Inside the AI Factory: How Electricity Becomes Intelligence
Nvidia calls them AI factories: buildings whose raw material is electricity and whose product is tokens. In 2026 the four biggest cloud companies plan to spend around $700 billion building them. This guide walks through the production line, the gigawatt campus, the 600-kilowatt rack, the power fight, the money and the backlash, and separates what is measured from what is marketing.
From data center to AI factory
A traditional data center stores and serves information: email, web pages, databases. Its servers sit mostly idle and its value is measured in uptime. An AI factory, in Nvidia's framing, does something different: it converts electricity into tokens, the units of text, code, images and actions that AI models produce. Jensen Huang's line at GTC 2026 was that data centers "are truly becoming AI factories", and his formula for their business was blunt: revenue = tokens per watt ร available gigawatts. At Nvidia's 2026 shareholder meeting he called every token "a profit unit".
The metaphor is useful because it changes what people measure. Instead of servers and square metres, operators talk about tokens per second, tokens per watt, cost per million tokens and how much of the time the chips are actually busy. It also explains why power has become the scarce input: if output scales with watts, the company with the most energy wins.
The token production line
Follow one kilowatt-hour through the factory. Click a stage to see what happens there and what it is measured by.
From electrons to tokens
Simplified. The moving dashes show energy and data flowing; the bottom row is where most energy leaves as heat.
Anatomy of a gigawatt campus
A gigawatt is roughly the output of a large nuclear reactor. A campus built around it looks less like an office building and more like a power plant with computers attached. Click the parts of this schematic site.
A schematic AI campus
Not to scale and not any specific site. Based on the subsystems named in Nvidia's Vera Rubin DSX reference design and partner announcements.
Nvidia's March 2026 Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint try to standardise this whole stack, with contributions from Schneider Electric, Siemens, Vertiv, Eaton, Trane, Jacobs and others. The blueprint is for "physically accurate digital twins": simulating the power, cooling and layout of a site before pouring concrete. Nvidia's stated reason is that energy is now the main constraint, citing more than $300 billion of equipment backlog and over 200 gigawatts of projects waiting in US grid interconnection queues.
The rack revolution
The single biggest physical change is how much power one rack draws. Air cooling runs out at a few tens of kilowatts per rack; everything above that needs liquid flowing directly to the chips.
Power per rack by Nvidia generation
Approximate figures from trade press and vendor-adjacent analyses. Vera Rubin is an estimate; Kyber is a target for 2027, not a shipping product.
- Liquid becomes mandatory. GB200 and GB300 racks use direct-to-chip liquid cooling. DCD reports that the Kyber design is "100 percent liquid cooling, no fans", with a sidecar rack for power and cooling.
- Buildings age fast. One analysis warns that a facility built in 2024 around 120 kW racks cannot host 600 kW Kyber racks without rebuilding power distribution, the cooling plant and possibly floor strength.
- Power distribution changes. To move that much current, the industry is shifting to 800-volt DC inside the building: Hitachi says its 800 VDC architecture is integrated into the Vera Rubin DSX design, and Flex has built an 800 VDC power rack with Nvidia.
The four networks inside the factory
A factory is only as fast as its slowest link. Networking is now described in layers, each with its own technology. Click a layer.
Scale-up, scale-out, scale-across, scale-in
"Scale-in" is a 2026 Nvidia addition that analysts describe as mostly vendor framing.
Power is the product
If tokens scale with watts, electricity is the true bottleneck. The International Energy Agency's 2025 Energy and AI report put data centers at about 415 TWh in 2024, roughly 1.5% of world electricity, heading to about 945 TWh by 2030, with AI-optimised data centers more than quadrupling. Its 2026 follow-up, as reported, says data center demand grew 17% in 2025 against 3% for electricity overall, and that AI-specific demand will triple by 2030.
Data center electricity use, IEA
TWh per year. 2024 baseline and 2030 projection from the IEA's 2025 report; the 2025 value applies the reported 17% growth to the baseline and is our approximation.
Where the power comes from
Click an option.
