The State of Data Centers 2026: A Complete Editorial Report on the Infrastructure Powering AI

Four companies will spend roughly $725 billion on capital expenditure in 2026, up 77% from a record $410 billion in 2025, and almost all of the increase is AI data centers.

Here's the problem. Roughly half of the US data center capacity scheduled to open this year is delayed or at risk of cancellation, and the reason isn't chips or capital. It's transformers, switchgear, and grid connections.

That inversion is the story of 2026. The binding constraint moved from silicon to electricity, and it's reshaping who builds, where they build, what the building looks like, and who pays.

The team at InsideDeepTech will break down the money, the power problem, the engineering redesign, the new players, the backlash, and the geopolitics, then tell you where it lands.

Key takeaways

• Capex has gone parabolic. Amazon, Microsoft, Alphabet, and Meta are guiding to about $725 billion combined in 2026, and Goldman Sachs now models $5.3 trillion across 2025 to 2030.
• Electricity is the bottleneck. The IEA projects global data center consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, and in the US, data centers are expected to account for about 50% of all electricity demand growth through 2030.
• The building is being redesigned around the rack. Nvidia's Vera Rubin racks draw roughly 190 to 230 kW in late 2026, the 2027 Kyber rack is specified at 600 kW, and both require 100% liquid cooling and 800-volt DC distribution.
• Ratepayers are now paying for it. PJM's capacity price went from $28.92 per MW-day for 2024/25 to the $333.44 cap for 2027/28, and the market monitor attributes 38% of the latest auction's $16.4 billion in charges to data centers.
• The public has turned. Gallup found 71% of Americans oppose an AI data center in their area, more than oppose a nearby nuclear plant.
• Two compute universes are forming. China is drafting a $295 billion national AI data center grid that mandates 80% domestic technology, locking Nvidia and AMD out.

The state of data centers in 2026 at a glance

The money: $725 billion and counting

Let's start with the number that anchors everything else.

Amazon, Microsoft, Alphabet, and Meta are collectively guiding to roughly $725 billion in 2026 capital expenditure, according to first-quarter earnings compiled by the Financial Times. That's a 77% jump over 2025's already record $410 billion.

Add Oracle's roughly $50 billion, a 136% increase over 2025, and the five largest US cloud and AI infrastructure providers are approaching $800 billion in a single year.

The guidance kept moving up through the year. Alphabet raised its ceiling to $205 billion at Q2 earnings, and Meta has raised guidance twice. Microsoft's CFO Amy Hood set calendar-year 2026 capex at $190 billion, attributing $25 billion of that to rising memory and component costs alone.

Where does the money actually go? Nvidia captures roughly 40% of hyperscaler capex, by Value Add VC's estimate, and memory alone will consume about 30% of hyperscaler data center spending this year, a fourfold increase since 2023.

The rest is the building, the power, and the cooling. And that's where the trouble starts, because you can order a GPU and receive it in months. You can't do that with a substation.

The skeptic's read

Investors aren't uniformly convinced. When Alphabet raised its 2026 capex forecast in July, shares fell 7% the next day, and Amazon, Meta, and Microsoft dropped alongside it.

The concern is simple: cash piles are dwindling and returns remain uncertain. Futurum notes that while OpenAI and Anthropic are growing revenue fast, their combined revenues remain a fraction of the infrastructure being built on their behalf.

The bulls point back at demand. Google Cloud revenue jumped 63% year over year to $20 billion in Q1 2026, and Microsoft says it expects to remain capacity-constrained through at least the end of 2026. Whichever side you take, the physical constraint below is what decides the timing.

The power problem is the whole problem

At first glance, data centers look like a small slice of the grid. The IEA puts global data center electricity consumption at 485 TWh in 2025, roughly 1.7% of global demand.

But it's not just the level. It's the slope, and where it lands.

The IEA's base case has that figure roughly doubling to 950 TWh by 2030, about 3% of global electricity. Consumption from AI-focused data centers triples in the same period, with electricity use in accelerated servers growing about 30% a year.

Here's why that matters. The growth is concentrated. The US and China together account for nearly 80% of projected global growth, and in the US, data centers are expected to make up about 50% of all electricity demand growth out to 2030.

Within PJM, the grid operator serving 65 million people across 13 states, the forecast is starker still. PJM attributes 94% of its projected 32 GW of peak load growth through 2030 to data centers, with that growth outpacing new generation by roughly two to one.

