized cash capex run-rate, based on the September 2026 quarter

Cash spent on property and equipment in the three months to Sep 30, 2026, multiplied by four, from each company's cash flow statement

Unit: US$ billions, annualized

Amazon

TTM capex $215.2B

309

Alphabet

TTM capex $159.8B

205.4

Microsoft

TTM capex $135.9B

157.5

Meta

TTM capex $104.4B

135.8

Oracle

TTM capex $75.7B

114

Two cautions on the arithmetic. First, the $1.4 trillion and the $1.2 trillion measure different things: the former is hardware and intellectual property bought across the whole chain, including the chips the hyperscalers buy, so the pieces cannot be added. Second, Goldman's own estimates move fast. In October, the firm's asset-management arm put US AI hyperscaler capex at 1.8% of GDP with a plausible range of 2.5% to 3.5%. Three weeks later the research forecast said nearly 4%. The gap between 3.5% and 4% of US output is half a point of GDP, and it is a forecast revision, not new information about the world.

  • Microsoft trimmed its fiscal 2027 capital spending plan to $170 billion from $175 billion, and its CFO said delays from third-party chip suppliers push available capacity into the first half of 2027.
  • Alphabet posted negative free cash flow of more than $9 billion in the September quarter and its shares fell about 10% after hours; its CFO said outside computing power came online only at the very end of the quarter.
  • Lenders are pulling back from neoclouds that lack investment-grade ratings as their cost of capital rises — the layer below the hyperscalers where the financing is thinnest.

The buyers

$691 Billion of Capex, $33 Billion of Free Cash Flow

Here is the part of the AI boom nobody is modelling: the five largest spenders are funding it with other people's money. Across the trailing twelve months they spent $691.0 billion on property and equipment against $723.8 billion of operating cash flow — combined free cash flow of $32.7 billion. They closed the gap with roughly $301.8 billion of net new debt and $57.0 billion of net new equity, while still paying about $76 billion of dividends and buying back stock. Every dollar of AI infrastructure bought in the last year converted into just 4.7 cents of free cash flow.

The five largest US hyperscalers: trailing twelve months to Sep 30, 2026, from company cash flow statements
CompanyCash capex% of revenueCapex / operating cash flowFree cash flowNet new debtNet debt / EBITDA
Microsoft$135.9B39.2%69%$60.0B-$7.9B0.51x
Amazon$215.2B27.0%128%-$46.9B$124.6B0.84x
Alphabet$159.8B33.9%89%$19.7B$96.6B0.27x
Meta$104.4B43.2%78%$28.7B$51.3B0.95x
Oracle$75.7B105%161%-$28.7B$37.2B3.17x
Combined$691.0B—95%$32.7B$301.8B—

The dispersion matters more than the total. Microsoft still runs capex at 69% of operating cash flow, carries net debt at half a turn of EBITDA — unchanged from a year ago — and repaid $7.9 billion of net debt while returning $48 billion to shareholders. Amazon is the mirror image: capital spending now exceeds operating cash flow by 28%, free cash flow is minus $46.9 billion, and total debt including $98.8 billion of finance leases rose $48.7 billion in a single quarter to $271.9 billion, leaving the current ratio slipped below 1.0. Oracle is the most stretched, with capex at 161% of operating cash flow and 3.2 turns of net debt to EBITDA.

  • Amazon has the clearest leverage story: net debt to EBITDA went from 0.40x to 0.84x over a year, and it now pays $3.1 billion a year in interest.
  • Meta nearly doubled its leverage, from 0.46x to 0.95x, while spending 43% of revenue on property and equipment.
  • Alphabet raised $96.6 billion of net debt and $51.0 billion of net equity even with $265 billion of cash and short-term investments on the balance sheet — pre-funding a build it does not need to borrow for yet.
  • Oracle funded capex above 100% of revenue with debt and preferred stock, leaving free cash flow at negative $28.7 billion.
The rate hike is not the problem for these five. A 25 basis point increase on roughly $571 billion of borrowings excluding leases costs about $1.4 billion a year — roughly 0.2% of their combined $607 billion of trailing net income; including the $275 billion of lease obligations, closer to $2.1 billion. What a December hike destroys is the return on the marginal data center, and the lenders who financed it.