The world's biggest AI factories
No single site runs a full gigawatt of IT power yet, according to Epoch AI's tracker, but several are close and the plans are far larger. Solid bars show estimated operating power; the lighter bar shows the planned figure.
Operating vs planned power (MW)
Epoch AI estimates as of October 2026 unless noted. Hyperion's 5 GW is a long-range goal; Epoch's own projection is 1,632 MW for 2028.
- xAI Colossus 2 (Memphis area): about 946 MW operating, 1,531 MW planned for 2027. A May 2026 report says xAI deployed 19 natural-gas turbines at the Southaven, Mississippi site.
- Anthropic and Amazon, New Carlisle (Indiana): about 910 MW operating.
- Microsoft Fairwater: Atlanta about 636 MW operating; Wisconsin is the largest planned site Epoch tracks at 2,263 MW by 2028.
- Stargate Abilene (OpenAI, Oracle, Crusoe): Oracle said six of eight buildings, 618 MW, had been handed over; the campus target is 1.2 GW, and Crusoe describes a path to 2.1 GW on the site. OpenAI says it has secured more than 10 GW of US capacity.
- Meta Hyperion (Louisiana): Meta said in July 2026 it is expanding the site toward 5 GW with more than $50 billion of investment.
Sovereign AI factories
Governments now treat AI factories as strategic infrastructure, like ports or power stations.
National and regional projects
| Where | What | Status |
|---|---|---|
| European Union | AI Gigafactories call for up to seven sites, each with at least 100,000 advanced processors (about four times Europe's largest today), backed by up to โฌ10 billion of EU and national money and expected to unlock at least โฌ20 billion of private investment. Letters of intent signed with AMD, Nvidia and Qualcomm. | Call opened late July 2026; closes 12 Nov 2026; awards expected early 2027 |
| Germany | Deutsche Telekom and Nvidia's Industrial AI Cloud in Munich, starting with 10,000 Blackwell GPUs; described as Europe's first industrial AI cloud. | Announced; exact launch date not confirmed in our searches |
| India | Reliance's Jamnagar campus with Nvidia Blackwell chips: 1 GW in recent reports, earlier reports said 3 GW and $20โ30 billion; Reliance has said it will finish within 24 months. | Under construction; scale figures vary by source |
| Portugal | Microsoft plans a $10 billion data center investment, cited as part of private gigafactory plans in Europe. | Announced |
One analysis expects Europe's first gigafactories to run mainly on US chips from Nvidia and AMD, which tempers the word "sovereign": the buildings and data are local, the silicon mostly is not.
The money
The spending is the largest private infrastructure build in modern history. Press reports of company guidance put 2026 capital expenditure for the four biggest US cloud companies at roughly $700 to $725 billion, up about 77% from around $410 billion in 2025.
2026 capex guidance by company (US$ billions, midpoints)
Microsoft's figure is disputed between sources (calendar vs fiscal year); Microsoft says about $25B of its total reflects component price inflation, not extra capacity. Amazon's includes its Leo satellite business.
- Cost per gigawatt: one widely shared estimate is about $38 billion per GW, with silicon about 57% of it. We found no authoritative single figure; treat it as an order of magnitude.
- Through 2030: McKinsey's base case, as summarised by one source, adds about 125 GW of AI capacity and $5.2 trillion of AI capex, around 60% of it chips and computing hardware.
- Efficiency gains: Nvidia says GB300 NVL72 delivers up to 50 times more throughput per megawatt and 35 times lower cost per token than Hopper, citing SemiAnalysis; one analyst estimates tokens per kilowatt-hour improved about 13 times from A100 to GB300. Better chips mean the same building produces far more tokens, which is also why old buildings and old chips lose value quickly.
Factory calculator: what must a token cost?
Work out the economics of a site yourself. The calculator asks what revenue each megawatt-hour of busy computing must earn to pay back the building and chips over their life, then turns that into a break-even price per million tokens using an efficiency figure you choose.
Break-even price of a token
Simplified: straight-line payback of capex, plus electricity and an operations allowance; no financing cost, taxes or profit. The tokens-per-kWh slider is an illustrative assumption: real efficiency varies by orders of magnitude with model size, hardware and batch settings.
