State of the AI Cloud
Who owns the GPUs, who rents them, and who gets paid — the compute-CAPACITY stack (hyperscaler cloud · neoclouds · crypto→AI-DC · the capacity-deal graph · orbital frontier) priced as toll-collectors, with the whole board a single AI-capex factor bet.
Own the best-capitalized, most-diversified toll-collectors of AI compute capacity — hyperscaler clouds with in-house accelerators (GOOGL/TPU, AMZN/Trainium) and the de-risked miner-pivots (IREN's Microsoft anchor, Google-backstopped WULF/CIFR) — and be selective-to-short on single-tenant concentration (ORCL ~54% OpenAI) and on cash burn priced as growth (SPCX spends 235% of revenue on capital; its sales multiple has already re-based from ~87x to ~56x). Rent the compute story, don't overpay for the concentration.
In 2026 the scarce asset is no longer the GPU — it is energized, contracted compute CAPACITY, and demand outstrips supply at every provider. Big-four hyperscaler AI capex is tracking ~$725B (up ~77% YoY; ~$700-900B incl. Oracle, Tom's Hardware/ValueAddVC Q2 2026), yet Microsoft still carries an ~$80B unfulfillable Azure backlog for lack of power, and even model-first Meta is standing up 'Meta Compute' to resell excess GPUs (reported 1 Jul 2026). CORRELATED-EXPOSURE WARNING: this whole board is secretly ONE bet — that the AI-datacenter capex super-cycle keeps compounding. Two consecutive hyperscaler capex guide-downs reprice every layer at once, hyperscalers to miners; diversifying across the five layers does NOT diversify this single factor. House call: own the toll-collectors with the deepest, best-capitalized tenant books and de-risked power; underweight/short single-tenant concentration and stretched multiples; treat the orbital frontier (Layer 5) as watchlist-only venture-exposure with NO investable public buy. US-first; every position carries a bull AND a bear; private names (OpenAI, Anthropic, Fluidstack, Crusoe, Lambda, Together, Reflection, Etched, Starcloud) are analyzed but flagged NOT ownable.
No single provider dominates the capacity market the way a foundry dominates the chip stack — capacity is a distributed oligopoly where the SAME names appear on both sides. The most durable positions are the hyperscalers that own a differentiated in-house accelerator (Google/TPU, Amazon/Trainium): they cut their own cost AND rent a Nvidia-alternative, funded by a profitable core, so they collect the capacity toll whoever wins the model war. Anthropic's up-to-1M-TPU + up-to-5GW-Trainium commitments are the cleanest proof the merchant-silicon renter base is real. The fragile positions are single-tenant-concentrated (Oracle ~54%/~$300B RPO on one unprofitable counterparty; every miner-pivot on one neocloud) and the cash-burning non-traditional provider (SpaceX, whose sales multiple has re-based to ~56x annualised revenue but which spent 235% of revenue on capital in the June 2026 quarter, on partly cancellable leases). Dominance here is a function of tenant-book quality and power de-risking, not of a single technology chokepoint.
Full-stack value chain
The AI compute-CAPACITY stack from the hyperscaler cloud down to the crypto→AI-DC power operators, the capacity-deal graph that binds them, and the orbital frontier. The marker shows where the capacity crunch is a tailwind, headwind, or mixed for each layer's margin pool. Tap any layer for the full read — and remember the whole board is one AI-capex factor bet.
The best-capitalized toll-collectors — hyperscalers with in-house silicon (GOOGL/TPU, AMZN/Trainium) and the miner-pivots on already-energized power (IREN, Google-backstopped WULF/CIFR) — get stronger as the capacity crunch bites, because energized MW and a diversified tenant book, not venture debt, set the price. The levered neocloud middle and the single-tenant-concentrated names (ORCL) do not. Layer 5 (orbital) is watchlist-only — no public buy.
Shift-point register
Ranked by margin-at-stake × demand-durability × evidence-strength. The flagged rows clear the bar and are promoted to a deep-dive.