Where does the power come from today? Natural gas supplies 26% of the electricity consumed by data centers, nuclear 15%, per the IEA. The agency expects gas generation for data centers to grow by another 175 TWh, much of it in the US.

The grid can't keep up

Sightline Climate tracked roughly 16 GW of US data center capacity slated for 2026 delivery across about 140 projects. Only around 5 GW is actually under construction.

The delays trace back to a handful of unglamorous components. Electrical infrastructure is less than 10% of total data center cost, yet a delay in any single piece halts the whole project.

Before 2020, a high-power transformer took 24 to 30 months to deliver. Today the wait can stretch to five years, per Sightline Climate. An AI data center's deployment cycle is under 18 months.

That means the scarce input in 2026 isn't capital and isn't GPUs. It's an energized megawatt with a transformer attached, and whoever holds one is a price-maker.

Bring your own power

The rational response to a five-year interconnection queue is to skip it. That's exactly what's happening.

Developers ordered so many gas turbines for on-site generation that US orders surpassed 14 GW in 2024, the highest since 2001, and Mitsubishi Heavy reports production slots booked through 2028. Of Stargate's seven US sites, at least three will use on-site natural gas plants, per Epoch AI's analysis.

Fuel cells are the other route. Oracle's Project Jupiter in New Mexico pivoted to a fully islanded Bloom Energy microgrid, with no grid interconnect, no diesel, and no turbines, according to Bloom's investor call.

On the flip side, behind-the-meter gas has its own problems. It's more expensive per megawatt-hour than grid power, it locks in emissions for decades, and it draws exactly the kind of local air-quality opposition that xAI's Memphis turbines triggered a NAACP lawsuit over.

The nuclear bet

Nuclear is the buildout's long-dated hedge. As of May 2026, the four major US hyperscalers had signed 13 nuclear deals totaling 9.8 GW, per the SMR Intel tracker.

Two things stand out in that table. First, the only nuclear electrons actually flowing to an AI data center today come from Amazon's deal with an existing conventional reactor, not a new build.

Second, the most advanced new deal, Microsoft's Three Mile Island restart, cleared its last major grid obstacle when FERC approved a transmission waiver on June 1, 2026, pulling the timeline forward to H2 2027. Nearly everything else is a 2030s story.

Of course, nuclear does nothing for a campus that needs power in 2027. It's a hedge on the decade, not a fix for the year.

Inside the AI factory: how the building changed

For twenty years, a data center rack drew 5 to 10 kW and you cooled it with air. That design is now obsolete for AI, and the reason fits in one number: 600.

Nvidia's rack power trajectory tells the story. A Hopper-era rack drew roughly 40 kW. Blackwell GB200 racks pushed to 120 to 130 kW. The Vera Rubin VR200 NVL72, shipping in volume in the second half of 2026, draws roughly 190 to 230 kW.

Then comes Kyber. The Rubin Ultra NVL576 rack, due in the second half of 2027, is specified at 600 kW, which DCD notes is beyond the limits of the vast majority of data centers today. Nvidia's roadmap points toward megawatt-class racks after that.

Three engineering consequences follow, and each one rewrites the building.

Liquid is mandatory

Vera Rubin requires 100% direct-to-chip liquid cooling at 45°C inlet water temperature. No air-cooled configuration exists. Nvidia's Charlie Boyle told DCD the Kyber design is 100% liquid cooled with no fans at all.

The warm inlet temperature is deliberate. It lets facilities reject heat to the outside air without evaporating water, which is why liquid cooling and the water debate are the same conversation.

Power delivery goes to 800 volts DC

Rubin racks run on an 800-volt DC power architecture, replacing the 48V in-rack distribution standard for over a decade and the 415/480 VAC systems that feed it. ModulEdge calculates that moves over 150% more power through the same copper and eliminates roughly 200 kg of busbar per rack.

That's a facility-level change. You don't retrofit a 480 VAC hall to 800 VDC over a weekend.

Stranded assets become a real risk

Nvidia's cadence is now explicit: Rubin in 2026, Rubin Ultra in 2027, Feynman in 2028 with co-packaged optics. Jensen Huang opened GTC 2026 citing roughly $1 trillion in orders through 2027.

