Editorial Notes

The DeepStack manifesto.

Ten notes on what we cover, how we argue and what we owe you. Read them once. Hold us to them every time.

Part III: the notes, tested against today’s data

The artificial-intelligence build-out is the most expensive bet of this generation. Amazon, Google, Microsoft and Meta booked $165 billion of capital spending in a single quarter. Lenders are pricing bonds on the promise that hyperscaler cash flow will roughly triple by 2030. The power grid is being repriced around the data center. And most of what is written about all of it reads like a press release or a prophecy.

DeepStack was built to read like neither. We are a publication with a point of view, and these are the rules that keep the point of view honest.

Part I

What we believe

  1. 01

    Every AI story is a financing story.

    Models make headlines. Balance sheets make cycles. A product launch tells you what a company can do; a $35 billion debt tranche tells you what it has promised. We start where the money moves, in capital spending, credit, contracts and power, because that is where this cycle will be decided, and where it will break, if it breaks.

    Tested live: Curve steepening
  2. 02

    The bubble question is the wrong question.

    Bubbles are diagnosed at the funeral. What can be measured today is who carries the risk and on what terms. The useful questions are narrower and harder: is the profit under the multiple durable, who holds the lease if the tenant walks away, and what happens to the collateral when scarcity ends?

    Tested live: Profits ahead of productivity
  3. 03

    Physics sets the schedule.

    Capital can be raised in a week. A substation cannot be built in one. Power, transformers, cooling and permits decide how fast a backlog becomes revenue, and a backlog cannot commission a site. When a forecast ignores the physical world, we discount the forecast.

    Tested live: Input costs rising
  4. 04

    Rates are the referee.

    With the 10-year Treasury above 5%, every long-dated promise competes with a riskless bond. Growth that was priced generously at 2% has to be earned at 5%. We read every valuation against that line, and we say so when the cushion is gone.

    Tested live: Above the 5% line
  5. 05

    Scarcity is a season. Debt is a contract.

    Today’s shortage of chips and power is real, and it is being pledged as collateral. Shortages end; loan terms do not. When lenders price scarcity as if it were permanent, we treat that as the story, not the footnote.

    Tested live: Real money is expensive

Part II

How we work

  1. 06

    A number without a date is a rumor.

    Every figure we publish carries its date and its source. Markets reprice by the minute; last month’s yield is a different fact. When we calculate a number ourselves, we show the arithmetic and label it as ours.

  2. 07

    Steelman, then strike.

    We give the other side its strongest form, attributed by name to the investor, analyst or company who makes it. Then we say where the evidence points. A piece that refuses to conclude is not balanced. It is unfinished.

  3. 08

    Opinion wears a label.

    Reporting and judgment never blur. Our verdict lives under one heading, The DeepStack view, and nowhere else is opinion dressed up as fact.

  4. 09

    Every verdict names its test.

    We say what would prove us wrong and when we expect to find out. Those dates go on The Docket. When a test arrives, we report the result, especially when it goes against us.

  5. 10

    Independent of the trade.

    Partners never see, shape or approve a verdict. Sponsored work is labeled on the page and in the inbox. We publish analysis and opinion, not investment advice, and we do not write to move a price.

Part III · Live

The notes, tested against today’s data.

Updated · Market and economic data: U.S. Treasury; BLS; Federal Reserve H.15, Federal Reserve Board, BLS, BEA, DOL, Census Bureau, University of Michigan, OMB, S&P Dow Jones Indices, Cboe, EIA and IMF via FRED; Trading Economics for household inflation expectations, European gas, European power, uranium, lithium and forecast paths

All 5 notes are under pressure. Each analysis takes one belief, sets it against the latest Treasury, inflation, energy and profit data, and names what would change our read. The numbers move; the test does not.

Note 04Rates are the referee.

Above the 5% line

The riskless bond pays more than the stock market.

