US AI infrastructure · overview refreshed 6 AM + 2 PM PT

AI demand outruns firm power by ~19 GW by 2030

Tokens pull capital, capital funds the buildout, the buildout lands on a grid that can't add firm power fast enough — and the gap keeps widening through 2030.

Decision implication Firm power — not chips or capital — is the binding constraint through 2030; treat secured megawatts and contracted offtake as the assets, merchant and single-counterparty exposure as the fragility. Editorial — not investment advice.

Read by role →

The AI infrastructure chain — where demand becomes constraint°

1 · Tokens3.2 quadrillion / moGoogle's all-surface inference — ~18× its billed API volume. Usage is the demand engine.driverwatch: NVIDIA Q2 FY27 · late Aug 2 · Capital$839B2026 capex chasing that demand · ~75% AI. Money isn't the constraint — over-committed books are (Oracle/CoreWeave ~10 yrs OCF).not bindingwatch: Q2 capex guides · late Jul–Aug 3 · Buildout97 GW queuedonly ~24 GW truly buildable · queued-not-secured MW carries the haircut. Queued ≠ buildable.high — gear-gatedwatch: LBNL Queued Up snapshot 4 · Grid~19 GW shortfirm power can't keep up — the long pole; PJM/N. Virginia loads most exposed.bindingwatch: FERC show-cause responses · ~Aug 5 · Rates & politicsat the cap, 3×PJM cleared AT the FERC cap three straight auctions ($325); costs land on ratepayers (VA +57% modeled°) — then on permits.risingwatch: White House pledge event · state PUC rulings

What breaks first — the constraint board

POWER↑ High^° ~19 GW firm-power shortfall by 2030; the base path widens as load compounds. Next: FERC show-cause responses — all six RTOs · ~Aug 2026 GRID GEAR→ High^° HV transformers 36–60 mo (vs ~24 pre-2020); gas-turbine slots sold out to 2029–30. Next: GE Vernova / Siemens / MHI order books INTERCONNECT QUEUE→ Elevated^° 97 GW queued, but ~78% historical withdrawal → ~24 GW credible; interconnect 48–84 mo. Next: annual LBNL Queued Up snapshot CAPEX↑ Elevated^° Oracle ≈10 yrs of operating cash flow pre-committed; ~$839B 2026 capex, ~75% AI. Next: Q2 2026 hyperscaler capex guides (late Jul–Aug) POLICY↑ Elevated^° TX SB2 large-load curtailment live (>75 MW); VA data-center tax-exemption rollback under review. Next: next state PUC rate-case ruling RATEPAYER↑ Elevated^° PJM cleared at the FERC cap three straight auctions ($329→$333→$325/MW-day); VA +57% residential by 2030 (modeled). Next: White House ratepayer-pledge event · coming weeks CHIPS / HBM↓ Watch^° GB200 NVL racks GA — allocation, not lead-time, now binds; CoWoS ~doubling (75k→130k wafers/mo). Next: NVIDIA Q2 FY27 supply commentary (late Aug)

Severity & trend (↑ worsening · → holding · ↓ easing) are editorial reads (marked °); every figure is sourced — tap a tile to see it on its chart. Worst-first: power is the long pole; chips already eased.

Where value accrues — and where it’s at risk^°

Value accrues to

Owners of firm megawatts — a secured build, not a queue position; power is the long pole. Contracted offtake — ~60% of ledger demand is take-or-pay, not merchant exposure. Owned / multi-sourced silicon — TPU · Trainium · MTIA cut merchant-GPU dependence.

Value at risk where

Obligations outrun asset life — chips depreciate in ~6 yrs; the leases run 12–25. Books are most pre-committed & accelerating — Oracle · CoreWeave ~10 yrs of OCF. One counterparty dominates — CoreWeave ~67% Microsoft; NVIDIA ~54% top-3. Cost lands on ratepayers / merchant supply — VA +57% (modeled); ~15–25% built merchant.