Two lessons fall out of playing with it. First, capex dominates: electricity is usually a small share of the cost per token, which is why operators pay almost anything to get power sooner. Second, payback period matters enormously: moving from six years to three roughly doubles the price each token must fetch, which is exactly what the depreciation debate below is about.
Training factories vs inference factories
Deloitte, as widely cited, estimates that inference (running models for users) is about two-thirds of all AI compute in 2026, up from a third in 2023 and half in 2025. Training still grows at the frontier, but the mix is shifting, and the two jobs want different factories.
| Training factory | Inference factory | |
|---|---|---|
| Shape | One enormous, tightly coupled cluster | Many smaller pools serving independent requests |
| Location | Wherever power is cheapest and largest; latency to users irrelevant | Closer to users and networks; latency matters |
| Network | Huge scale-out and scale-across bandwidth; one failure can stall the job | Scale-up within a rack matters most; jobs fail independently |
| Power profile | Synchronised swings as thousands of chips start and stop together, hard on the grid | Follows user demand through the day |
| Business | A cost centre whose payoff is a better model | Sells tokens directly: the "factory" metaphor fits best here |
The bear case: depreciation, debt and circular deals
The build-out has serious critics, and their arguments are mostly about money rather than technology.
Bull vs bear, point by point
| Question | ๐ Bull view | ๐ป Bear view |
|---|---|---|
| How long do GPUs last? | Hyperscalers depreciate servers over about 5 to 6 years; older chips keep earning on inference. | Michael Burry (Nov 2025) argues the real economic life is 2 to 3 years and estimates about $176 billion of understated depreciation in 2026โ28 (his estimate). Goldman calculates that a 3-year life would raise 2026โ31 depreciation from about $3 trillion to nearly $4 trillion. |
| Does accounting matter? | Depreciation is non-cash; free cash flow is the same whatever schedule is used. | Reported profits, and the valuations built on them, look better than the cash economics. |
| How is it financed? | The biggest builders are among the most profitable companies in history. | Morgan Stanley expects about $3 trillion of spending through 2028 with internal cash covering roughly half; analysts cite about $121 billion of hyperscaler debt in a year, special-purpose vehicles that keep debt off balance sheets, and Oracle's roughly $248 billion of lease commitments not yet on its balance sheet. |
| Is demand real? | Inference demand and token volumes are growing fast; capacity sells out. | "Circular deals", in which chipmakers, clouds and labs invest in each other, can make demand look stronger than end-customer spending; Moody's has flagged such arrangements. |
| What if growth slows? | Spare capacity gets absorbed as models and users grow. | Analyst Gil Luria warns that if demand merely levels off, the industry could be left with overbuilt capacity and stranded debt. |
The neighbours: bills, water and moratoriums
The most immediate limit on AI factories may be local politics. Opposition in 2026 has spread across many US states and across party lines, driven by electricity bills, water use and secrecy around deals.
PJM capacity price, US$ per MW-day
PJM is the largest US grid market (13 states and DC). Capacity prices are paid by all customers; data center demand is the main reason cited for the jump.
- Bills: PJM's independent market monitor estimates that data centers have cost its 67 million customers about $29 billion over roughly two years. One estimate puts the added cost for Maryland households at about $168 to $216 a year.
- Opinion: a Reuters/Ipsos poll found only 14% would support a data center being built in their own community.
- Restrictions: cities in Texas, North Carolina, Kentucky and New Jersey have pursued bans or moratoriums; Austin is drafting rules; Maine's legislature passed a statewide moratorium that the governor vetoed; Pennsylvania now requires annual water and energy reporting from large data centers.
- Responses: Oregon created the first dedicated data center electricity rate class; major AI companies backed a "Ratepayer Protection Pledge", which Brookings argues needs state enforcement.
- Water: one study found power plants serving data centers in seven states withdraw about 3.4 trillion gallons of freshwater a year; closed-loop liquid cooling reduces on-site water use, but power generation still draws water.
