The AI Buildout, Three Ways: Scenarios and Example Wheel Candidates (as of 2026-07-01)

Updated 12 October 2026: a new scenario analysis by sub-sector sits at the end of this article. The original July analysis below is unchanged.

As of 2026-07-01. Educational and analytical only. Not investment advice. Numbers below come from a reproducible fact pack (yfinance, 2026-07-01, ET); the scenarios are point-in-time judgment on top of that data. Nothing here is a buy or sell instruction. Where a cash-secured put is mentioned, its max loss and breakeven are stated.

This article was generated by AI — levelbox's deep-dive tooling — and lightly reviewed. The data is point-in-time and can be wrong; if you spot a mistake, please tell us.

The AI buildout is a capex story before it is a demand story. Google, Amazon, Microsoft and Meta have guided to roughly $725B of capital spending in 2026, up 77% on 2025, and the constraint they all describe is supply rather than demand. Power is the binding input. Memory is the second. Underneath both sits the question nobody has settled: does the revenue show up before the depreciation does.

That tension is the whole thing. Capex you can see today. Return you can only project. So this piece walks three ways the next one to two years can go, then uses a few names as worked examples of how the wheel lens scores a candidate — and at what risk.

Three scenarios

Base — the supply-constrained supercycle grinds on · 50%

Capex holds near guidance and tracks toward the >$1T that Evercore and BofA pencil in for 2027. The market stays supply-constrained. Power and memory stay tight.

The drivers: 2026 capex around $725B (AMZN ~$200B, MSFT ~$190B, GOOGL $175-185B, META $115-135B); hyperscalers reporting they are supply-constrained rather than demand-constrained; nuclear PPAs (Constellation-Microsoft, Vistra-Amazon/Meta) locking multi-year power demand; data centres now taking more than 50% of DRAM/NAND. Winners cluster in NVDA, MU, GEV, plus AVGO, CEG, VST, VRT. The laggards are low-moat suppliers with no pricing power.

What would break this read: a hyperscaler trims 2027 capex guidance on an earnings call, or cloud backlog growth slows two quarters running. Watch for those rather than the headlines.

Bull — acceleration, >$1T pulled forward · 25%

Inference demand inflects, 2027 capex is guided above $1T sooner than expected, and power and memory deals accelerate. The whole enabler stack re-rates up. The tell here is neocloud backlog: CoreWeave reported a $99.4B backlog in Q1'26. Winners broaden to MU, NVDA, VRT, CRDO, ALAB, CRWV, NBIS.

Invalidation: capex intensity moderates toward historical norms, or frontier-model training orders visibly slow.

Bear — digestion / air-pocket · 25%

The ROI gap and depreciation drag force a pause. One hyperscaler trims, overbuild fears surface, the memory cycle rolls over. A sharp, sentiment-driven derating, even if long-run demand is intact.

The numbers behind the worry: roughly 95% of enterprise AI projects show no measurable ROI; GPUs have a 3-5 year life and depreciate about 20% a year — near $400B of annual expense, above combined hyperscaler profits; capex intensity around 34% of revenue, roughly double the 1990s internet peak, with cash generation turning negative. Relative winners in this case are the names with the strongest balance sheets and steadier demand — MSFT, GOOGL, AVGO, CEG, VST. Hit first: SMCI, CRWV, NBIS, MU, WDC.

Invalidation: capex reaccelerates, or enterprise-AI ROI evidence broadens clearly.

Three scenarios, and two of them are not the happy one. Keep that in view before the wheel part, because the wheel does not save you from the bear case. It only changes the price you enter at.

The wheel angle: three worked examples

The wheel is simple. You sell a cash-secured put on a stock you'd accept owning at the strike, and get paid to wait for your price. If assigned, you own it. If not, you keep the premium and repeat. The risk is not exotic: it is owning the stock, minus the premium collected. So the only names worth wheeling are ones you'd hold through a drawdown.