10-year Treasury
5.28%
+50 bp in a month
S&P 500 earnings yield at 19.0x
5.26%
FactSet multiple, Oct 2
Cushion over Treasuries
−2 bp
Distance to 5.50%
−22 bp
Where cycles buckle

The 10-year Treasury yields 5.28%, up 50 basis points in a month and up 114 basis points over the year. At the 19.0 times forward earnings FactSet recorded on Oct 2, the S&P 500 earns 5.26% on its price. A riskless bond now pays 2 basis points more.

Note 04 is a hurdle, not a forecast. Above 5%, every long-dated promise in The Buildout, from hyperscaler bonds to 15-year data center leases, is priced against this yield. Equity investors are paying for growth with no margin for error: earnings have to rise just to match what the Treasury already pays.

What would change our read. Our read eases if the 10-year closes below 5.00%, or if forward earnings rise about 10% at today’s prices. That would cut the multiple to 17.3x and restore a 50-basis-point cushion.

U.S. Treasury; Federal Reserve H.15 via FRED (as of Oct 8, 2026, 6:30 a.m. ET). Earnings multiple: FactSet, as cited in Part 10. The cushion and the distances are DeepStack arithmetic.

Part 10: The Bubble Is in the DenominatorPart 07: The Fed Hiked. The Long End Still Climbed.

Note 01Every AI story is a financing story.

Curve steepening

The long end is doing the tightening.

Fed funds target
3.75–4.00%
Raised Sep 16
2-year since the hike
+3 bp
Now 4.77%
30-year since the hike
+32 bp
Now 5.67%
2-year to 10-year spread
+51 bp
+27 bp on hike day

The Fed raised its target range to 3.75–4.00% on Sep 16. Since that day’s close, the 2-year yield has moved +3 basis points, the 10-year +27 and the 30-year +32. The gap between 2- and 10-year yields has widened from 27 to 51 basis points.

That is the bond market charging for time, not for policy. The front end follows the Fed; the long end follows supply, deficits and the savings the buildout has to borrow. With federal debt at 123% of GDP and a deficit of 5.8% of GDP in 2025, hyperscaler bonds, data center leases and GPU-backed loans compete with the Treasury for the same long money.

What would change our read. If the long end falls while the Fed holds, the market is pricing weaker growth rather than heavier supply. Cheaper long money would ease the financing math in Parts 03 and 08, and we would say so.

U.S. Treasury; Federal Reserve H.15 via FRED; Federal Reserve Board via FRED; OMB via FRED (as of Oct 8, 2026, 6:30 a.m. ET). Changes since the hike and the spread are DeepStack arithmetic on daily closes.

Part 07: The Fed Hiked. The Long End Still Climbed.Part 03: The Machines Will Work. The Debt Might Not.Part 08: AI Lenders Are Financing the End of Scarcity

Note 03Physics sets the schedule.

Input costs rising

The cheap electron is American.

US natural gas
$3.03
−9% in a year
European gas (TTF)
€79.16/MWh
+142% in a year
European power, five markets
+68%
Average change in a year
Copper, monthly average
$6.14/lb
+39% in a year, to Jul 2026

American natural gas trades at $3.03 per MMBtu, down 9% on the year. European gas is up 142%, and wholesale power in Germany, France, Spain, Italy and the UK is up 68% on average. Brent is $125.44 a barrel, up 87%.

Note 03 says physics sets the schedule. Prices say where. A data center is a long contract for electricity, and the cheapest long contract on offer is written on an American grid. That gap is the quiet subsidy behind the US buildout, and it is why the binding cost in Part 06 is grid capacity, not chips. Copper, the metal in every transformer and transmission line, is up 39% in a year: the bottleneck has a commodity price.

What would change our read. If US gas climbs toward European levels, or copper breaks lower on slowing grid orders, the scarcity premium in Part 06 is moving, and our read moves with it.

