Datacenters & Digital Infrastructure — Research Frontier

The picks-and-shovels of the AI buildout. The founding thesis was: compute is bottlenecked by power, power is bottlenecked by grid + equipment, and the whole stack gets financed through capex no generalist is modelling end-to-end. The 2026-07-22 compile made the first clause plural; the 2026-07-23 compile put a quantified backbone under the third — capex now has primary Q2 prints, and the master unit (capex/MW), the queue, the rack, and the packaging gap all have numbers.

Evidence quality — read this first

Eighteen sources, zero papers — but now five company filings. The 2026-07-23 compile added the topic's first primary filings: three Q2 2026 8-Ks (GOOGL, TSLA, TXN) and CoreWeave's Q1 2026 IR release, joining the Fed MPR and NVIDIA (7 technical-reports, 11 analysis). Practical consequence:

• The strongest evidence is now the Q2 capex/FCF prints (first-party 8-Ks) and the Fed MPR + ConstructConnect macro pair — strong to moderate.
• Several claims remain weak: CoWoS capacity (TrendForce, secondary — no TSMC filing), capex/MW (single-operator Crusoe pricing, as-of ~Apr 2025), Tesla's Cortex-specific detail (500 MW / Megapacks — blog-sourced; the print governs on conflict).
• The GB300 kW/rack figure (135 kW) is vendor-primary — authoritative for the spec, not independent.
• 2026 capex aggregates remain forecasts (no summed actual until the year closes) — but the quarterly Q2 prints for GOOGL/TSLA/TXN are actuals.

2026-09-24 update: 24 sources now, 10 of them technical-report. The three added this compile are an intergovernmental forecast (IEA — moderate), a self-reported operator survey (Uptime — moderate for direction across a large sample, never for any single site), and an operator's own efficiency page (Google — moderate, best case by design). None was measured by anyone outside the organisation that published it.

Active Frontiers (compiled 2026-07-23; 7 added 2026-09-24)

1. The constraint went plural — and is now quantified

Status: Rapid change · Key sources: Carbon Direct, Lenovo GB300, TrendForce, TSMC Q2, Samsung · Key concepts: Power · Advanced Packaging · Rack Power Density

Packaging, memory and power bind simultaneously — capital is the only abundant input. This compile put numbers on each: the grid queue is 3–4 years in DC load zones (>300 GW across ~1,500 projects), and the binding constraint there is queue throughput/deliverability, not raw MW (the queue holds enough "on paper" through 2030). CoWoS's gap narrows 20% → ~10% by end-2026 but does not close — packaging stays binding through year-end even as industry capacity quadruples toward ~200K wafers/mo. And the rack is 135 kW (GB300 NVL72, ~90% liquid), the physical unit everything upstream and downstream is denominated against. Power still has the strongest evidence base; what's new is that all three constraints now have sourced magnitudes rather than a qualitative "sold out."

Open problems: sequencing (which binds first, for whom); who holds the booked CoWoS queue; HBM allocation by supplier; whether behind-the-meter genuinely relieves the queue or just skips it for the well-capitalised.

2. The demand side has stopped self-funding

Status: Rapid change · Key sources: GOOGL Q2, TSLA Q2, TXN Q2, Futurum, Value Add VC · Key concept: Hyperscaler Capex Cycle

The first primary Q2 prints answer this page's central financing question for two buyers. Alphabet's Q2 capex ($44.9B, calendar-Q2 cash PP&E) outran operating cash flow ($39.1B) → free cash flow negative $(5.9)B, plugged from capital markets ($49.6B equity + $20.3B notes; LT debt doubled to $98.2B). Tesla did the same ($5.79B capex > $4.7B OCF → FCF $(1.1)B) but self-funded from a $1.2B cash drawdown. TXN is the supply-side mirror — a seller into the buildout whose own capex is falling as its fab wave crests, sending FCF +271%. The seam: buyers' funding gaps widen while the equipped, subsidised supplier harvests the cash. The aggregate is still a contested band ($610–725B, the whole gap being Microsoft's ~$70B), but two names now print actual quarterly capex — and both exceed internal cash generation. Alphabet's FY capex dollar guide and RPO are absent from the 8-K (call/10-Q items) — unresolved by the print, not invented.

Open problems: whether Amazon/Microsoft/Meta show the same signature when they print (late Jul); how long negative FCF can run before it constrains the buildout; resolving Microsoft's ~$70B aggregate discrepancy with a primary source.