In-house accelerator rental (TPU / Trainium)
The second axis of the capacity market opens — GOOGL and AMZN rent DIFFERENTIATED in-house silicon (TPU, Trainium) as a Nvidia-alternative, and Anthropic's giant commitments prove the external renter base is real
Google and Amazon are the only hyperscalers renting a self-designed accelerator at scale — Google's TPU (8th-gen) and Amazon's Trainium — which both cuts their own cost and lets them undercut GPU clouds. The landmark validation is Anthropic: up to 1M Google TPUs (>1GW, tens of $B, tied to Google's ~$40B investment, fact) AND a $100B / up-to-5GW AWS Trainium deal through 2036 (~2.5-3M XPUs; Claude already runs on >1M Trainium2 chips, fact). GCP revenue +~63% YoY with cloud backlog ~$462B (nearly doubled sequentially); AWS re-accelerated to ~28% YoY (fastest in 15 quarters) with a >$225B Trainium revenue-commitment book, Trainium3 nearly fully subscribed (fact, Q1 2026).
Opportunity board
Own the toll-collector, not the tenant — the best-capitalized, most-diversified capacity providers; be selective-to-short on single-tenant concentration and stretched valuations. Every position carries a bull AND a bear; orbital is venture-watch only.
Durable compounder
2Undervalued / high-potential
4Short / avoid
1Views & the Voices
The strongest bull and the strongest bear case, the US-listed ETF expressions of the theme, and where the tracked QAI Voices roster nets out — each stance stamped trackable vs inferred.
Bull case
the longOwn the toll-collector, not the tenant — in 2026 the scarce asset is energized, contracted compute CAPACITY, demand outstrips supply at every provider, and the best risk-adjusted longs are the providers with the deepest, best-capitalized, most-diversified tenant books, funded by proven core businesses rather than venture debt. The safest capacity-toll margin sits with the hyperscalers that own a differentiated in-house accelerator (GOOGL/TPU, AMZN/Trainium) — they cut their own cost AND rent a Nvidia-alternative — and with the miner-pivots that convert already-energized, interconnect-queued power into 10-15yr credit-backstopped leases (IREN's Microsoft/Nvidia anchor, Google-backstopped WULF/CIFR). The scarce asset is the SITE'S POWER, not its ASICs.
- The capacity scarcity is real and power-gated: Microsoft carries an ~$80B unfulfillable Azure backlog for lack of power, GCP cloud backlog nearly doubled to ~$462B in one quarter, and Oracle RPO hit ~$523-553B — demand outstrips supply at every provider, and energized MW is the binding ceiling (fact, Q1 2026).
- In-house accelerators are the emerging second axis: GOOGL (TPU) and AMZN (Trainium) both cut their own cost AND rent a Nvidia-alternative, and Anthropic's up-to-1M-TPU (>1GW) + $100B/5GW-Trainium commitments prove the external merchant-silicon renter base is real — the cleanest validation on the board (fact, 2026).
- The miner-pivot re-rating is a genuine mispricing: where power was already energized and interconnect-queued, IREN (direct Microsoft ~$9.7B + Nvidia ~$3.4B, owned-GPU AI Cloud) and the Google-backstopped pair (WULF ~360MW + ~14% Google equity, plus a 20-yr ~401MW/~$19B lease signed directly with Anthropic at Justified; CIFR ~207MW of contracted critical IT load at Barber Lake) convert volatile hash-price revenue into contracted 10-15yr leases at far richer multiples (fact, 2026).
Bear case
the short / avoidThe whole board is secretly ONE bet — that the AI-datacenter capex super-cycle keeps compounding — and it is financed against a revenue base that doesn't exist yet, underwritten by a handful of private, unprofitable tenants, with Nvidia's circular financing looping through the entire graph. RPO/backlog is a commitment, not cash: heavily back-end-loaded to 2027+ and partly cancellable (the SpaceX Colossus-2 leases are 90-day-terminable after 2026-12-31). The fragile middle (GPU-only neoclouds) is levered to a depreciating asset, the most-watched re-ratings rest on single-tenant concentration (ORCL ~54% OpenAI), and diversifying across the five layers does NOT diversify the one factor that can hit them all at once.
