Bubble Watch · The Walk
The Walk
Fifteen stations. Simple on one face, the filed receipt on the other.
source: bubble-watch engine spine · data as of 2026-07-14
Bubble Watch · The Walk
The Two Cases
Some say it's the best lemonade stand the town will ever have; others say they built a marble palace before selling a single cup. Both can't be right. Walk the twelve signs and decide.
Two paths, one start — walk the signs to see which one fits
The Two Cases
Every argument about the AI boom reduces to a race between two clocks. Bull case: today's capex is an option premium on owning the next platform — as the marginal cost of intelligence falls, TAM expands into every vertical and the funding gap is retired within ~5 years. Bear case: if scaling laws bend (energy/compute stops buying proportional capability), the ~$700B–$1.4T annual funding need becomes un-financeable and spend stops abruptly.
The desk does not adjudicate. The twelve signals that follow are the count that separates the cases, and their sum is the Divergence.
Source: the engine, two-clocks model.
The race — the payoff clock against the runway clock, live from the engine.
LIVE RENDERER · spine as of 2026-06-30
STATION 01 · Bubble Watch · The Walk
The Tape
The excitement is real and you can measure it. Price is the only signal that can't be estimated — it simply is what it is. The ticket costs half again its usual price — the show had better be great.
SOXX above its 200-day trend — real number: +50.5%
The Tape — +50.5%
Receipt. SOXX closed 566.32, ≈ +50.5% above its own 200-day trend; the desk's FP1/FP2 reversion lines sit at 376 (−33.6%). Price is the only input that is not an estimate. A trend line is not a target; this measures how much optimism is already capitalized, not direction.
Source: index tape (SOXX), 2026-07-07.
Falsify: price mean-reverting toward trend — multiple compression via earnings growth, not drawdown.
The market lens — the chip index against its own 200-day trend.
STATIC EXHIBIT · as of Jul 2, 2026 · goes live: prices feed (Task #18)
STATION 02 · Bubble Watch · The Walk
The Spend
The builders are spending wildly — two-thirds more than last year — a bet that huge sales are coming.
Hyperscaler capex: FY2024 vs FY2025 — real number: +67%
The Spend — +67%
Receipt. Hyperscaler capex stepped in unison, straight from FY2025 10-Ks: Amazon $78B→$128B, Alphabet $53B→$91B, Meta $37B→$70B, Microsoft $44B→$65B, Oracle $6.9B→$21B — aggregate growth >+67% YoY. Filed figures, primary source, not guidance.
Falsify: capex growth decelerating toward the growth rate of the revenue that services it.
Filed capital spending, FY2024 → FY2025 — every major spender at once.
STATIC EXHIBIT · as of Jul 2, 2026 · goes live: capex ledger
STATION 03 · Bubble Watch · The Walk
The Demand
They've spent over a trillion dollars building — but the cups actually sold run almost half short of what all that building assumes. The first crack.
What was built vs what demand supports — real number: ~46% short
The Demand — ≈10× gap
Receipt. One year of AI capex (~$1.04T committed) against realized AI end-user revenue is an order-of-magnitude gap — realized revenue ~46% short of what the spend's own return math requires. The gap closes from one side or the other: revenue rises (boom) or spend falls (bust).
Source: one-time analysis, funding-need vs realized-revenue.
Falsify: filed enterprise-AI revenue compounding fast enough to close the gap from demand.
What AI earns today against what the spending needs it to earn.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 04 · Bubble Watch · The Walk
The Financing
To keep building this fast they need about $650 for every $50–150 they really have. The rest is borrowed or just promised.
Funding need vs real revenue — real number: $650B needed, ~$100B have
The Financing — $650B needed
Receipt. Sustaining the current build requires ~$650B/yr of real end demand (JPM frames the annual funding need at $700B–$1.4T by 2030); honest realized end-user revenue today is ~$50–150B. The delta is financed — equity, debt, or vendor-funded customers — until demand arrives.
Falsify: filed revenue disclosures closing the order-of-magnitude promise-vs-receipts gap.
Realized end-user revenue against the annual funding need.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 05 · Bubble Watch · The Walk
The Recycling Ratio
The builders keep buying from each other — the same ten dollars goes around about fifteen times, so the street looks fifteen times busier than the real money in it.
Recycling ratio on funded-cash basis — real number: 15.5×
The Recycling Ratio — 15.5×
Receipt. Committed lab compute ÷ outside equity funded into the labs, reported on every basis because the honest range is the point:
| Basis | Equity in | Compute out | Ratio |
|---|---|---|---|
| Funded cash | $34.8B | $539.5B | 15.5× |
| Filed | $103B | $539.5B | 5.2× |
| Filed + reported | $151B | $539.5B | 3.6× |
| PV-adjusted | $34.8B | $450.8B | 13× |
The legs, filed: Microsoft $11.8B funded → OpenAI's $250B Azure commitment; Amazon $15.0B → $138B AWS (existing $38B + $100B/8yr); Amazon holds $8B of Anthropic. ~96% of the compute routes to the same two firms that are the labs' largest equity holders.
