The Gigawatt Buildout

· Jai An

The AI labs have committed to a datacenter buildout that, on the high-end arithmetic, costs a single-digit percentage of US GDP — every year, for years. This essay works that claim from the biggest possible frame down through the unit economics of a single gigawatt to the one piece of electrical equipment currently gating the whole thing. Every number is sourced; the full source notes live in the appendix.

If you only remember five numbers

The biggest frame first

Somewhere between $3 trillion and $7.5 trillion is going to be spent building AI datacenters over the next five years, depending on whose scope you take. The low end is the BG2 headline — ~$3T of total AI capex the industry is trying to absorb this cycle. The high end is Jamin Ball's arithmetic in The Neocloud Boom: ~$50B to build a gigawatt, times ~150 GW of new capacity by 2030, equals $7.5T. David Cahn at Sequoia splits the difference by reframing the question: 100 GW is AI's "$8 trillion question," 250 GW is the "$20 trillion question."

Numbers that size stop being a tech-sector story. Spread over four and a half years, the high-end path is ~$1.7T per year against a ~$32T US economy — about 5% of GDP, every year. For calibration (Ball's comparison set): the railroad buildout of the 1880s peaked around 6% of GDP, telecom in the early 2000s was ~1.2%, and the Manhattan Project was ~0.4%. We are watching something between telecom and the railroads, and it is being financed mostly by five companies' operating cash flow.

This post walks that claim from the top down: how much compute exists today and how fast it's growing, what one gigawatt actually costs (and why published numbers disagree by 4–5×), who's paying the bill and with what instruments, what a deployed gigawatt is worth once it's live, what physically gates the build, and what breaks if the curve stops paying.

Where the world is in 2025

Roughly 20–25 GW of AI compute exists globally — that's Elon's framing on Cheeky Pint ("the world has maybe 20, 25 gigawatts of compute"), and it's consistent with Dylan Patel's numbers. About 10–15 GW of that was built in 2025 alone (Dario, on Dwarkesh), with ~20 GW of incremental US capacity added in 2025 (Dylan). The US has ~5,400 datacenters — more than every other major country combined, roughly 10× Germany at #2 (Cloudscene / Kobeissi). Hyperscalers are already the largest growth vector, but the clean current-capacity number is ~44–48% of global datacenter capacity in 2025, with ~60% projected around 2030.

In other words: the installed base is small relative to the ambition, and the growth rate is the entire story. A fifth to half of everything that exists was built in 2025 alone.

The trajectory to 2030

The most-cited path is Dario's curve — industry build roughly tripling every year:

Year — Industry GW (built that year) — Silicon-only @ $10–15B/GW (Dario) — Full system @ $40–60B/GW (Cahn/Jensen)
2025 — 10–15 GW — $100–225B — $400–900B
2026 — 30–40 GW — $300–600B — $1.2–2.4T
2027 — 60 GW (BG2 normalization) — $600–900B — $2.4–3.6T
2028 — 100 GW — $1.0–1.5T — $4.0–6.0T
2029 — 300 GW — $3.0–4.5T — $12–18T
2030 — up to 200 GW (ASML hard ceiling, Dylan) — $2.0–3.0T — $8–12T

Two dollar columns because the per-GW figure depends on what you count (next section). Dario's $10–15B is silicon only; Cahn/Jensen's $40–60B is the full powered shell + system. The full-system column blows past most published forecasts after 2028 — which is why the 2029–30 GW path is the load-bearing assumption to stress-test, not the dollar figure.

The ~200 GW number is a hard physical ceiling Dylan derives from EUV-tool throughput: ~3.5 EUV tools per GW of AI silicon, against ~700 EUV tools in the world by 2030 (ASML constraint). Sam Altman's stated target of ~52 GW/yr would consume a quarter of global lithography capacity for one lab. And note the contradiction hiding in the table: Dario's ~300 GW in 2029 is above Dylan's 200 GW ceiling for 2030. Either Dario is rounding the curve up, Dylan's EUV math is conservative, or the path bends before 2029. They can't all be right.

Per lab, the best estimates ("—" = not publicly disclosed):

Lab — End 2025 — End 2026 — End 2027 — Stated 2028–2030 ambition — Share of 2025 industry
OpenAI / Stargate — 2 GW — 6 GW — 12 GW — 50 GW (Sam target); $150B capex 2030 — 8–10%
Anthropic (rented, AWS+Google) — 1.5 GW — 4.5 GW — 8 GW (3.5 GW Google TPU online) — 10+ GW — 6–8%
xAI / Colossus — 0.5–1 GW (Memphis) — 2 GW (Memphis target) — 3+ GW (Colossus 2, 1M+ GPUs) — — — 3–5%
Meta (Hyperion + Prometheus) — 2 GW — 3+ GW — 5+ GW — multi-GW commitments — 8–10%
Microsoft — 4–5 GW (incl. OpenAI co-located legacy) — 6+ GW — 8+ GW — — — 20%
Google — 5+ GW (largest aggregate, undisclosed) — 7+ GW — 10+ GW — — (signed 1 GW just for demand response in early 2026) — 25%+
Industry (sum) — 20–25 GW — 30–40 GW — 60 GW — 100–300 GW/yr — 100%

Caveats: Anthropic and OpenAI numbers are consumed compute (rented). Google and Microsoft mix AI + cloud and don't publicly separate. Percentage shares are illustrative — they wouldn't sum to 100% if you double-count the OpenAI/Microsoft and Anthropic/Google cloud relationships.

This is also where Ball's 150 GW comes from, by a different route: OpenAI has talked publicly about 30 GW by 2030; assume Anthropic has similar plans and the two labs alone bring ~55 GW; if they're 30–40% of new builds, the industry adds >150 GW in 4.5 years. That lands comfortably inside Dario's path and well above the realistic construction floor (see the transformer section below) — a useful midpoint between the optimists and the pipeline data.

