The Information Machine

This story was merged into Microsoft, Meta Use Neocloud Deals to Obscure AI Infrastructure Cost →

MergedRan 31 daysin the edition since 7 Oct 20264 sourcesupdated 6 Nov 2026

AI infrastructure capex and the revenue gap

The gist

Latest: Hyperscaler Capex Nears $1.9T; Cloud Revenue Trails $300B Breakeven

The gap between what hyperscalers are spending and what they need in AI revenue to break even sets the central question for the investment cycle. Goldman Sachs calculates that reaching the $300 billion breakeven would imply roughly $1 trillion in annual end-user AI spending economy-wide, a threshold the industry has not yet demonstrated.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

Combined capital expenditure estimates for major hyperscalers across 2026-2027 stand at nearly $1.9 trillion, roughly 66% above where they were at the start of 2026, an increase of approximately $750 billion. Goldman Sachs calculates that hyperscalers need to generate roughly $300 billion in annual AI revenues to break even on those investments; as of Q2 2026, cloud revenue growth had accelerated to 48% year-over-year and was annualizing about $70 billion above pre-AI trends, still well short of that threshold. Goldman Sachs and Visible Alpha project the eight-company group reaches $1.75 trillion in annual capex by 2030.

To finance the buildout, every major hyperscaler has simultaneously tapped capital markets. Meta raised $15 billion in bonds that drew $58 billion in orders, with 30-year notes priced roughly 250 basis points over Treasuries; S&P cut Meta's credit outlook to negative. Meta's total equity program runs roughly $97.5 billion across a $25 billion Class A common offering, $12.5 billion in mandatory convertible preferred, and a planned $60 billion at-the-market program, with demand for equity tranches running roughly five times available shares and the 15 largest investors taking about 70%. Nvidia committed $3 billion and AMD $1 billion, making Meta's chip suppliers co-investors in their largest customer; Meta also holds a warrant for up to 160 million AMD shares. Meta now ranks third among Big Tech bond issuers since 2025 at roughly $70 billion total, behind Alphabet and Amazon. Meta's Q3 EPS of $3.83 missed the $6.80 analyst consensus by roughly 44% even as revenue of $64.6 billion beat the $63.4 billion estimate, and the company guided 2026 capex at $135-145 billion.

Microsoft issued at least $25 billion in investment-grade bonds, its first US deal since 2017, with quarterly capex at $45.2 billion and the next quarter guided above $50 billion. Oracle filed a $10 billion equity offering and a $10 billion bond offering on the same day, fulfilling the full $20 billion outside-financing ceiling management pledged at its investor day. Oracle carried approximately $236.1 billion in consolidated debt as of August 31 and S&P rates it BBB-minus, one notch above junk, with leverage expected above 4x for two years. Oracle's remaining performance obligations grew from $138 billion to $638 billion in a single year, and the company plans to scale compute capacity from roughly 1 GW in FY25 to 10 GW by FY28, projecting revenue exceeding $225 billion by FY2030. Big Tech has collectively issued nearly $430 billion in investment-grade bonds since last year.

Amazon CEO Andy Jassy said AWS capacity for 2026 is fully booked, 2027 is largely sold out, and a significant portion of 2028 has been committed. Memory, photonics, and power are among the biggest bottlenecks. Amazon's trailing free cash flow swung from a $14.8 billion inflow to a $42.4 billion outflow over 12 months as quarterly equipment spending hit $77.3 billion, while the company maintained its roughly $220 billion 2026 capex plan and guided 2027 spending above 2026 levels.

On the enterprise demand side, a Menlo Ventures report found that the average company spent 4.7 times its planned 2025 AI budget against an anticipated 1.9 times increase, with the median company overshooting more modestly at 2.2 times versus a planned 1.8 times; a small group of heavy AI agent users drove most of the surge. Announced revenue backlogs across hyperscalers exceed $1.5 trillion. Some analysts viewed the synchronized borrowing across all major hyperscalers as a bullish signal about AI's commercial potential.