First published at 26×; a line-item re-verification of the filings on 2026-07-02 (Amazon $15B funded + $35B commitment letter, EDGAR accession 0001018724-26-000014) raised real funded equity and pulled the ratio to 15.5× — corrected in public, dated on the receipts ledger.
Sources: Microsoft 10-Q 0001193125-25-256321, 0001193125-26-191507; Amazon 10-Q 0001018724-26-000014. Filed dollar amounts, not press figures.
Falsify: outside, non-vendor cash rising until the ratio compresses toward low single digits.
Committed compute over outside funded cash — live from the filings ledger.
LIVE RENDERER · spine as of 2026-07-14
STATION 06 · Bubble Watch · The Walk
The Loops
Thirty-three 'you buy from me, I buy from you' deals. Circles like that make a quiet street look like a party.
33 filed financing/procurement loops — real count: 33 edges
The Loops — 33 edges
Receipt. 33 filed financing/procurement edges form closed rings: e.g. MSFT→OpenAI→Azure, AMZN→Anthropic→AWS, the CoreWeave backstop, the AMD warrant. Vendor-funds-customer deals inflate apparent market size and share downside — when one edge breaks the loop unwinds rather than shrinks. The chart shows filed-vs-reported provenance of each edge.
Falsify: edges settling into arms-length, audited third-party revenue outside the ring.
The financing loops, combined — who funds whom, who buys from whom.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 07 · Bubble Watch · The Walk
The Depreciation
They say the machines last six years. Really it's three or four. Pretend longer and your books look richer than your stand is.
Real hardware life vs booked — real numbers: 3.5y real, 5.5y booked
The Depreciation — 3–4y real vs 5–6y booked
Receipt. Effective economic life of AI accelerators ≈ 2.5–4y (obsolescence + utilization); several giants book 5–6y useful life. The stretch flatters current earnings — desk estimate $40–70B earnings at risk across named firms (not filed), ~+$10.52B named. Depreciation is where the build's true cost is disclosed honestly or deferred.
Falsify: booked life moving toward demonstrated life, or write-downs taken openly.
Real economic life against booked life — and the earnings at risk.
STATIC EXHIBIT · as of Jul 2, 2026 · goes live: depreciation ledger
STATION 08 · Bubble Watch · The Walk
The Debt
More than sixty-five cents of every dollar in the biggest builders is borrowed — fine while cups sell, frightening if they stop.
AI infrastructure spend: borrowed vs own — real number: >65% financed
The Debt — >65%
Receipt. >65% of this year's AI infrastructure spend is financed — bonds, loans, leases, off-balance-sheet vehicles — vs paid from operating cash flow. On/near-balance-sheet named debt totals $255.7B; ~$330B of tech-linked HY + leveraged loans + BDC exposure matures through 2028. Debt is what turns a disappointment into a crisis: it reprices whether or not revenue arrived.
Falsify: the build funded from operating cash flow; issuance reverting to its pre-2025 pace.
The maturity wall — what comes due for refinancing through 2028.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 09 · Bubble Watch · The Walk
Insider Selling
The people who own the most famous stand quietly sold sixteen billion dollars of their own share — and bought nothing. When the owners drift to the door, look.
Named insider transactions, Form-4 filed — real: $15.6B sold, $0 bought
Insider Selling — $15.6B sold, $0 bought
Receipt. Named insiders at the center of the trade sold $15.6B of their own stock (roster includes Nadella/MSFT, Su/AMD, Thiel/PLTR, Samueli/AVGO); open-market purchases at Nvidia over the same window = $0. Discretionary sold ~$3.75B; AI-core open-market buys ≈ $0. Insiders can be wrong but are never uninformed — a one-way flow at a boom top is a late-cycle marker, filed under oath.
Falsify: meaningful open-market buying by the same insiders — the flow turning two-way.
The sellers, by name — from their own Form-4 filings.
STATIC EXHIBIT · as of Jul 2, 2026 · goes live: insider ledger (Form-4 feed)
STATION 10 · Bubble Watch · The Walk
The Power Draw
This stand drinks electricity — soon nearly a tenth of the whole town's power. Nobody can fake the electric bill.
US data-center grid share — real: ~4% today, ~8–12% projected by 2030
The Power Draw — 3–4% → 8–12%
Receipt. US data-center electricity draw is 3–4% today; announced build implies 8–12% of the grid by 2030 (~15%/yr, ~2.3× by 2030, ~4× faster than grid buildout). PJM capacity clears ~9.3× prior, with ratepayer socialization. Power is the physical ceiling — interconnection queues gate whether the compute switches on at all.