- The correlated factor is the whole risk: two consecutive hyperscaler capex guide-downs reprice every layer simultaneously — hyperscalers to miners — so the book's real independent risk is far below its gross, and the diversification across five layers is partly illusory (estimate, 2026).
- Single-tenant concentration is existential: ~54%/~$300B of Oracle's RPO is OpenAI/Stargate alone, and OpenAI is private + unprofitable, so ability-to-pay is unobservable; every miner-pivot rests on one neocloud counterparty (Fluidstack, CoreWeave) — a slip at one hub hollows the most-concentrated books first (fact/estimate, 2026).
- Circular financing is systemic: Nvidia is investor, supplier AND (via sale-leaseback at Lambda/Together) customer, and the same tenants underwrite most backlogs — io-fund flags the loop as fragile; capital intensity at ~45-57% of revenue + ~$1.5T projected 2027 debt issuance is the overhang (estimate, 2026).
Bullish-but-bifurcating — net long the best-capitalized capacity toll-collectors (hyperscalers with in-house silicon GOOGL/AMZN; de-risked energized-power miner pivots IREN/WULF/CIFR), short/avoid the single-tenant concentration (ORCL ~54% OpenAI) and the cash-burning, cancellable-backlog names (SPCX ~56x annualised revenue with capital spending at 235% of it), and underweight the levered neocloud middle (CRWV the canary).
Dispersion: Moderate-to-high. The operators + 13F smart money (Nadella, Jassy, Coatue, Baker, Aschenbrenner-long-leg) cluster bullish on the capacity toll-collectors and confirm the rotation into providers + power. The dispersion is on the FINANCING and CONCENTRATION legs: the forensic/macro bears (Chanos, io-fund, Cembalest, Kedrosky) attack GPU depreciation, circular financing, hyperscaler leverage, and OpenAI ability-to-pay, while the balancers (Cahn on the capex-revenue gap, Forbes on OpenAI dependence) flag the specific fragilities. Aschenbrenner uniquely encodes the split as a position — long the physical capacity layer, put-hedged on the chips.
The consensus is NOT 'buy the neocloud'; it is 'own the best-capitalized, most-diversified toll-collector — the hyperscaler with in-house silicon and the miner-pivot on already-energized power — and be selective-to-short the single-tenant concentration and the stretched, cancellable-backlog valuations.' Where bulls and bears actually agree: capacity demand outstrips supply at every provider (Microsoft's ~$80B unfulfillable backlog is the shared fact), and energized power is the binding ceiling. They split on whether the ~$725B+ capex against tens-of-billions of AI revenue is buildable from the FCF buffer (bull) or a credit-and-concentration event waiting on the neocloud refinancing wall + OpenAI ability-to-pay (bear). Highest-conviction shared call: CoreWeave is the levered, depreciating-collateral canary; GOOGL/AMZN (in-house silicon) + the de-risked miner pivots are the defended longs; ORCL/SPCX are the concentration/valuation shorts.
US-listed ETF expressions
7The QAI Voices
stance · trackable / inferredPrediction matrix
Directional calls across Sep'26 / Dec'26 / Jun'27 / Jun'28, confidence decaying high → low over the horizon. Each cell is the call; click a row for the full reasoning, leading indicator, and falsifier.
Across Sep'26 → Jun'28 the AI compute-capacity market stays a demand-pull, supply-constrained regime: big-four AI capex tracks ~$725B in 2026 (up ~77% YoY), analysts project $1T+ in 2027, and energized power — not GPUs — is the binding ceiling (Microsoft's ~$80B unfulfillable Azure backlog). The safest capacity-toll margin sits with the hyperscalers that own in-house silicon (GOOGL/TPU, AMZN/Trainium) and the de-risked energized-power miner pivots (IREN, Google-backstopped WULF/CIFR). The near-term swing factors are (a) whether the FY2027 hyperscaler capex-guide season (late-2026/early-2027) confirms or breaks the demand-pull, (b) whether the SpaceX Colossus-2 leases survive the 90-day cancellation window after 2026-12-31 (Google's 2026-09-30 delivery-or-kill clause is the first tell), and (c) whether OpenAI/Anthropic ability-to-pay holds under the concentration the whole graph rests on. Base case: backlogs and toll-collector economics hold through Jun'27; by Jun'28 the question shifts from 'can they book capacity' to 'does the correlated AI-capex factor keep compounding without a digestion air-pocket' — the regime most likely to crack the bull thesis at all five layers at once.