Here's the problem for anyone building. A new rack-scale platform with a materially different facility envelope arrives every twelve months, while a data center takes 18 to 36 months to design and build. Each step, as ModulEdge puts it, invalidates the previous step's electrical design.

Colocation operators with AI-ready halls are already capturing rent premiums over conventional space, per CBRE's H2 2025 report. The corollary is that halls built for 40 kW air-cooled racks in 2023 are, for frontier AI, already legacy.

What it costs to build now

Industry benchmarks put a standard data center at $8 million to $12 million per MW in 2026 and an AI-optimized build at $15 million to $20 million-plus per MW, against roughly $7 million per MW in 2020.

Electrical and power infrastructure eats 40% to 45% of that budget. The physical structure is a small slice.

A worked example: what a gigawatt campus actually takes

Abstractions hide scale. Stargate's flagship site in Abilene, Texas, makes it concrete.

Now multiply. Stargate alone is projected to exceed 9 GW across seven US sites by 2029. Meta's Louisiana campus is planned to scale to 5 GW. CoreWeave alone added nearly 500 MW of active power in a single quarter.

Each gigawatt is another million homes' worth of demand landing on a grid that added transmission at a fraction of its historical pace.

Who's building: hyperscalers, neoclouds, and the new landlords

The hyperscalers write the biggest checks, but they're no longer the only builders that matter. A second tier has emerged, financed largely with debt.

The neocloud model, stress-tested

CoreWeave posted $2.575 billion in Q2 2026 revenue, up from $1.212 billion a year earlier, while expanding active power to 1.5 GW across 51 data centers. Its contracted backlog stands at roughly $108 billion.

It also posted a $626 million net loss for the quarter, guides to $35 billion to $39 billion of 2026 capex against $12.4 billion to $13.2 billion of revenue, and carried $25.1 billion of debt at the end of Q1. Interest expense is the main drag on profit.

Nebius took a different route. Its Q2 revenue grew 454% to $582.3 million, and roughly 70% of its signed deals include customer prepayments covering 50% to 60% of capex, about $9 billion in 2026. That's the neocloud model with the financing risk pushed back onto the customer.

The whole cohort trades as one rates bet. When the 30-year Treasury hit 5.32% in August, CoreWeave fell 12% in a day and Nebius and TeraWulf fell with it, on a quarter that was operationally strong.

That means the neocloud story isn't about demand. Demand is contracted. It's about whether debt-financed capacity can be energized fast enough to convert backlog to revenue before financing costs eat the margin.

The colocation market has never been tighter

For everyone who isn't a hyperscaler, capacity comes from colocation, and there's almost none available. CBRE reports primary-market vacancy in North America fell to a record low of 1.4% at year-end 2025 even as supply grew 36% to 9,432 MW.

Northern Virginia, the world's largest market, hit 0.3% vacancy in Q1 2026. Asking rents there ran $190 to $235 per kW per month, with Chicago highest at $200 to $230.

Net absorption across primary markets set a record of 2,497.6 MW in 2025, and Northern Virginia alone absorbed 1,102 MW. Most 2026 capacity is already committed, pushing preleasing into 2027 and beyond.

Globally, inventory grew fastest in Latin America (41.3%) and North America (33%), while Singapore commands roughly $403 per kW per month, the highest in APAC.

If you're an enterprise trying to lease 15 to 20 MW of contiguous space in a Tier 1 market for 2027 delivery, you're already late.

The backlash: ratepayers, water, and the politics of siting

In 2024, data center opposition was a zoning story. In 2026, it's a national political issue with poll numbers.

The electricity bill argument

The most potent grievance is the power bill, and it has a paper trail. PJM's capacity auction, which sets the price utilities pay to guarantee supply three years ahead, cleared at $28.92 per MW-day for 2024/25. It hit $269.92 for 2025/26, $329.17 for 2026/27, and the $333.44 regulatory cap for 2027/28.

PJM's independent market monitor, Monitoring Analytics, attributed 63% of the 2025/26 increase, or $9.3 billion, to data centers. In the most recent auction, data centers accounted for $6.3 billion of $16.4 billion in charges, and the 2027/28 auction cleared short of PJM's reliability requirement for the first time in the market's history.