To show how the wheel lens reads a candidate, here are three names from this one fact pack — quality plus enough premium at ~0.10 delta, ~30 DTE. Treat them as worked examples rather than a buy list:

Ticker	Verdict	CSP max loss	Breakeven	Strike / credit
NVDA	wheelable	≈ $16,896	$168.96	$170 / $1.04
MU	wheelable	≈ $80,628	$806.27	$825 / $18.73
GEV	wheelable	≈ $99,170	$991.70	$1,000 / $8.30

NVDA is the demand centre of the buildout, with a full-stack moat spanning CUDA, interconnect and systems, well beyond the silicon. The other side: revenue is concentrated in the same hyperscalers whose capex could pause, custom silicon (TPU, Trainium, MTIA, MAIA) chips away at merchant-GPU share, and implied vol around 0.38 says the market already prices big moves. Max loss on the put is about $16,896, breakeven near $168.96.

MU has the richest premium in the sleeve (IV ~0.95), because HBM/DRAM is a genuine bottleneck and memory is deeply cyclical commodity at the same time. It rolls over first in a digestion scenario. A put at the $825 strike ties up about $82.5k of collateral against a ~$1,154 stock; max loss about $80,628, breakeven near $806.27. Own-it-at-the-strike discipline matters most here; the fat premium is secondary.

GEV is levered to the number-one bottleneck, power. Quality and premium clear the bar, but the share price near $1,175 means large collateral: max loss about $99,170, breakeven near $991.70. Given the collateral, this calls for sizing down.

The method matters more than these three

Three names is a snapshot rather than a shortlist. Run the same screen on a different day, or with a different delta, account size or DTE, and other candidates surface — plenty of AI-buildout names (MSFT, GOOGL, META, AMD, ASML, WDC, LRCX among them) can read well once premium, collateral and quality line up for you. The useful part is the lens, applied to names you'd genuinely own.

One pattern worth keeping: high implied vol pays more premium and assigns harder. The fat premium is usually the market quoting the risk back to you, so the steady move is a quality underlying you'd hold through a drawdown, at a lower delta, over premium-chasing the most volatile name on the board.

For the current, full list — scored live with fresh numbers, at your own delta and account size — run the levelbox.ai screener.

The one line to remember

Wheel names you'd own outright, at the strike, for years. Everything above is scenarios and math to help you decide which those are. It is not a recommendation to sell any put, and the bear case has a one-in-four weight for a reason. Verify the figures against primary filings before acting.

Sources

Fact pack: deep-dive-analytics (yfinance, as of 2026-07-01, ET) — the wheel scores, IV, and CSP max-loss/breakeven figures. Narrative figures are read via reputable outlets; the capex and backlog numbers trace to company earnings guidance reported by those outlets.

• Big-tech 2026 capex ~$725B (+77%), and 2027 >$1T (Evercore/BofA): Tom's Hardware
• Hyperscalers supply-constrained, per-company 2026 capex, custom silicon: ValueAddVC
• Power / nuclear PPAs (Constellation, Vistra): 24/7 Wall St
• Memory TAM (data centres >50% of DRAM/NAND): Intellectia
• CoreWeave Q1'26 backlog $99.4B: io-fund
• ROI gap (~95% no measurable ROI) and GPU depreciation (~20%/yr): Goldman Sachs; Guinness Global Investors
• Capex-to-revenue gap (~34% of revenue): Forbes

Update, 6 August 2026: the market began answering the capex-versus-demand question in the weeks after this was written. It briefly sold the hyperscalers for raising capex, then reversed hard once AWS and Google Cloud showed accelerating revenue alongside the spending — while the hardware layer fell even on record results. See The AI Buildout's Second Phase: From Capex to ROI Capture.

Update, 6 September 2026: the bear case above assumed a demand pause. The likelier mechanism turns out to be a financing one, and it has already fired twice without any lab being in trouble. AI Value Chain Risk: What Actually Breaks works through who absorbs it.

Update, 12 October 2026: scenarios by sub-sector

Three months on, prices have split the buildout into parts. Companies that use AI or protect it have risen. Companies that sell the capacity have stalled, and storage and memory have fallen. The original scenarios treated the buildout as one trade. This update breaks it into twelve sub-sectors and runs each one through three scenarios for the next six to twelve months.