Valuation scenarios
Every target is scenario-conditional with a probability; the verify produced zero outright buys. Tap a name for its full bull / base / bear ladder.
Regime calls
The four cross-cutting forces and when each bites across the Sep'26 → Jun'28 horizon. Tap any force or modulating risk to read the full call.
Compute
Demand pulls hard, energized power stays the binding ceiling, and the toll-collectors with the deepest tenant books (GOOGL/AMZN) + de-risked power (IREN/WULF/CIFR) collect the capacity rent. Own the toll-collector, not the tenant; the neoclouds are the fragile, levered middle. Holds through Jun'27 with high confidence; by Jun'28 the question is whether the correlated AI-capex factor keeps compounding.
AI
Hyperscalers cut FY2027-28 capex on AI-ROI doubt or efficiency gains — every layer de-rates at once, hyperscalers to miners, because diversifying across the five layers does NOT diversify this one factor. The macro tail that breaks the demand-pull premise; watch two consecutive hyperscaler capex guide-downs and CoreWeave book-to-bill as the leading tells.
Concentration/Counterparty Break
A slip at a private demand hub (OpenAI ability-to-pay, Anthropic renegotiation) hollows the most-concentrated books first — ORCL (~54% OpenAI), then the miner-pivots on one neocloud — and cascades through the circular-financing loop (Nvidia on every side). Triggered by an OpenAI funding wobble or a Stargate scale-back; the sharpest single-name drawdown (ORCL).
Backlog
The SpaceX Colossus-2 leases cancel at the 2026-12-31 window (Google's 2026-09-30 delivery-or-kill first), back-end-loaded RPO slips, and names priced for flawless conversion (SPCX still ~56x annualised revenue after a ~28% fall, NBIS +167% YTD) de-rate. Hits the valuation-stretched + cancellable-backlog names specifically, leaves the profitable-core anchors relatively unscathed.
Financing
An AI-capex sentiment turn closes the refinancing window — the levered middle (CRWV the canary, the miner-pivots) faces GPU-obsolescence-on-leverage + a refinancing wall and is forced to deleverage. Hits the neocloud/miner sleeve specifically; the profitable-core hyperscalers (GOOGL/AMZN/MSFT) fund from FCF and are the ballast.
How they modulate the book
AI-capex air-pocket (the single correlated factor)
The dominant risk across all five layers. The whole board is secretly ONE bet that the AI-datacenter capex super-cycle keeps compounding — two consecutive hyperscaler capex guide-downs or an AI-capex-digestion regime reprice every layer at once, hyperscalers to miners. FY2026 capex tracked ~$725B (still rising); the binding window is the FY2027 capex-guide season (late-2026/early-2027). Diversifying across the five layers does NOT diversify this factor, so the book's real independent risk is far below gross. LEADING INDICATORS: the four hyperscalers' FY2027 capex guides, and CoreWeave book-to-bill/utilization (the neocloud canary that pre-dates provider-income-statement weakness). Express via a portfolio-level AI-capex factor budget, not name-by-name sizing.
Circular financing / vendor-financed demand
Nvidia is investor, supplier AND (via sale-leaseback at Lambda/Together) customer, and the same tenants (OpenAI, Anthropic) underwrite most provider backlogs — io-fund and others flag this loop as systemic. Capital intensity at ~45-57% of revenue and ~$1.5T projected 2027 debt issuance is the macro overhang. A break in the loop (a demand air-pocket at one hub) cascades through the whole graph faster than the back-end-loaded contracted terms suggest. Sized down by trimming gross exposure and holding cash, not by a single-name trade.
