A holistic read across AI-exposed equities, credit and private markets — live data where it exists, curated tables where it doesn't, always labelled which is which. Descriptive only. Not a signal, not advice.
27 instruments across 13 layers: AI compute, custom silicon, foundry, memory (Micron plus Korea's two HBM co-leaders, Samsung and SK Hynix), hyperscalers, networking, AI applications, power and grid infrastructure, cloud compute, one public AI-cloud pure-play, China's low-cost-model layer, and two broad credit ETFs standing in for the debt side of the buildout. Rule-based: breadth is a moving-average count over the equity subset only, credit ETFs excluded from that stat.
Why a custom basket rather than just SOXX or IGV: a single ETF hides overlap. GOOGL sits in "Hyperscaler" for capex and cloud revenue, but its TPU programme quietly competes with NVDA in Compute. AMZN's AWS is an AI infrastructure business classified outside tech ETFs by sector. VRT's power and cooling exposure isn't captured by either SOXX or IGV at all. The ETF column below shows the closest proxy per name — not where BK/IQ thinks it "belongs."
Why credit is here at all: S&P Global Ratings has flagged rising capex/revenue intensity at the largest hyperscalers as a credit question, and Oracle's ~$50bn 2026 debt raise (tied to its OpenAI cloud commitment — see circular deals below) is a direct example. LQD and HYG are broad proxies, not issuer-specific instruments; a widening in HY relative to IG while AI capex commentary intensifies is the kind of divergence worth noticing, not a precise attribution.
Why VRT is here: US data-center grid-power demand rose roughly 22% in 2025 alone (+11.3 GW, to ~61.8 GW) and is projected to reach ~134.4 GW by 2030 — more than double today's draw (S&P Global Commodity Insights, Oct 2025). Power and cooling capacity is the physical constraint behind every capex number elsewhere on this page.
| vs. off-the-shelf ETFs | 1D | 1M | 3M | Vol (21D ann.) |
|---|---|---|---|---|
| Loading… | ||||
Same calculation applied to the basket's equal-weighted average and to SOXX/IGV/QQQ — any gap is the basket's different membership, not a different formula. Answers "does a custom basket actually diverge from just buying the sector ETF."
Loading…
Each layer's line above its table is a plain mean/breadth count across that layer's own names — not the LLM. Vol and 90D drawdown in the tables are each name's own risk figures, not the basket aggregate. P/E and beta are per-name fields from Yahoo Finance directly (not BK/IQ-computed) — a "—" means the field wasn't reported for that name this run, the company is loss-making (P/E specifically), or it's LQD /HYG, where these fields aren't meaningful. The basket-wide P/E and beta stats above are medians, not averages, so one extreme name (a near-breakeven name can print a P/E in the thousands) can't distort the headline number — the per-name table still shows its real figure. Swipe a table sideways on mobile to see all columns.
No live pricing exists for these — valuations are last-round post-money marks, not tradable quotes, and move in discrete jumps rather than daily. Updated by hand when a new round or filing lands.
| Company | Last valuation | Note |
|---|---|---|
| Anthropic | ~$965bn | Series H, May 2026, ~$47bn ARR reported. Filed for IPO in early June 2026 — not yet listed. |
| OpenAI | ~$852bn | $122bn round, Mar 2026, ~$25bn ARR reported. Filed for IPO June 8, 2026, reportedly targeting ~$1tn at listing — not yet listed. |
| xAI | ~$230bn | Series E, Jan 2026. Since merged with SpaceX; combined-entity figures vary by source and are harder to isolate. |
| Databricks | ~$188bn | Coatue-led round, Jul 2026, ~$5bn ARR reported. |
Figures compiled from press reporting (CNBC, TechCrunch, Bloomberg and others), not primary disclosure, and different outlets show mild variance in timing and exact size. The more important fact than any single number: two of the four largest private AI companies have now filed to go public. If either lists, it moves from this table into the live equities section above — worth revisiting this table on that news rather than treating it as fixed.