The realistic ladder, year by year

"3× per year" invites a tempting misreading: the installed base tripling — ~25 GW today, 75 GW in 2026, 225 GW in 2027. Nobody is claiming that. Dario's 3× is the build rate (GW constructed per year), and even that is the aspiration curve. What the grid physically delivers grows slower than both. Since the three ladders differ by hundreds of gigawatts within three years, they're worth laying side by side:

Year — "Stock ×3" misread (GW installed) — Dario's build curve, summed (GW installed) — Realistic power ladder (GW installed)
2025 — 25 — 22 — 20–25
2026 — 75 — 60 — 35–45
2027 — 225 — 120 — 55–75
2028 — 675 — 220 — 90–120
2029 — 2,000 — 520 — 140–190
2030 — — — 720 — 220–300

The realistic column grows ~1.6–1.8×/yr — doubling roughly every 18 months, a power-sector Moore's law. Three reasons it sits below the other two:

1. Compute grows faster than power by construction. Each chip generation does more compute per watt — Epoch AI measures ML-GPU energy efficiency doubling every ~3 years at FP32 (~1.26×/yr), with bigger step-changes from tensor cores and lower-precision formats (tensor-FP16 alone is ~10× FP32 throughput). SemiAnalysis measured total AI compute capacity growing 50–60% per quarter through 2023–24 while power grew far slower. So a 3×/yr compute story never required 3×/yr gigawatts; roughly 2×/yr of power carries it.

2. The grid delivers about half of what's scheduled. Goldman Sachs Research (on Aterio's facility-level tracker): US datacenter additions actually realized were 6.4 GW in 2024 and 8.5 GW in 2025; the construction schedule says 13.6 GW (2026) and 36.3 GW (2027), but historically only ~72% of capacity scheduled within four quarters goes live on time, and Goldman risk-adjusts to ~60% materializing on a one-year view, ~50% on two — supply chains, labor, and the 18–24-month build cycle eat the rest. That's the same gap Sightline's 11-of-12-GW-stalled number measures from the transformer side.

3. The model-based forecasts agree. SemiAnalysis's energy-dilemma model had AI datacenter critical IT power crossing 10 GW in early 2025 and reaching ~40 GW by 2026 (within global all-datacenter demand going 49 GW in 2023 → 96 GW in 2026) — steep doubling off a small base, decaying as the base grows. Goldman's headline is US datacenter power doubling 2025→2027, not tripling.

The dollars confirm the middle can't hold a 3× stock curve: the misread's 2027 step (+150 GW in one year) at $40–60B/GW all-in is $6–9T of capex in a single year — ~10× actual Big-4 spend (~$600B/yr). The realistic ladder's 2027 step (~+20 GW) is ~$0.8–1.2T/yr — right on Dylan's ~$1T full-ecosystem run-rate. And the 2030 landing zone (~220–300 GW installed) is exactly the range the money people are already debating — above Ball's ~150 GW of new capacity, spanning Cahn's "100 GW = $8T question, 250 GW = $20T question," and safely under Dylan's 200 GW/yr ASML ceiling on annual chip output.

So when someone says "AI compute triples every year," the year-by-year translation for the power ladder is: ~25 GW now → ~40 in 2026 → ~65 in 2027 → ~100 in 2028 — not 75 and 225. Compute can triple; gigawatts double at best, every 18 months, transformer queue permitting.

What one gigawatt costs

Published cost-per-GW figures range $10–60B, and the spread is mostly about what's included in the lens, not real disagreement:

Lens — What's counted — Number
Silicon / chips per delivered GW — GPUs + HBM only (not the building, not the power) — Dario: $10–15B/GW
Powered shell + facility per GW — Land, permits, building, transformers, switchgear, cooling, on-site power, networking — but not the compute itself — Chamath's project escalation traces this lens: $4–5B → $50B
Full AI system capex per delivered GW — Silicon + powered shell + networking + service margin + memory; the headline build cost — Cahn: $40B/GW · Jensen: $50–60B/GW with Vera Rubin

These stack rather than conflict: ~$10–15B of silicon riding on ~$25–45B of powered shell and infrastructure equals ~$40–60B all-in per delivered GW.

Ball's decomposition of the ~$50B adds line-item granularity for a Blackwell-era build: ~70% chips/compute (50% GPUs, 10% networking, 5% CPUs, 3% storage, 2% assembly), ~20% datacenter shell + power infrastructure (substations, transformers, switchgear, backup generation, HV interconnect), ~5% cooling, and ~5% land, permitting, and other.

One genuine disagreement worth flagging: Ball puts ~70% of the $50B on chips/compute — roughly $35B/GW — where Dario's silicon-only lens says $10–15B/GW. That's not a rounding difference; it inverts which slice is "big." The likely reconciliation is that Ball's 70% counts full GPU systems at market price — servers, networking, NVIDIA's margin baked in — while Dario is counting something closer to bare silicon cost. But it's the one place in this post where two careful sources are actively measuring the same lens and getting different answers, so it's flagged in the divergence table at the bottom.

The total bill, and who pays it

Multiple capex lenses converge on the same shape — steep through 2026, flattening after 2028:

Year — Mag 5 capex (BG2) — Big-4 hyperscaler (Dylan) — Total ecosystem (Dylan, incl. supply chain)
2023 — $156B — — — —
2024 — $241B — — — —
2025 — (jump) — $600B — $1.0T
2026 — $379–450B — rising — rising
2028 — $535B — — — —
2030 — $568B — — — —

The scopes nest — BG2's Mag 5 ⊂ Dylan's Big 4 + supply chain ⊂ Dario's whole industry — so the 1.5–2× gaps between them are definitional, not contradictions. Cumulative 2025–2030, the industry is trying to absorb ~$3T of AI capex on the narrow scope and $7.5T+ on the full-system × 150 GW scope. NVIDIA alone: ~$160B/yr of revenue today, consensus ~$350–400B/yr in five years.

