The Infrastructure of Intelligence

Commands like [map.highlight] are Stagehand directives — they control the map renderer and pass through schema validation before any visual effect reaches the public output.

[camera.establish_globe lat=20 lon=10 zoom=1.6]
[map.basemap kind=dark]

[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=intro eyebrow="CLIO AGIS" title="The Infrastructure of Intelligence" subtitle="$700 billion. 1,000 terawatt-hours. The machines that power AI need a place, a grid, and a supply chain."]
[mark.clip type="hook"]

There is a capital spending race underway that does not fit inside a quarterly earnings headline.
In 2026, five companies — Microsoft, Alphabet, Amazon, Meta, and Oracle — are on track to spend somewhere between six hundred and seven hundred billion dollars on AI infrastructure.
That is not the total over a decade. That is a single calendar year.
This episode maps where the money is going, what the machines physically require, where the chips come from, which nations are treating compute as sovereign infrastructure, and why the fusion wildcards matter to the power equation.

[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

// ==========================================
// ACT I · THE GEOGRAPHY OF COMPUTE
// ==========================================

[map.clear annotations]
[map.basemap kind=light]
[camera.establish_globe lat=38.9 lon=-77.5 zoom=8]
[chat.say source="hyperscaler_capex_2026"]

The machines need a place.
Not just any place — a place with cheap land, access to water, fiber backbone, and enough grid capacity to run a small city's worth of power.
[source.show id="hyperscaler_capex_2026" text="Microsoft, Alphabet, Amazon, Meta, Oracle: combined 2026 capex guidance $660–$690B, the vast majority directed at AI compute, data centers, and networking." confidence=0.85]
Northern Virginia is the densest cluster of data centers on earth.
Loudoun County alone holds roughly two hundred operational facilities — more than any other county in the world.
[map.view lat=33.45 lon=-112.0 zoom=9]
Phoenix, Arizona: nearly two hundred more, with chip fabrication plants rising alongside them.
[camera.center target="city:dublin_irl" zoom_level=regional padding=110]
[map.highlight entity="country:ireland" color="#38bdf8" opacity=0.35]
Dublin and the Irish midlands: where American hyperscalers found European land, cool maritime air, and favorable tax treatment.
[camera.center target="city:amsterdam" zoom_level=regional padding=110]
[map.highlight entity="country:netherlands" color="#38bdf8" opacity=0.35]
Amsterdam: the fiber crossroads of Europe, hosting capacity that routes half the continent's traffic.
[camera.center target="city:singapore" zoom_level=regional padding=110]
[map.highlight entity="country:singapore" color="#38bdf8" opacity=0.35]
Singapore: gateway to Southeast Asian demand, increasingly capacity-constrained by land and water limits imposed by the city-state itself.
[camera.establish_globe lat=20 lon=10 zoom=1.6]
[map.highlight entity="country:united_states" color="#38bdf8" opacity=0.15]
[map.highlight entity="country:ireland" color="#38bdf8" opacity=0.25]
[map.highlight entity="country:netherlands" color="#38bdf8" opacity=0.25]
[map.highlight entity="country:singapore" color="#38bdf8" opacity=0.25]
These clusters are not accidents.
They are the geographic sediment of two decades of real-estate, regulatory, and grid arbitrage.
And they are all straining at once.

// ==========================================
// ACT II · THE POWER EQUATION
// ==========================================

[map.clear annotations]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=chapter eyebrow="CLIO AGIS" title="The Power Equation" subtitle="17% surge in 2025. 1,000 terawatt-hours by 2026. The grid was not built for this."]
[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

[camera.establish_globe lat=38 lon=-98 zoom=3.8]
[map.basemap kind=dark]
[layer.on admin1_regions]
[chat.say source="iea_datacentre_electricity_2025"]

Data center electricity use surged seventeen percent in 2025.
AI-focused data centers climbed even faster — well outpacing global electricity demand growth of three percent.
[source.show id="iea_datacentre_electricity_2025" text="IEA: global data center electricity consumption projected to exceed 1,000 TWh by end of 2026 — equivalent to Japan's entire annual electricity use. AI-focused data centers are set to triple power use by 2030." confidence=0.92]
By the end of 2026, the International Energy Agency projects total data center consumption will exceed one thousand terawatt-hours — the equivalent of Japan running for a full year on electricity.
In the United States alone, total data center power demand is on track to nearly double — from eighty gigawatts in 2025 to one hundred and fifty gigawatts by 2028.
[layer.off admin1_regions]

The hyperscalers are not waiting for the grid to catch up.
They are buying the power plants.

