Datacenters & Digital Infrastructure
The picks-and-shovels of the AI buildout — hyperscaler capex, neoclouds, power/cooling supply chain, and the power-vs-GPU bottleneck. A commercial-layer tracking beat at the seam of hardware, electrification, and markets.
Datacenters & Digital Infrastructure
The picks-and-shovels of the AI buildout. The founding thesis: compute is bottlenecked by power, power is bottlenecked by grid + equipment, and the whole stack is financed through capex no generalist is modelling end-to-end. This is the seam where the hardware (compute), energy/electrification, and markets beats meet.
Third compile, 2026-07-23 — 8 new sources (18 total), the topic's first primary company filings: three Q2 2026 8-Ks (GOOGL, TSLA, TXN) plus five discovery sources. The spine finding: the demand side has stopped self-funding. Alphabet's Q2 capex ($44.9B) outran operating cash flow ($39.1B) so its free cash flow went negative $(5.9)B, funded by a $49.6B equity raise plus $20.3B of notes; Tesla's capex ($5.79B, +142%) did the same, funded by a $1.2B cash drawdown. TXN is the supply-side mirror — capex rolling off, FCF +271%. Four long-standing gaps closed: kW/rack (GB300 NVL72 = 135 kW — the topic's oldest gap), grid-interconnect queue (3–4-yr waits, constraint is throughput not raw MW), capex/MW (~$59M all-in, single-operator), CoWoS capacity (gap narrows 20%→10% but doesn't close). And the topic got its first neocloud primary (CoreWeave Q1: rev $2.08B, $99.4B backlog) — which also surfaced the backlog-double-counting edge: Meta's $21B commitment sits inside CoreWeave's backlog while Meta is itself a tracked hyperscaler.
Second compile, 2026-07-22 — 7 new sources → 8 concepts + 8 entities. The constraint went plural (packaging + memory + power bind at once); the capex aggregate became contested ($610–725B, the whole gap being Microsoft); and the demand-side backlog thread opened.
Evidence health: improved this compile — the topic now has company filings. 18 sources: zero papers, but now five first-party 8-K/IR filings (GOOGL, TSLA, TXN Q2 2026; CoreWeave Q1 2026) plus the Fed MPR and NVIDIA — 7
technical-reports and 11analysispieces. The load-bearing capex/FCF/backlog claims are now primary-sourced for the names that have printed; several other claims (CoWoS capacity, capex/MW, Tesla's Cortex-specific detail) remain single-analysis or vendor-secondary and are flagged inline. 2026 capex aggregates remain forecasts; the Q2 quarterly prints are actuals. Promotion to a deep beat should still be gated on primaries, not page count — but this compile is the first that added them.
Frontier — What's Moving Now
- The demand side stopped self-funding. For the two buyers that have printed, Q2 capex now exceeds operating cash flow — Alphabet plugs the gap from capital markets, Tesla from a cash drawdown.
- The constraint stack is now quantified. kW/rack (135 kW), grid queue (3–4 yr, throughput-bound), CoWoS gap (20%→10%, doesn't close) and capex/MW (~$59M) all now have numbers — the buildout can be modelled end-to-end for the first time.
- The backlog stack has a double-counting edge. CoreWeave's $99.4B backlog includes Meta's $21B while Meta is itself tracked — the same end-demand books at two layers; overlap unquantified.
- The capex aggregate is still contested at $610–725B, with a ~$70B Microsoft discrepancy driving the dispersion — but two names now have actual Q2 prints.
- It's a macro variable and has a fourth ownership category — first-party builders (Tesla Cortex, now disclosed in MW, >205 MW onsite).
See frontier.md for the full treatment, the six tradable debates, and the Knowledge Gaps list.