3. Who pays the rent — backlog, now with a double-counting edge

Status: Rapid change · Key sources: CoreWeave Q1, Futurum, GOOGL Q2 · Key concept: Contracted Backlog & RPO · Key entities: CoreWeave · Oracle

The KB now holds a backlog figure at the neocloud layer: CoreWeave's $99.4B (Q1 2026), beside Oracle's $523B RPO and Alphabet's ~$462B cloud backlog. And it holds, for the first time, both ends of one double-counting edge: CoreWeave names a $21B Meta commitment inside its $99.4B backlog — and Meta is itself a tracked hyperscaler. The same end-demand books at two layers; the magnitude of overlap across the hyperscaler → neocloud → model-builder stack is not yet quantified by any source. The disclosure is thin at every layer — no duration, cancellability or customer concentration for any of these figures, including the CoreWeave primary. Alphabet's Q2 8-K carries no RPO line, so the $240B-vs-$462B conflict stays open.

Open problems: duration and counterparty quality of the $523B RPO and $99.4B CoreWeave backlog; the size of inter-layer double counts; reconciling Alphabet's $240B → ~$462B.

4. The buildout is a macro variable

Status: Steady · Key sources: Fed MPR July 2026, ConstructConnect · Key concept: Data-Center Construction as a Macro Variable

The best-evidenced thread. The Fed reports business fixed investment at +11% annualized in Q1 2026 vs 5.5% in 2025, attributing most of the strength to AI infrastructure; ConstructConnect independently measures data centers at >20% of all US nonresidential building starts and YTD spending of $58.1B, >4x the 2025 record pace. The buildout subtracts ~0.4pp from GDP via AI-related imports, and the productivity payoff has not arrived. As of mid-2026 the AI economy's macro footprint is the construction of capacity, not its use — the "who pays the rent" question restated in national accounts. The negative-FCF prints (Frontier 2) are the corporate-finance version of the same tension.

Open problems: sizing the GDP downside of a capex air-pocket; why five states capture ~60% of starts; when the productivity channel becomes visible.

5. A fourth ownership category — vertically-integrated compute (now printed in MW)

Status: Early stage · Key sources: TSLA Q2, Tesla Cortex 2 · Key concept: Vertically Integrated Compute · Key entity: Tesla

Beyond hyperscaler / neocloud / colo sits the first-party builder. Tesla's Q2 8-K makes its compute figures first-party and switches the unit to MW: Cortex 1 >90 MW, Cortex 2 >115 MW (both Production), >205 MW combined onsite ("more than doubled" in H1) — so the blog's 500 MW is a full-buildout envelope, not the printed figure. Cortex pairs GPU clusters with on-site Megapack storage, the live test of "who owns the power" — and Carbon Direct now supplies the why: on-site generation is attractive because it skips the 3–4-year grid queue. Tesla is also integrating deeper (an in-house Austin fab for its own logic/memory chips). xAI's Colossus 2 (500,000+ GPUs) may belong to this category but is still unclassified.

Open problems: grid-vs-behind-meter split at Cortex; Cortex-specific capex per MW; replicability without an in-house storage business.

6. The supply-chain re-rate — the rack finally has a number

Status: Steady · Key sources: Lenovo GB300, NVIDIA/DOE · Key concept: Rack Power Density · Key entity: Vertiv

GB300 NVL72 rack-scale density pushes air cooling toward liquid, re-rating the vendors NVIDIA names (Vertiv, Eaton, GE Vernova, Schneider, Tesla, Trane) and lifting the ~$85B networking/cooling capex slice. The topic's oldest gap is closed: the rack is 135 kW TDP / ~155 kW peak with a ~90% liquid / ~10% air split (Lenovo vendor spec, corroborated by HPE at 132–140 kW) — roughly 18–19× a legacy ~7 kW rack, and ~7.4 racks/MW, the sourced mechanism behind "liquid cooling becomes mandatory." NVL72 deployment is also supply-gated by CoWoS.

Open problems: GB200-vs-GB300 delta and sustained-vs-peak draw; whether supply-limited racks push operators toward even higher density per delivered rack.