An independent cross-check from the profitability-tracker side: Is AI Profitable Yet? tracks cumulative AI spend vs. AI revenue per company (May 2026 snapshot). Its Big-4 annual burn — Google $155B, Amazon $174B, Microsoft $153B, Meta $142B — sums to ~$620B/yr, right on Dylan's ~$600B Big-4 capex figure, and its ~$1.1T of cumulative Big-4 AI spend since 2022 is consistent with the BG2 capex curve. Same shape, measured from the spend ledger instead of the buildout.

The chart that first stopped me, from the BG2 "$3 Trillion CapEx" segment — Mag 5 capex as a share of operating cash flow, 2015 → 2030E:

Mag 5 capex as a share of operating cash flow — the share roughly triples from 27% in 2015 to ~75% by 2030 (data: BG2, Oct 2025)

The share roughly triples in fifteen years. The Mag 5 now spend the majority of every operating dollar on infrastructure, most of it AI. Whether that's the new normal or peak overshoot is the trillion-dollar question — but either way it's a different valuation regime from "asset-light platform business." These are utilities with a software business attached.

How the bill actually gets financed is at least as interesting as its size, because the instruments are getting exotic:

Five percent of GDP

Take the full-system arithmetic seriously for a moment: $50B per gigawatt × 150 GW = $7.5T over ~4.5 years ≈ $1.7T per year ≈ 5% of US GDP (~$32T/yr). Ball ran exactly this calculation and pulled the historical precedents that make it legible:

Buildout — Peak share of US GDP
Railroads, 1880s — 6%
AI datacenters, 2026–2030 (projected) — 5%
Telecom, early 2000s — 1.2%
Manhattan Project — 0.4%

So: bigger than telecom by ~4×, bigger than the Manhattan Project by an order of magnitude, just shy of the railroads. High, but with precedent — and the precedent cuts both ways. The railroads built infrastructure that compounded for a century and wiped out a generation of investors on the way; telecom's dark fiber eventually carried the internet, after the equity went to zero. A buildout at this share of GDP changes interest-rate sensitivity, energy markets, and industrial labor allocation simultaneously — it's already a macroeconomic event regardless of how the AI bet itself resolves. ($3T over five years is also ~4–5× the entire global VC market in its 2021 peak year of roughly $640B, for a sense of how far outside venture-scale this is.)

Dario's macro view bookends it from the demand side: he expects 10–20% real GDP growth per year post-AGI — against compute itself growing ~300%/yr. Even on the optimistic case, compute outruns the economy that's supposed to absorb it.

What a live gigawatt is worth

Capex is only half the ledger. The other half — and the reason the spend might be rational — is what a gigawatt is worth once it's energized and rented.

Ball's yardstick from the public neoclouds: CoreWeave and IREN both trade at roughly $90B of enterprise value per live gigawatt (Nebius meaningfully higher). He's upfront that this is "fun math" on a made-up multiple — but as a unit of account it does real work. If $90B of EV per deployed GW held across 150 GW of new capacity, that's $13.5T of implied enterprise-value creation in 4.5 years. Even discounting hard — much of the build is hyperscaler-internal and never sellable, so say neoclouds capture only ~20% — that's still $2.5T+ of equity value for the sector serving the buildout.

The revenue side roughly cross-checks the multiple. Dylan's rental math puts a GW at ~$10B/yr of compute revenue (Anthropic's incremental $40B of compute spend ≈ 4 GW of inference), Sarah Frier's earlier rule of thumb was "every gigawatt is about $10B in revenue," and CoreWeave's own guidance (~$18.5B annualized run-rate against ~1.7 GW by end of 2026) lands at ~$11M/MW — call it $10–11B of revenue per live GW per year. $90B of EV per GW is then ~8–9× revenue, which is rich but not insane for infrastructure with locked five-year contracts; H100 economics ($2.40/hr signed vs. ~$1.40/hr build cost) show 35%+ gross margins for whoever locked in early.

Numbers this large can't be served by a few players — which is Ball's actual thesis: expect a neocloud boom, a handful of very large players plus a long tail of independents (down to single-site operators), the way the independent power producer market evolved. Expect private equity to play a big role. And the SpaceX/xAI datapoint suggests the revenue ceiling per MW may be higher than the comps: per the SpaceX S-1, Anthropic is paying ~$15B/yr for roughly 500 MW of Colossus capacity — ~$30M/MW, ~3× the neocloud norm — which reads as a compute-constrained buyer stretching on a cancellable contract, and as a measure of how much desperation pricing exists in the current market.

What actually gates the build

None of the trajectory math survives contact with the supply chain unless the physical bottlenecks clear. Ranked by how binding they look right now:

1. High-voltage transformers — the active constraint. Sightline Climate's 140-project tracker (via Aakash Gupta) shows the announced-vs-built gap on 2026 US capacity is brutal: 12 GW announced, ~11 GW of it stalled at the announcement stage with no physical progress, and only ~5 GW separately under construction. A quarter of announced projects haven't disclosed a power strategy at all — vapor capex with a press release attached.

America's data-center boom is getting stuck in the pipeline — of 12 GW announced for 2026, ~11 GW is stalled and only ~5 GW is actually under construction (separate bucket)

The gating equipment is specific: high-voltage transformers, switchgear, grid-tie batteries. Pre-2020 lead time on a high-power transformer was 24–30 months; today it's ~5 years. Electrical equipment is less than 10% of datacenter cost and 100% of the bottleneck. When a hyperscaler announces $50B of capex, the Street books it as compute online in 18 months — but if the transformer order wasn't placed in 2022, that money sits as commitment without capacity. You cannot pay for a transformer that doesn't exist. The winners under this regime locked PPAs and electrical-equipment orders 3–4 years ago; everyone else is in line. Meanwhile GPUs bought against stalled shells depreciate on NVIDIA's annual cadence while paying carrying costs — every dark quarter is compounding waste. The "six months from running out of compute" panic is really "five years from running out of transformers." Capital fixes one; capital cannot manufacture the other.

