The AI‑capex bet behind SOXL, TQQQ and the semis sleeve, tracked against the credit, rate,
power and demand signals that would crack it. Live, updated through the session — this is the page we
read before we change a position, not after.
The thesis says what to believe; this strip says when and how much. Trend sets the timing, volatility sets the size. A 3x product can bleed 20–40% in a sideways chop while the thesis is 100% right — so we read the tape and the decay before sizing the leverage. Live.
🎯 The Thesis & The Gamble
The whole leveraged book (SOXL 3x semis, TQQQ 3x Nasdaq, MSTR) rides one bet. Here it is in plain terms.
The Bull Case why we're long
Hyperscalers (Google, Microsoft, Amazon, Meta, Oracle) are spending an unprecedented amount building AI datacenters. That capex flows straight to chips (NVDA/AVGO/TSM), the boxes (DELL), and the power/cooling (VRT, VST, CEG). Demand is supply-constrained, not demand-constrained — the bottleneck is power and HBM memory, not orders. If monetization keeps pace, semis run for years.
The Gamble the real risk
To fund this, the hyperscalers are burning free cash flow and starting to borrow — Google just posted its first-ever negative FCF while raising capex again. The bet: AI stays wildly profitable long enough to earn a return on hundreds of billions of spend. If the debt gets expensive or the payback slips, the whole trade re-rates.
How It Breaks exit signals
Watch three cracks: (1) credit — spreads blow out / bond issuance stalls, so borrowing gets costly; (2) cost — power & memory prices spike, crushing the ROI on capex; (3) demand — cloud growth decelerates or capex gets cut. Any one turning is the tell to de-risk the leverage.
🚨 Credit & Rates — cracks in the debt bet
The single most important cluster. This is where the "borrowing to build" gamble shows stress first — before it hits the stocks. Green = calm, amber = watch, red = stress.
🏦 Hyperscaler Capex & Debt — the borrowers
The five names making the bet. Stock prices are live; capex/FCF update at each earnings print — that's the number that moves the thesis.
📅 Capex & FCF — the number that matters GOOGL verified Jul 22others: next print
Name
Latest capex
FCF signal
Next watch
GOOGL
$44.9B/qtr · FY guide $195–205B
−$5.9B (first-ever negative)
Is the raise funded by debt?
MSFT
watch next print
Azure capex intensity
Late-July Q2 report
AMZN
watch next print
AWS capex vs FCF
Late-July Q2 report
META
watch next print
Full-year capex guide
Late-July Q2 report
ORCL
watch next print
RPO / backlog growth
OCI datacenter build
Why it matters: As long as record capex is funded from operating cash flow, the bet is self-financed and low-risk. The moment it's funded by new debt issuance — and Google going FCF-negative is the first sign of that shift — the trade's survival depends on the credit market above. That's why the two cards are read together.
💰 Path to Profitability — AI revenue vs the spend
The core question of the whole trade: hyperscalers are spending ~$600B+ in 2026 on AI infrastructure against under ~$50B of identifiable AI revenue. Nobody breaks out "AI revenue" cleanly — it's buried inside Cloud and ad-targeting. Below is the best sourced estimate per name, and the gap the bulls have to close. Research snapshot, sourced — not live.
Where each borrower stands reported figuresAI-rev = estimate
Cleanest AI disclosure; OCF still > capex — self-funded
AMZN
AWS $37.6B, +28%; AI >$15B run-rate
~$200B
TTM $1.2B (was ~$26B)
AWS margin 37.7%, compressing on AI depreciation
META
AI = ad uplift ($55B ad rev, +33%)
$125–145B
+$12.4B < capex
41% op margin; ROI via ads is the clearest payback
ORCL
OCI $5.8B, +93%; RPO $638B
$55.7B (FY27 ~$90B+)
−$23.7B (−$42B FY27e)
GPU-rental margin only ~14–16%; S&P cut to BBB−
The self-funding era is ending. The three clearest tells: Alphabet's first-ever negative FCF, Amazon's trailing FCF collapsing from ~$26B to $1.2B, and Oracle's S&P downgrade to one notch above junk. Microsoft and Meta still generate positive FCF but are compressing fast, and both tapped record debt / off-balance-sheet financing (Meta's ~$30B Blue Owl SPV keeps Hyperion off the books). The unresolved question everywhere is margin at scale before the GPU depreciation wave hits.