Concept Map
Concepts
| Concept | Key Claim | Sources | Evidence | Last Updated |
|---|---|---|---|---|
| Hyperscaler Capex Cycle | The demand side has stopped self-funding — Q2 capex > OCF for GOOGL & TSLA; aggregate band $610–725B (contested) | 8 (3 first-party 8-Ks) | Moderate→strong on the prints; 3 unresolved aggregate conflicts | 2026-07-23 |
| Power: The Binding Constraint | Power gates the buildout — one of three simultaneous constraints; queue is 3–4 yr, throughput-bound | 8 (0 papers) | Moderate | 2026-07-23 |
| Rack Power Density | GB300 NVL72 = 135 kW/rack (~155 kW peak), ~90% liquid — oldest gap now closed | 4 (1 vendor-primary) | Moderate | 2026-07-23 |
| Data-Center Construction as a Macro Variable | DC construction moves US national statistics; productivity payoff absent | 2 (0 papers) | Moderate — best-evidenced in topic | 2026-07-22 |
| Advanced Packaging (CoWoS) Bottleneck | CoWoS gap narrows 20%→10% by end-2026 but doesn't close; ~200K industry wafers/mo | 2 (0 papers) | Weak→moderate — two analysis sources, no TSMC filing | 2026-07-23 |
| HBM & DRAM Supply Tightness | Memory scarcity evidenced by pricing power, not by supply data | 1 (0 papers) | Weak — single source | 2026-07-22 |
| Contracted Backlog & RPO | Backlog is the demand-side test; CoreWeave $99.4B holds Meta's $21B — the double-counting edge | 4 (1 neocloud primary) | Weak→moderate, one unresolved conflict | 2026-07-23 |
| Vertically Integrated Compute | A fourth ownership category; Tesla now discloses compute in MW (>205 MW onsite) | 2 (1 first-party 8-K) | Weak→moderate — compute figures now primary | 2026-07-23 |
Entities
| Entity | Type | Subsegment | Confidence | Key Fact |
|---|---|---|---|---|
| Hyperscaler Buyers | Group (MSFT/GOOGL/META/AMZN) | demand side | Low | 2026 capex band $610–725B; Microsoft figure contested by ~$70B |
| Alphabet | Company (GOOGL) | hyperscaler | High | Q2 actual: capex $44.9B > OCF $39.1B → FCF $(5.9)B; funded by $49.6B equity + $20.3B notes |
| Oracle | Company (ORCL) | hyperscaler | Low | $523B RPO against $50B 2026 capex (+136% YoY) |
| TSMC | Company (TSM) | foundry / packaging | Low | CoWoS + N3 sold out through end-2026; gap 20%→10% (TrendForce) |
| Samsung | Company (005930.KS) | memory / foundry | Low | Preliminary Q2 op. profit ~₩89.4tn (~19x YoY); full results 2026-07-30 |
| Tesla | Company (TSLA) | first-party builder + cooling vendor | High | Q2 actual: AI compute now in MW (>205 MW onsite); capex $5.79B → FCF $(1.1)B (cash-funded) |
| CoreWeave | Company (CRWV) | neocloud / GPU operator | High | Q1 actual: rev $2.08B, $99.4B backlog, capex $7.70B (~3.7×), net loss $(740)M |
| Vertiv | Company (VRT) | power & cooling | Medium | NVIDIA-named cooling vendor; the "cooling re-rate" name |
Timeline
See timeline.md — chronological developments, plus a clearly separated forward calendar (SK Hynix 2026-07-29 · Samsung full results 2026-07-30; Amazon/Microsoft/Meta Q2 prints late Jul).
Sources
Commercial Layer
- Capex — hyperscaler capex, capex per MW (~$59M, added), neocloud build cost (CoreWeave), the Q2 prints, financing
- Bottlenecks — power/grid, grid-queue (3–4 yr) and rack density (135 kW) now added, transformers, cooling
- Supply chain · Market structure · Positioning
Note: capex.md and bottlenecks.md were refreshed this compile (2026-07-23) with the capex/MW, neocloud-build-cost, grid-queue and rack-density figures. supply-chain.md, market-structure.md and positioning.md still date to the 2026-06-24 compile and should be refreshed before any /kb report datacenters run.
Research Frontier
See frontier.md — six active frontiers, six tradable debates, and the Knowledge Gaps list feeding the next discovery cycle.