Enterprise AI factories
The term has also become a product category for companies that want AI on their own premises. Dell AI Factory 2.0 (May 2026) is built on Nvidia Blackwell Ultra, with up to 256 GPUs per rack in Dell's IR7000; a secondary report says Dell named Eli Lilly, Honeywell, Samsung and Ascension as reference customers. HPE sells Private Cloud AI, co-developed with Nvidia, and put networking at the centre of its 2026 pitch. Cisco builds servers around Nvidia's Vera Rubin platform. These are factories in the small: a few racks rather than a gigawatt, bought for data control, compliance and predictable costs rather than scale.
What to watch to 2030
โก Power
Grid interconnection queues, gas-turbine delivery dates, the first small modular reactors at data centers (mostly 2030s), and whether 800 V DC becomes standard.๐ฐ Money
Whether 2027 capex guidance keeps rising, depreciation-schedule changes in annual reports, credit spreads on data center debt, and token prices per million.๐๏ธ Politics
FERC and state rules on who pays for grid upgrades, local moratoriums, the EU gigafactory awards in early 2027, and export controls on chips.Reading the numbers: where sources disagree
| Claim | What the sources say |
|---|---|
| "50ร more throughput per megawatt" | Nvidia's figure for GB300 NVL72 vs Hopper, citing SemiAnalysis; workload-dependent and vendor-selected. |
| 2026 hyperscaler capex | About $700โ725B in most roundups; one June report says about $520B because it uses Microsoft's fiscal year. Microsoft's calendar-year figure ranges from $110โ120B to about $190B across sources. |
| Cost per gigawatt | About $38B per GW from one analysis; we found no authoritative figure. "$50B per GW" circulates but was not supported. |
| Stargate Abilene "1.2 GW" | A target. Oracle reported 618 MW handed over in September; Epoch had expected completion in Q4 2026. |
| Meta Hyperion "5 GW" | Meta's stated expansion goal; Epoch projects 1,632 MW by 2028. |
| Reliance Jamnagar size | 1 GW in recent reports; 3 GW and $20โ30B in earlier ones. |
| Kyber 600 kW per rack | A 2027 target described consistently in trade press; "1 MW later" is a roadmap aim. |
| Inference = two-thirds of compute | One Deloitte estimate repeated by many sites; lifetime-spend figures of 80โ90% measure something different. |
| Burry's $176B | His own estimate of understated depreciation, not an accounting finding. |
| IEA 2026 update figures | Seen only in press coverage, not the IEA report itself. |
๐ References
- Nvidia and partners
- Nvidia: AI Factories, the new infrastructure of intelligence
- Data Center Frontier: Jensen Huang maps the AI factory era at GTC 2026
- Nvidia DSX platform ยท coverage of the Vera Rubin DSX design
- Hitachi: 800 VDC power simulation with Omniverse DSX ยท Flex 800 VDC power rack
- DCD: Nvidia on the future of rack density
- SDxCentral: what "scale-in" means ยท Nokia on scale-across networking
- Sites, power and policy
- Epoch AI: largest AI data centers by power ยท Stargate: where the US sites stand
- IEA projection via APPA ยท IET on the IEA 2026 update
- NAM: behind-the-meter AI projects surge ยท Jones Walker: turbines at the fence line
- EU AI Gigafactories call ยท Polytechnique Insights on the gigafactories
- TechCrunch: Reliance's Jamnagar plan
- Money and the bear case
- Hyperscalers hit $700 billion in 2026 spending plans
- Creative Strategies: Gigawattonomics
- How fast does an AI chip depreciate? ยท The $176 billion accounting question
- NPR: why AI bubble concerns are bigger than ever ยท The Register: the $3T infrastructure binge
- Communities and grid prices
- IEEFA: PJM capacity prices up by a factor of 10
- NPR: how much are ratepayers on the hook for?
- Brookings: the ratepayer pledge needs enforcement
- Data Center Frontier: community opposition as gatekeeper
- On this site
- Data Centers in Space
- AI chips and compute infrastructure
- The compute frontier, explained