Sources: company Q1–Q2 2026 filings; Google Q2 (Jul 22 2026); MSFT/AMZN/META Q1 2026 (Apr 29); Oracle FY26 Q4 (Jun 2026); The Information (GPU margins); S&P (Oracle, Jul 9). AI-revenue splits are run-rate claims / analyst estimates — not GAAP-disclosed.
Identifiable pure-play AI revenue (2026)under ~$50B
OpenAI ARR ≈ share of 2026 capex~$20B ≈ 3%
Sequoia (D. Cahn): annual spend needing ~$3T revenue to justify~$1.5T
Bain: new annual revenue needed by 2030 (≈$800B shortfall)$2T
Mgmt-guided return inflection2027
Read it plainly: the industry is in the capex-dominant, pre-payback phase. Bulls are underwriting hundreds of billions against a demand curve (OpenAI $100B revenue by 2029) that isn't proven yet. The dashboard exists to catch the moment the market stops giving that curve the benefit of the doubt — the leveraged book is the first thing to cut when it does.
📊 The AI Scissors — hyperscaler capex vs. identifiable AI revenue ($B, annual aggregate)
Big 5 aggregate: MSFT + GOOGL + AMZN + META + ORCL. AI revenue = analyst estimate (cloud AI + managed services) — no hyperscaler cleanly discloses this; use for orders-of-magnitude comparison, not precision. 2026 = annualized guidance run-rate. Sources: company Q2 2026 filings; Sequoia Capital "AI's $600B Question"; The Information (GPU margins). The gap is the central risk the dashboard watches — the moment the market loses patience with the capex-ahead-of-revenue bet, the leveraged book is the first thing to cut.
🏗️ Datacenter Buildout — projection vs reality
Where the capex actually lands: gigawatts of compute. The bet is power-gated, not demand-gated — announced pipelines dwarf what can actually be energized. Tracking what's online, what's building, and what's slipping. Research snapshot, sourced — not live.
🌐 Global AI-specific power — 2026~40 GW online of 80–100+ GW announced
■ Online (~40 GW)■ Building (~35 GW)■ Delayed/paper (~25 GW)
🏗️ Stargate (OpenAI/SoftBank)0.3 GW online of 9+ GW planned
🏗️ Meta (Prometheus + Hyperion)~1 GW of 6 GW
🏗️ Microsoft (Fairwater + self-build)~1 GW of 3.5 GW target (1.5 frozen)
🏗️ xAI Colossus (Memphis)2 GW online + scaling
PJM interconnect queue: ~40 months · N. Virginia vacancy: 0.5% (record low) · Sources: SemiAnalysis, Epoch AI
Major projects — status & schedule announced capacities are volatile
Project
Operator
Location
GW
Status
Stargate Abilene
OpenAI/Oracle
Abilene, TX
1.2
~0.3 online
Stargate (all sites)
OpenAI/SoftBank
multi
>9 by '29
building
Doña Ana
Oracle/STACK
New Mexico
2.2
slipped to 2029
Prometheus
Meta
New Albany, OH
~1.0
online end-2026
Hyperion
Meta / Blue Owl
Richland, LA
5.0
2GW '30, 5GW '32
Fairwater
Microsoft
Mt Pleasant, WI
→2.0
building
Colossus 1+2
xAI
Memphis, TN
2.0
online, scaling
Self-build / leases
Microsoft
US/EU
~1.5+2
frozen / walked
What's slipping: Stargate New Mexico pushed to 2029 (gas pipeline + permitting); Abilene expansion capped, grid interconnect delay >1yr; Microsoft froze ~1.5 GW of self-build and walked ~2 GW of leases in 2025. Important caveat: SemiAnalysis explicitly rebuts the "half of 2026 is cancelled" narrative — the cuts are concentrated in early-stage, non-binding LOIs; their 6-month forecast moved only ~1% (hyperscaler) / <5% (colo). The constraint is real power delivery, not demand.