This also reframes the on-site generation stories. xAI's ~15 gas turbines in Memphis and OpenAI's gas-fired Stargate site aren't just labs bypassing the grid for speed — they're bypassing the transformer queue, which binds grid-tie regardless of generation source. And even that path has a cap: gas-turbine vanes have 12–18 month lead times with only three casting companies worldwide able to make them (Elon's own flag).

2. Power and interconnect generally. Substations take 4–7 years; transmission upgrades take longer; ERCOT, MISO, and PJM interconnect queues run multi-year. The US grid grows ~2%/yr against China's ~30%/yr, and China's electricity output is already roughly 2.3–2.4× the US. Whoever locks interconnect wins the decade. The nuclear-restart era — Three Mile Island for Microsoft, Palisades, a stack of SMR commitments — is happening with AI load as one of the main demand catalysts.

3. Lithography. The ~200 GW/yr ceiling from EUV throughput, above. This is the constraint that bites last but absolutely.

4. Memory. HBM3e/HBM4 capacity is sold out through 2027; ~30% of 2026 big-tech capex is going to memory (Dylan). The downstream effects reach consumer electronics: DRAM spot up ~3× ($3–4/GB → $12/GB), iPhone BOM up ~$250, smartphone unit projections collapsing from 1.4B toward 500–600M as HBM crowds out low/mid-end memory. New fabs don't land until late 2027/2028. This is the most under-covered second-order effect of the buildout.

5. People and water. Datacenter electricians, HVAC techs, and power engineers gate how fast a 1 GW facility gets wired. Liquid cooling is table stakes at GW scale, so site selection increasingly turns on water rights and ambient temperature.

The honest synthesis: Dario's 30–40 GW (2026) is what the curve does if the supply chain unblocks; Sightline's ~5 GW actually under US construction is the floor of what's physically in progress. The truth for 2026 lands between them, and the gap is measured in transformers.

The power seller's view

Everything above looks at the buildout from the compute buyer's side. Dan Dreyfus of Bornite Capital told the same story from the seller side — pitching Talen Energy at All-In's best-ideas competition (Liquidity Summit 2026) — and his grid-level numbers slot into this post's frame almost exactly.

The demand snapback is visible in plain grid data. US annual power demand growth by decade (Grid Strategies / EIA, from his slides):

US power demand growth by decade — 8.8% in the 1950s decaying to 0.6% by the 2010s, then snapping back to 2.9% in 2024–25 (data: Grid Strategies / EIA, via Dreyfus)

Two decades of effectively zero demand growth — efficiency gains plus offshoring the power-hungry industry (aluminum smelters and the like) to China — and the grid's supply chain atrophied to match. Then the 2020s: 0.6% (2022) → 0.9% (2023) → 2.9% (2024) → 2.9% (2025). Demand growth roughly quintupled in three years, back to 1980s rates. His sharpest framing: AI doesn't create the tight power market — "AI demand just turbocharges" a cycle that was turning anyway on electrification and manufacturing reshoring. Even with zero datacenters, the efficiency well was dry.

PJM is the case study. PJM's own 2026 long-term load forecast says the one grid region (Pennsylvania–Jersey–Maryland, where Talen operates 12 GW) needs 106 GW of gross new capacity over 10 years — 66 GW of demand growth plus 40 GW of scheduled thermal retirements. For scale, that's more than Japan's entire average power draw (~99 GW). One US grid region needs to add more than Japan consumes, in a sector where (his line) infrastructure happens in geological time. His two glosses: the coal "retirements" mostly won't happen ("we ain't retiring those coal plants"), and — the part that rhymes with the transformer thesis above — PJM's merchant power price is still too low to stimulate new capacity: "the math still doesn't work." The scarcity premium isn't showing up in the spot market; it's concentrated entirely in bilateral hyperscaler deals. Which brings us to:

The panic is measurable in the PPAs. His "Buyers Panic" slide — hyperscalers paying premium prices for decade-plus power contracts:

Deal — Date — Size — Term — Pricing
Talen × AWS — Susquehanna nuclear plant, PA — Jun 2025 (expanded) — 1.92 GW of nuclear capacity — 17-year PPA through 2042 — Above-market; $18B contract value, full ramp by 2032
Constellation × Microsoft — Three Mile Island Unit 1, PA — Sep 2024 — 835 MW — 100% of plant output — 20-year fixed offtake — Above-market; restart 2028

On the Three Mile Island economics, Dreyfus is specific: market power runs ~$50/MWh; Microsoft offered ~$100/MWh guaranteed for 20 years — a 2× premium to restart the reactor whose 1979 meltdown gave nuclear its 30-year bad name. This is the power-market twin of the ~$30M/MW Colossus desperation pricing above: when compute is the constraint, buyers pay 2–3× market on decade-long commitments and consider it cheap. The regulatory pushback is already here — "if you're taking all this power off the grid, how are we going to heat the homes?" — and his working compromise is the same one Google's 1 GW demand-response deal gestures at: sign the PPA, but force the datacenter to put batteries and peakers behind it to survive the few genuinely stressed hours a day.

A datacenter is a refinery. His commodities-investor translation: a refinery takes crude in and ships gasoline out; a datacenter takes electricity in and ships tokens out. Same shape — big capital-intensive box, "$50 billion per gigawatt," with power as the feedstock. That's a fourth independent arrival at the full-system number (after Cahn, Jensen, and Ball).

And the market values the refinery ~30× the power plant feeding it. Talen owns ~8 GW of baseload (2 GW nuclear + 6 GW gas) at a ~$25B enterprise value — roughly $3B of EV per GW of generation, against the ~$90B of EV per live GW the neoclouds command (CoreWeave/IREN, above) and below its own ~$45B replacement cost. The grid asset trades below what it would cost to rebuild, while the thing it feeds trades at multiples of it. That spread is either the buy signal Dreyfus is pitching, or a market verdict that nearly all the value accrues to compute over electrons — probably both. (Caveat on this whole section: it's a stock pitch from a fund manager long Talen. The slide data carries primary sources — PJM, EIA, Grid Strategies — but the Talen-specific math is an interested party's.)