The financing

One Third of 2027 Capex Will Be Borrowed Money

Goldman Sachs Research's credit desk has been explicit about how this gets funded: roughly 33% of hyperscaler capex was debt-financed in 2026, and the share peaks at about 35% in 2027, on the logic that internal cash alone cannot carry a build of this size. The same desk counts nearly $500 billion of AI-related debt issuance so far in 2026, of which hyperscalers issued $194 billion, up from $108 billion in all of 2025. AI-related paper now makes up roughly 18% of US investment-grade supply, against 7% in 2025 and 1% in 2024.

Share of 2027 hyperscaler capex debt-financed

~35%

Goldman Sachs Research credit desk, Aug 5, 2026; roughly 33% in 2026

Hyperscaler AI-related debt issued in 2026 to date

$194B

Goldman Sachs Research, Aug 5, 2026; $108B in all of 2025

AI-related share of US investment-grade issuance

~18%

Goldman Sachs Research, Aug 5, 2026; 7% in 2025 and 1% in 2024

Data center project finance expected in 2027

$300B

Goldman Sachs Research, Aug 5, 2026; separate from direct hyperscaler supply

The reason this matters for rates is duration, not volume. Goldman estimates hyperscalers have raised about $400 billion in the US investment-grade market — roughly 20% of gross issuance — with maturities three to four years longer than the broad investment-grade index, which pushes upward pressure on the long end of the curve. The squeeze is already visible: the AI leadership basket has widened from 74 basis points at its tights to nearly twice that, and insurance clients placing $50 million-plus orders on long-end AI paper fell from 15 in the first quarter to about half that number by the end of the second. Inside the same stack, 40% of all 15-year-plus investment-grade issuance in 2026 came from AI companies or companies funding the AI theme.

The sub-investment-grade layer is where a December hike actually bites. Lenders have started pulling back from neoclouds that lack investment-grade ratings as their cost of capital rises. These are the companies turning $1.4 trillion of intended spending into data centers, and they have no balance sheet to absorb a higher discount rate.

The policy layer

The Fed Named AI Capex in Its Own Minutes

This is not a variable the Fed is ignoring. The minutes of the September 15–16, 2026 meeting, released November 3, record that all participants supported raising the target range by a quarter point to 3.75%–4.00%, and that staff and participants cited surging AI-related investment among the pressures on inflation, alongside energy prices. The staff put total PCE inflation at 3.8% and core at 3.4% on a twelve-month basis in August. Participants also cited business investment as support for real GDP growth outpacing potential through 2028.

What the Fed projected on Sept 16, 2026, and what November brought
Fed projection, Sept 16, 2026November reality
Policy rate, end-20264.1% median3.75%–4.00% now; all 101 economists in the Reuters poll expect 4.00%–4.25% on Dec 9
PCE inflation, 20263.7% medianOctober headline CPI rose 0.4% on the month
Core inflation3.4% core PCE median for 2026Core CPI rose 0.3% in October, a third straight 0.3%
Real GDP, 20272.4% medianGoldman attributes 0.6 points of that to AI investment
Longer-run funds rate3.2% medianThe 2027 median is also 4.1% — no easing assumed

Two wrinkles make December 9 harder than the projections imply. The October CPI release cannot print a year-over-year figure at all: the Bureau of Labor Statistics did not collect October 2025 data during the appropriations lapse, so the Fed judges a fourth straight 0.3% core print with no twelve-month anchor. Second, Governor Waller has put a half-point increase on the table if inflation keeps accelerating, making a jumbo hike an acknowledged possibility rather than a tail risk. An independent read lands in the same place: ING estimates AI-related investment accounts for about 37% of US real GDP growth in 2027.