The power gap — base / bull / bear: Base ~19 GW short by 2030 · Bull — on EPRI’s low case the gap closes · Bear — LBNL high end ~100+ GW

Top 5 watch signals Hyperscaler Q2 capex guides · late Jul–Aug FERC show-cause responses · ~Aug 2026 Oracle 10-Q RPO print LBNL Queued Up snapshot Next state PUC ruling

What would change our mind: FERC show-cause responses (~Aug 2026) · Q2 hyperscaler capex guides · next LBNL Queued Up · Oracle RPO print. Structural read — editorial.

Where the next entrant wins — sweet spots by constraint^°

Constraint layer	Binding today	Relief window°	Incumbent density	Next entrant’s edge°
Grid gear	HV transformers 36–60 mo	turbine slots booked to 2029–30	high — Vertiv · Eaton · GE Vernova	capacity expansion, retrofit & speed plays
Firm power	grid interconnect 48–84 mo	on-site first-power 3–18 mo (fuel cells / BTM gas)	forming — utilities + co-located developers	own the energize date — power-first site development
Memory · HBM	HBM capacity spoken for into 2026	US fab buildout underway (Micron plan, TSMC AZ) — output timing not yet cited here	very high — Micron · SK hynix · Samsung	thin near-term — capacity is the moat; packaging & test niches
Compute resale	1.4% vacancy; single-buyer books (CoreWeave ~67% MSFT)	~60% of ledger demand already take-or-pay — spot stays thin	crowded at the top — hyperscalers + neoclouds	diversified-counterparty capacity with GPU allocation secured

Figures trace to charts on this dashboard (links); Relief° and Edge° are editorial reads, archetype-level; not investment advice.

Read it by role — what a CAIO, infrastructure chief or VC should take away^°

CAIOChief AI Officer

Capacity risk is geographic, not vendor — concentrated in power-short PJM / N. Virginia; ~78% of queued GW historically withdraws. queue →
Token growth becomes power risk at the grid gear, not the GPU — HV transformers 36–60 mo, turbine slots sold to 2029–30. lead times →
Multi-source silicon + counterparty — owned silicon (TPU/Trainium/MTIA) cuts merchant-GPU risk; single-buyer providers (CoreWeave ~67% MSFT) pass it to you. players →

Decision implication Weight sourcing toward secured firm power + diversified silicon over the best-looking queue or price.

Watch: hyperscaler Q2 capex guides · late Jul–Aug

CIOChief Infrastructure Officer

Pick the power path before the site — on-site first-power in 3–18 mo (fuel cells / BTM gas, modeled bands) vs 48–84 mo grid interconnect. time-to-power →
Order long-lead gear early — HV transformers 36–60 mo, turbine slots sold to 2029–30; gear, not land, sets the energize date. lead times →
Re-weight the site scorecard to your build profile — Speed / Cost / Water presets re-rank the same cited market scores; check the rate-revolt map before committing. scorecard →

Decision implication Back-plan from the target energize date: power path first, long-lead gear second, market third.

Watch: FERC show-cause responses — all six RTOs · ~Aug 2026

VCVenture partner

Value migrates from compute abundance to the binding constraints — durable pools sit with secured-MW + contracted-offtake owners, not merchant compute. power gap →
The exposed names are visible — Oracle/CoreWeave ~10 yrs pre-committed, CoreWeave ~67% MSFT, NVIDIA ~54% top-3; circular financing on a ~6-yr-chip vs 12–25-yr-lease mismatch. quadrant →
Truest picks-and-shovels = power/grid gear — transformers, turbines, cooling (Vertiv, Eaton, GE Vernova) gate every campus, whichever lab wins. suppliers →

Decision implication Favor the binding-constraint layer; treat merchant compute + single-counterparty operators as the fragile end.

Watch: NVIDIA Q2 FY27 supply commentary · late Aug

Connecting the dots — what links up

Verified edges between players — one deal in two filings, a loop between the same names. Tap a card for the evidence; the badge says how firm the link is.