The space wildcard

The most uncertain claim in the whole space is Elon's: ~100 GW/yr of AI in space within ~5 years, scaling toward ~1 TW/yr before fuel supply binds (Cheeky Pint). The arithmetic underneath: 100 GW to orbit ≈ ~10,000 Starship launches/yr ≈ one launch per hour with 20–30 reusable Starships; ~1M tons/yr to orbit at ~100 kW/ton payload density; SpaceX and Tesla both building toward 100 GW/yr of solar-cell production.

The unit economics are less crazy than they sound — energy is only 10–15% of datacenter TCO today, and solar in space runs ~5× ground efficiency (no atmosphere's ~30% loss, no day/night cycle), so the marginal dollar to launch isn't fighting the dominant cost line. The execution risk is everything else: SpaceX flew ~165 Falcon 9 launches in 2025, so one launch per hour is a ~50–60× cadence step-up on a vehicle that hasn't completed operational orbital cargo at scale; "20–30 reusable Starships" assumes Falcon-class reuse cycles not yet demonstrated; and 100 GW/yr of solar-cell production is no longer the whole global deployment rate, but it would still require internalizing ~15–20% of annual world solar manufacturing against recent global additions of >600 GW/yr.

Reasonable base case: a few GW/yr in space by 2030, not 100. Elon's number is a ceiling, not a forecast. Worth tracking, not pricing in.

The regulation wildcard

The space wildcard is a supply question. The next two are demand questions — and demand is what the whole buildout is borrowing against. The first one is a brake.

Every gigawatt in this post is underwritten by an unstated assumption: that the frontier model it serves stays deployable. That assumption is softer than the financing treats it. On June 12 2026, the US Commerce Department invoked national-security export controls to bar Anthropic from giving any foreign national — including foreign nationals inside the US, and its own foreign-national staff — access to its two most capable models, Fable 5 and Mythos 5. The directive was too broad to comply with selectively, so Anthropic disabled the models for everyone; only the less capable Claude Opus 4.8 stayed live. A frontier model — one whose safeguards tested stronger than anything previously shipped — went from released to dark in a day, by government directive, not by market.

That's the channel the capex curve doesn't price. The build is justified by demand pacing capability — but demand only shows up if the capability is allowed to diffuse. A government that can kill-switch the top model can also gate which models reach which markets, slow the release cadence the revenue curve assumes, or fragment the frontier along national lines. Any of those bends the AI-diffusion timeline, and the diffusion timeline is exactly what converts a gigawatt of installed compute into the ~$10B/yr that pays for it.

Base case: this is friction, not a wall. Export controls reroute demand more than they destroy it — the compute still gets used, often by the next model down. But it adds variance from a new direction. The buildout's load-bearing risk used to be purely economic (does the curve keep paying); a political variable now sits beside it (is the curve allowed to keep paying), and it's the one with the least historical data to underwrite.

The recursive self-improvement wildcard

The regulation wildcard is the demand brake. This is the demand accelerant — and it's the one that would make 250 GW look conservative.

Anthropic's framing of recursive self-improvement — AI systems capable enough to design and train their own successors — carries one line that matters more than the rest for this essay: in that regime, "the pace of progress in AI development becomes determined entirely by the availability of compute." Compute stops being one input among several (talent, data, ideas) and becomes the input. The trend lines underneath aren't hypothetical: Claude already authors over 80% of merged production code at Anthropic, the length of task a model can finish autonomously is doubling roughly every four months, and one code-optimization benchmark went from a ~3× speedup in May 2025 to ~52× by April 2026.

If that curve holds, compute allocation flips from a cost center into the rate-limiter on intelligence itself — and the rational response to a rate-limiter is to buy all of it you can. Every lab in the per-lab table is already racing for inference capacity to serve today's demand; an RSI regime stacks a second, larger claim on the same gigawatts, because training the successor becomes a compute-bound race where being one generation ahead compounds. That's the scenario where Dario's 300 GW in 2029 stops looking like it overshoots Dylan's ASML ceiling and starts looking like the reason someone tries to break it.

Base case: unproven, and Anthropic is explicit that RSI is "not inevitable" and that it's "genuinely unclear whether today's training methods and architectures could unlock that capacity." But it's the asymmetric bet sitting under the whole buildout. If it doesn't arrive, the curve flattens and the impairment case below applies. If it does, no published GW forecast is high enough — and the transformer queue, not capital or ambition, decides who's standing in the right place when it hits.

What breaks it, and what it means

The demand side, so far, is keeping up. The open question that looked scariest a year ago — does revenue accelerate fast enough to justify the curve — looks provisionally answered by the Opus 4.5 capabilities jump and the step-up in Anthropic revenue that followed ($1B → ~$9–10B in 2025, another few billion added in January 2026 alone, targeting 20–30×/yr growth). Dylan's framing makes the coupling explicit: if Anthropic adds another $60B of revenue in ten months, it needs ~4 GW of incremental inference capacity just to serve the growth. Demand is pacing capex — for now.

But the structural risks compound:

1. If scaling laws plateau, the impairment cycle is historic. $568B/yr of Mag 5 capex assumes the curve keeps paying. If returns flatten before 2028–29, the write-downs are bigger than telecom dark fiber, bigger than 1999. 2. Labs without hyperscaler parents carry infrastructure risk that's hard to underwrite. Anthropic's ~$20B ARR implies someone laid out ~$50B of capex for them. xAI more so. If revenue growth misses, the impairment lands somewhere — and it's not obvious where. 3. Concentration risk is systemic. NVIDIA on track to control >70% of N3 wafer capacity by 2027. Three turbine-casting companies. One country refining ~98% of gallium. One EUV supplier whose entire ecosystem is ~10,000 people. Any single-point failure is a systemic event. 4. Attribution and sustainable funding matter more, not less. The next decade of AI economics has to amortize against $3–7.5T of capex. The bigger the bill, the worse a bad attribution model gets.