The awkward part of December is that the central bank is expected to raise rates into a spending boom its own minutes identify as a source of price pressure, with the boom financed by debt priced off the very yields the decision sets — raising rates into the boom it calls an inflation pressure.

The chain

The Money Moves Downstream, the Debt Moves Upstream

The chain is not symmetric, and the asymmetry is the trade. NVIDIA is the toll booth: trailing revenue of $303.0 billion, a 74.7% gross margin, $127.0 billion of free cash flow and just $7.4 billion of its own capital spending — 2.4% of revenue. It monetizes $691 billion of customer capex with almost no capital of its own, carries net debt at 0.07 times EBITDA, and trades on 21.9 times trailing earnings. Broadcom grew revenue 85.5% year on year to $29.6 billion in its September quarter, converts 68.8% of revenue to gross profit, and has taken net debt to EBITDA from 1.41 turns down to 0.68. TSMC grew third-quarter revenue 42.8% to NT$1,493 billion, is spending NT$1,654 billion a year on capacity, and still runs a net cash balance sheet.

The memory layer is the most profitable and the most contested. Micron earned an 80.7% gross margin on $133.2 billion of trailing revenue — 86.7% in its September quarter alone — and trades at about 10.9 times trailing earnings with a 6.4% free cash flow yield. SK hynix reported quarterly operating profit of 83.3 trillion won, more than seven times the year-ago 11.4 trillion won, on 104.6 trillion won of revenue, with a 74.5% trailing operating margin and net cash. Its shares, though, sit at 1,254,000 won against an all-time high of 2,987,000 won set on June 24, 2026, down 58% while profits hit a record.

Where the AI spending lands: trailing twelve months to Q3 2026
LayerCompanyRevenue (TTM)MarginCapex intensityBalance sheet
AcceleratorsNVIDIA$303.0B74.7% gross2.4% of revenueNet debt 0.07x EBITDA
Custom silicon / networkingBroadcom$89.1B68.8% gross1.4% of revenueNet debt 0.68x EBITDA
FoundryTSMCNT$4,944B58.0% operating34.6% of revenueNet cash
MemoryMicron$133.2B80.7% gross23.1% of revenueNet cash
MemorySK hynix₩269.3T74.5% operating13.4% of revenueNet cash
Power and coolingVertiv$13.1B38.1% gross4.0% of revenueNet cash
Electrical infrastructureEaton$32.0B17.3% operating3.5% of revenue$9.2B net new debt

Then there is the layer that sells the bottleneck. Vertiv supplies power and thermal management for data centers: September-quarter revenue of $4.31 billion, up from $3.27 billion in June, trailing revenue of $13.1 billion at a 38.1% gross margin and $3.1 billion of free cash flow, with net cash. Eaton, which absorbed $11.1 billion of acquisitions in its March quarter, generated $32.0 billion of trailing revenue and $3.8 billion of free cash flow. These two are paid on installed megawatts, not on tokens, and the constraint investors keep hearing about is power itself: Goldman Sachs' October discussion of the AI trade described a power spread of about 20 at 40 to 50 gigawatts for next year, while Microsoft's CFO said capacity slips into the first half of 2027.

  • The return test: hyperscalers need roughly $308 billion of annual AI revenue just to break even on 2026–27 average capex, and $417 billion to $526 billion for a 10% to 20% return, per Goldman Sachs Research.
  • NVIDIA and Broadcom get paid before the revenue arrives; Amazon and Oracle pay for it before it does.
  • SK hynix is committed to 1,100 trillion won of memory capacity through three clusters — 600 trillion at Yongin, 400 trillion in the southwest, 100 trillion at Cheongju — with the fourth Yongin fab pulled forward to 2033 from 2045.