This week in US AI infrastructure

When each constraint binds — power is the long pole

Power is the multi-year long pole; this timeline shows when each narrower constraint binds along the way — GPUs in 2023, packaging & HBM in 2026 — but firm power and grid gear stay the gate underneath all of them.

Capital

Where the money goes — capex, vacancy, returns, M&A. For capital allocators, VCs, corp dev, public-market investors.

↑ the 19 GW gapMoney isn't the constraint — power is. $839B is chasing a market that's ~19 GW short on firm power by 2030; the winners own the megawatts, not the capex.

So what

$839B can buy compute—but not firm power. Value accrues to megawatts already controlled; risk clusters where long obligations depend on one counterparty.

The most committed balance sheets also carry the most concentrated demand° X + capex primary Y coverage-biased · colour editorial

Who cannot back down (years of operating cash flow already pre-committed) vs what they’ve actually secured (named live ledger power) · bubble = 2026 capex · colour = silicon strategy · the six operators with a filed commitment book

So what · 2nd order

Capital quality beats headline capex — and the risk is correlated. The two names deepest in the “over-committed, power-short” corner, Oracle (~10 yrs of cash flow pre-committed) and CoreWeave (~10 yrs, no owned build), are the same two leaning hardest on a single concentrated counterparty.^°

CoreWeave books ~67% of FY2025 revenue from Microsoft (primary); Oracle’s 10-K credits its RPO surge to “large-scale AI contracts” (press reads: OpenAI — never named in the filings; analyst); both ride NVIDIA, top-three ≈54% of revenue (primary). OpenAI anchors 3 of 5 closed loops — correlated, not diversified, risk.

Decision implication Capital quality beats headline capex — the spread that matters is secured megawatts vs pre-committed obligations, not who spends most.

Buyer posture — the same numbers, from the buyer’s chair^°

Early contracting beats spot — 1.4% vacancy, asking rents +42% since 2022 (CBRE); spot capacity is the expensive residual. vacancy →
Your provider’s balance sheet is your availability risk — Oracle & CoreWeave run ~10 yrs of OCF pre-committed; single-counterparty providers pass that risk through. commitment book →
Allocation, not list price, binds the newest SKUs — GB200 racks GA yet allocation-gated; NVIDIA’s top-three ≈54% of revenue. quadrant →

Scarcity rewards early commitment — waiting preserves flexibility, not allocation°

The four paths to AI capacity, joined from this dashboard’s cited anchors · native units on purpose — they don’t reduce to one honest number · Editorial° synthesis; not procurement advice

2026 capex by operator

Largest US AI-infra operators · hyperscalers (~75% AI) + AI-native challengers (xAI, Nebius — est.) · USD billions

AI-attributed share of capex° modeled

Build-cost anatomy — what a megawatt costs (~$42M all-in, GPUs dominate) — lives on Buildout →

Combined hyperscaler capex — the curve actuals + ’26 guidance

Capex vs operating cash flow — the crossover watch actuals + ’26 guidance

The commitment book — years of cash flow pre-committed primary

Undiscounted future lease payments (commenced + signed-not-yet-commenced) plus disclosed purchase / construction commitments, per operator, against annual operating cash flow · the per-row figure = years of OCF already spoken for — the single best read on who cannot back down · bases and as-of dates differ per company (tooltip + method note)

The three clocks — assets, obligations, and power run on different time

Chips depreciate on filed useful lives of ~5.5–6 years · the leases financing them run 12–25 years · new firm power arrives in ~3–7 · the tenor mismatch is the structural risk under every take-or-pay signature

Where the capex flows — 2026

~$799B across the six operators with disclosed spend buckets — the ~$839B headline adds xAI + Nebius (~$40B) · totals are each company's 2026 guidance (primary) · the bucket split is modeled · select a company to drill in

swipe the chart to explore →

Offtake coverage — is 2026 capex demand-pulled?

Bar = modeled share of the named GW pipeline with signed offtake; ledger = multi-year $ backlog (primary) — not comparable to annual capex.