What I actually believe

Having walked all of it — the trajectory math, the cost lenses, the financing instruments, the bottleneck rankings — here's where I land.

The build is real, and it lands somewhere near 250 GW by 2030. Not the 700 GW you get by summing the aspiration curves, and nothing like a stall — the realistic ladder's ~220–300 GW, doubling every eighteen months. Compute keeps tripling anyway, because chip efficiency carries the difference. Every independent check in this post — the grid data, the capex arithmetic, the chip supply — points at the same landing zone.

Transformers are the real clock. Not capital: the largest balance sheets in the world are already committed. Not chips: the EUV ceiling doesn't bind until ~200 GW a year. A five-year queue for high-voltage transformers decides which announced gigawatts become live ones, and the winners were decided years ago, when somebody did or didn't place an electrical-equipment order. Capital can buy chips; it cannot manufacture time.

And the equity is the risky layer, not the infrastructure. That's the railroad precedent read correctly: the rails compounded for a century while a generation of investors got wiped out, and telecom's dark fiber carried the internet after the equity went to zero. If scaling returns flatten, the write-downs will be historic — but the gigawatts, substations, and interconnects stay built and get used. Whichever way the AI bet resolves, the country comes out the other side with the power grid it should have started building a decade ago.

This essay started as raw notes on BG2, Dwarkesh, 20VC, All-In, and Cheeky Pint episodes; the per-source notes live in the appendix below.

Where sources diverge

The headline numbers in this post come from people measuring different things. The biggest disagreements, and what to read into them:

Metric — Year — Sources & figures — Spread
Industry GW (built / stock / incremental) — 2025 — Dario 10–15 GW built · Dylan 20 GW US incremental · Elon 20–25 GW world stock · BG2 60 GW total (incl. replacement) — 4–6× depending on lens
Industry GW path — 2029–30 — Dario 300 GW (2029) · Dylan 200 GW ASML ceiling (2030) · Sam 52 GW/yr target · Ball 150 GW cumulative new by 2030 — Dario > physical ceiling
US additions, scheduled vs realized — 2026–27 — Goldman/Aterio scheduled 13.6 / 36.3 GW · historically 72% on time, risk-adjusted 60% / 50% materialize — 2× announce-vs-build gap
Cost per GW — — — Dario $10–15B · Cahn $40B · Jensen $50–60B · Ball $50B · Chamath project escalated $4–5B → $50B — 4–5× spread
Chip share of a full-system GW — — — Ball 70% of $50B ($35B) for chips/compute · Dario $10–15B silicon-only — 2–3×; systems-at-market-price vs. bare silicon
Mag-N / Big-N capex — 2026 — BG2 Mag 5 $379–450B · Dylan Big-4 $600B (+ $1T full ecosystem) — 1.5–2× (different scopes)
Mag-N / Big-N capex — 2028 — BG2 Mag 5 $535B · Dario industry $1.0–1.5T/yr — 2–3× (different scopes)
Revenue per live GW — — — Dylan $10B/yr rental · Frier $10B · SpaceX S-1 implies Anthropic paying $30M/MW (3× comps) — 3× at the desperation margin
NVIDIA revenue — this yr — Dylan $160B · BG2 $200B — 25%
NVIDIA revenue — 2030 — Dylan $350–400B · BG2 $350B — converge within 10%
China power vs US — today — Elon 2.3–2.4× (electricity output) · Dreyfus 3× (generation capacity) — output vs nameplate capacity — China's huge solar fleet inflates capacity more than output

What to read into this:

1. GW numbers aren't comparable without a denominator. "World stock" (Elon, 20–25), "US incremental" (Dylan, ~20), "industry-built that year" (Dario, 10–15), and "total normalized incl. replacement" (BG2, ~60) are four different metrics. Any single GW headline without the lens is hand-waving. 2. Cost per GW depends on what you count. Dario's $10–15B is silicon-only. Cahn's $40B, Jensen's $50–60B, and Ball's $50B include the powered shell, networking, and next-gen silicon. Chamath's $4–5B → $50B is the same delta in motion on one project. The chip-share row is the only live disagreement. 3. Dario's 2029 path breaks Dylan's ASML ceiling. 300 GW > 200 GW hard cap. They can't all be right. 4. The capex scopes nest (BG2 Mag 5 ⊂ Dylan Big-4 ⊂ Dario industry), so those aren't contradictions. The shape they agree on: steep through 2026, flattening after 2028.

---

<summary>Appendix: source notes — casual readers, you're done. Expand for the raw per-source notes the essay is built from.

The raw, per-source notes the essay above is built from. Kept verbatim for reference-chasing; half-life on accuracy is short.

The capex curve (BG2 chart)

The BG2 "$3 Trillion CapEx: Are We Overbuilding AI?" segment, Oct 14 2025 — Mag 5 capex vs. operating cash flow, 2015 → 2030E:

Big lab compute footprints (current)

Numbers here are rough — half from press, half from podcasts.

### OpenAI / Stargate (via Microsoft + Stargate LLC) - End of 2025: ~2 GW of operational compute capacity (≈ output of 2 average nuclear plants). - Multi-site, headline ~10 GW total commitment, Abilene TX as anchor, Oracle as compute partner, Microsoft no longer exclusive. - Bloomberg / SoftBank-Stargate: 10 GW center described by Son + Trump officials would be among the largest in the world. Site = former uranium enrichment complex owned by US Energy Department. Powered by ~$33B of natural-gas-fired electricity. - Vast majority of compute reportedly signed on five-year deals. - Incremental added compute is where cost explodes. - Dylan Patel: "OpenAI went from 600 MW to 2 GW last year, 2 GW → 6+ GW this year, 6 → 12 GW next year." [Dwarkesh × Dylan Patel, ch. "ASML constraint" 34:34] - Sam wants 50 GW. - Stated 30 GW by 2030 (OpenAI Newsroom).