How it developed
5 November 2026

Goldman Sachs and Visible Alpha projected combined capex across eight major tech companies growing from $0.46 trillion in 2025 to $1.75 trillion by 2030, extending a buildout whose 2026-2027 estimate for the five largest hyperscalers had grown 66% to nearly $1.9 trillion in under nine months.

Meta disclosed a deal to buy up to $60 billion of AMD chips over five years, with Helios rack systems under testing, putting Meta's projected AI spend at more than $740 billion over three years. AMD CEO Lisa Su projected data-center chips at 80% of 2027 revenue.

First citedRohan Paul on X
4 November 2026in the 6 Nov edition

Goldman Sachs tracked the revenue side of the hyperscaler capex buildout, finding cloud revenue growth at 48% year-over-year by Q2 2026 and revenues annualizing about $70 billion above pre-AI trends, against a roughly $300 billion annual breakeven Goldman estimates is needed to justify the investment, a gap the firm says implies roughly $1 trillion in end-user AI spending economy-wide.

A Menlo Ventures report found enterprises spent 4.7 times their planned 2025 AI budgets on average against an anticipated 1.9 times increase.

3 November 2026in the 4 Nov edition

SK Hynix reported an 80% operating margin on quarterly revenue of 104.6 trillion won tied to AI memory chip demand, and AMD posted a record $8.55 billion in Q3 data-center chip revenue, up 97% year-over-year.

In bond markets, Meta raised $15 billion drawing $58 billion in orders as S&P cut its credit outlook to negative, and Microsoft raised at least $25 billion in the same week, bringing collective Big Tech investment-grade bond issuance to nearly $430 billion since last year.

2 November 2026in the 3 Nov edition

Bloomberg's tally of five hyperscalers' future data center commitments crossed $3 trillion, with roughly 70% of $1.3 trillion in leases added in the last four quarters.

Q3 earnings from Google, Amazon, and Microsoft confirmed cloud revenues growing while free cash flow turned deeply negative: Google Cloud's deficit widened past $9 billion and Amazon's swung $57 billion over 12 months. Oracle on November 2 filed a $10 billion equity and $10 billion bond offering simultaneously, completing pledged outside financing, with S&P rating it BBB-minus and leverage above 4x for two years.

1 November 2026in the 2 Nov edition

SpaceX disclosed a $40 billion binding commitment to lease Nvidia GPUs through an Apollo-managed financing vehicle, which SpaceX describes as the largest committed GPU financing, exceeding a $36 billion notes deal used to lease GPUs to Anthropic, with leases beginning as GPUs arrive in the first half of 2027.

OpenAI separately announced plans to rebrand its data centers as 'AI compute campuses' and offer community compacts proposing state legislation to give residents input on construction and benefits, launching during the lame-duck session after the midterm elections, as data center construction has drawn opposition from both parties.

30 October 2026in the 31 Oct edition

Bloomberg reported October 30 that five hyperscalers hold over $3 trillion in forward data center commitments, with Alphabet alone exceeding $1 trillion, nearly eight times its level a year earlier, and Meta at $810 billion.

Sundar Pichai said Google Cloud is foregoing short-term revenues to prioritize frontier AI compute and that no company correctly forecast AI demand two years ago. Andy Jassy said AWS is sold out through 2026, mostly through 2027, and partly into 2028, with 2027 spending set to exceed the current $220 billion capex plan.

First citedThe Neuron
29 October 2026

Wall Street's combined 2027 capex estimates for the five largest hyperscalers rose from $495 billion to $1.13 trillion in roughly 12 months.

Google Cloud Q3 free cash flow fell to a deficit exceeding $9 billion as AI data center spending intensified despite revenue doubling year-on-year to $30.7 billion, and Microsoft Cloud surpassed $63 billion with 30% year-over-year growth. Gartner found AI data center owners would need $8.5 trillion in cumulative revenue through 2030 to cover announced capex, against its forecast of only $3.6 trillion in attributable AI revenue, placing more than $4 trillion in asset value at write-down risk through 2032.

in the 30 Oct edition

Microsoft moved to debt markets for at least $25 billion, its first US bond deal since 2017, as Azure grew 48% and capex reached $45.2 billion.

