Datacenters & Digital Infrastructure

Where inference runs, and what stops it running. This rung covers the building, the megawatts behind it and the queue to get those megawatts energised — the capex that is committed, the capacity that actually gets built, and the physical inputs that decide which of those two numbers is smaller. Power is not a footnote here; it is most of the subject.

What this rung prices

A token price is quoted per million tokens, but a token has to be produced somewhere, and the somewhere costs money before a single request arrives. This rung supplies the denominator: how many dollars of capital sit behind one megawatt of serving capacity, how long that megawatt takes to energise, and how much of the capital is idle while it waits. One operator's breakdown puts an all-in AI datacenter at roughly $59M per MW — about $30M of IT and $29M of shell, generation and plant — against about $15M/MW/year of lease revenue and roughly a four-year payback (Economics of a Megawatt of AI Data Center). Change the interconnection wait, the rack density or the packaging lead time and that payback moves, and the floor under the price of a served token moves with it. That is why the rung's central question is not "how much are they spending" but "how much of the spend has turned into something that can serve a request" — and the gap between those two is where the price of a finished task is actually set.

The load-bearing findings

• The shortage is queue throughput, not megawatts. Interconnection waits in data-center load zones run three to four years, against a combined PJM + ERCOT queue of over 300 GW across roughly 1,500 projects — enough capacity on paper to meet projected data-center demand through 2030. Renewables and storage make up 77% of PJM's active queue and 87% of ERCOT's, while gas clears "substantially more quickly"; monthly gas entries to ERCOT's queue rose about 150% after the One Big Beautiful Bill Act (AI Meets the Grid). That single distinction — deliverability rather than generation — is what makes on-site power worth its siting cost, because it skips the queue entirely (Power 2026).
• The buyer's cash flow has already inverted. Alphabet's Q2 2026 capex was $44.9B against $39.1B of operating cash flow, so free cash flow printed negative $(5.9)B; the gap was plugged with a $49.6B June equity raise explicitly earmarked for AI-infrastructure capex plus $20.3B of new senior notes, and long-term debt more than doubled to $98.2B in six months (Alphabet Q2 2026 8-K Ex-99.1). The build is no longer funded out of the ad business. Tesla's own print shows the same shape one order of magnitude down — capex $5.79B, up 142% year on year, free cash flow negative $(1.1)B, self-funded from cash (Tesla Q2 2026 8-K Ex-99.1) — see hyperscaler capex cycle.
• Three physical inputs bind at once, and capital is the only abundant one. TSMC's CoWoS advanced packaging and the N3 node are reported sold out through end-2026 with lead times into 2027 (TSMC Q2 2026); industry packaging capacity is heading to roughly 200,000 wafers/month in 2026 yet the supply-demand gap only narrows from about 20% to about 10% by year end (TrendForce on CoWoS); and Samsung's preliminary Q2 2026 guidance points to operating profit roughly 19× higher year on year on HBM and DRAM tightness — a price signal, not a supply measurement, and preliminary rather than final (Samsung Q2 2026 preliminary). See advanced packaging bottleneck and HBM supply tightness.
• A modern rack draws what a small building used to. The GB300 NVL72 is specified at 135 kW rack TDP, up to about 155 kW peak, with 72 B300 GPUs and 36 Grace CPUs, and heat removed roughly 90% by liquid and 10% by air (Lenovo GB300 NVL72 product guide) — a vendor datasheet, authoritative for the spec but not independent. At 135 kW nominal that is about 7.4 racks per megawatt, which is the arithmetic that ties rack counts to the capex-per-MW and grid-queue figures (rack power density).
• The construction is now big enough to show up in the national accounts. The Federal Reserve reports business fixed investment growing 11% annualised in Q1 2026 against 5.5% in 2025 and attributes most of the strength to AI-services infrastructure, while noting AI-related imports took roughly 0.4 percentage points off growth through net exports and that the productivity payoff has not arrived — labour productivity about 2.1% annualised since late 2019, with only a "modest" AI contribution (Monetary Policy Report, July 2026). Measured from the ground, data-center construction spending reached $58.1B year-to-date through May, more than four times the 2025 record pace, and data centers are now over 20% of all US nonresidential building starts (ConstructConnect, July 2026). This is datacenter construction as a macro variable.
• The demand behind the buildout may be counted more than once. CoreWeave's Q1 2026 print shows $2.08B of revenue against $7.70B of quarterly capex — roughly 3.7× revenue — an operating loss of $(144)M, a net loss of $(740)M, a $99.4B revenue backlog and over 3.5 GW contracted power, with a $21B Meta commitment named (CoreWeave Q1 2026). That backlog is contracted largely to the same hyperscalers whose own cloud backlogs sit one layer up, so the same end-demand can be booked twice.