### xAI / Colossus (Grok) - Memphis, TN. Rapid local infra build. Scaled to ~200K H100s in record time, on-site gas turbines because the grid couldn't deliver. - Approval for 300 MW (0.3 GW) from Tennessee Valley Authority. - ~15 mini gas turbines on-site, hundreds of additional MW. - Second facility underway, third building purchased. - Plan: total power needs across Memphis region → 2 GW. - Colossus 2 reportedly targeting 1M+ GPUs. - Per the SpaceX S-1 (via Ball): SpaceX (post-xAI acquisition) rents Colossus capacity to Anthropic at ~$15B/yr for ~500 MW (Colossus 1 ~300 MW + assumed ~40% of Colossus 2's operational ~500 MW) — ~$30M/MW vs. $10–12M/MW neocloud comps, with a 90-day out clause. Cursor also has announced capacity. SpaceX's full-year 2025 revenue was $18.7B; the Anthropic deal alone nearly matches it.

### Anthropic / AWS Project Rainier - Doesn't build its own physical datacenters — secures cloud compute from Google + AWS. - Trainium2 cluster, multi-site, "hundreds of thousands" of accelerators. Anthropic's Claude training stack. - 1 GW of compute coming online in 2026 via Google Cloud agreement. - Just signed massive expansion with Broadcom + Google: another 3.5 GW of Google TPU capacity starting 2027. - Total secured footprint: ~4.5 GW over the next year or two. - Dylan Patel: Anthropic needs to get well above 5 GW by end of this year (tough but possible). - Anthropic ARR ~$20B. Margins sub-50% (per Information). → ~$13–14B compute rental → ~$50B of CapEx someone laid out for them.

### Meta / Hyperion + Prometheus - Louisiana (Hyperion, ~2 GW), Ohio (Prometheus). Zuck has been public about multi-GW build commitments.

### Google - Total global capacity arguably the largest of the group, dozens of datacenters worldwide. Distributed across existing hyperscale footprint, TPU-heavy, less single-site headline-grabbing but probably the largest aggregate. - Don't publicly separate AI vs. cloud GW. - Accounted for a large chunk of the 20+ GW of corporate clean-power agreements signed by big tech in 2025. - Early 2026: signed 1 GW of capacity just for demand response — enough AI workloads to dial back 1 GW to stabilize local grids.

### Microsoft (Satya, 2025) - Wisconsin, plus the OpenAI co-located footprint pre-Stargate. - "Just last year, we added over 2 gigawatts of new capacity — roughly the output of 2 nuclear power plants." - Reportedly stepping back from some leases as OpenAI diversifies.

Cost / unit economics

### David Cahn @ Sequoia, 20VC, Oct 27 2025 - "Sam Altman now talks about gigawatts every day. He's not talking about dollars anymore… we're moving away from dollars and we're moving toward gigawatts." [20VC, "Physical Data Centres is a Moat" 5:04] - Rough math: 1 GW = $40B to build out. - Jensen: $50–60B with Vera Rubin. - 100 GW buildout (what people are talking about now) = AI's $8 trillion question. - 250 GW = AI's $20 trillion question.

### Jamin Ball, Clouded Judgement 5.22.26 — The Neocloud Boom - ~$50B total cost to build 1 GW (Blackwell-dominated): ~70% chips/compute (50% GPUs, 10% networking, 5% CPUs, 3% storage, 2% assembly), ~20% shell + power infra, ~5% cooling, ~5% land/permitting/other. - ~150 GW of new capacity in next 4.5 years (from OpenAI 30 GW by 2030 + Anthropic assumed similar = ~55 GW, at 30–40% share of new builds). - $50B × 150 GW = $7.5T total ≈ $1.7T/yr ≈ 5% of US GDP (~$32T). - Historical precedents: railroads 1880s ~6% of GDP, telecom early 2000s ~1.2%, Manhattan Project ~0.4%. - ~$90B enterprise value per live GW (CoreWeave, IREN; Nebius higher). 150 GW × $90B = $13.5T implied EV creation; ~20% neocloud capture ≈ $2.5T+. - CoreWeave ~6× NTM revenue, Nebius/IREN ~10×; EBITDA: CoreWeave ~10×, IREN/Nebius ~20×. - Neocloud market structure prediction: a handful of very large players + long tail incl. single-site independents, like the IPP market; PE plays a big role. - SpaceX S-1: Anthropic paying ~$15B/yr for ~500 MW of Colossus (~$30M/MW vs. $10–12M/MW comps; CoreWeave's guidance implies ~$11M/MW), 90-day out clause.

### Jaime Sevilla (@sevillamol), Jan 5 - "Bitcoin arguably has much more compute power than Google" — disputes this. - Bitcoin hashing infra worldwide ≈ $30B ≈ same as cost of one 1 GW AI datacenter. - Google produced and installed TPUs worth more just in 2025 Q3.

### Chamath - Greenlit a project he thought was $4–5B → went to $10B → $15B → $20B → upwards of $50B for the powered shell, land, permits, infrastructure, people, all of it. - Sarah Frier (~year ago): every gigawatt is about $10B in revenue.

Dylan Patel (SemiAnalysis) on Dwarkesh

Source: Dwarkesh × Dylan Patel — full episode. Chapters: H100 economics 0:00 · ASML constraint 34:34 · Memory crunch 1:16:01 · Power scaling 1:42:34.

Per-hyperscaler capex (this year → next): [~1:05:37] - Google: $180B → $300B. - Amazon: ~$200B. - NVIDIA revenue: ~$160B/yr (~$40B/qtr).

Recent lab raises: OpenAI just raised $110B; Anthropic just raised $30B. Anthropic adding $4–6B revenue/month.

TSMC capex: $30B + $30B + $40B = ~$100B over 3 years.

Per-GW silicon recipe (back-of-envelope from Dylan): - ~55,000 wafers of 3nm + ~6,000 of 5nm + ~170,000 of DRAM per GW of AI capacity. - ~2M EUV passes per GW (~1.1M for logic alone). - EUV tool ≈ $300–400M × 3.5 tools = ~$1.2B per GW just in litho. - NVIDIA on track to control >70% of N3 wafer capacity by 2027. - ASML supply chain ≈ 10,000 people worldwide.