The commitment flywheel — where capital comes back as “demand”

Vendors and private credit put equity + GPU-backed debt into labs and neoclouds, which commit capital back as compute purchases. Default = the true two-way loops only; edge width = disclosed $ size; line style = instrument (solid equity · dashed compute · dotted debt); faint/dash-dot = no public $ or analyst-tier leg.

Demand is partly recycled. NVIDIA takes equity in OpenAI (up to $100B), CoreWeave (~7%) and Anthropic ($10B) — who then commit to buy NVIDIA compute. Some of the "demand" is the vendor's own capital cycling back.

One node anchors the whole loop. OpenAI alone carries ~$1T+ of disclosed compute commitments (Oracle $300B, Microsoft $250B, Broadcom ~$350B, AMD ~$90B, NVIDIA ~$100B, AWS $38B) — a stumble there cascades to every supplier.

The chips are becoming collateral. Private credit — Apollo's $35B Broadcom XPV, Blackstone's $8.5B CoreWeave loan — now lends against the GPUs themselves, a new financialization layer beneath the equity and compute deals.

swipe the chart to explore · hover a node to pull its thread →

Vertical integration — when the buyers build their own chips scenario

OpenAI's Broadcom-built "Jalapeño" joins Google TPU and Amazon Trainium moving inference off merchant NVIDIA GPUs. Good or bad for the industry? Three sourced lenses — market-structure analysis, not investment advice.

Only premium tokens pay back a megawatt before its chips die°

Power-to-revenue yield: the same $42M/MW, monetized as a wholesale facility lease vs token sales, on one payback scale. Inputs are cited; payback is modeled arithmetic, not a return forecast · green = pays back inside the filed 5.5–6-yr server life (tenor clocks).

Capital & deals tracker

Major financings, partnerships & M&A shaping the buildout · past 30-60 days · each row cited

Public-market plays

Listed companies grouped by thesis · not investment advice

1.4% vacancy leaves buyers with almost no spot leverage

Primary-market vacancy · year-end · the supply-tightness story

Vacancy compressed from ~9% in 2019 to 1.4% today — landlord pricing power has shifted decisively. Average asking rents are up ~42% since 2022 and ~64% from the 2021 trough (CBRE, below).

Avg asking rate · 250–500 kW · N. America primary wholesale · $/kW-month · CBRE H2 series (analyst)

Build your IRR scenario

100 MW build · adjust assumptions · IRR recomputes live

Lease price $160/kW-mo

Power cost $55/MWh

Delivery delay 0 months

Project IRR

—

15-yr hold · 5% terminal cap · $42M/MW build

Tokens → $/token compression compresses inference revenue per token. Open the unit-economics deep dive →

Buildout

How capacity actually gets built — queues, lead times, cost stack, site selection. For infra leaders and supply-chain planners.

↑ the 19 GW gapQueued MW is not buildable MW. 97 GW in the queue collapses to ~24 GW real — and transformers + interconnection are what hold the ~19 GW power gap open.

So what · 2nd order

A 97 GW queue is not a 97 GW asset. Only ~24 GW looks buildable; the scarce asset is power-deliverable land.^°

Named-build reality stack derived from reported status

Every rung comes from a real projects.json status: announced universe → active after verified retreats → construction + operational → operational. No invented “secured” stage; no queue-wide extrapolation.

The queue is 16× larger than active construction

Four separate snapshots of US data-center grid capacity — not one cohort moving through stages · GW · LBNL Queued Up 2025 + CBRE + Goldman Sachs

Only about one-quarter of queued demand appears buildable near term

Phantom-load reconciliation — the headline queue net of duplicate and speculative requests · GW · haircuts modeled

Methodology & sources

Generation interconnection queue by ISO — credible vs. phantom analyst

Generation + storage queue by ISO (LBNL Queued Up) — the pipeline DC load competes within, not a DC-only queue; credible = post-withdrawal (78% = historical withdrawal rate, applied as proxy).