[map.highlight entity="region:pennsylvania" color="#f59e0b" opacity=0.45]
[map.label entity="region:pennsylvania" text="Pennsylvania · nuclear power corridor"]
Pennsylvania is now the epicenter of AI nuclear procurement.
Microsoft signed a twenty-year agreement with Constellation Energy to restart Unit One at Three Mile Island — rebranded the Crane Clean Energy Center — at eight hundred and thirty-five megawatts, dedicated entirely to Microsoft's data center load.
[source.show id="iea_datacentre_electricity_2025" text="Microsoft-Constellation PPA: 835 MW from Three Mile Island restart. Amazon-Talen Energy: 1,920 MW from Susquehanna through 2042. Meta-Constellation: 1.1 GW from Clinton Clean Energy Center, Illinois." confidence=0.88]
Amazon expanded its agreement with Talen Energy to one thousand nine hundred and twenty megawatts from the Susquehanna Steam Electric Station through 2042.
[map.clear annotations]
[map.view lat=40.5 lon=-89.0 zoom=7]
Meta signed a twenty-year deal for one-point-one gigawatts from the Clinton Clean Energy Center in Illinois.
These are not renewable energy credits.
They are dedicated capacity agreements — the hyperscalers locking in baseload carbon-free power for decades, because inference workloads run twenty-four hours a day and solar does not.

// ==========================================
// ACT III · THE COOLING PROBLEM
// ==========================================

[map.clear annotations]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=chapter eyebrow="CLIO AGIS" title="The Cooling Problem" subtitle="1.6 billion gallons. Loudoun County, Virginia, 2023. Potable water."]
[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

[map.view lat=38.95 lon=-77.5 zoom=9]
[map.basemap kind=satellite]
[chat.say source="iea_datacentre_electricity_2025"]

Power is only half the physical constraint. The other half is water.
A typical one-hundred-megawatt AI data center uses between one-point-five and three million cubic metres of water per year for evaporative cooling.
[source.show id="iea_datacentre_electricity_2025" text="Loudoun County, Virginia: potable water use grew 250% between 2019 and 2023, reaching 1.6 billion gallons in 2023. Two-thirds of data centers built since 2022 are located in water-stressed regions." confidence=0.85]
In Loudoun County, Virginia — the epicenter of the Northern Virginia cluster — potable water use grew two hundred and fifty percent between 2019 and 2023, reaching one-point-six billion gallons.
That is the same water supply that serves municipal drinking needs.
[map.view lat=33.45 lon=-112.0 zoom=9]
Phoenix, Arizona hosts nearly two hundred data center and chip fabrication sites.
Projections suggest metropolitan Phoenix data center water consumption could grow tenfold — to three-point-eight billion gallons per year.
In a city already managing a Colorado River allocation deficit, that is not a footnote.

The industry knows this.
Average rack power density rose from eight kilowatts in 2021 to seventeen kilowatts by 2024.
By early 2026, AI-driven racks frequently exceed fifty kilowatts per rack — a sixfold increase in five years.
At those densities, air cooling fails.
Liquid cooling — direct-to-chip or full immersion — reduces direct water consumption by seventy to ninety percent.
But the transition takes capital and time, and the water bills are already due.

// ==========================================
// ACT IV · THE CHIP CORRIDOR
// ==========================================

[map.clear annotations]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=chapter eyebrow="CLIO AGIS" title="The Chip Corridor" subtitle="Taiwan. South Korea. Mexico. Five countries. One supply chain. One chokepoint."]
[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

[camera.center target="corporation:tsmc" zoom_level=regional padding=100]
[map.basemap kind=light]
[map.highlight entity="country:taiwan" color="#ef4444" opacity=0.5]
[map.label entity="country:taiwan" text="TSMC · N4 node · Blackwell wafers"]
[chat.say source="nvidia_gb200_supply_chain_2026"]

Every GPU that powers an AI cluster begins here.
Taiwan Semiconductor Manufacturing Company fabricates the silicon wafers for NVIDIA's Blackwell line on its N4 process node.
[source.show id="nvidia_gb200_supply_chain_2026" text="TSMC produces NVIDIA Blackwell wafers on N4 process. NVIDIA holds 70%+ of TSMC's CoWoS advanced packaging capacity. CoWoS lines: ~75,000 wafers/month in 2025, expanding to 95,000 by 2026." confidence=0.82]
NVIDIA holds over seventy percent of TSMC's CoWoS advanced packaging capacity — the bonding technology that stacks GPU die and high-bandwidth memory into a single compute module.
CoWoS lines ran at full saturation through 2025.
Meaningful capacity relief is not expected until the second half of 2026.

