AI Worldview: Start Here
The truth is probably a circle, and we piece it together from different angles. This page is the entrance to the whole research program, five layers top to bottom: stance, core thesis and reading principle; the golden triangle (how three lines check against each other); the breakeven timeline (when we will know); the pieces of each line; and the warning lights (if we are wrong, where it shows first). Every piece links to its data page. Grey ones are still being prepared.
StanceLong-term bullish, short-cycle vigilant. Compute overbuild is possible, just not yet. Small-cycle record: sell signal May 2022 (electronics down 24.5%), buy signal 3Q23, mild sell signal May 2025.
Core thesis (at least the next 18 months)
Compute stays tight and gets tighter past mid-2027, until we reach the breakeven line where a chip pays for itself the moment it is bought. On that day the AI story changes completely: the market may start believing the investment pays off in the long run. Bubble history says that belief is where the risk begins, but that is a problem for the next stage.
Reading principle: sort into tiers firstAI is already a very large market and no longer fits one story. As the industry grows it splits into tiers (pricing, credit, model layers), and the market often misreads one tier moving for the whole thing turning.
The golden triangle: how the three lines check against each other
ADemand ↔ Supply and Spending
Supply can't keep up, so is demand really above supply?
- Start on the supply side: data centers are not being built fast enough, supply falls short from now through next year, and it is tightest around mid-2027.
- Then check demand: usage held after price increases (DeepSeek +13%), and earnings calls say the shortage runs to end-2027 and gets more valuable. Demand really is above supply.
- Reverse check: if power equipment can't deliver capacity, less supply makes it tighter still (P warning lights).
BSupply and Spending ↔ Financing
Spending keeps rising and keeps moving off the income statement. Can the money keep up?
- four hyperscalers all-in investment (CapEx + new leases) runs at $726B annualized. Leases, outsourcing and off-balance-sheet commitments keep growing, so reported CapEx no longer shows the whole picture (Microsoft's apparent cut was a lease reclassification).
- The quantity of money is fine: next year's refinancing is small and debt-only coverage runs to end-2028 at 1.07x. The problem is price: tiers are getting pricier, and off-balance-sheet deals plus NVDA backstops concentrate the risk in a few hands.
- Reverse check: once money gets expensive, can the data centers still be built?
CDemand ↔ Financing
Can revenue cover the interest and depreciation?
- AI revenue (five companies, $105.8B ARR) is climbing in a step-function, but still below what was hoped. Whether it covers interest and depreciation and reaches the breakeven line is the question for the stretch to end-2027.
- After 2028 supply comes online in bulk. Whether revenue keeps up with it is the real test.
- Reverse check: if demand disappoints, work backwards to see whose runway hits the band first (META first, ORCL negative in every cell).
Breakeven timelineHow much business each unit of compute earns in a year, and when it meets the cost line
Now (first 3Q26 numbers)$3,087 annualized revenue per H100e (Q2: $2,445). Five-company ARR $105.8B.
End of 2026Breakeven line at $4.5k needs ARR of $191B or more. The base path reaches $138B, so it depends on whether a new step arrives in the second half.
Mid-2027Tightest supply window. The base path touches the breakeven line around the second half. The test begins: does the next round of big compute deals get signed?
End of 2027Sweet spot at $5.6k needs ARR of $398B or more (base case $367B, $5.2k). Hurdle rate for cloud growth: 53% to 55% YoY.
2028: the testSupply comes online in bulk. Do applications take off, and does revenue keep up with the compute that arrives?
The pieces of the three lines
DemandIs demand real? Hypothesis → earnings calls → prices → revenue
1Demand hypothesis: compute is tightweekly
Supply and demand are tightest around mid-2027; the compute cost curve crosses the cost band a year out.
Data center construction can't keep up with demand. Re-run the simulation every six months.
2Earnings check: what the calls saidquarterly
2Q26 earnings calls: the shortage runs to end-2027 and gets more valuable.
The four hyperscalers signed big deals and backlogs jumped. Microsoft's "cut" was a lease reclassification.
4Price check: pricing tiersmonthly
Price cuts are not a retreat, they are tiering: the frontier costs more, old models approach free.
Same-model repricing: 23 up, 14 down. DeepSeek volume held after a price increase, up 13%.
5Revenue: ARR and productsmonthly
Growth is a step-function: a new product opens a new customer group, and that is the next step.
Five-company ARR $105.8B (Aug 28). New ARR per new chip $5,782 versus an average of $2,446.
5bEnterprise adoption: AI inside companiesmonthly
Breadth shows in adoption rates and job postings; depth shows in who gets the enterprise dollars.