This section is an AI-assisted view, not financial advice. The probabilities are judgement, and the tickers are examples that define each group, not picks.

Where each part stands

Typical (median) member of each basket, closes to 9 October 2026:

Sub-sector	Basket	1 week	3 months	From 52-week high	Above 200-day	Median forward P/E
Cybersecurity	CRWD, PANW, ZS, FTNT, NET, S	+4.3%	+37.8%	−0.1%	6 of 6	71
AI data and analytics software	PLTR, SNOW, DDOG, MDB	+8.0%	+25.8%	0.0%	4 of 4	94
GPUs and accelerators	NVDA, AMD	−3.0%	+13.3%	−5.3%	2 of 2	27
Hyperscalers	MSFT, GOOGL, AMZN, META	+2.9%	+7.8%	−7.6%	4 of 4	23
Foundry	TSM	−4.1%	+7.8%	−6.7%	1 of 1	21
Networking equipment	ANET, CSCO, CIEN	+5.5%	+1.0%	−8.2%	3 of 3	38
Power and cooling	GEV, VRT, CEG, VST	+8.5%	−0.7%	−24.4%	3 of 4	24
Neoclouds and AI hosting	CRWV, NBIS, ORCL	−8.4%	+5.0%	−42.6%	1 of 3	loss-making / 13 (ORCL)
Custom silicon and connectivity chips	AVGO, MRVL, CRDO, ALAB	+0.1%	−5.6%	−26.5%	3 of 4	30
Chip equipment	LRCX, AMAT, KLAC, ASML	−5.8%	−7.5%	−28.1%	4 of 4	28
Memory	MU, SK hynix, SNDK	−8.0%	+9.8%	−15.2%	2 of 2 measured	5
Hard-disk storage	WDC, STX	−6.1%	−18.7%	−37.5%	1 of 2	13

The Nasdaq-100 (QQQ) rose 0.2% in the week and 5.7% over three months.

• Software and security are at their highs. Ten of the ten names in those two baskets are above their 200-day averages, and most are within 1% of a 52-week high. They are also the most expensive groups in the buildout by some distance.
• Hardware has cooled without collapsing. Memory has a median forward P/E of 5 because its earnings are at a cyclical peak. Quarterly revenue at MU, SK hynix and SanDisk is running between 3.6 and 4.8 times its level a year earlier. Hard-disk storage fell after Toshiba said on 2 October that it plans to double its drive capacity by its 2027 financial year (Nikkei Asia), which is the kind of supply response that usually ends a pricing upswing. Chip equipment is down 28% from its highs, though every name is still above its 200-day average.
• The debt-funded part is where the damage is. Oracle is 54% below its high and CoreWeave 43%. The financing risk set out in our September post now shows up in prices.

What changed since July

• Rates. The Fed raised rates on 16 September. The 10-year Treasury yield touched 5.36% on 7 October, its highest since 2002. Higher rates make borrowed capex dearer and lower what investors pay for long-dated profits.
• Questions about who pays. On 8 October the Financial Times reported OpenAI's annual revenue run rate at about $50 billion, below the roughly $68 billion reported earlier. Much of the gap is gross against net accounting, and Bloomberg reported a $70 billion-plus year-end target, but Broadcom, Oracle and CoreWeave fell that day (CNBC).
• Security spending. Cybersecurity stocks jumped on 14 September, after Anthropic chief executive Dario Amodei published an essay on the risks of frontier AI. Company news, including Zscaler reaffirming its guidance on 9 October, kept them rising.
• The deal window narrowed. Global M&A by value fell 41% from the second quarter to the third, and several US IPOs were postponed near the end of September, with AI-spending worries among the reasons given (LSEG via Reuters; Renaissance Capital).

Three scenarios for the next 6 to 12 months

1. Monetisation: the rotation continues (about 45%). Hyperscaler capex holds near guidance, and the revenue shows up in cloud, software and security. Hardware growth slows from a peak, and memory and storage prices level off as new supply arrives. The market keeps paying for the companies that turn AI into revenue. What would confirm it: the hyperscalers' third-quarter calls in late October holding capex while cloud growth accelerates.

2. A second capex leg: hardware leads again (about 20%). The third-quarter calls guide 2027 capex above $1 trillion and the OpenAI revenue worry fades. The rotation reverses. Chip equipment, memory and custom silicon recover first, and the expensive software names pause. What would confirm it: capex guidance raised on the calls, and TSMC (15 October) raising its own spending or AI revenue outlook.

3. Financing squeeze (about 35%). Yields stay above 5% and credit spreads keep widening. The IPO window stays shut. Spending paid for with borrowed money slows first: neoclouds, Oracle and private AI labs. Hardware orders follow. Software demand may hold, but at forward P/Es of 70 to 95 the multiples fall on higher rates alone. Cash-rich hyperscalers hold up best. What would confirm it: a hyperscaler or a large lab trimming plans, high-yield spreads moving past their April levels, or a neocloud struggling to refinance.

Against July's weights (50% base, 25% bull, 25% bear), we have moved weight towards the bear case. The mechanism it depends on, financing, is now visible in prices, and rates have risen. These weights are judgement.

How each sub-sector fares

Sub-sector	Monetisation	Second capex leg	Financing squeeze	Holds up in
Hyperscalers	Strong	Mixed: up if revenue outgrows capex, down if not	Holds up best	2 of 3 (3 if revenue keeps pace)
Cybersecurity	Strong	Neutral	Demand holds, valuation at risk	2 of 3
Networking equipment	Good	Strong	Weak	2 of 3
Power and cooling	Good	Strong	Mixed (contracted power holds, new builds slip)	2 of 3
AI data and analytics software	Strong	Neutral	Weak (highest multiples)	1 of 3
GPUs and accelerators	Neutral	Strong	Weak	1 of 3
Foundry	Neutral	Strong	Weak	1 of 3
Custom silicon and connectivity chips	Neutral	Strong	Weak	1 of 3
Chip equipment	Weak	Strong	Weak	1 of 3
Memory	Weak (peak earnings)	Strong	Weak	1 of 3
Hard-disk storage	Weak (new supply)	Good	Weak	1 of 3
Neoclouds and AI hosting	Neutral	Strong	Hit hardest	1 of 3

Sub-sectors of interest

The groups worth a closer look are the ones that hold up in more than one scenario:

• Hyperscalers hold up in the monetisation and financing-squeeze cases. They fund most of their capex from cash flow and collect the cloud revenue. The second capex leg is less clear for them, because they are the ones paying for it (see the test below). At a median forward P/E of about 23, they are priced well below the software groups.
• Cybersecurity has the most durable demand of the strong groups. Security spending tends to hold up in a slowdown, and AI gives attackers new tools as well as defenders. The risk is the price. CRWD trades at about 51 times sales.
• Networking equipment and power benefit as long as the buildout keeps going, in any form. They sit between the software winners and the hardware cycle.

The hardware groups (memory, storage, chip equipment) only do well if the second capex leg arrives. At about 20%, that is the least likely of the three cases. The sharp fall in storage has not by itself made those names cheap: their earnings are at a peak, so their low P/Es could rise if earnings fall.

The test for the hyperscalers: does revenue outgrow capex?

A second capex leg is good news for the companies that sell the equipment. For the companies that buy it, the market's reaction has depended on whether revenue grew fast enough to make the spending look small. In July 2026 it did: Amazon rose 15.3% on its results day as AWS growth picked up alongside the spending, and Alphabet recovered a 7.1% results-day fall within a month (covered in our August post). When spending rose without that revenue, the reaction went the other way. Meta fell 24.6% on its results day in October 2022.

So far capex is the heavier side. In the latest quarters, capex grew several times faster than revenue at all four:

Latest quarter	Revenue growth (year on year)	Capex growth (year on year)	Capex as % of revenue (a year earlier)
Alphabet, Q2 2026	+24.2%	+100%	37.5% (23.3%)
Microsoft, Q2 2026	+17.7%	+110%	39.8% (22.3%)
Meta, Q2 2026	+28.0%	+82%	49.5% (34.8%)
Amazon, Q1 2026	+16.6%	+77%	24.4% (16.1%)

Alphabet gives one reference point. Against the same quarter a year earlier, its revenue rose $23.4 billion and its capex $22.5 billion, roughly dollar for dollar.