The fair counterpoint: PJM lost more than 9 GW of generating capacity between 2023 and 2025, and consultancy E3 estimates data-center-driven load growth explains only about half the price increase, with retirements and market design explaining the rest. Data centers exposed a weak grid as much as they broke it.

Either way, households feel it. NRDC estimates the average PJM household faces about a $70 monthly increase by 2028. That's the number behind every moratorium floor speech.

The public has decided

Gallup's first-ever poll on the subject, published May 13, 2026, found 71% of Americans oppose an AI data center in their local area, 48% strongly. Only 7% strongly favor one.

That's more opposition than a nearby nuclear plant draws (53%). The concerns are concrete rather than abstract: half of opponents cite resource use, split evenly between water and electricity, and about a fifth cite utility bills and cost of living.

The opposition has no demographic signature. Gallup found no meaningful differences by age, race, education, income, or urbanicity, with the Midwest (76%) and South (75%) most opposed.

From protest to policy

Opposition is now a line item in project attrition. ITIF counted more than $130 billion in delayed or abandoned projects in the first quarter of 2026 alone.

The policy response runs in three directions:

• Make data centers pay their own way. Utilities established at least 38 specialized large-load tariffs between 2018 and early 2026, 30 of them in 2025 and 2026, using take-or-pay provisions, minimum bills, and exit fees. Oregon created the first dedicated data center rate class.
• Curtail or exclude them. PJM's board proposed that new large loads that don't bring their own supply be curtailed near emergency conditions, and FERC's chairman warned it would impose reforms if PJM failed to act by September.
• Pause them. At least six states introduced construction moratoriums and seven moved to repeal or restrict tax incentives. In Congress, Rep. Ro Khanna's Data Center Bill of Rights, introduced August 6, 2026, would let local governments block construction and shield those decisions from state override.

Morningstar DBRS now calls stakeholder opposition a potential material credit factor for data center project debt. When the ratings agencies price the protest, it's no longer just a protest.

The water fight

Water is the second flashpoint, and the industry's answer is architectural. Microsoft CEO Satya Nadella told Build 2026 that the company's new closed-loop designs are filled once and then use roughly as much water annually as a single restaurant, against tens to hundreds of millions of gallons a year for legacy evaporative hyperscale sites.

The trade-off is real but small: dry and closed-loop systems cut evaporative losses to near zero but raise electricity use by roughly 1% to 1.5%. Given the power constraint above, even that isn't free.

Disclosure is the weak point. Google's 2025 water consumption reportedly reached 10.9 billion gallons, up 34%, while replenishing about 78% of the freshwater it consumed against a 120% target for 2030. Texas found 83% of its 341 data centers had not complied with mandatory water reporting. Treat the aggregate estimates as reported, not audited.

The geopolitics: two compute universes

Data centers used to be real estate. In 2026 they're industrial policy.

China is drafting a five-year, 2 trillion yuan (about $295 billion) plan to link a nationwide web of AI data centers into a single computing grid by 2028, operated by China Mobile and China Telecom and financed by ultra-long sovereign bonds. Folding in grid upgrades could push the total past 5 trillion yuan.

The critical clause: at least 80% of the underlying technology, AI chips included, must come from domestic suppliers such as Huawei. That builds on 2025 rules requiring 50% local chips and then barring foreign accelerators from state-funded projects entirely.

Huawei shipped roughly 812,000 Ascend chips in 2025 and projects about $12 billion in processor revenue for 2026. But domestic high-bandwidth memory supply is the binding constraint, and analysts estimate Chinese suppliers will cover only about 76% of domestic AI chip demand by 2030.

For Nvidia, China was $19.7 billion of revenue in fiscal 2026, about 9% of the total and shrinking. The plan formalizes what export controls started: two separate AI infrastructure stacks, US-aligned and China-aligned, each with its own silicon, its own software, and its own power problems.

The US, meanwhile, is running the world's largest privately financed infrastructure program on public grids that weren't built for it. ConstructConnect tracked $46.5 billion in US data center construction starts through March 2026, versus $7.3 billion in the same period of 2025, putting the year on pace for roughly $121 billion.

Both models have a weakness. China's is chips. America's is the wire.

Where it breaks: the risks nobody should skip

A report with no downside is marketing. These are the failure modes the team at InsideDeepTech considers live, not hypothetical.

• Energization risk. Sightline Climate expects 30% to 50% of 2026 US openings to slip or die. Analysts project $150 billion to $200 billion of capex sliding from 2026 into 2027 and 2028, which compresses the construction surge and recreates the same bottleneck later.
• Revenue-to-capex gap. Roughly $725 billion of 2026 spending sits against AI-native revenues that Futurum describes as a fraction of the infrastructure investment. The hyperscalers can absorb a miss. Debt-financed neoclouds can't absorb many.
• Rates. The neocloud cohort repriced as a block when long yields rose in August. A sustained 200 basis point move in long rates changes the math on every leveraged buildout.
• Stranded facilities. Halls designed for 40 kW air-cooled racks can't take Rubin, and halls designed for Rubin may not take Kyber's 600 kW and 800 VDC without rework. Annual platform cadence versus multi-year build cycles is a structural mismatch.
• Political and regulatory reversal. With 71% local opposition, curtailment proposals at PJM, a federal bill to empower local vetoes, and ratings agencies flagging opposition as a credit factor, the permitting assumptions baked into 2024-era pro formas no longer hold.
• Behind-the-meter lock-in. Gas turbines ordered today run for 30 years. If grid capacity catches up or carbon rules tighten, on-site gas becomes a liability, not an asset.
• Component and workforce shortages. Transformers and switchgear are the headline, but the qualified electrical engineers and skilled trades to install them are at least as constrained. Solving the queue without solving the labor pipeline just relocates the bottleneck.

InsideDeepTech's take

Here's where the team at InsideDeepTech lands.

The 2026 buildout is not a bubble in demand. It's a bubble in schedules. Contracted backlogs at CoreWeave, Nebius, and Oracle are real money under real contracts. What's fictional is the delivery timeline that most announcements assumed, because nobody ordered the transformers in 2022.

That distinction matters for how you should read the next two years. Expect announced gigawatts to keep outrunning energized gigawatts by a wide margin through 2027, and expect the gap to show up as slipped revenue at leveraged builders rather than as cancelled demand at the hyperscalers.

The second position is less comfortable for the industry. The ratepayer argument is correct and the industry lost it. Data centers didn't build PJM's fragility, but they exposed it, and 71% of the public has already decided who's responsible. Large-load tariffs that make data centers carry their own capacity costs are the right fix and the industry should stop fighting them, because the alternative is moratoriums.

Third: the building is now a consumable. When the rack envelope changes every twelve months and the building takes two years, the facility becomes the thing that depreciates, not just the chips. Operators who design for 800 VDC and warm-water liquid cooling in 2026, even at a premium, will own the only halls that can take 2028 silicon. Everyone else will be leasing from them.

Where this doesn't apply: if you're running enterprise workloads at 10 kW per rack, none of the Rubin-era redesign is your problem, and the record colocation rents are simply a tax you're paying for sitting next to the AI boom. Move to a secondary market.

What to watch through 2027

The next twelve months have a short list of tells.

Q4 2026: whether Stargate Abilene actually reaches 1.2 GW, and whether Vera Rubin VR200 racks ship in volume on schedule. Both are tests of whether the physical layer can hit the roadmap.

September 2026 onward: what FERC forces on PJM if the grid operator misses its reform deadline, and whether the curtailment-for-large-loads model spreads to ERCOT, MISO, and the Southeast.

Through 2027: the first sustained wave of Q3 and Q4 2026 delays turning into restated capex guidance, and whether the Big Four's $1 trillion-plus 2027 projections survive the delivery gap.

H2 2027: Three Mile Island Unit 1 sends its first nuclear electrons to Microsoft, and Kyber's 600 kW racks arrive. One shows the power side catching up. The other shows the demand side pulling away again.

If you're building, ordering, leasing, or regulating, the practical instruction is the same one the market spent 2026 learning the hard way. Secure the power first, design for the rack that doesn't exist yet, and assume the public is watching your water meter.

Frequently asked questions

How much are companies spending on AI data centers in 2026?

Amazon, Microsoft, Alphabet, and Meta are guiding to roughly $725 billion in combined 2026 capital expenditure, up about 77% from $410 billion in 2025, according to company guidance compiled by the Financial Times. Most of the increase funds AI data centers, GPUs, custom silicon, and power. Adding Oracle's roughly $50 billion brings the top five US providers close to $800 billion. Analysts expect combined Big Four capex to exceed $1 trillion in 2027.