Where the capacity sits — primary US markets

Pipeline view · bubbles sum to ≈ 97 GW interconnect queue (LBNL Queued Up 2025, allocation modeled) · color = market status · hover for detail

US Markets

Power-constrained Major growth Emerging

Low High Modeled · state/ISO inputs · regional screen, not a site confirmation.

Power path—not project size—separates a build from a press release°

Announced data-center campuses + the behind-the-meter / colocated power securing them · MW = announced campus or dedicated-generation capacity · power-procurement model classified editorially (BTM = on-site generation that bypasses the interconnect queue · Colocated = adjacent to existing nuclear/gas · Grid = utility interconnection) · sorted by MW · each row cited to operator / utility filings & press (not third-party trackers)

firm-typed announced / ultimate target · soft figure shelved · disputed (excl. GW) fill = evidence firmness · colour = status

MW = announced/ultimate targets (some disputed); total-power rows include cooling & generation, not IT load; graveyard excluded from totals; undisclosed MW named, never estimated. Tap a bar for its cited row.

Decision implication Underwrite to the buildable base, not the press release — a signed-power, firm-typed build is the asset; queued or announced-target MW (~78% historical withdrawal) is the residual.

Data table + per-record citations

Graveyard & stalls — verified retreats

Announced projects later paused, stalled, or cancelled — the counterweight to announcement bias, from the same cited open dataset (status changes carry their own sources; the stated reason is tiered separately from the status fact). Excluded from the headline GW totals above.

The site decision kit — path, market, gear^°

Grid power takes years; on-site bridges buy months°

The binding 2026 question isn't whether power is scarce — it's which path energizes a site fastest, at what trade-off · months from decision to first power · firmness tag + one-line trade-off · grid duration cited (LBNL); on-site gas / recip bands modeled (labeled)

Power by 20XX — which paths still make it^°

Modeled°: months to Jan 1 of the target year vs the time-to-power bands above; the grid path also checks HV transformers (36–60 mo). A path makes it when even its slow end fits; tight when only its fast end does.

Site-selection scorecard

Per-factor cited scores across top US markets · 1-10 (higher = better for builds) · Readiness = modeled weighted composite · pick a weight profile to re-rank

The transformer order date—not the land—sets energization°

Months from order to delivery · 2026 industry estimates

Decision implication Transformer and turbine orders belong at site-control, not at permit — the order date, not the land, sets the energize date. Editorial.

More buildout analysis — cost stack, cadence, efficiency & supply chain

Time to deployment by region

Years from greenfield to energized · ranked by lead time

Where every $1M goes

Build cost breakdown per MW · facility vs compute layers

Self-built capacity: operational vs pipeline

Top 5 hyperscalers · MW · modeled from IR + analyst sources

Compute-per-watt trend — the efficiency offset the demand forecasts assume away

Every demand-gap projection implicitly assumes efficiency gains don't outrun demand. This is the only place a reader sees why those GW forecasts could run high. Bars/line = peak dense FP16/BF16 TFLOPS per watt by GPU generation (one pinned metric), indexed to A100 = 100. Dashed markers = a separate modeled "effective inference per-watt" basis (FP4/NVFP4 + NVL72 rack-scale) — shown to explain why real deployments see multi-fold gains the single-precision line does not.

Buildability — what's moving

Dated, individually-cited regulatory / queue / generation events that move a market's path to the next gigawatt. Direction (easing / tightening / stalled) is an editorial read of travel vs. the prior event — not a metric. Curated monthly · trend dots also annotate the map above.

AI infrastructure stack — where the binding constraint sits

Each layer of the buildout has its own choke point. As one loosens the constraint cascades upstream. Status reflects current public commentary; every row carries a citation.

Upstream substrate bottlenecks

Supply concentration beneath the transceiver layer

Optical interconnect roadmap

Bandwidth generations + CPO transition

Tokens → Tokens per MW-hour. Energy efficiency curves for inference. Open the energy bridge →

Grid

The energy + policy story — demand vs generation, rate impact, regulation, utility responses. For utility planners, regulators, energy reporters, policy analysts.