Sources: SemiAnalysis (AI Datacenter Energy Dilemma; Microsoft's Datacenter Freeze); Epoch AI (Stargate sites); CNBC (Meta LA $50B, Jul 13 2026); OpenAI. Capacity figures are estimates — no official registry; treat ±20%.
⚡ The power bottleneck — where it physically jams the hard ceiling
Chokepoint
Reading
What it means
N. Virginia vacancy
0.5–0.94%
Record low; 96% of '26 supply pre-leased. No room.
PJM interconnect queue
~40 months
76% of queued projects never reach agreement
Gas turbines
110 GW orders / 60–70 GW capacity
Prices +195% since '19; in-service slips 2+ yrs
GE Vernova backlog
100 GW (Q1 '26)
Up from 46 GW a year ago — the power-gen scramble
Key AI-power deals (our VST/CEG/URA thesis): Microsoft–Constellation (CEG) 837 MW, Three Mile Island restart · AWS–Talen 1,920 MW nuclear · Meta–Vistra (VST)/Oklo up to 6.6 GW incl. an SMR option. These PPAs are the compute-cost signal — when the price of a datacenter power contract jumps, it shows up here first.
⚡ Compute Cost & Power — the input bill
AI's real bottleneck is power and memory, not chips. Rising power/fuel is demand confirmation but also a cost that eats the capex ROI — track both directions.
🧠 Compute-cost signals with no clean live ticker manual watch
HBM memory (high-bandwidth memory): the genuine supply bottleneck for AI accelerators. Track via MU (above) and SK Hynix / Samsung headlines — a memory price spike raises every GPU's cost. GPU rental / spot pricing: falling H100/B200 cloud rental rates would signal oversupply (bearish for the buildout); rising = tight. No free feed — watch CoreWeave / Lambda / cloud pricing pages. Datacenter power availability: the hard ceiling. Grid interconnect queues and PPA prices (VST/CEG deals) are the tell that supply — not demand — caps the boom.
📈 Demand & Monetization — does revenue keep pace?
The bet only pays if AI spend converts to revenue. Cloud growth is the scoreboard; the semis complex is the market's real-time vote.
☁️ Cloud growth — the monetization scoreboard GOOGL verifiedothers: next print
Platform
Latest growth
Read
Google Cloud
+82% YoY → $24.8B
Accelerating; 35.6% margin — funds the capex
Azure (MSFT)
watch next print
Is AI contribution still expanding?
AWS (AMZN)
watch next print
Re-acceleration = thesis intact
The tension to watch: capex is growing faster than cloud revenue right now. That's fine while demand is supply-constrained — but if cloud growth decelerates two quarters running while capex keeps climbing, the ROI math breaks and the leveraged book is the first thing to cut.
📊 Fundamental Pulse — the monthly heartbeat
Between quarterly earnings the tape is all you have. These are the monthly hard-data reads on the whole AI-chip supply chain — TSMC revenue, Korea chip exports, memory/HBM pricing. They lead the quarterly prints. No free API exists, so each figure is curated from the primary release with its source and date stamped. refreshed ~monthly
⏳ Loading fundamental pulse…
🧠 Demand Durability — does open-source break the ROI?
The deepest bear case on the whole board: if open-weights models (DeepSeek, Kimi, Qwen) and raw efficiency keep crushing the price of intelligence, does that crack the hyperscaler capex ROI — or expand the market (Jevons) and justify more capex? Falling token price is the weather, not the signal. The signal is Price×Quantity, where paid volume routes, and who cuts flagship price defensively. Research snapshot, fully sourced — not live.
⏳ Loading demand-durability module…
⚠️ Bubble Watch — the bear case, tracked honestly
The disciplined way to stay long is to know exactly what would prove us wrong. These are the credible red flags — the things that predict the problem before the tape does. Research snapshot, sourced — not live.