7. Efficiency is not the lever — and most of the queue is not real

Status: New 2026-09-24 · Key sources: Uptime 2026, Google PUE, Microsoft PUE/WUE, IEA Electricity 2026 · Key concepts: Power Usage Effectiveness · Grid Connection Queue

Two ways out of the power constraint that are often assumed turn out narrow. Efficiency: the industry average PUE is 1.52 and has barely moved in seven years; Google (1.09) has almost nothing left to cut, and Microsoft (1.17) got slightly worse as AI load rose. Better buildings will not free up much power on the timescale of this buildout. The pipeline: the IEA counts at least 150 GW of data-centre projects queued at advanced stages, but only about 20% of US requests turn into load, and utilities now charge for the right to wait (Dominion's proposed 14-year, 60%/85% commitment; ComEd's $1M deposit). Both findings point the same way: the power limit binds harder and longer than headline capacity suggests, which supports the three power calls.

Open problems: the PUE of a dedicated AI hall at full load; how fast the load queue clears (no source gives a wait); whether take-or-pay grid terms show up in lease prices; a global data-centre TWh forecast on this rung (the IEA's 485 → 950 TWh is in a different, un-ingested report).

Recent Developments

Date	Development	Source type	Evidence
2026-07-28	Uptime 2026 survey: industry average PUE 1.52 (1.54 in 2025), seven years of near-stasis; typical rack passes 11 kW; 24% of sites have racks ≥30 kW	Tech-report (self-reported survey)	Moderate
2026-07-22	Alphabet Q2 (8-K): capex $44.9B > OCF $39.1B → FCF $(5.9)B; funded by $49.6B equity + $20.3B notes; Cloud $24.8B (+82%) at 35.6%	Primary	Strong
2026-07-22	Tesla Q2 (8-K): capex $5.79B (+142%) → FCF $(1.1)B (cash-funded); AI compute switched to MW (>205 MW onsite)	Primary	Strong
2026-07-22	TXN Q2 (8-K): revenue $5.46B beat; capex falling $4.94B→$3.31B TTM; FCF +271% — the supply-side mirror	Primary	Strong
2026-07-16	TSMC: CoWoS + N3 sold out through end-2026; record Q2 (NT$1.27T, +36%)	Analysis	Weak
2026-07-15	Lenovo/NVIDIA GB300 NVL72 spec: 135 kW/rack TDP, ~155 kW peak, ~90% liquid — closes the kW/rack gap	Tech-report (vendor)	Moderate
2026-07-10	Fed: business fixed investment +11% annualized, attributed to AI infrastructure	Tech-report	Moderate
2026-07-07	Samsung preliminary Q2: ~₩89.4tn operating profit (~19x YoY) on HBM/DRAM tightness	Analysis	Weak
2026-06-30	Google Q2 2026 PUE: 1.10 quarterly, 1.09 TTM — flat since Q4 2024; 2025 fleet 1.09 vs Microsoft FY25 1.17	Tech-report (operator)	Moderate
2026-06-25	ConstructConnect: DC construction >20% of US nonresidential starts; $58.1B YTD	Analysis	Moderate
2026-06-15	TrendForce: CoWoS ~200K industry wafers/mo; supply-demand gap 20% → ~10% by end-2026 (narrows, doesn't close)	Analysis	Weak–moderate
2026-05-07	CoreWeave Q1 (first neocloud primary): rev $2.08B, $99.4B backlog, >3.5 GW contracted, Q1 capex $7.70B, net loss $(740)M; Meta $21B named	Primary	Moderate
2026-05-02	NextBigFuture/Crusoe: AI datacenter ~$59M/MW all-in, ~4yr payback — fills the capex/MW gap (as-of ~Apr 2025)	Analysis	Weak
2026-05-01	Carbon Direct: PJM+ERCOT queue >300 GW / ~1,500 projects, 3–4 yr waits — fills the interconnect-queue gap	Analysis	Moderate
2026 (early)	IEA Electricity 2026: data centres ~half of US demand growth to 2030; ≥150 GW queued at advanced stages; only ~20% of US requests materialise	Tech-report (forecast)	Moderate
2026-02-12	Futurum: $660–690B five-company 2026 capex; Oracle $50B / $523B RPO	Analysis	Weak
2026-01-08	Tesla Cortex 2: 500MW first-party AI datacenter with on-site Megapack storage	Analysis	Weak

All 24 sources are now compiled into concept/entity pages (staleSources 0, as of 2026-09-24).