↑ the 19 GW gapThis is where the ~19 GW gap stops being a chart and becomes rates + policy — unmet load lands on ratepayers: two states model >25% rate hikes (five total sit at ≥15%).

So what

The bottleneck turns political before the grid runs out: Virginia (+57%) and Texas (+28%) already model rate impacts above 25%.

Rate revolt starts before the grid runs out—Virginia and Texas cross +25%°

Political-failure risk by state across four dimensions — rate-increase risk, utility exposure, data-center load concentration, regulatory posture. Unmet load + overbuild risk get socialized onto ratepayers, so the binding constraint shifts from engineering (queues) to politics (PUCs, governors, legislatures). No blended score — the four cells stay separate; read a red Rate cell as a bill a PUC must approve and a governor must defend.

Geographic pressure map

Where rate risk is priced — and where load arrives first

State-by-state evidence matrix

Decision implication Ratepayer backlash can stop the AI buildout before physical scarcity does — the fight turns political (PUCs, governors) while the grid still has headroom.

State risk table + attribution

Who demands what — large-load rules by utility & regulator

Requirements already on the books for big loads, joined from the dated movements + ledger rows on this dashboard · each row cited · — = not tracked (never assumed)

Regulator · utility	State	Requirement for large loads	Next date	Source
FERC — all six RTOs	federal	show-cause: justify or rewrite large-load interconnection tariffs	responses ~Aug 2026	FERC / White & Case · primary
PUCT / ERCOT	TX	SB2 registration + curtailment >75 MW · ride-through rules (stay connected through disturbances) · draft SB6 interconnection standards	SB6 final Dec 2026	PUCT / ERCOT / Utility Dive · primary + analyst
Virginia SCC	VA	>25 MW rate class — 14-yr contracts assign costs to data centers	in effect (Jan 2026)	Virginia SCC · primary
Arizona ACC — APS / SRP	AZ	large-load docket open · APS proposed ~45% XL-load increase · SRP E-67 80% minimum billing	docket open	Arizona Corporation Commission · primary
Illinois ICC — ComEd	IL	higher interconnection deposits ≥50 MW · consumer-cost probe	—	Illinois Commerce Commission · primary
PUCO — AEP Ohio	OH	large-load tariff: 85% take-or-pay for >25 MW	—	ledger keyRisk (AEP Ohio) · analyst
LPSC — Entergy	LA	dedicated-generation path fast-tracked for Hyperion (7 added gas units)	final vote Dec 2026	Entergy / LPSC · primary

Decision implication Engage the regulator before the site: the requirement set (take-or-pay %, deposits, curtailment, ride-through) now varies more by state than power price does. Editorial.

Firm power never gets ahead — the unfilled gap peaks at +7 GW in 2027°

Bar height = DC demand added per year · green = new firm gen committed to DC · red = unfilled · GW · 2024–2030 · modeled

Healthy today — until the announced ledger lands° derived

Nameplate-minus-peak proxy (EIA-930 + EIA-860), not accredited headroom · ● today (derived) · ○ after live ledger load mapped to each BA (modeled°)

What power costs, by state primary

Industrial retail electricity price ($/MWh) across the AI data-center corridor · a rough proxy for the power-cost term in siting decisions (large DCs often negotiate below-tariff rates) · fetched 2× daily in CI

Two states cross the rate-revolt line° — Virginia at 2× the field

Modeled residential rate impact by 2030 under high-DC-load scenarios vs. 2024 · dashed line = +25% revolt threshold° · red ring = also pays >$90/MWh industrial today — the corridor’s top price band (EIA)

PJM: 11× in four auctions — then pinned at the FERC cap°

Base Residual Auction RTO clearing price by delivery year (auction timing has been irregular) · data-center load is the primary driver of the run-up · the 2028/29 auction (results July 14, 2026) cleared AT the extended cap at $325/MW-day UCAP, with the entire RTO short of its reliability requirement for the first time

