The capex is already contracted. The supplier is paid first.
Multi-year datacenter commitments are signed long before the revenue they are meant to produce arrives. The cash reaches whoever sells the capacity well before it reaches the buyer's earnings.
Follow in this browser. Revisit Following for evidence updates and call results.
So once again, someone explain to me how exactly AI capex is all a bubble? $NBIS got a $20B deal for 5 years. Their market-cap was $15B at end of day. Microsoft, a hyperscaler, already is spending massive amounts on capex. The AI bears will tell you the entire capex lifecycle is done and peaked because $NVDA is down $10 after earnings, as if a company never fell after earnings. Then Nadella whips out $20B after hours for more datacenter/cloud/compute buildout… Zucks is ready to “conservatively” spend $600B on capex by the end of 2028. Amazon, Microsoft, Meta and Google are expected to spend $465 billion on capex in 2026, up 21% YoY. What are we talking about? How can someone reasonably say it’s all a bubble? At this point it feels like people wanting to just say AI is pointless for the sake of saying it to “call the top” vs actually analyzing the rate of growth, spend, and ROI that we are seeing from AI. $AMZN $META $MSFT $GOOGL
amit ·
Inspired by amit · Basket by Thesis
The basket and call are our interpretation. No author endorsement.
The time-bound call
Fixed at publicationThis basket will outperform the S&P 500 over 90 days.
- Started
- Deadline
- Last updated
How the call is scored
A fixed model basket versus SPYx, each initially quoted from $150 USDC. The original raw token quantities never rebalance. Returns compare their USDC sell quotes with the starting sell quotes. The call succeeds only if the basket's return is strictly higher. Equal returns are a tie. Resolution uses the first complete quote observation started on or after the deadline, within 48 hours. If that window has no valid observation, the result is unresolved. Quotes include route price impact but exclude wallet execution and network costs. These are model returns, not your investment results.
Prices come from live Jupiter quotes and refresh once a day. A price that looks unreliable is skipped rather than guessed.
The deadline only scores the idea. It never sells your holdings and pays no prize. A new version of the thesis starts a new call.
Quote source: Jupiter ↗The argument
Why this idea. Why these businesses.
Amit argues the spending is contracted rather than speculative, and lists the commitments to make the point. Our reading is narrower than his: whether or not the buildout pays off for the buyers, the cash leaves their balance sheets on a schedule already agreed, and it lands on whoever sells the hardware and the capacity. That is a claim about sequence, not about whether AI earns its cost. He named four spenders; he did not choose these assets, set these weights, or make our timed call.
Inside the basket
- 40%
NVIDIA
The largest single recipient of committed datacenter spending, and so the clearest expression of a claim about who is paid first.
The tradeoff. Concentration in one product cycle and one customer set. Those same customers are funding alternatives, and a supplier priced for scarcity is the first thing repriced when scarcity ends.
About this tokenized stock
Underlying: NVDA. Issuer: Backed Finance. Backed Finance issues these tokens. The issuer can move tokens out of any wallet (permanent delegate) and can freeze all transfers (pausable). Balances rebase for dividends and splits, so your share count can change without a trade.
Issuer terms (opens in a new tab) - 30%
Microsoft
Sits on both sides: it commits the capex and rents the resulting capacity out, which makes it a partial hedge against the supplier leg rather than a second copy of it.
The tradeoff. Depreciation at this scale lands on its own margins, so being on both sides cuts in both directions.
About this tokenized stock
Underlying: MSFT. Issuer: Backed Finance. Backed Finance issues these tokens. The issuer can move tokens out of any wallet (permanent delegate) and can freeze all transfers (pausable). Balances rebase for dividends and splits, so your share count can change without a trade.
Issuer terms (opens in a new tab) - 30%
Amazon.com
The second large committed buyer named in the post, and the one whose earnings are least dominated by AI, which stops the basket being one bet expressed three times.
The tradeoff. Retail and the rest of AWS can move this holding for reasons unrelated to the claim.
About this tokenized stock
Underlying: AMZN. Issuer: Backed Finance. Backed Finance issues these tokens. The issuer can move tokens out of any wallet (permanent delegate) and can freeze all transfers (pausable). Balances rebase for dividends and splits, so your share count can change without a trade.
Issuer terms (opens in a new tab)
Why these weights. Weighted toward the supplier because the claim is about who receives the money first, not about who ultimately profits from it.
The strongest case against
Being paid first is not the same as being paid durably. Contracted orders can be renegotiated, delayed, or filled by a competitor, and a supplier priced for permanent scarcity is priced for a cycle that has always ended. If the buyers conclude the returns are not there, the next round of contracts simply is not signed.
What would change the thesis?
Two consecutive quarters in which a major hyperscaler guides capex down, or a large contracted order is publicly cancelled or repriced.
Read the evidence. Make up your mind.
Sources supporting the idea, including the ones that challenge it.
- Hyperscaler capex has quadrupled since GPT-4's release (opens in a new tab)
Epoch AI
Measures combined capex at Alphabet, Amazon, Meta, Microsoft and Oracle growing about 72% a year since Q2 2023. Epoch states its own limits: finance leases may be understated, operating leases are excluded, companies define capex differently, and the share that is specifically AI is not disclosed. It establishes that the spending is large and rising; it does not establish who profits from it.
- Hyperscaler capex on track to overtake operating cash flow (opens in a new tab)
Epoch AI
Projects aggregate cash capex overtaking operating cash flow around Q3 2026, with capex growing about 70% a year against 23% for cash flow. Epoch calls these simple extrapolations, notes the crossover moves between Q2 and Q4 2026 depending on the fitting window, and does not model whether the returns justify the spending. This is the strongest published case against both baskets: spending that outruns cash generation is the mechanism by which the buildout stops.
- Amit argues AI capex is contracted, not speculative (opens in a new tab)
amitisinvesting on X
The post that prompted this basket. He names four spenders and argues the commitments are real; he does not recommend this allocation, choose these weights, or make our timed call.
Updates and version history
An update appends dated evidence to this argument. A new version changes the argument or the allocation itself. Neither one touches a position you already hold.
No updates since publication. When the author adds one, it appears here with its date and its sources — it never changes what is written above.
- Version 1
5d2859b2f5Current
Written by the Thesis team. The team holds no position in these companies or their tokenized shares, and is paid nothing by any of them.
Tracking begins at publication. This thesis has no established performance history. Tokenized stocks carry issuer and market risk.