The market's answer

Equities Have Started Pricing the End. Credit Has Not

Goldman Sachs' own October review of why investors were trimming AI exposure is the most useful counterweight to its capex forecast: the trade had collapsed into one position — long AI, short everything else — and was unwinding as the economy moved from a savings glut to a spending glut, pushing the cost of capital up almost everywhere. Investors were asking whether backlog converts into revenue, whether free cash flow break-even arrives before credit markets get uncomfortable, and whether power can actually be delivered.

What the market pays for the boom: trailing multiples on the Nov 11, 2026 close, calculated from reported results
CompanyPrice / earningsEV / free cash flowFree cash flow yieldNet debt / EBITDA
NVIDIA21.9x33.1x3.0%0.07x
Microsoft27.8x66.4x1.5%0.51x
Broadcom42.4x42.7x2.4%0.68x
TSMC25.7x38.6x2.5%Net cash
Micron10.9x15.0x6.4%Net cash
SK hynix4.7x10.0x10.0%Net cash
Vertiv36.3x25.4x3.9%Net cash
Oracle20.1xNegative-7.4%3.17x

The pattern is legible. The layers that sell picks and shovels trade at 22 to 42 times earnings; the layers that own the physical build and the debt trade at 5 to 11 times; and the one name whose free cash flow is most stretched, Oracle, trades on no free cash flow at all. SK hynix is 58% below its June 2026 peak and Micron is about 35% below its 52-week high of $1,255 — both in record profit quarters. The equity market has already priced a cycle that does not last, and credit markets, with AI paper around 18% of investment-grade supply, have not.

Goldman's forecast and the market's pricing are not in conflict. If AI capex really reaches $1.4 trillion and the $308 billion of annual AI revenue does not follow, the assets bought with $302 billion of new debt areworth less than the debt. That is the scenario equities are discounting today.

Horizons

What to Watch, and When

  • Days to quarters — December 9, 2026: the FOMC is expected to lift the target range to 4.00%–4.25%, and a half-point increase stays live if core inflation keeps printing0.3% monthly.
  • January and February 2027: fourth-quarter results are the first real test of 2027 capital budgets, and the first guidance cut — Microsoft to $170 billion — has already landed.
  • Credit supply: roughly $250 billion of direct hyperscaler issuance plus about $300 billion of data center project finance is expected in 2027, against AI paper already at 18% of the investment-grade market.
  • One to three years — the revenue test: roughly $308 billion of annual AI revenue breaks even on 2026–27 average capex, and $417 billion to $526 billion buys a 10% to 20% return.
  • One to three years — the growth arithmetic: with AI capex at 4% of output, the Fed's median for 2027 growth is still 2.4% and Goldman's AI contribution is 0.6 points, which is why a hiking Fed can coexist with the boom.
  • One to three years — the memory build: SK hynix has committed 1,100 trillion won of capacity through 2033 and TSMC is spending NT$1,654 billion a year; the supply being designed today is the oversupply that would end the cycle.

The verdict

The Risk Is a Discount Rate, Not a Demand Shock

Goldman is right about the scale and imprecise about the worry. At $1.4 trillion of annualized AI spending by end-2027, the buildout is a macro variable: large enough that the Fed's own minutes cite it as an inflation pressure, and that a quarter-point hike lands on the same balance sheets funding it. But the filings say the hyperscalers absorb the hike easily, because the risk is a discount rate, not a demand shock. What cannot absorb it is the marginal project funded with borrowed money, the neocloud without an investment-grade rating, and the investor underwriting the long end of a curve that AI issuance has already made longer.

So the trade is a position on where the financing sits, not on whether AI spending continues. Near term, own the toll booths — NVIDIA, Broadcom, TSMC — that monetise the boom without funding it, and the bottleneck suppliers selling megawatts and thermal capacity. Over one to three years, watch memory: SK hynix and Micron earn extraordinary margins at multiples that assume the cycle ends, and the 1,100 trillion won of Korean capacity now being designed is the mechanism that would end it. Own the toll booths, not the borrowers.