Transmission constraints by ISO

Large-load interconnection queue depth, by region

Regulatory tracker

State + federal actions shaping the buildout

Utility M&A + capacity additions

Who's adding what, for whom

Demand response & flex load

The grid's pressure-release valve

Short ~19 GW by 2030 — or zero, or ~108: the published bracket°

Standing GW shortfall: running DC demand added minus running firm generation for DC · national · 2024-2030 · toggle the demand path to published low/high anchors (EPRI ’24 low · LBNL ’28 high) — scenario arithmetic, labeled

Gas-turbine slot reservations by OEM

Firm equipment backlog vs. deposit-backed slot reservations · GW · Q1 FY2026 · earliest hard signal of firm generation

Powering the load — new generation by source

Additional annual generation committed to US DC load, by source · TWh · near-term vs 2030

Tokens → Tokens × energy: how much grid load does AI inference actually create per unit of usage? Open the adoption-vs-energy curve →

Tokens

The AI economy underneath the buildout — how much is run, what it costs, what it consumes.

↑ the 19 GW gapUsage growth is the demand engine. ~330× token volume in 24 months is what keeps data-center demand outrunning firm power toward that ~19 GW 2030 gap.

The thesis

Usage—not token price—is setting the power curve. The modeled provider series rises ~22× from Q1 ’24 to Q2 ’26; Google’s disclosed all-surface volume rises ~330× in 24 months.

Modeled token volume rose ~22× while the cheapest frontier price fell 73%—power followed usage° price cited · usage modeled

Each bubble is one provider-quarter · Y = modeled tokens/quarter (log scale) · bubble area = implied power contribution at 200M tokens/MWh · price ribbon = cheapest closed-frontier flagship · select a provider to isolate its story; hover any bubble for details.

Industry total: 100T → 2,180T tokens/quarter · implied continuous power 0.23 → 4.98 GW.

Cheapest closed-frontier flagship · $ / M output tokens

$30$21$15$8

Decision implication Cheaper tokens don’t mean cheaper work — efficiency delays the infrastructure wall, it doesn’t remove it. Editorial.

What would change the thesis: a step-change in energy-per-token, sustained low GPU utilization, materially slower enterprise adoption, or workloads shifting to far smaller/cached models.

Model choice is now a power decision—one prompt to one gigawatt

The 5-stage executive chain — prompt → tokens → energy → continuous MW → capex / power source — labelled single-prompt unit economics → platform-scale extrapolation → industry-buildout comparison. Each rung sourced or [modeled]. Go deeper for the 10-step technical walkthrough.

Decision implication AI-adoption planning is now infrastructure planning — model choice and inference architecture directly set your power exposure.

Industry token volume by provider

Trillions of externally-billed API tokens per quarter · stacked area · modeled from provider revenue + Epoch AI + SemiAnalysis — magnitudes are estimates, read the shape not the decimals. Excludes in-product inference (Search AI Overviews, YouTube, Workspace) — see Disclosed totals below.

Reasoning models give back much of the $/token gain°

Sticker $/token fell ~100–300×, but reasoning/agentic models burn far more tokens per task · cost to COMPLETE a fixed task ($/task, log Y), held-constant basket · 2023 vs 2026 frontier · $/token cited, tokens-per-task & $/task modeled

Decision implication Power exposure is set by AI strategy: a frontier reasoning-heavy stack burns far more tokens (and watts) per task than efficient enterprise inference — model mix, routing, caching and context length are power decisions. Editorial.

Rent prices fall—allocation still gates the newest GPUs primary · 2× daily

Published on-demand list prices by provider × GPU · negotiated / reserved / spot rates differ · freshness is per provider (a provider that fails to parse keeps its last good date — shown) · levels today; the trend accrues in data/history

Decision implication Anchor build-vs-buy against the buy-side $/GPU-hr level — falling list prices favor renting; allocation still gates the newest SKUs. Editorial.

Disclosed total tokens processed — the bigger picture

When platforms disclose total inference across all surfaces (API + in-product), the numbers are materially larger than monetized API volume alone.

