August 20, 202619 nodes#AIInfrastructure#PowerGrid#MemoryPrices#LaborMarket#CostPassThrough

Who Gets the Bill

Money spent on the AI buildout does not settle inside the data center. This map lays out the three places the cost actually lands: the power contract that recovers what someone paid to secure electricity, the memory that reached the data center first, and the single word written into a layoff announcement.

The brief, in full

The money that builds an AI data center is not settled inside that building. Whoever secures the electricity first passes the price to a tenant, the memory that reached the data center first shows up on a console price tag, and a shrunken payroll budget gets summarized as one word in an announcement. This map lays the three paths side by side and adds what to check at each one.

Capital moved from equity into power plants#

Fewer deals, each one larger

Samjong KPMG counts $1 trillion of global private equity investment in the first half of this year across 9,294 deals. The amount held while the deal count fell to its lowest in more than five years, which means each deal grew. Private equity attaches to real assets with a visible path to recovery, and the asset it picked this time is the data center and the electricity behind it.

open_in_newstartupxo.com/ko/news/2026/08/pe-ai-infra-power-secured-sites

$354.7 billion and $149.2 billion#

Tech, media and telecom, then energy

By sector, $354.7 billion went to technology, media and telecom, $154.0 billion to industrial manufacturing, and $149.2 billion to energy and natural resources. Software equity and generation assets sitting in the same table is the shape of this half year. Over the same period Korean domestic private equity shrank to $5.6 billion across 69 deals.

The shopping list Helix wrote down#

Generation and transmission on the line after server buildings

The largest deal of the half was Helix Digital Infrastructure, launched by KKR with more than $10 billion of long-duration capital. Its investment scope names hyperscale data center development and operations, then baseload and flexible generation, transmission and distribution, and fiber. Anchor investors include the Kuwait Investment Authority, NVIDIA, and Vistra, which runs a 50 GW power portfolio.

Whoever secures it recovers from whoever uses it#

The long-term contract is the channel

A company that secures electricity first has one way to get that money back. It collects from the people who use that electricity, spread across a long-term contract. Once a line item that never appeared in an AI cost model becomes a contract clause, the escalation term decides your spending two years out more than the headline rate does.

522 in the queue, ten approved#

In Korea the price already shows up in rent

CBRE Korea counts 522 applications totaling 33,592 MW for the first-stage technical review of grid impact assessment in the Seoul metropolitan area, cumulative through March 2026. Of those, 279 were found unservable and only 24 reached full review, ending in 10 final approvals for 1,010 MW. That is 1.9 percent of applications and 3 percent of requested capacity.

open_in_newreputo.net/ko/jobs/data-center-engineer/specializations/grid-interconnection

What a site with power already attached costs#

140,000 won per kW became 250,000 won

When attaching new power gets hard, the sites that already have it get expensive. Data center rent in the Seoul metropolitan area rose from about 140,000 won per kW in 2019 to about 250,000 won in 2025, up 70 percent over six years. For a team drawing 100 kW, a monthly bill of 14 million won became 25 million won. Vacancy sits below 5 percent, and global cloud providers and large domestic tech platforms take 88 percent of demand.

10 MW is the review line#

Five months that start after you file

Article 23 of Korea's Distributed Energy Act requires a grid impact assessment from anyone newly drawing 10 MW or more. Whether you are exempt is answered within 60 days of filing, and if you are not, the assessment itself typically runs about five months. The time spent producing that filing sits outside those five months.

Memory arrives as a console price tag#

Added output goes to data centers first

When Samsung and SK Hynix route added output to high bandwidth memory for data centers first, the commodity memory that goes into consoles and PCs waits behind it. Consoles are sold at a loss and recovered through games and subscriptions, so as memory and storage take a larger share of the bill of materials, the room to absorb that loss shrinks. This is the corridor through which the AI buildout crosses into the price of a game machine.

open_in_newstartupxo.com/ko/news/2026/08/next-gen-console-price-install-base

$649.99 to $799.99#

The multiple the company wrote into its own notice

From August 1 the Xbox Series X went from $649.99 to $799.99, and the Series S from $399.99 to $499.99. Microsoft wrote in the notice that console storage and memory prices have risen more than 2.5 times and that it expects another doubling by the fall of 2027. That is not an analyst estimate but a sentence published under the name of the company buying the parts.