Bottom line: the AI buildout is not a growth story the Fed has to rescue, and it is not a bubble the Fed has to puncture. It is a debt-financed spending boom that turns December into a decision about the long end of the curve — andthe burden falls on the marginal project, not on the megacaps.

The investable map

NNVIDIANVDA--
--Vol --
-
Bullish
  • Through 2027, capex of $7.4 billion on $303.0 billion of trailing revenue lets it monetize $691 billion of customer capex at 2.4% capital intensity.
  • A 74.7% trailing gross margin and net debt at 0.07x EBITDA mean a December rate hike does not touch its cost of capital.
  • Near term, 21.9x trailing earnings with a 3.0% free cash flow yield prices in neither a cycle peak nor a bust.
MMicrosoftMSFT--
--Vol --
-
Bullish
  • Through 2027, trailing capex of $135.9 billion is covered 1.4 times by operating cash flow, and net debt to EBITDA has not moved from 0.51x in a year.
  • Free cash flow of $60.0 billion still funds $48 billion of dividends and buybacks, so the December hike is immaterial to the plan.
  • Risk: a 1.5% free cash flow yield and 66.4x EV to free cash flow leave no cushion if Azure monetization slips into 2028.
AAmazonAMZN--
--Vol --
-
Bearish
  • Capex now exceeds operating cash flow by 28%, with free cash flow at minus $46.9 billion and net debt to EBITDA at 0.84x.
  • Total debt jumped $48.7 billion in one quarter to $271.9 billion while shareholder returns went to zero — no buybacks, no dividend.
  • A 25 basis point hike adds roughly $300 million a year in interest on new debt, marginal against earnings but decisive for the last data center.
GAlphabetGOOGL--
--Vol --
-
Mixed
  • Into 2027, $265.1 billion of cash and short-term investments can fund the build, yet the company raised $96.6 billion of net debt and $51.0 billion of equity anyway.
  • September-quarter free cash flow of minus $9.1 billion and a 10% after-hours fall showed the market will not fund negative cash flow indefinitely.
  • The swing factor is whether $195 billion to $205 billion of 2026 capex converts into revenue before the December move lifts the discount rate.
ABroadcomAVGO--
--Vol --
-
Bullish
  • Into fiscal 2027, revenue grew 85.5% year on year to $29.6 billion in the September quarter, with capital intensity of just 1.4%.
  • Net debt to EBITDA fell from 1.41 turns to 0.68 while $39.4 billion of free cash flow funded $20.6 billion of dividends and buybacks.
  • Risk: 42.7x EV to free cash flow prices in the custom-silicon win and leaves nothing for a second disappointment.
TTSMCTSM--
--Vol --
-
Bullish
  • Third-quarter revenue rose 42.8% to NT$1,493 billion on NT$442 billion of quarterly capex, the physical response to $1.4 trillion of intended spending.
  • Through 2028, a net cash balance sheet funds capacity expansion without touching the credit markets.
  • Near term, 25.7x trailing earnings with a 2.5% free cash flow yield still leaves room for another capex step-up.
0SK hynix000660.KS--
--Vol --
-
Mixed
  • Record third-quarter operating profit of 83.3 trillion won and a 74.5% trailing operating margin come with 4.7 times trailing earnings.
  • Shares sit 58% below the June 24, 2026 peak of 2,987,000 won, so the cycle is being priced as if it ends next year.
  • The 1,100 trillion won capacity commitment running to 2033 is the mechanism that would deliver the oversupply the market is discounting.
VVertivVRT--
--Vol --
-
Bullish
  • September-quarter revenue of $4.31 billion rose from $3.27 billion in June, as the power and cooling constraint binds.
  • Over the next 12 to 18 months, net cash and $3.1 billion of trailing free cash flow mean it is paid for the build without financing it.
  • Risk: 36.3x trailing earnings leaves the multiple exposed if permitting or power delivery slips the 2027 order book.