Falsify: generation + transmission build keeping pace with the queue; capacity prices cooling.
Data-center electricity demand — the physical ceiling on the whole thesis.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 11 · Bubble Watch · The Walk
The Corroboration
Several smoke detectors, in different rooms, wired by different people, are all beeping at once. It doesn't prove there's a fire — but you don't ignore three independent detectors.
BoE, BIS, IMF + the desk: different sources, same warning
The Corroboration — the referees
Receipt. Independent official-sector bodies — Bank of England, BIS, IMF — have each published warnings that rhyme with this board: stretched valuations, concentration, circular financing. Different incentives reading the same strain lowers the odds this is one desk's error. Corroboration is error-checking, not proof.
Falsify: the official-sector ladder going quiet — warnings withdrawn or walked back.
The official-sector warnings, in order — who said what, when.
STATIC EXHIBIT · as of Jul 2, 2026 · revised on editorial cadence
STATION 12 · Bubble Watch · The Walk
The Composite
Roll all twelve signs into one dial: 49 out of 100 — eyes open, not run.
Composite divergence score — real number: 49/100 (needle at real position)
The Composite — 49/100
Receipt. All signals weighted into one 0–100 reading: 49 today. Strain concentrates in capex-vs-demand (65) and demand durability (57); depreciation integrity (37) and energy (40) sit calmer. No single signal is dispositive — convergence is the warning. The live trace is the engine: market clock M(t), filings ground truth G(t), and D(t) between them, by quarter.
Falsify: components de-correlating — capex strain easing while demand durability improves.
The honest close: We are not forecasting a crash. Three falsifiers, if they trip, retire the thesis loudly: (1) filed revenue catches the excitement (D(t) compresses toward 0); (2) the loops unwind into arms-length demand (recycling → low single digits); (3) insiders start buying open-market. Each is already tracked on the Falsifier Watch. That's the exit, stated before we'd ever need it.
The engine behind the number — the market clock, the filings clock, and the gap.
LIVE RENDERER · spine as of 2026-07-14
Bubble Watch · The Walk
The Memory
The town made a giant bet like this once before. Here's how today rhymes with then — and the honest places it doesn't. History rhymes; it doesn't promise.
AI cycle vs dot-com cycle — real position: month 44 of 48 (~92% of the way)
The Memory — month 44 of 48
Receipt. The AI cycle overlaid on the telecom/dot-com cycle — both NASDAQ, both indexed to 100 at their boom's first month. Telecom peaked at month 48; the AI cycle is at month 44. The overlay shows shape, not prophecy — where we sit on a curve the market has walked before, with the honest differences named (no CDS-style credit blowup yet; earnings base real, if concentrated).
Source: FRED NASDAQCOM, monthly close, indexed to boom start = 100.
FRED NASDAQCOM (monthly close) · both cycles indexed to their boom start (=100).
LIVE RENDERER · authored · data updated on engine refresh
Bubble Watch · The Walk
The Record
We're not hoping for a crash. Here are the three things that would prove us wrong — and if they happen, we'll say so, loudly. We've already corrected our own biggest number once, in public. That's the deal: the meter beats the story. Even ours.
as of the board · as of Jul 2, 2026
The Record — the honest close
The desk's only conclusion is a measurement, so the exit is stated before it's needed. Three falsifiers, tracked live on the Falsifier Watch; any one tripping retires the thesis loudly:
- The receipts catch the excitement — filed AI revenue compounds until D(t) compresses toward 0 (today: +4.06σ, widest on record).
- The loops unwind — outside, non-vendor cash rises until the recycling ratio falls toward low single digits (today: 15.5× funded-cash, $539.5B/$34.8B).
- The owners turn buyers — meaningful open-market insider buying at the AI-core names; the flow turns two-way (today: one-way; ~$0 open-market buys at Nvidia).
The precedent that we mean it: the flagship measurement was restated in the open, 26× → 15.5×, on an EDGAR line-item re-verification (accession 0001018724-26-000014, revision 2026-07-02). The old figure is superseded and blocked at build time.
Falsify (meta): this card falsifies the desk itself — that is the point.
Links: Receipts · the Falsifier Watch · EDGAR accession 0001018724-26-000014
authored · falsifiers updated on editorial cadence
The condition · the record
Corrected in public, on purpose
This page's flagship measurement was restated in the open: an EDGAR line-item re-verification (Amazon's $15B funded stake plus a $35B commitment letter) widened the outside-cash denominator, and the recycling ratio moved from the widely-quoted 26× to ~15.5× on the funded-cash basis. The old figure is superseded and blocked at build time. That is the desk's posture in one act: the meter beats the narrative — including ours.
EDGAR accession 0001018724-26-000014 · revision 2026-07-02 · the full ledger of calls, restatements, and open falsifiers lives on Receipts and the Falsifier Watch.