[map.clear annotations]
[camera.center target="country:south_korea" zoom_level=regional padding=100]
[map.highlight entity="country:south_korea" color="#f59e0b" opacity=0.5]
[map.label entity="country:south_korea" text="SK Hynix · HBM3e memory"]
The memory stack comes from a different country.
SK Hynix in South Korea produces the HBM3e high-bandwidth memory that bonds to every Blackwell chip.
Samsung raised HBM3e supply prices nearly twenty percent for 2026 contracts, reflecting demand pressure across the entire stack.

[map.clear annotations]
[map.view lat=20.67 lon=-103.35 zoom=7]
[map.highlight entity="city:guadalajara_mex" color="#22c55e" opacity=0.8 pulse=true]
[map.label entity="city:guadalajara_mex" text="Foxconn · GB200 assembly · world's largest"]
Final assembly of the GB200 superchip — the server rack that anchors hyperscaler AI clusters — is happening in Guadalajara, Mexico, where Foxconn is building the world's largest GB200 assembly facility.

[map.clear annotations]
[map.fit entities="country:taiwan,strait:taiwan_strait,country:south_korea,country:japan" padding=100 maxZoom=4]
[map.highlight entity="strait:taiwan_strait" color="#ef4444" opacity=0.75 pulse=true]
[map.label entity="strait:taiwan_strait" text="Taiwan Strait · AI supply chain chokepoint"]
[map.arrow from="country:taiwan" to="country:south_korea" color="#f59e0b"]
[map.arrow from="country:taiwan" to="country:united_states" color="#38bdf8"]
The Taiwan Strait is the most consequential chokepoint in the global AI supply chain.
A disruption to TSMC production — through conflict, natural disaster, or political closure — does not merely affect chip prices.
It halts the production of every advanced GPU on the planet for months to years.
The hyperscalers know this.
TSMC's Arizona fabs are beginning N4 production, and NVIDIA is now producing its first US-made Blackwell wafers there.
But those chips still travel back to Taiwan for final CoWoS packaging.
The decoupling is partial.
The exposure is real.

// ==========================================
// ACT V · SOVEREIGN AI
// ==========================================

[map.clear annotations]
[flow.clear]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=chapter eyebrow="CLIO AGIS" title="Sovereign AI" subtitle="States are treating compute as critical infrastructure. The race is not just corporate."]
[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

[camera.establish_globe lat=24 lon=54 zoom=5]
[map.basemap kind=light]
[chat.say source="sovereign_ai_gcc_2026"]

The race for AI compute is no longer only a corporate story.
It is becoming a state story.
[map.highlight entity="country:uae" color="#f59e0b" opacity=0.6 pulse=true]
[map.label entity="country:uae" text="UAE · 5 GW AI campus · 10 square miles"]
[source.show id="sovereign_ai_gcc_2026" text="UAE-US AI Campus: 5 GW capacity, 10 square miles — the largest AI infrastructure campus outside the United States. UAE data center capacity projected to triple to 3.3 GW by 2030." confidence=0.82]
The UAE is constructing a five-gigawatt AI campus — ten square miles of infrastructure — the largest AI facility outside the United States.
Microsoft has committed fifteen-point-two billion dollars to UAE investment through 2029.

[map.clear annotations]
[map.highlight entity="country:saudi_arabia" color="#ef4444" opacity=0.45]
[map.label entity="country:saudi_arabia" text="HUMAIN · $100B PIF · 6 GW by 2034"]
Saudi Arabia's sovereign AI firm HUMAIN — backed by over one hundred billion dollars in planned investment from the Public Investment Fund — is targeting eleven data centers with a combined capacity of twenty-two hundred megawatts, with a roadmap to six gigawatts by 2034.
[source.show id="sovereign_ai_gcc_2026" text="Saudi HUMAIN: $100B+ PIF backing, 2,200 MW target, 6 GW by 2034. GCC region data center investment accelerating to $5–7B in 2026." confidence=0.80]
Two large campuses are already under construction.