The tradable debates

1. Which constraint binds first — power, packaging, or memory? All three now have sourced magnitudes (queue 3–4 yr; CoWoS gap 20%→10%; rack 135 kW) but the sequencing for a given project is still unresolved.
2. Do neoclouds earn their cost of capital, or is it a depreciation mirage? CoreWeave is the test case — now with numbers on one side ($2.08B rev vs $7.70B capex, $(740)M net loss) but the answer hinges on the still-undisclosed GPU depreciation schedule.
3. Does liquid cooling become mandatory (Vertiv-et-al a secular winner) or stay niche? Sharpened: at 135 kW/rack with a 90/10 liquid split, mandatory is now the sourced default at NVL72 scale.
4. Will hyperscaler capex inflect down, and what is the leading indicator? The nearest tell just printed: Alphabet's Q2 FCF went negative while capex kept rising — the buildout is now balance-sheet-financed, so the leading indicator may be financing capacity, not demand.
5. Who owns the power? Utilities, IPPs, or buyers building behind the meter. Tesla Cortex (now >205 MW, MW-disclosed) is the live case; the queue's 3–4-yr wait is the reason behind-the-meter pays.
6. Is the backlog real — and is it double-counted? Advanced 2026-07-23: the KB now holds one concrete double-counting edge (Meta $21B ⊂ CoreWeave $99.4B), but duration/concentration are absent at every layer.

Knowledge Gaps

Areas where the KB needs more sources. These feed the next discovery cycle. The 2026-07-23 compile CLOSED six gaps (see "Filled" below) and opened/sharpened several others.

Still open (highest priority first)

• Backlog duration, cancellability and customer concentration for Oracle RPO ($523B), Google Cloud backlog (~$462B), Microsoft's unfulfilled Azure backlog, and CoreWeave's $99.4B — the most decision-relevant unknown in the topic, and still undisclosed even in the CoreWeave primary — suggested search: "Oracle remaining performance obligations duration disclosure 10-Q"
• Amazon, Microsoft and Meta Q2 2026 capex/FCF — do the other three big-four buyers show the same "stopped self-funding" signature GOOGL and TSLA printed? They report late July — suggested search: "Amazon Microsoft Meta Q2 2026 capital expenditures free cash flow 10-Q"
• The magnitude of backlog double-counting — the KB now holds one edge (Meta's $21B inside CoreWeave's $99.4B while Meta is a tracked hyperscaler) but no source quantifies the overlap across the hyperscaler → neocloud → model-builder stack, nor nets out capacity hyperscalers rent from neoclouds — suggested search: "AI cloud contract backlog double counting neocloud reseller GPU commitment"
• Microsoft 2026 capex, from a primary source — resolves the ~$70B conflict driving the aggregate dispersion; MSFT prints ~2026-07-29 — suggested search: "Microsoft FY2026 capital expenditures 10-Q quarterly report"
• Alphabet FY2026 capex dollar guide + RPO/cloud backlog — the Q2 8-K is compiled but both are call/10-Q items absent from the exhibit; needed to resolve the guide (watched $180–190B) and the $240B-vs-$462B backlog conflict — suggested search: "Alphabet Q2 2026 earnings call capex guidance remaining performance obligations"
• CoreWeave GPU depreciation schedule + customer concentration — the Q1 primary gives revenue/backlog/capex but not the depreciation life or the Microsoft/OpenAI share — the two inputs debate #2 turns on; Q2 2026 prints mid-Aug — suggested search: "CoreWeave 10-Q GPU useful life depreciation customer concentration 10-K"
• HBM bit supply, capacity and allocation by vendor — the KB has the price signal (Samsung's margin) and zero volume data — suggested search: "HBM3E HBM4 supply capacity allocation SK Hynix Samsung Micron 2026"
• SK Hynix — the largest HBM supplier has no source in this KB at all; reports 2026-07-29 — suggested search: "SK Hynix Q2 2026 results HBM capacity"
• A primary TSMC source on CoWoS — capacity trajectory is now sourced (TrendForce, secondary) but the sellout + wafer numbers want a TSMC filing/transcript — suggested search: "TSMC Q2 2026 earnings call transcript CoWoS advanced packaging capacity"
• US deliverable power-capacity shortfall (today vs 2028) — Carbon Direct says the queue holds "enough on paper through 2030," but the deliverable shortfall path is still unquantified — suggested search: "US data center power capacity shortfall gigawatts 2028 deliverable forecast"
• xAI Colossus 2 ownership classification — 500,000+ GPUs, uncategorised between neocloud and first-party — suggested search: "xAI Colossus 2 Memphis ownership power supply financing"
• GB200-vs-GB300 rack-power delta + sustained-vs-peak draw — the 135 kW figure is GB300-specific and vendor-primary; an NVIDIA/independent-lab number on sustained draw and the GB200 delta would deepen it — suggested search: "NVIDIA GB200 NVL72 rack power sustained draw datasheet"