What to look for in the late-October results:

	Revenue outgrowing capex	Capex heavier
Capex as % of revenue	Flat or lower than last quarter	Higher again
Revenue growth	Accelerating	Flat or slowing
Cloud segment growth	Accelerating	Slowing
2027 capex guidance	Held, or raised alongside faster cloud growth	Raised on its own

The likeliest outcome is some of each, which is why the second-leg column reads "mixed" for the hyperscalers.

Optionality the capex test cannot see

Some of the hyperscalers' value sits in businesses that barely show up in quarterly revenue yet.

• Waymo (Alphabet). Raised $16 billion in February 2026 at a $126 billion valuation, with Alphabet supplying most of the capital. It reported 400,000 paid rides a week in February and targets 1 million a week by the end of 2026. The valuation is about 3% of Alphabet's market value, and its revenue is still small.
• Willow (Alphabet). A 105-qubit quantum chip whose error rate fell as it scaled, a first reported in December 2024. It has no revenue and works on a timescale of years.
• Warehouse robotics (Amazon). Amazon has committed up to $25 billion to robotics-led warehouses. Morgan Stanley estimates savings of about $10 billion a year by 2030 if 30% to 40% of US orders go through them. That is an analyst estimate, but it bears directly on the test above: lower fulfilment costs raise operating cash flow, which makes the same capex easier to carry.

None of these changes the next quarter's numbers. They are reasons some investors hold these companies through a capex-heavy stretch, and each one could reprice the parent on news.

For put sellers

A rally and a well-paid put are different things. On our 10 October readings, implied volatility was below recent realised volatility on CrowdStrike (0.76), Zscaler (0.71) and Snowflake (0.83). Selling puts on names that have just run up hard is often the underpaid side of the trade. The falling hardware names carry richer premium, but several sit in falling trends, which is the risk discussed in value traps. Breadth across the whole market is covered in our S&P 500 breadth update.

The levelbox screener updates trend, quality and premium for each name daily, at your own delta and account size.

Dates that will move these weights: TSMC results (15 October), the hyperscaler results (around 28 to 29 October), the Fed decision (28 October), and Micron's next guidance.

Sources for this update

• Prices, returns, 200-day averages: daily closes to 9 October 2026, our own calculation. Forward P/E and revenue growth: point-in-time data-provider snapshot, 10 October 2026; treat as approximate.
• Treasury yields: US Treasury daily par yield curve; NBC News (7 October 2026). Fed decision: Federal Reserve statement (16 September 2026).
• OpenAI revenue report and the 8 October moves: CNBC (8 October 2026).
• Toshiba drive capacity: Nikkei Asia (2 October 2026).
• Cybersecurity rally: Yahoo Finance / 24/7 Wall St. (14 September 2026).
• M&A and IPOs: Reuters via Zawya, LSEG data (October 2026); Renaissance Capital third-quarter 2026 IPO review (1 October 2026).
• Implied versus realised volatility: levelbox wheel-candidate gather, 10 October 2026.
• Hyperscaler revenue and capex by quarter: company quarterly filings via a data-provider snapshot (10 October 2026); treat as approximate. Results-day price moves: our own calculation from daily closes.
• Waymo funding and rides: Wards Auto via Yahoo Finance and electrive (3 February 2026). Willow: Google's December 2024 announcement as reported by Maginative. Amazon robotics: Benzinga (February 2025) on the $25 billion commitment and the Morgan Stanley estimate.

This update is analytical and educational and is not financial advice. Sub-sector baskets and tickers are examples used to define each group, not recommendations. Probabilities are AI-assisted judgement. Figures are a point-in-time snapshot and can contain errors. This update was generated with AI assistance. If you spot a mistake, tell us.

Point-in-time narrative on a reproducible fact pack. Full citations with tiers live in analysis.json and factpack.json.