Memory crunch downstream effects: [1:16:01] - DRAM spot: $3–4/GB → $12/GB (≈3x). - iPhone memory cost: $50 → $150 (+$100); total iPhone BOM up ~$250. - Smartphone unit projections crash from 1.4B → 1.1B → 800M → 500–600M as HBM crowd-out gutters the low/mid-end. - New memory fabs don't come online until late 2027 / 2028.

Power-growth gap (China): [Power scaling 1:42:34] - US grid: ~2% power growth/yr vs. China ~30%/yr. - Amazon's fastest datacenter build start-to-online: ~8 months (vs. 2–3 yrs typical).

Dario (Dwarkesh)

Source: Dwarkesh × Dario Amodei — "We are near the end of the exponential". The compute-trajectory quote is in chapter "If AGI is imminent, why not buy more compute?" 46:20.

Anthropic revenue ladder + economics [~12:36 + ~58:49]: - 2023: $0 → $100M. 2024: $100M → $1B. 2025: $1B → ~$9–10B. Jan 2026 alone added "another few billion." - Targeting 20–30x/yr growth (vs. current ~10x). - Compute split: roughly 50/50 train vs. inference today; inference gross margin >50%. - Profitability target: ~2028. - Headcount: ~2,500 people. - Computer-use benchmark (OSWorld): ~15% → 65–70% in a year. - Macro: Dario expects 10–20% real GDP growth/yr post-AGI — vs. compute itself growing ~300%/yr. (i.e. compute outruns the economy that's supposed to absorb it.)

Elon (Cheeky Pint w/ John Collison + Dwarkesh, Feb 5 2026) — AI in space

Source: Cheeky Pint Ep. 21 — full episode. Most quotes here come from chapter "Space GPUs" 0:23; some recur in "Space GPUs redux" 2:34:58.

Space-GPU unit economics (the part that actually makes this less crazy): - Energy is only ~10–15% of datacenter TCO today — i.e. the marginal $ to launch isn't fighting the dominant cost line. - Solar in space: ~5x ground efficiency; atmosphere costs ~30% loss + day/night cycle. - Solar cells in China today: $0.25–0.30/W. - 100 GW to orbit ≈ ~10,000 Starship launches/yr ≈ 1 launch/hour, doable with 20–30 reusable Starships. - Logistics target: ~1M tons/yr to orbit in 3–4 years, at ~100 kW/ton payload density. - Cooling overhead in space adds ~40% to power; servicing reserve adds 20–25%. - Ground equivalent: 1 TW solar = 4 TW of panels at 25% capacity factor ≈ ~1% of US land area.

Industrial bottlenecks Elon flagged: - Gas-turbine vanes/blades have 12–18 month lead times, with only 3 casting companies worldwide capable of making them — a hard cap on how fast on-site gas generation can scale. - China accounts for ~98–99% of primary low-purity gallium production/refining. - China's electricity output is currently closer to ~2.3–2.4× US output, not 3× yet, though the gap is still widening.

Total Musk-co headcount (Tesla + SpaceX + xAI + …): ~200,000 people — for context on the labor scale of any 100 GW/yr ambition.

BG2 Pod, Oct 14 2025 — "$3 Trillion CapEx: Are We Overbuilding AI?"

Source: BG2 — "AI Bubble, Stablecoin Boom, and Runnin' Down a Dream". Relevant chapters: "The AI CapEx Bubble?" 7:00 · "$3 Trillion CapEx: Are We Overbuilding AI?" 19:15 · "MAG5 CapEx Surge" 21:45 · "OpenAI's Race for Escape Velocity" 25:30.

Kobeissi Letter, Oct 28 2025

Aakash Gupta on the transformer bottleneck (Sightline Climate data)

Source: [Aakash Gupta on X]. Direct evidence for the "power not GPUs" thesis — and worse than I'd been pricing in. Sightline Climate's 140-project tracker shows the announced-vs-built gap is now ~50% on 2026 capacity, with grid transformers (not silicon, not permits) as the active bottleneck.

The Sightline Climate numbers (2026 US capacity, 140 projects tracked):

The actual gating constraint — high-voltage transformers, switchgear, grid-tie batteries:

Dan Dreyfus (Bornite Capital) — Talen Energy pitch @ All-In Liquidity Summit

Source: All-In — "Best Ideas Pitch Competition", Dreyfus segment from 13:09. Slide data transcribed from the stream; quotes checked against the YouTube transcript.

Power Cycles slide (sources on slide: Grid Strategies "Strategic Industries Surging: Driving U.S. Power Demand", Dec 2024; EIA Electric Power Monthly) — US annual power demand growth by decade:

PJM Capacity Forecast slide (sources on slide: PJM 2026 Long-Term Load Forecast, Jan 2026; PJM 2025 Regional Transmission Expansion Plan, Apr 2026; OCCTO Japan FY2024; IEA World Energy Investment 2025):

Buyers Panic slide (hyperscalers paying premium prices for decade-plus power contracts):

The Talen pitch itself (interested party — he's long):

Goldman Sachs / Aterio — realized vs scheduled US additions

Source: Goldman Sachs Research — "US data center power demand projected to double by 2027" (Wei, Struyven, Dart; Aterio facility-level activation data: locations, permitting, construction status, satellite imagery).

Epoch AI — chip efficiency trend

Source: Epoch AI — "Trends in machine learning hardware" (47 ML accelerators 2010–2023 + 1,948 GPUs).

SemiAnalysis — AI datacenter energy model (Mar 2024)

Source: SemiAnalysis — "AI Datacenter Energy Dilemma" (3,500+ North American datacenters tracked + accelerator-model power demand; free half).

Sources

Primary podcasts (with YouTube + chapter timestamps inline above):

Written:

Other: