Asymmetric Short Thesis · July 29, 2026

The AI infrastructure
bubble is forming.
We know how it ends.

Hyperscalers are committing trillions to AI infrastructure. The ROI hasn't materialised — and the supply side just confirmed it. Broadcom missed AI revenue forecasts on June 6, 2026 — the first infrastructure earnings miss of this cycle. The correction has started. We're positioning into it.

$1T+
AI capex committed by the 4 hyperscalers in 2025–26
7.2/10
Avg AI ROI language score — executives still bullish, customers less so
$144M
Discretionary insider sells in AI names over the last 60 days
Aug 2026
SPCX lock-up selling begins — staggered through Dec 9, 2026
See the thesis ↓
How the thesis works

Three things have to be true.
All three are.

This isn't a bet against AI. It's a bet that the infrastructure buildout is ahead of reality — and that the gap will close painfully for whoever bought at the top.

01

The overbuild is real

Meta, Microsoft, Google and Amazon are collectively spending $300B+ per year on AI infrastructure. NVIDIA's data centre revenue grew 93% YoY in early 2025 — but is decelerating to 77%. The demand curve is bending. The capex curve isn't — yet.

02

ROI hasn't materialised — and the supply side just confirmed it

Hyperscalers score 6–7/10 on AI ROI language — optimistic but hedging. Their customers are capping spend (Uber burned a year's AI budget in 4 months). On June 6, 2026 Broadcom — the AI backend supplier — missed its AI revenue forecast. The demand-side slowdown has now propagated to the infrastructure layer. This is the Cisco parallel in real time.

03

There's a specific date that triggers the unwind

SpaceX (SPCX) IPO'd at a $1.75T valuation. The lock-up is staggered — the first tranche of insider selling begins in August 2026, with tranches every 15 days through full unlock on December 9, 2026. $200B+ in stock becomes eligible for sale over that window. This is the kind of supply shock that starts avalanches in overvalued sectors.

Historical precedent

This has happened twice.
We know the pattern.

Every infrastructure overbuild ends the same way — the leader collapses, not because the technology fails, but because the valuation was pricing in a future that arrives much later than expected.

⚡ Live — Jun 6, 2026

Broadcom Q2 2026: first AI infrastructure earnings miss vs forecasts. $300B wiped in minutes. Demand-side signals (Uber cost caps, Microsoft building in-house models) have propagated to the supply side. This is the Cisco moment — June 2000 was Cisco's first miss. The 86% decline followed over the next 30 months. Thesis score +15pt in one week.

CISCO PEAK (2000)
$500B
38× revenue · internet picks-and-shovels
CISCO BOTTOM (2002)
−86%
30 months · never recovered to peak
CITIGROUP PEAK (2006)
$270B
finance-as-infrastructure · 3× book value
CITIGROUP BOTTOM (2009)
−94%
36 months · bailed out at $1/share
NVIDIA TODAY
$2.5T
19× revenue · AI picks-and-shovels
━━ Cisco (dot-com) — fell −86% over 30 months ╌╌ Citigroup (2008) — fell −94% over 36 months

Both charts show % drawdown from the all-time peak. Day 0 = the top.

Live signals — updated daily

The data is already
telling the story.

We track 162 real-time signals across news, earnings calls, datacenter announcements and SEC filings. Here's what's moving right now.

📡
News signals (30 days)
162
Articles scored by AI as bearish for the thesis. 84 scored 8/10 or higher — the strongest conviction signals.
🏛
Insider sells
$144M
Discretionary open-market sales by ARM, MSFT and NVDA executives in the last 60 days. ARM's Chief Legal Officer sold his entire position to zero.
📋
Earnings language
7.2/10
Average AI ROI score from NVIDIA, Meta, Google, Microsoft, Amazon earnings calls. High conviction language from management while customers cap spend. Watch for this to drop next earnings cycle — that's the capitulation signal.
🏗
DC cancellations
We track every announced vs cancelled datacenter project. Cancellations are rising. Each one is a hyperscaler quietly reducing their bet.
How we play it

Long-dated put options.
Defined risk. Asymmetric upside.

We buy put options expiring January 2028 — giving us 18 months past the December 2026 lock-up catalyst. The maximum we can lose is the premium paid. The potential upside is 5–10× if NVIDIA corrects like Cisco did.

Example position — NVDA $140 put Jan 2028
NVDA today $197
Strike price $140 (−29% from today)
Premium paid $11.85/share = $1185/contract
Break-even at expiry $128.15 (-35%)
3× return if NVDA falls to $104.45 (-47%)
Why this structure?

Max loss is known upfront

Unlike shorting stock, we can never lose more than the premium paid. No margin calls.

18 months of runway

January 2028 expiry gives the thesis time to play out past the December 2026 catalyst.

Cisco fell 86%

If NVIDIA follows a similar path from $205 to ~$80, a $140 put returns ~5× on premium.

Key risks
🏛
Government equity stake in AI — active, Jun 2026
Trump/White House discussing buying a US stake in OpenAI. If announced, creates a political floor — pension funds follow. Could push prices temporarily higher before the eventual correction. Does not fix the cost problem, but can delay it 6–12 months. Response: roll the expiry date forward.
📈
Inference volume outpaces efficiency gains
If AI usage grows 10–20× faster than models get cheaper, GPU demand stays high. Our max loss is still the premium paid — position expires worthless but no other downside.
🔒
Lock-up selling absorbed without market impact
SPCX unlock gets absorbed by ETF inflows. Historical base rate: ~70% of overvalued IPO lock-up expirations cause a drawdown. 30% don't.
Next steps

The entry window is open.
The first infrastructure miss just happened.

We're currently in the pre-entry monitoring phase — tracking signals daily, identifying the optimal entry window (low IV + confirming data). The December 2026 lock-up is 18 months away. The longer we wait, the more expensive the options become.

Live signal monitor
Everything updating in real time
View Signal Monitor Full Dashboard
Entry vehicle
Long-dated puts
Jan 2028 expiry
Catalyst window
Aug → Dec 2026
SPCX staggered lock-up · first tranche Aug, full unlock Dec 9
Max loss
Premium only
No margin, no unlimited downside
Cisco scenario
5–10×
return on premium if −60% correction