What we do not know yet

• Nothing on this rung was measured here. Every figure above is read, not run: no site visited, no rack metered, no filing audited beyond reading it. The rack-power number comes from a vendor's own product guide, the capex-per-MW number from a single operator's contract pricing republished by an analyst, and the memory tightness from a preliminary guidance release rather than final segment results.
• How much HBM exists, and who gets it. The rung holds the price consequence and none of the supply: no bit-supply, wafer-allocation, HBM4 qualification or per-stack pricing data appears in any source (Samsung Q2 2026 preliminary). Allocation is the tradable question and it is entirely unanswered.
• What power for AI actually costs. The two price anchors on the rung — about $271/MWh implied for the Anthropic–TeraWulf lease without GPUs, against roughly $5,000/MWh for the SpaceX–Reflection arrangement with GPUs ready to go — are both the author's own derivations, one of them labelled napkin math, and neither is a market print (Power 2026). The gap between selling electricity and selling compute is the right shape; the numbers are not quotable as prices.
• Whether the neocloud model earns its cost of capital. CoreWeave's release omits the GPU depreciation schedule, and omits backlog duration, cancellability and customer concentration (CoreWeave Q1 2026) — which is exactly the disclosure the question turns on.
• Where the guidance actually landed. Alphabet's printed 8-K exhibit carries neither a full-year capex dollar guide nor a remaining-performance-obligation figure; both live in the call and the 10-Q, so the watched $180–190B guide and the $462B cloud backlog can be neither confirmed nor broken from the primary print (Alphabet Q2 2026; Alphabet Q2 2026 preview, which was fetched only partially behind a paywall).
• How the aggregate capex number is built. Two sources on the same year differ by roughly a hundred billion — around $725B for 2026 in one (AI Hyperscaler Capex 2026) against a $660–690B range in another (The $690B Infrastructure Sprint) and "in excess of $600B" in a third (Hyperscalers in 2026). The concept page reconciles the spread to guidance vintage, fiscal-versus-calendar basis and basket membership rather than to disagreement, but no single stated basis exists to quote.
• What the pipeline converts at. There are 770 future hyperscale facilities in the pipeline and 36 or more projects worth $162B blocked or delayed (Hyperscalers in 2026), and no source here tracks what share of an announced pipeline has historically energised, or with what lag.
• Whether owning your own power and compute works. Tesla's Cortex is the rung's one worked example of the vertically integrated path — a 500 MW campus paired with on-site storage (Tesla Cortex 2) — and it is a single site with no disclosed cost per megawatt to compare against the leased alternative (vertically integrated compute).

Read next

• carbon-direct-pjm-ercot-interconnection-queue-2026 — the queue numbers that turn "power is the constraint" into a measurable wait.
• nextbigfuture-ai-datacenter-capex-per-mw-2026 — the per-megawatt unit that makes the whole buildout modellable, caveats and all.
• googl-q2-2026-earnings — the primary filing where the buyer's free cash flow goes negative and the funding shifts to capital markets.