A recurring 2026 pattern: a chip or cloud vendor invests in an AI lab, the lab commits to buying compute from cloud providers, and those providers use the revenue to buy chips from the same vendor. Reported commitments across the network are estimated by various analysts at well over $800 billion in aggregate.
| Counterparty | Reported commitment | Role in the loop |
|---|---|---|
| Oracle | ~$300bn | Five-year cloud deal with OpenAI; Oracle in turn buys Nvidia chips to build the capacity it owes. |
| Nvidia → OpenAI | ~$100bn | Planned Nvidia investment into OpenAI, which is also Nvidia's largest disclosed AI-lab customer. |
| Microsoft | ~$250bn | Azure compute commitment from OpenAI; Microsoft is also an OpenAI investor and leases back Nvidia capacity via CoreWeave. |
| CoreWeave | ~$22bn | OpenAI cloud commitments; Nvidia holds a large equity stake in CoreWeave and is also its primary chip supplier. |
| AMD | ~$90bn | Multi-year compute deal with OpenAI structured with equity-like warrants tied to deployment milestones. |
| Broadcom | ~$10bn+ | Custom AI silicon commitment, one piece of OpenAI's broader stacked infrastructure spend. |
These figures are press- and analyst-reported, not uniformly confirmed by audited disclosure, and several are explicitly conditional or staged. OpenAI's 2026 revenue was reported near $20bn against a committed spend stack analysts put near $1.4tn over the contract lives — the gap those two numbers describe is the thing worth tracking, not any single deal in isolation. Vendor financing is common in capital-intensive industries and isn't inherently improper; the distinction is whether compute gets used by paying end-customers or mainly sustains the loop itself.
Recurring patterns in how AI-related results get reported and read — not specific to any one company or newsletter.
Vendor-defined metrics aren't audited revenue.
Figures like "agentic work units" or "% of new bookings driven by AI" are company-defined, not standardised or reconcilable to a GAAP line. They can move independently of what actually shows up in revenue.
Backlog growth is a demand signal, not cash.
Remaining performance obligations and bookings show intent to pay, not payment. Financing terms and contract structure can separate booked demand from realised cash flow for several quarters.
Watch the support ratio, not just the growth rate.
If supply commitments, receivables, or vendor financing are growing faster than the revenue they sit behind, growth may be increasingly financed rather than organically funded. Compare the two rates, not just the headline number.
Efficiency gains can shrink demand even as usage grows.
Cheaper inference means the same workload can run on less hardware. Rising application revenue doesn't mechanically imply proportionally rising compute demand upstream — the two can decouple.
One company's print is not a sector's proof.
A single strong or weak result is often used to confirm a whole thematic layer. Small samples and concentrated customer bases make individual names weak proxies for broad claims.
A rebound can be starting-valuation, not new information.
A stock recovering after being deeply out of favour can reflect how low expectations had fallen, rather than a genuine shift in fundamentals. Check the prior drawdown before reading a rally as a regime change.
Unverified trade-press claims shouldn't move position sizing.
Reports of new chip designs, spec changes, or roadmap shifts often precede or substitute for primary company disclosure. Treat a single sourced claim as a hypothesis to track, not a confirmed input.
The same names sit on multiple sides of the trade.
A handful of hyperscalers are simultaneously the largest customers, largest capital spenders, and — via investment stakes — financiers of the same suppliers. Exposure that looks diversified across "layers" can share one underlying counterparty. See the circular deals map above for named examples.
Rising capital intensity is a credit question, not just an equity one.
Rating agencies read hyperscaler capex differently than equity investors do. S&P Global Ratings has flagged capex/revenue for the largest hyperscalers rising sharply through 2026–2027, funded partly by debt — a trend that pressures credit headroom regardless of how the equity story is framed.
Falling token prices can erode the assumptions behind today's guidance.
Inference pricing has been declining as competition and efficiency both increase. Revenue models built on current cost-per-token or price-per-query levels may not hold if those unit economics keep compressing faster than volume grows to offset them.
Model quality is a moving target, not a fixed moat.
Gaps between frontier models narrow with each release cycle, and open-weight alternatives close in faster than pricing power usually adjusts. A competitive advantage stated in terms of "our model is better" is harder to underwrite than one stated in terms of data, distribution, or workflow lock-in.
Low-cost Chinese models test whether Western pricing power holds.
Chinese developers have shipped models claiming frontier-comparable benchmark results at a fraction of reported training and inference cost. Whether or not the cost claims fully hold up, the credible threat of much cheaper substitutes changes the pricing and margin assumptions embedded in Western AI application and model valuations — a competitive input, not just a geopolitical one.
Physical build-out timelines are the real constraint, not capital availability.
Committed spend can be announced faster than power, cooling and grid capacity can be built. With roughly 90 large US data centers in operation today needing to triple to around 280 by 2030 to match planned projects, the binding constraint on how fast committed capex actually converts to usable compute may be construction and permitting timelines, not the money itself.