Filled at the 2026-09-24 compile (now in the wiki)

• ✅ PUE — a fleet-wide average and two hyperscalers on stated boundaries: industry 1.52 (Uptime, self-reported), Google 1.09, Microsoft 1.17 → power-usage-effectiveness.
• ✅ How much of the load queue is real: only ~20% of US data-centre connection requests materialise; ≥150 GW queued at advanced stages globally (IEA) → grid-connection-queue. The deliverable shortfall path gap above stays open: the IEA gives conversion, not a GW-by-year shortfall.
• ✅ Rack density across the industry, not just the NVL72: typical rack >11 kW, 24% of sites with racks ≥30 kW (Uptime) → rack-power-density.

Opened at the 2026-09-24 compile

• Global data-centre electricity forecast (TWh) — the IEA's 485 TWh (2025) → 950 TWh (2030) is in Key Questions on Energy and AI (April 2026), not in Electricity 2026; propose that report on its own — suggested search: "IEA Key Questions on Energy and AI data centre electricity 2030"
• PUE of AI-dedicated halls — none of the three PUE sources separates them — suggested search: "liquid cooled AI data center measured PUE full load GB200"

Filled at the 2026-07-23 compile (now in the wiki)

• ✅ kW/rack for GB300 NVL72 — the topic's oldest gap, closed: 135 kW TDP / ~155 kW peak, ~90% liquid (Lenovo GB300) → rack-power-density. Vendor-primary.
• ✅ Grid-interconnect queue wait times — closed: 3–4 yr waits, >300 GW combined PJM+ERCOT queue, constraint is throughput not raw MW (Carbon Direct) → power-the-binding-constraint.
• ✅ Capex per MW — closed (analysis-tier): ~$59M/MW all-in, ~half IT (NextBigFuture/Crusoe) → capex.md. Single-operator, as-of ~Apr 2025.
• ✅ CoWoS wafers/month capacity & end-2026 gap — closed (secondary): ~120–140K TSMC + 50–60K OSAT; gap 20%→~10% (TrendForce) → advanced-packaging-bottleneck. TSMC-primary still wanted (kept open).
• ✅ Neocloud financials (CoreWeave) — closed with the first neocloud primary (CoreWeave Q1): rev $2.08B, $99.4B backlog, capex $7.70B, net loss $(740)M → CoreWeave dossier. Depreciation + concentration kept open.
• ✅ Q2 2026 buyer prints (GOOGL, TSLA) + supply-side (TXN) — closed: the demand side has stopped self-funding (capex > OCF → negative FCF) (GOOGL, TSLA, TXN) → hyperscaler-capex-cycle. Alphabet FY guide + RPO absent from the exhibit (kept open).

Still to compile

Entities: Microsoft (blocked on the capex conflict; prints ~Jul 29) · Amazon · Meta (Q2 prints late Jul) · Nebius (NBIS) · Crusoe · Lambda · Eaton (ETN) · nVent (NVT) · Schneider Electric · Super Micro (SMCI) · Dell · Equinix (EQIX) · Digital Realty (DLR) · Vistra (VST) · Constellation (CEG) · Talen (TLN) · GE Vernova (GEV) · SK Hynix · Micron · NVIDIA · xAI

Concepts: Power Usage Effectiveness (PUE) (compiled 2026-09-24) · Direct-to-chip vs air cooling · Grid interconnection queue (compiled 2026-09-24 as grid-connection-queue) · Capex per MW (currently in commercial/capex.md + benchmarks — could graduate to its own concept) · Power Purchase Agreements and behind-the-meter generation · The neocloud business model (GPU depreciation, contract duration, customer concentration — seeded via the CoreWeave dossier)

Tracking tier

Datacenters remains a tracking beat (frontier monitoring + fast-takes + entity pages), not yet in the daily /dispatch rotation. The 2026-07-23 compile is the first to add primary company filings (five: GOOGL/TSLA/TXN Q2, CoreWeave Q1) and to quantify the constraint stack end-to-end. Evidence quality genuinely improved this compile — the earlier "analysis-only" characterization no longer holds for the capex/FCF/backlog spine, though CoWoS and HBM volume data remain secondary. The first dossier target (CoreWeave) is now seeded with a Q1 primary; the deep-dive proper awaits the 10-Q (depreciation + concentration) and the mid-Aug Q2 print.