Single-tenant concentration + ability-to-pay
Oracle (~54%/~$300B RPO on OpenAI) and, to a lesser degree, AWS (Anthropic as marquee Trainium tenant) and every miner-pivot (one neocloud counterparty) carry existential exposure to one counterparty's ability AND willingness to pay. OpenAI's unprofitability is the biggest question mark — and it is private, so ability-to-pay is unobservable directly. The board expresses this as a short/underweight sleeve (ORCL) and by preferring diversified tenant books (GOOGL/AMZN) over concentrated ones.
Backlog is a commitment, not cash — and often cancellable
RPO/backlog is heavily back-end-loaded to 2027+, and the marquee SpaceX leases carry 90-day termination rights after 2026-12-31 plus GPU-delivery conditions (Google additionally has a 2026-09-30 delivery-or-kill clause) — so 'contracted' revenue can shrink WITHOUT a demand collapse. CONTRACTED-VS-CANCELLABLE for Colossus: only the Anthropic/Colossus-1 lease ($1.25B/mo) is effectively locked; the Google ($920M/mo) + Reflection ($150M/mo) Colossus-2 leases are the cancellable slice of the '>$76B through 2029'. The filed picture is smaller than the headline: SpaceX's AI segment booked $2.561B of revenue in the June 2026 quarter and total-company backlog is $47.5B incl. $14.3B of deferred revenue across all three segments. Mark backlog-priced names against the filed numbers, not against headline RPO.
Balance-sheet strain + GPU obsolescence on leverage
CRWV ($31-35B) and NBIS ($20-25B) 2026 capex vastly exceed current revenue, and the miners post large GAAP losses NOW (WULF ~$428M Q1'26; APLD ~$101M Q3 FY26) even as contracted revenue scales — financing risk if AI-capex sentiment turns. Owned-GPU models (IREN, neoclouds) additionally carry depreciation/technology-obsolescence risk on leveraged fleets: a faster Nvidia refresh cadence or a shift in accelerator economics strands capital. CRWV is the live tell — the refinancing wall is the leading stress indicator for the whole levered middle.
Geography / framing caveats (US-first, no China)
US-first per the mandate: Layer 1 is all five US-listed providers. NBIS (Netherlands, restructured ex-Yandex — explicitly NOT China), IREN (Australia, US-asset-heavy at Childress/Sweetwater TX), and SFTBY (Japan, named for analysis only) are non-US and flagged as such. No mainland-China-listed equity is framed as ownable or shorted anywhere on this board. The private demand hubs (OpenAI, Anthropic, Reflection) and marquee operators (Crusoe, Lambda, Together, Etched, Starcloud) are US companies but PRIVATE — proxy exposure (MSFT/ORCL/NVDA for OpenAI; AMZN/GOOGL for Anthropic) is diluted across giant businesses and carries its own dilution risk.
Orbital frontier is venture-watch, NOT investable (Layer 5)
Every Layer-5 pure-play (Starcloud, Axiom, Sophia, Lonestar, OrbitsEdge) is private and pre-revenue; public exposure is indirect and immaterial (GOOGL Suncatcher optionality, NVDA arms-dealer, small-caps SIDU/LUNR already on the space-launch board). The economics are launch-cost-gated (~$200-500/kg needed vs ~$1,500/kg Falcon 9 today) and MW-scale radiative cooling (~1,200 m²/MW in vacuum) is unsolved. Manufacturing a 'public proxy' to force exposure is the trap — treat any 'operational orbital/lunar datacenter' framing as aspirational; no position on this board.
Premise pressure-test
The six named 2026-Q3 catalysts the thesis rests on, probability-weighted. Click any premise for the if-true / if-false split.
AI-datacenter capex keeps compounding (THE factor)
holdingPower — not GPUs — is the binding ceiling
holdingIn-house accelerators become a real second capacity axis
holdingThe demand hubs (OpenAI, Anthropic) can pay
at-riskBacklog converts — 'contracted' isn't cancelled
at-riskDebt-fueled capex stays financeable
holdingUS-first exposure holds without offshore reach
holding