Oracle's October 2026 Investor Day revealed plans to grow compute tenfold to 10 gigawatts by FY28, with remaining performance obligations reaching $638 billion. Anthropic's daily token usage on OpenRouter tripled from roughly 0.6 trillion to above 2.2 trillion by October 24, 2026, following new model releases, as TSMC prepared CoWoS price increases of 10 to 20%.

27 October 2026in the 28 Oct edition

Meta closed a $37.5 billion equity raise, the opening tranche of an infrastructure plan now totaling approximately $97.5 billion, with demand running five times the available shares and the top 15 investors taking 70%.

Nvidia committed $3 billion and AMD $1 billion to the deal, and Nvidia broadened its partnership to cover CPUs, Spectrum-X networking, and BlueField-4 alongside GPUs. Q3 earnings released alongside the closing showed revenue of $64.6 billion beating analyst estimates of $63.4 billion, but EPS of $3.83 missed the $6.80 consensus by roughly 44%.

26 October 2026in the 27 Oct edition

Meta filed with the SEC to raise $95 billion in new equity for AI data centers and superclusters, with free cash flow negative against 2026 capex guidance of $135 to $145 billion and chip suppliers Nvidia, AMD, and Broadcom indicating interest in up to $5 billion of the shares.

SemiAnalysis confirmed via satellite imagery that more than 24 GW of 2027 data center capacity is already in advanced construction, backed by roughly $1 trillion in non-cancellable leases signed in 2026. Gartner added that debt markets have largely cut off financing for neoclouds without substantial independent revenue.

First citedRohan Paul on X
25 October 2026

SemiAnalysis forecast 51.2 GW of new data center capacity in 2027, more than double 2026, with over 24 GW already confirmed under construction and roughly $1 trillion in non-cancellable leases signed in 2026.

A16z data shows H200 GPU rentals up about 160% since the start of 2026, the B200 up approximately 120%. Investor Chamath Palihapitiya argued roughly 90% of AI spending is concentrated in 1% of companies, about 80% from unprofitable VC-backed startups recycling capital, questioning whether it would hold if venture funding slowed.

added 26 Oct

Gartner found AI data center owners will need $8.5 trillion in cumulative revenue through 2030 but will receive only $3.6 trillion, putting over $4 trillion in asset value at write-down risk and forecasting more than 80% of neocloud firms will cease independent operations by 2030.

Bain calculated the industry needs $6 trillion in annual revenue by 2031 against $1.2 to $1.8 trillion from existing sources, with the $4.2 trillion gap dependent on categories including autonomous systems and products that do not yet exist.

24 October 2026

TikTok is reportedly targeting as much as 4 gigawatts of computing capacity by the early 2030s on Oracle Cloud Infrastructure, extending the infrastructure buildout to consumer platforms beyond frontier AI labs.

Amazon CEO Andy Jassy confirmed AWS is completely sold out for 2026 and mostly for 2027, even at $220 billion in annual spending, while TrendForce projects server CPU shipments growing more than 40% in 2027 as AI agent inference tightens the CPU-to-GPU ratio from 1:8 toward 1:1.

First citedRohan Paul on X
23 October 2026

Morgan Stanley's October forecast raised combined hyperscaler and neocloud spending projections to $1.58 trillion in 2027 and $1.81 trillion in 2028, up from July figures of $1.2 trillion and $1.4 trillion.

Bank of America found capex running 50% above prior forecasts against revenue only 20% above, and FastGPU and TrendForce data showed GPU rental prices rising across 81 providers and lead times running two to four times normal. Kazakhstan entered the buildout as Firebird secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs, with a second phase planned to approximately 500 megawatts.

22 October 2026

Sundar Pichai said in a public interview that no company accurately forecasted AI compute demand two years ago

21 October 2026in the 22 Oct edition

Google CEO Sundar Pichai confirmed that 2027 compute capacity is effectively already committed and that Google Cloud is forgoing short-term revenues to redirect compute toward frontier AI development.

Nvidia CEO Jensen Huang added unit economics: a 1 GW AI factory costs roughly $50 billion to build and can generate approximately $40 billion annually, with each month of deployment delay costing about $3.3 billion in foregone potential revenue. Bank of America put hyperscalers' weighted average cost of capital at 10%, up from 4% in 2023.