39 million units is what $300 costs#

An install base 38 percent smaller

Ampere Analysis projects that if next-generation consoles launch at $1,000 rather than $700, roughly 39 million fewer units sell over the first five years and the install base ends up 38 percent smaller. Player spending across the generation's first three years falls by $3.4 billion, or 12 percent. For a project that breaks even at a million copies, the same million becomes a much harder target.

The opposing call is $600 to $800#

Built on the parts crisis clearing

Kagan, part of S&P Global Market Intelligence, puts next-generation pricing between $600 and $800. It holds that view while forecasting 33.9 million console shipments this year, down 19.5 percent from last year. The call rests on component pressure easing enough by 2028, and Microsoft has already published a sentence that contradicts that premise.

One word in the announcement#

The reason is not a finding but the explanation a company chose

Challenger, Gray & Christmas compiles the job cuts announced by US employers each month. The tally does not investigate what caused the cuts; it counts the reason each company wrote into its own announcement. So the sentence that AI leads the list measures the explanation employers picked before it measures anything about causation.

open_in_newstartupxo.com/ko/news/2026/08/ai-layoff-attribution-vs-hiring-signal

33,429, the lowest in two years#

The January to July total is 41 percent below last year

July announcements in the US came to 33,429 cuts, down 27 percent from June's 45,849 and 46 percent from 62,075 a year earlier. The January through July total of 477,033 sits 41 percent below the 806,383 recorded over the same period last year. It runs opposite to how the weekly layoff headlines feel.

The numerator fell to a quarter, the share held#

38,579 in May became 10,970 in July

Cuts attributed to AI fell from 38,579 in May to 10,970 in July, a quarter of the level in two months. The share barely moved, at 31 percent in June and 33 percent in July. May's total was 97,006 and July's was 33,429, so the denominator shrank faster and the quotient held. Read the share alone and you get the opposite story.

The other half of the same report#

16,095 hiring plans in July

The same release carries 16,095 announced hiring plans for July. That is 47 percent above June's 10,933, and the year-to-date total of 107,500 is 25 percent above last year. If the roles cut and the roles opened are the same function it is a reassignment, and if they differ the business changed direction, and one word in a reason field separates neither.

open_in_newreputo.net/ko/jobs/data-scientist/specializations/workforce-planning-analytics

Where to read the numbers#

Clauses over rates, counts over shares

Each of the three paths hides the same kind of trap. In power, applied and approved get used as if they were one word. In hardware, the sticker price covers the cost base. In labor, the share covers the total. What follows is one thing worth opening at each stop.

The build price and the monthly meter#

The meter follows the quote

The cost of adding an AI feature has the same shape. The build price goes on the quote, and the meter that spins with every request follows behind it. A colocation contract and a chatbot bolted onto a page ask for the same check. Whether that number tracks a published rate or stays fixed for the term decides what you actually spend two years from now.

open_in_newtheweple.com/insights/ai-feature-cost-drivers

What no contract says yet#

How far down does owning generation reach

How much of the cost advantage a company like Helix gets from owning generation reaches the tenant rate card cannot be answered from anything published today. The lease terms of the first campus built around its own generation will make that comparison possible. Until then, the move is to ask, when someone offers megawatts opening next year, whether that project is filed, past first review, in full review, or approved.

Sources & related

Want this built for real?

Once the brief is read, the next question is usually how to actually build it. Weple is an outsourced development studio that takes structures like the one on this page and ships them in short, scoped pieces. Send this page along with your inquiry.

Ask about building it