Supply and SpendingIs the money turning into compute? Spent → macro → shipments → built
3Spending: CapEx structurequarterly
Reported CapEx understates the real spend more and more; use all-in investment (CapEx + new leases).
four hyperscalers all-in investment $726B annualized. All three sets of books laid out, plus the lease share.
3bSupply baseline: Epoch trackingweekly
How many chips shipped, how much compute has accumulated, and whether prices and monetization line up.
Performance vs Price FrontierChip Prices vs MonetizationCompute Map and Economics2027 Supply-Demand Check and S Warning Lights · Coming soon
9Macro cross-check: equipment investment itselfmonthly
It shows up in the macro accounts: US equipment investment $798B annualized, up 31.6%, consistent with the four hyperscalers's $726B all-in investment.
Computer and electronics unfilled orders at a record $157.4B. Nominal orders still below the June 2000 peak.
8Supply chain check: Taiwan, Korea, USmonthly
The money did turn into shipments, confirmed at four levels.
Taiwan basket +60%. Korea semiconductor exports +155%. US server imports +64%.
Taiwan Supply Chain Basket · Coming soonMonthly Revenue by Segment · Coming soonUS Server Imports · Coming soonKorea Semiconductor Exports · Coming soonNVDA Breakdown: Customer Mix and Networking · Coming soonSupply-Side AI Revenue vs CSP CapEx · Coming soon
7Physical side: power and equipment bottlenecksmonthly
It is not that they don't want to build, equipment can't keep up: an execution bottleneck, not weakening demand.
Transformer lead times around 130 weeks, prices 95% above 2019. Data center construction $68.3B annualized.
Construction and Supply · Coming soonPower Pipeline and Demand · Coming soonScarcity Price Signals · Coming soonEquipment Bottlenecks · Coming soonGW Three-Tier Check · Coming soon
FinancingCan the money keep up? Where it comes from → who it concentrates in → where we are in the cycle
6Financing: the chain and credit tiersdaily data
Next year's refinancing is small: on a debt-only basis the runway reaches end-2028 at about 1.07x.
2027 external funding need, base case $190B. Borrowing capacity at 1.5x leverage: $806B.
CSP Cash Flow and HeadroomCSP Debt, Issuance and BuybacksWho's Buying the Bonds, and at What CostSpread Tiers and CDS · Coming soonBond Issuance, Deal by Deal · Coming soonFive Charts on Balance-Sheet Health: Does It Hold? · Coming soon
6bConcentration: off-balance-sheet, circular deals, NVDA backstopsevent-driven
Quantity is fine. Pricier tiers, deals moving off balance sheet and NVDA backstops add up to concentrated risk.
NVDA mobilizing $500B of third-party capital. A Broadcom special-purpose vehicle of up to $100B under discussion.
Circular Deals MapOff-Balance-Sheet Commitments and RPO Concentration · Coming soonHow Collateral Evolved, and How the Money Loops · Coming soon
6cThe 30-year anchor and who's buying the bondsdaily data
The 30-year at new highs is money moving, not money running out: AI bond deals are still oversubscribed.
What's Moving the 30-Year · Coming soonSupply: Treasuries Step Back, Corporates Step In · Coming soonDemand: Who's Taking It Down · Coming soonTreasuries as the Competing Price · Coming soonTwo Stories, One Scoreboard, and R Warning Lights · Coming soon
10Historical frame: where we are in the cyclebackground
A five-year investment cycle, and cycles die on the financing side.
Analogues: PCs in 1975 to 1980, the internet in 1995 to 2000. Use the 2000 analogy with care.
PMI Sell Signals · Coming soon
The outer ring: warning lightsIf we are wrong, where it shows first
SSupply-demand checkThe gap between ARR and compute supply. Light on = re-test the demand side.2027 Supply-Demand Check and S Warning Lights · Coming soon
FFinancing chainSpread tiers, issuance terms, off-balance-sheet commitments. Light on = money getting expensive or shrinking.NVDA Exposure and F Warning Lights · Coming soon
PPower (reversed)Construction, lead times, rents and vacancy. Light on = supply loosening = evidence against the shortage thesis.Nuclear Deal Ledger and P Warning Lights · Coming soon
TTaiwan and Korea supply chainMonthly revenue, ten-day exports, US imports. Light on = CapEx not turning into shipments.Taiwan Supply Chain Basket · Coming soon
MNew ARR per new chipHow much new business each new chip carries. Marginal below average = early sign the yardstick has peaked.New ARR per New Chip →
Numbers on this page are snapshots; the data pages are authoritative. Sources: company filings, earnings calls, public statistics and open datasets (Epoch AI under CC BY 4.0); consensus and forecast figures are analyst estimates (FinSight compilation).