$/token compression — the 100× price drop

$ per 1M output tokens at launch · log Y axis (data spans ~125× from frontier to hosted open models — linear would hide the cheap end) · lower-right = newer + cheaper

Tokens × energy bridge — from AI economy to grid load

The link that ties this tab to the rest of the dashboard

Inference vs training split

Where the compute actually goes · late-2025 industry estimates

Players

The same cited numbers, indexed by entity instead of theme — who is doing what, under which binding constraint. Every figure on a card is a live join from its home chart (click through); nothing is re-keyed.

↑ the 19 GW gap Positions differ because constraints differ: some players own megawatts, some own obligations, some own the silicon — the cards below say which.

Provider posture at a glance — four ways resilience shows up^°

A CAIO’s counterparty screen and a VC’s market-structure screen on the same scale: power independence · capital room · silicon control · counterparty resilience. Longer = more strategic room; hatched = not honestly known, never zero.

Decision implication Provider risk is multi-dimensional: negotiate portability where any one lane is weak, because a balance-sheet, power, silicon, or customer-concentration failure can become the same availability failure.

Player dossiers

Each card leads with a Constraint Fingerprint — a 6-lane silhouette (capex · committed-yrs-OCF · secured-MW | silicon · counterparty · grid-risk), each lane a live join from its home chart · reads are editorial, tier-tagged

Decision implication Winners aren’t the biggest spenders — they hold the best constraint portfolio: secured power + capital quality + diversified silicon + low counterparty concentration. Editorial.

Watchlists — thesis support (not full dossiers)^°

Dossiers explain the thesis; these support it. Chips link to existing sourced rows; grey = name-only, not yet tracked (never zeroed or invented).

Neocloud / developer

Nebius $22B capex · PA 1.2 GW Applied Digital Delta+Polaris Forge · ~$36B take-or-pay IREN Childress live · Sweetwater Crusoe Stargate Abilene (developer) Fluidstack · Lambda · TeraWulf

Power bank (utilities / IPPs)

Talen Susquehanna (w/ AWS) NextEra Duane Arnold (w/ Google) Constellation Crane/TMI restart · MSFT PPA Vistra · Entergy

Equipment bottlenecks

GE Vernova Vertiv Eaton Siemens Energy · Mitsubishi · Schneider · Quanta

Capital nodes

Apollo $35B Broadcom chip SPV Blackstone CoreWeave DDTL · Google TPU JV Blue Owl Project Blue (stalled) Brookfield

The value chain — silicon, memory, power & grid gear

The picks-and-shovels that pace the buildout: HBM memory, NAND storage, precision timing, advanced PCBs/substrates, burn-in test, challenger & custom silicon beside NVIDIA, and the power/cooling/electrical/turbine gear behind the power gap. Identity + tier-tagged role reads only (no capex/commitment joins — these aren't data-center operators); their SEC filings ride the Filings watch below.

The power bank — who holds the megawatts

Named-build MW from the source-verified ledger, by power-procurement model (BTM = on-site generation · Colocated = adjacent nuclear/gas · Grid = utility interconnect) · announced targets, not current draw · shared campuses (e.g. Stargate Abilene) count once per named operator · toggle to plot GW against 2026 capex

Player × constraint grid

The dossier numbers as one scannable matrix · every cell is the same live join (nothing re-keyed) · dashes = not applicable or not disclosed

Neo-cloud scoreboard nullable

The AI-native challengers, cited-figures-only · blank ≠ zero — cells fill as filings passes land · concentration and debt cells live canonically in the counterparty sidebar on Capital

Filings watch primary

Latest disclosure-relevant SEC filings per tracked player (10-K / 10-Q / 8-K; 20-F / 6-K for foreign filers) · direct EDGAR links · a periodic filing dated after the commitment book’s last review is flagged — that is the signal to re-verify the filed numbers on Capital

Who did what this week

The weekly feed, indexed by player · auto-tagged by the twice-daily refresh

Go deeper

Evidence lives on the theme tabs: the commitment book · capex by operator · the named-project ledger & Graveyard · the gap & its scenarios · token economics · the Europe dashboard ↗ · open dataset ↓