Circular financing high
Nvidia invests ~$100B into OpenAI → OpenAI rents from Oracle → Oracle buys Nvidia GPUs. Analysts estimate >$800B of these looping arrangements; OpenAI's total infra commitments hit $1.15T across 7 vendors. Revenue booked at each hop while cash moves in a circle — the sharpest dot-com echo.
OpenAI cash burn high
The anchor tenant for the whole buildout is on track to lose ~$14B in 2026 (nearly 3× 2025) and — per HSBC — needs +$207B of funding by 2030. If OpenAI's demand curve stalls, the leases underwriting Stargate/Oracle lose their foundation.
Hidden depreciation medium
Hyperscalers depreciate GPUs over 5–6 years; Burry argues real economic life is 2–3. That understates depreciation by an estimated ~$176B (2026–28), inflating Oracle/Meta operating income >20% above reality. GPUs lose ~64% of rental value in 18 months (Morgan Stanley).
Debt divergence medium
Allianz pegs the capex-vs-revenue divergence at ~46% — worse than the 32% of the 2001 telecom bust. Morgan Stanley sees hyperscaler borrowing topping $400B in 2026 (2×+ 2025) and ~$1.5T of new tech debt over three years. Ties straight back to the credit card at the top.
The Microsoft blink medium
Microsoft froze ~1.5 GW of self-build and walked ~2 GW of leases in 2025 — bears read it as the first hyperscaler flinch. Counter: SemiAnalysis calls it portfolio optimization of non-binding LOIs, not a retreat. Watch whether others follow at earnings.
Counter-signal (bull) bullish
Against all the above: datacenter vacancy is near zero (0.5% NoVA), demand is outpacing supply per Alphabet's CFO, and the "50% cancelled" meme is debunked. The constraint is power delivery, not demand — which is why our power sleeve (VST/CEG/URA) is the honest way to play it.
📅 Catalyst Calendar — what to watch & when
The dated events that move SOXL / TQQQ / NVDA / MSTR. ⭐⭐⭐ = major mover. earningsmacroaistructural. Live: earnings dates are IR-verified and cross-checked against Finnhub each cycle — a date that drifts vs Finnhub is flagged verify rather than trusted. Past events auto-strike; the next 14 days show a countdown.
⏳ Loading calendar…
Policy watch (no fixed date): the AI OVERWATCH Act (2-yr ban on Blackwell-class sales to China) cleared committee but has no floor vote scheduled — direct NVDA/AMD risk. The BIS H200/MI325X China rule (in effect since Jan 15 2026) has ~400k approved-but-unshipped H200s frozen; each licensing decision is a catalyst.
🔔 Alert Log — what fired & when
Every threshold cross the background monitor caught, newest first — the push layer so a crack isn't missed on a day the page isn't open. Refreshed every 15 minutes; critical alerts also ping our Telegram. Each alert carries the number and the pre-committed action, never a bare ping.
⏳ Loading recent alerts…
🎯 Positioning & Our Book
Our actual leveraged exposure plus the sentiment gauges. High vol + wide spreads = size down; calm + tight = the bet is working.
✅ Daily Watch Checklist
The 60-second morning routine before touching the HR book.
1
Credit spreads (HY OAS). Under ~3.5% = calm. Popping toward 5% = the debt bet is cracking — de-risk leverage.
2
Bond vol (MOVE) & the 10Y. A spike in MOVE or a 10Y push past ~4.75–5.0% raises the cost of the whole buildout and hits everything rate-sensitive.
3
Semis tape (SMH / NVDA). The market's live vote on demand. A washout on forced selling is a buy; a break on weak guidance is not.
4
Power & memory (VST / CEG / MU). Rising confirms demand but flags cost. A memory price spike is a direct hit to capex ROI.
5
Hyperscaler capex & FCF (at earnings). Raised guide funded by cash = bullish. Funded by debt, or a capex cut = thesis change.
6
Our vol (VIX) & BTC. VIX >30 = size down the 3x book. BTC is the MSTR tell.
Educational monitoring tool — not financial advice. Live data via Yahoo Finance & FRED (St. Louis Fed), keyless and delayed. Thresholds are our own framework, not signals. Past performance does not guarantee future results.