How much electricity do data centers use, and how fast is it growing?

Data centers consumed about 485 TWh of electricity globally in 2025, roughly 1.7% of world demand, per the International Energy Agency. The IEA's base case projects that roughly doubling to 950 TWh by 2030, about 3% of global electricity, with AI-focused data center consumption tripling over the period. In the United States, data centers are expected to account for about half of all electricity demand growth through 2030.

Why are so many US data centers delayed in 2026?

Roughly 30% to 50% of US data center capacity scheduled to open in 2026 is delayed or at risk of cancellation, according to Sightline Climate data reported by Bloomberg. The bottleneck is electrical infrastructure, not chips or capital. High-power transformer lead times have stretched from 24 to 30 months before 2020 to as long as five years, switchgear is effectively sold out through 2028, and grid interconnection in markets like Northern Virginia, Phoenix, and Dallas takes four to seven years.

What is Stargate and how far along is it?

Stargate is the $500 billion AI infrastructure program led by OpenAI, Oracle, and SoftBank, targeting 10 GW of US capacity. Its flagship campus in Abilene, Texas, is designed for 1.2 GW and more than 450,000 Nvidia GB200 GPUs across eight buildings. Epoch AI estimated about 0.3 GW was operational in April 2026, with the full 1.2 GW expected in Q4 2026, and projects the seven US sites will exceed 9 GW by 2029. At least three sites plan on-site natural gas generation to bypass grid queues.

How much power does an AI rack use in 2026?

Nvidia's Vera Rubin VR200 NVL72 rack, shipping in volume in the second half of 2026, draws roughly 190 to 230 kW, up from about 120 to 130 kW for Blackwell GB200 and about 40 kW for Hopper-era racks. The Rubin Ultra Kyber rack due in the second half of 2027 is specified at 600 kW. These racks require 100% direct-to-chip liquid cooling and 800-volt DC power distribution, which most existing data centers cannot support.

Are data centers raising electricity bills?

In the PJM grid region serving 65 million people, yes. PJM's capacity auction price rose from $28.92 per MW-day for 2024/25 to the $333.44 cap for 2027/28, and the independent market monitor attributed 63% of the 2025/26 increase to data centers. In the latest auction, data centers accounted for $6.3 billion of $16.4 billion in charges. Analysts note plant retirements and market design explain roughly half the increase, but NRDC estimates the average PJM household faces about a $70 monthly increase by 2028.

Do Americans support building data centers?

No. A Gallup poll published May 13, 2026, found 71% of Americans oppose an AI data center in their local area, including 48% strongly opposed, with only 7% strongly in favor. That is more opposition than a nearby nuclear plant draws (53%). Opponents most often cite water and electricity use, utility bills, and quality of life. The opposition shows no meaningful difference by age, race, education, income, or urban versus rural residence.

Which nuclear plants will power AI data centers?

As of May 2026, US hyperscalers had signed 13 nuclear deals totaling about 9.8 GW, per SMR Intel. Amazon already receives power from Talen's Susquehanna plant under a 1,920 MW agreement. Microsoft's 835 MW deal to restart Three Mile Island Unit 1 is expected to deliver power in the second half of 2027. Meta has committed up to 6.6 GW across Constellation, Vistra, TerraPower, and Oklo, and Google has agreements with Kairos Power and Elementl. Most small modular reactor capacity is not expected before 2030.

What is China's $295 billion data center plan?

China is drafting a five-year, 2 trillion yuan (about $295 billion) plan to build a nationwide AI data center network linked into a single computing grid by 2028, operated by China Mobile and China Telecom and funded by sovereign bonds, according to Bloomberg reporting. At least 80% of the technology, including AI chips, must come from domestic suppliers such as Huawei, effectively excluding Nvidia and AMD. Including grid upgrades, the total could exceed 5 trillion yuan.

How much does it cost to build a data center in 2026?

Industry benchmarks put a standard data center at $8 million to $12 million per megawatt and an AI-optimized facility at $15 million to $20 million or more per megawatt, up from roughly $7 million per megawatt in 2020. Electrical and power infrastructure accounts for 40% to 45% of construction cost. A 1 GW AI campus therefore implies $15 billion to $20 billion for the facility alone before GPUs, which can cost $3 million to $4 million per rack for Rubin-class systems.