20 October 2026

Bank of America research published October 20 finds AI capital spending running 50% above forecasts this year while revenue is only 20% above.

The bank also reports that hyperscalers' weighted average cost of capital has more than doubled, from 4% in 2023 to 10%. A demand projection in the same research puts 212GW of new data center capacity needed between 2026 and 2030 at $9.5 trillion, $2.1 trillion above the current $7.4 trillion capex forecast.

in the 21 Oct edition

FastGPU data published October 20 shows the NVIDIA B300 GPU rental price up 52.8% and the H100 up 30% over 60 days.

A TrendForce report the same day found HDD lead times at 50 weeks against a normal 16 and ABF substrate lead times at 48 to 56 weeks against a normal 12, with hyperscalers locking in supply through 2028 and still expecting shortages. Both corroborate Bank of America's finding that AI capex runs 50% above forecasts while revenue is 20% above.

19 October 2026

Jassy disclosed to Horowitz that Amazon has no available compute through at least 2028

17 October 2026

ARK Invest CEO Cathie Wood argued that AI could replicate the Industrial Revolution's fivefold GDP growth multiplier, potentially lifting global real GDP growth to roughly 15% from a pre-Industrial Revolution rate of about 0.6%.

The argument adds an optimistic counterpoint to a debate framed by a $4.2 trillion projected revenue gap that Bain & Company attributes to 'new products and uses that don't exist today,' and Jefferies strategist Chris Wood's stated base case of 'massive capital destruction in the US'.

in the 18 Oct edition

Wall Street's combined 2027 capex estimate for Amazon, Google, Meta, Microsoft, and Oracle has jumped from $495 billion to $1.13 trillion over roughly 12 months, a $636 billion increase for the same forward year.

Google's 2027 estimate alone rose from $103 billion to $335 billion. The 2026 estimate nearly doubled over the same period, from $465 billion to $892 billion, bringing the combined 2026-2027 projection to nearly $1.9 trillion.

16 October 2026

IMF downside scenario: global GDP 1.1% below baseline by 2028 if AI investment collapses; flags $300B circular financing

15 October 2026

Jefferies strategist Chris Wood called 'massive capital destruction in the US' his base case October 15, citing Chinese open-source models processing nearly five times the token volume of top US models on OpenRouter in a single week ending July 19 as evidence that credit markets could cut off AI funding.

Updated estimates put the four largest US hyperscalers at about $695 billion in capex in 2026 and $870 billion in 2027, while UBS projects global AI rack capacity growing roughly 9x by 2030, with Nvidia's share of racks falling from about 65% to 48% as custom ASICs scale.

14 October 2026

October 14 reporting puts combined 2026-2027 capital expenditures for Alphabet, Microsoft, Amazon, Meta, and Oracle at nearly $1.9 trillion, up roughly 66%.

Over $90 billion in new chip debt has been arranged to bridge the gap between paying for hardware and receiving revenue, including Broadcom seeking more than $50 billion for OpenAI's custom silicon and SpaceX arranging $40 billion for Nvidia hardware. Bain projects the industry needs a $6 trillion annual market by 2031, with $4.2 trillion attributed to products that do not yet exist.

13 October 2026

The IMF published October 13 forecasts assuming no additional productivity growth from AI, with economy-wide data confirming little effect on wages, jobs, or measured productivity.

Its models put the downside at nearly twice the upside: 1.1% of world output lost by 2028 in a bust against 0.6% gained by 2029 in a boom. A $42 billion financing facility tied to Anthropic was also reported, showing Broadcom financing both Anthropic and OpenAI simultaneously through separate debt arrangements.

12 October 2026

Broadcom, SpaceX, and Oracle turned to Wall Street debt on October 12 for a combined $90 billion or more in AI hardware financing, with Broadcom seeking over $50 billion for custom chips for OpenAI, SpaceX arranging $40 billion for Nvidia chips, and Oracle in talks with Apollo and Goldman Sachs to bridge chip payments and cloud revenue.