EIA via FRED; Trading Economics (European gas); IMF via FRED; Trading Economics (European power) (as of Oct 8, 2026, 6:30 a.m. ET). Year changes and the five-market average are DeepStack arithmetic. Copper is a monthly average (latest Jul 2026); its changes compare monthly averages, not weeks.

Part 06: Power Doesn’t Depreciate. Its Premium Moves.

Note 02The bubble question is the wrong question.

Profits ahead of productivity

Profits are running ahead of productivity.

Corporate profits, Q2 2026
$3.88T
+7.7% on the quarter
Profits since Q4 2022
+39%
The last quarter before ChatGPT
Output per hour since Q4 2022
+9.1%
+1.4% annualized in Q2 2026
Payrolls, September 2026
+29k
Unemployment 4.2%

Since the last quarter before ChatGPT, US corporate profits have risen 39%, to an annual rate of $3.88 trillion in Q2 2026. Output per hour has risen 9.1%. Employers added 29,000 jobs in September, against 133,000 the month before, and unemployment is 4.2%.

This is the denominator from Part 10, measured from the other side. Profits are real money, but they have grown about 4 times faster than the productivity that would make them permanent. Note 02 says the bubble question is the wrong question. The right one is who carries that gap if productivity does not catch up. Today it sits with shareholders paying 19 times earnings and with the lenders financing the spending behind those earnings.

What would change our read. Productivity growth above 2% a year for two straight quarters would start to close the gap. A fall in profits with productivity flat would tell us the denominator was borrowed.

BEA via FRED; BLS via FRED (as of Oct 8, 2026, 6:30 a.m. ET). Growth since Q4 2022 and the ratio are DeepStack arithmetic.

Part 10: The Bubble Is in the Denominator

Note 05Scarcity is a season. Debt is a contract.

Real money is expensive

Debt is a contract written in real terms.

10-year real yield (TIPS)
2.92%
+49 bp in a month
Inflation the bond market prices
2.36%
10-year breakeven
Households expect, 5 years
3.0%
NY Fed survey, September 2026
Headline CPI, August 2026
3.4%
Core 2.4%

The 10-year inflation-protected Treasury yields 2.92%. That is the real price of a decade of patience, before any credit spread. The bond market prices 2.36% annual inflation over that decade. Households surveyed by the New York Fed expect 3.9% over the next year and 3.0% over five, and consumer prices rose 3.4% in the year to August.

Note 05 says scarcity is a season and debt is a contract. A lender financing GPUs or a data center today locks in a real coupon of nearly 3% plus a spread, on collateral whose scarcity premium can fade within the life of the loan (Part 08). If inflation runs at what households expect rather than what bonds price, the real burden on borrowers falls and the lender absorbs the difference. Either way, the contract outlives the shortage.

What would change our read. A 10-year real yield back below 2% would cut the real cost of the buildout’s debt. Breakevens rising toward 3% would mean the bond market has started to believe the households.

U.S. Treasury; Trading Economics (household inflation expectations); BLS (as of Oct 8, 2026, 6:30 a.m. ET). The breakeven is DeepStack arithmetic.

Part 08: AI Lenders Are Financing the End of ScarcityPart 03: The Machines Will Work. The Debt Might Not.

Analysis and opinion, not investment advice. Live figures are refreshed from U.S. Treasury; BLS; Federal Reserve H.15, Federal Reserve Board, BLS, BEA, DOL, Census Bureau, University of Michigan, OMB, S&P Dow Jones Indices, Cboe, EIA and IMF via FRED; Trading Economics for household inflation expectations, European gas, European power, uranium, lithium and forecast paths when the site is rebuilt; each card shows the time of its data. Earnings multiples are FactSet’s, as cited in Part 10.

Corrections. When we get a fact wrong, we correct it in place and say so at the top of the story, with the date of the change.

DeepStackEditorial Notes, first edition. October 2026.

Read The Buildout Run the numbers on the Data Desk