[map.clear annotations]
[camera.center target="country:india" zoom_level=regional padding=100]
[map.highlight entity="country:india" color="#22c55e" opacity=0.35]
[map.label entity="country:india" text="Reliance · 1 GW Gujarat · $20–30B"]
Reliance Industries is constructing a one-gigawatt data center in Gujarat, India, powered by NVIDIA Blackwell processors, at an estimated investment of twenty to thirty billion dollars.

[map.clear annotations]
[camera.establish_globe lat=24 lon=54 zoom=3.5]
[map.highlight entity="country:uae" color="#f59e0b" opacity=0.45]
[map.highlight entity="country:saudi_arabia" color="#ef4444" opacity=0.35]
[map.highlight entity="country:india" color="#22c55e" opacity=0.3]
McKinsey projects global demand for data center capacity will reach two hundred and twenty gigawatts by 2030.
AI workloads alone are expected to account for one hundred and fifty-six gigawatts of that figure, up from forty-four gigawatts in 2025.
The nations that secure compute capacity in this decade are positioning themselves for a structural advantage in every intelligence-intensive domain that follows.

// ==========================================
// ACT VI · THE FUSION WILDCARD
// ==========================================

[map.clear annotations]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=chapter eyebrow="CLIO AGIS" title="The Fusion Wildcard" subtitle="Commonwealth Fusion. Helion. TAE Technologies. The long-run answer to the power equation."]
[scene.title kind=clear]
[scene.fade opacity=0 duration=500]

[map.view lat=42.3 lon=-71.1 zoom=9]
[map.basemap kind=dark]
[chat.say source="commonwealth_fusion_sparc_2026"]

If the power equation has a long-run structural solution, it is not gas plants or even conventional nuclear.
It is fusion.
[source.show id="commonwealth_fusion_sparc_2026" text="Commonwealth Fusion Systems SPARC: first plasma 2026, net energy gain demonstration targeted 2027. Italian firm Eni signed a $1B+ power offtake agreement in late 2025. Google-backed CFS." confidence=0.80]
Commonwealth Fusion Systems, headquartered outside Boston, is building SPARC — a compact tokamak designed to achieve first plasma in 2026 and demonstrate net energy gain in 2027.
Italian energy company Eni signed a power offtake agreement with CFS in late 2025 valued at over one billion dollars.

[map.view lat=47.7 lon=-122.2 zoom=9]
Helion Energy, headquartered outside Seattle and backed by Microsoft with a fifty-megawatt power purchase agreement targeting 2028, reported the first deuterium-tritium plasma operation by a privately funded fusion device in January 2026 — reaching temperatures of approximately one hundred and fifty million degrees Celsius.

[map.clear annotations]
[camera.establish_globe lat=30 lon=0 zoom=2]
TAE Technologies announced it would proceed directly to its first power-plant prototype — Da Vinci — targeting commercial electricity production in the early 2030s.
Google has backed TAE in its latest funding round.

The fusion companies are receiving their first data center power purchase agreements because the hyperscalers understand the arithmetic.
Data centers need power continuously.
The grid is constrained.
Nuclear restarts take years.
Fusion — if it delivers on the 2028-to-2033 window that multiple firms are now targeting — would be dispatchable, carbon-free, and geographically flexible.
The timeline mismatch is real: the capex is being committed today, and fusion power is five to ten years away in the most optimistic projections.
But the offtake agreements are being signed now precisely because no one wants to be last in line.

// ==========================================
// OUTRO
// ==========================================

[map.clear annotations]
[camera.establish_globe lat=20 lon=10 zoom=1.6]
[map.basemap kind=dark]
[scene.fade color="#020617" opacity=0.32 duration=500]
[scene.title kind=outro eyebrow="CLIO AGIS" title="Who controls the compute corridor controls the intelligence corridor." subtitle="Subscribe for the next briefing."]

[chat.say source="hyperscaler_capex_2026"]
By 2030, the five largest hyperscalers will have added approximately two trillion dollars in AI-related assets to their balance sheets.
Global data center capacity will reach two hundred and twenty gigawatts.
[source.show id="hyperscaler_capex_2026" text="Five hyperscalers — Amazon, Google, Meta, Microsoft, Oracle — targeting ~$2T in AI-related balance sheet assets by 2030. 2026 capex alone: $660–$690B." confidence=0.78]
The machines will consume more electricity than entire G7 nations, pull water from aquifers across three continents, and depend on a supply chain that runs through one island in the Taiwan Strait.
Clio maps the plumbing.
The power lines. The chip corridors. The water pipes. The offtake agreements.
Everything the headline skips.
Subscribe, and bring your questions.