One source described chip costs as too large to absorb on balance sheets. Aravind Srinivas said some Meta engineers spend around $10 million a year on AI coding tools and certain Perplexity Computer users spend upwards of $10,000 a month running agent loops.

11 October 2026

A16z data showed agents burn nearly 5 times the tokens humans do, up 14 times in six months, pointing to shift toward consumption-based metering

9 October 2026

Breakdown of NVIDIA backstop financing and specialized GPU debt facilities published; USD.AI $34M NexGen Cloud deal cited as example

8 October 2026in the 9 Oct edition

An a16z State of Markets report published October 8 found that 69% of S&P 500 companies cite a live AI deployment but only 2% disclose a metric they track over time, adding a demand-side account to Bain's earlier finding of a $4.2 trillion annual shortfall by 2031 tied to products that do not yet exist.

A separate October 2026 analysis argued that AI capital spending is 'nearly immune' to higher borrowing costs, because sufficiently high real returns on AI investment push real interest rates up regardless of Federal Reserve policy.

7 October 2026

SpaceX reported seeking $40 billion led by Apollo to purchase Nvidia chips, structured as $10 billion in bank loans and $30 billion in investment-grade debt, targeting close in 2027

First in the edition · 7 October 2026
6 October 2026

Aravind Srinivas cites $10M/year/engineer AI costs at Meta; some Perplexity users spending $10K/month on agent loops

5 October 2026

Epoch AI analysis found median OpenAI researcher spent $601/day on coding-agent inference by mid-August 2026, up from under $1/day in January, doubling roughly monthly

4 October 2026

a16z report: only 2% of S&P 500 disclose tracked AI metrics; 2.2% of U.S. households pay for AI; 2027 capex forecast $1.1T

2 October 2026

Reports from Bain and McKinsey warn of rising total AI bills despite falling per-token prices

1 October 2026

a16z State of Markets report found 69% of S&P 500 companies cite a live AI deployment but only 2% disclose a tracked metric; diffusion described as extremely uneven

30 September 2026in the 7 Oct edition

The combined 2026-2027 capital expenditure for five major hyperscalers stands at nearly $1.9 trillion, up about 66% since the start of 2026, while Bain & Company projected September 29 that even after crediting all identified revenue streams, the AI industry faces a $4.2 trillion annual gap by 2031 tied to products that do not yet exist.

Jensen Huang told CNBC a 1-gigawatt NVIDIA AI factory recovers its $50 to $60 billion build cost in roughly one year through rental revenue.

28 September 2026

Analysis published finding debt-funded non-Nvidia chip clusters likely cost more to finance than they save on hardware, with interest premium reaching 2.5 percentage points at B+ rating

26 September 2026

Agentic AI workloads identified as shifting compute demand beyond GPU inference to full data-center stack

25 September 2026

Combined hyperscaler capex estimates for 2026-2027 revised up roughly 66%, from about $1.14 trillion to nearly $1.9 trillion, an increase of approximately $750 billion.

20 September 2026

Requesty platform data showed open weights at 54.6% of tokens but only 14.9% of spend, noting token share is the most flattering metric for cheap models

14 July 2026

Morgan Stanley projects Big Five cloud capex at $1.2 trillion in 2027 and $1.4 trillion in 2028

13 July 2026

Morgan Stanley raises 2027 hyperscaler capex forecast to ~$1.2 trillion and 2028 to ~$1.4 trillion

15 June 2026

GitHub Copilot shifted from flat subscription to metered, token-based billing

12 June 2026

Financial analysis published on circular neocloud financing, showing Microsoft and Meta shift hundreds of billions in infrastructure costs off cash flow statements via operating expense structures

2 June 2026

Meta signs 20-year PPA with Constellation Energy for 1,121 MW of clean energy credits from the Clinton plant; also backs 1.2 GW Oklo reactor campus in Ohio targeting 2034

23 February 2026

Wells Fargo projects hyperscaler capacity doubling from 49 GW to over 125 GW by 2028, with $2.47 trillion in total spending 2026-2028

6 October 2025

Data centers begin signing long-term nuclear PPAs; Meta and Microsoft each sign 20-year deals with Constellation Energy

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy