Skip to content
TEN Brief Ten verified stories a day 2026.09.08 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Tech · 3 min read · Reference

What AI capex is — how big tech spending turns into Korean export figures

Capex, short for capital expenditure, is money a company spends on assets it will use for years — buildings, land, equipment — as distinct from opex, the day-to-day cost of running the business. AI capex is the part going into data centers: the shell, the power feed, the cooling, and above all the accelerators inside. This number matters well beyond the firms spending it, because AI capex is unusually concentrated: a handful of buyers commit multi-year budgets, and a large share of each lands on a short list of suppliers. Goldman Sachs estimated on September 4, 2026 that US big tech capital spending for next year would be revised up from about 800 billion dollars to 1.2 trillion, and kept a KOSPI target of 12,000 on that basis. The route from that estimate to Korean export statistics runs through hardware — budget buys racks, each rack carries dozens of GPUs, each GPU carries HBM, and HBM comes from two Korean makers and one American one. That is how a spending line in California appears in a customs release in Daejeon

A data center construction site in morning light, the steel frame of a long low building with cranes against a blue sky

The three lines

  • Definition — capex is money for multi-year assets, distinct from operating cost. AI capex is the data-center share of it
  • Scale — Goldman estimates US big tech capex next year rising from 800 billion to 1.2 trillion dollars (September 4, 2026)
  • Route — budget to rack to GPU to HBM to Korean exports. The chain is short and concentrated, so one budget revision moves an index

Key questions

What is capex
**Money spent on assets a company will use for years.** From capital expenditure. Corporate spending splits in two. | Type | Meaning | Examples | |---|---|---| | **Capex** | Buying **assets** used over several years | Data-center buildings, land, servers and GPUs, substation equipment | | **Opex** | **Operating** cost for the year | Electricity, salaries, rent, cloud subscriptions | **The accounting treatment differs, and that matters.** Opex hits the income statement in full in the year it is spent. Capex is booked as an asset and then becomes expense gradually through **depreciation**. **So a large capex year dents current profit less, and future years more.** Which is where an accounting judgement starts moving reported results — change how many years a GPU is depreciated over and the same spending produces different profits (「What GPU depreciation is」). **AI capex** is the data-center portion. It breaks roughly into four parts — **building and land / power feed and substation / cooling / servers and accelerators** — and in the current cycle **accelerators take the largest share.**
Why does everyone watch this number
**Because there are few buyers and even fewer places the money goes.** Capital spending in an ordinary industry is spread across thousands of firms; one company changing plan does not move the market. **AI capex is different.** A handful of US big tech firms account for a large share, and each announces **multi-year budgets at once.** The list of suppliers those budgets land on is short — accelerators are effectively Nvidia-centred, and the high-bandwidth memory inside them comes from **two Korean makers and one American one** (「What HBM is」). **A short chain transmits signals fast and hard.** Three consequences. **① The forecast is the share price.** Goldman Sachs kept a KOSPI target of 12,000 on September 7 on the strength of an estimate that US big tech capex would rise from **800 billion to 1.2 trillion dollars** next year (「Goldman keeps its KOSPI target at 12,000 (September 7)」). **② Cancellation is also the share price.** A delayed plan travels the same chain in reverse. **③ The bottleneck migrates.** However large the budget, if power, land and cooling cannot keep up, the money does not get spent that year (「What the data-center power crunch is」).
How does that money reach Korea
**Budget, rack, GPU, HBM, customs data — five steps.** | Step | Content | |---|---| | ① **Budget** | Big tech fixes annual data-center capex | | ② **Rack** | The purchase unit is now the rack, not the card. A GB200 NVL72 rack holds **72 GPUs** (「What Grace Blackwell NVL72 is」) | | ③ **GPU** | Each Blackwell GPU carries **HBM** | | ④ **HBM** | Made by SK hynix, Samsung Electronics and Micron | | ⑤ **Customs data** | Appears in Korean semiconductor export totals | **The figures confirm the route.** Korean semiconductor exports from January to August 2026 came to **281.2 billion dollars**, up **169.6 percent** year on year and equal to **40.6 percent** of all exports. In August alone the share was **47.5 percent** (「Korea's exports pass 709.4 billion dollars (September 5)」). **But the path has lag and friction.** Quarters pass between an announced budget and an actual order (「How long semiconductor capex takes to reach prices」), and tariffs bend the route itself (「US weighs second-round semiconductor tariffs (August 27)」). **And capex is a commitment, not a payment.** Announced amounts are not guaranteed to be spent — nor are they capped. The estimate that 800 billion will become 1.2 trillion is itself evidence of how much this number moves.

A spending line in California turns up in a customs release in Daejeon. The chain between them is five links long, which is why this number becomes a Korean market catalyst.

1. Capex and opex

TypeMeaningExamplesAccounting
CapexBuying multi-year assetsData-center buildings, land, GPUs, substationsBooked as an asset, expensed via depreciation
OpexOperating cost for the yearElectricity, salaries, rentFully expensed in the year spent

The distinction moves reported results. A large capex year dents current profit less and future years more. And how many years to spread it over is a company choice — change a GPU's depreciation schedule and the same spending produces different profits (「What GPU depreciation is」).

AI capex is the data-center portion, in roughly four parts.

ComponentContent
Building and landThe shell and the site
Power feed and substationThe practical bottleneck in this cycle
CoolingAt 120 kW per rack, liquid is assumed
Servers and acceleratorsThe largest share today

2. Why this one line is watched

Ordinary industrial capex spreads across thousands of firms. One company's change moves nothing.

AI capex inverts that.

PropertyConsequence
Few buyersOne firm's revision moves the whole market
Multi-year budgets announced at onceThe number itself becomes a demand signal
Short supplier listSignals transmit fast and hard

So the forecast becomes the share price. On September 4, 2026, Goldman Sachs estimated US big tech capex for next year rising from 800 billion to 1.2 trillion dollars and kept a KOSPI target of 12,000 (「Goldman keeps its KOSPI target at 12,000 (September 7)」).

The reverse travels at the same speed. A delayed plan runs the same chain backwards. And however large the budget, if power, land and cooling cannot keep up, it is not spent that year (「What the data-center power crunch is」).

StepContent
BudgetBig tech fixes annual data-center capex
RackThe unit of purchase is the rack. NVL72 holds 72 GPUs (「What Grace Blackwell NVL72 is」)
GPUEach Blackwell GPU carries HBM
HBMSK hynix, Samsung Electronics, Micron
Customs dataKorean semiconductor export totals

The figures confirm it. Korean semiconductor exports for January to August 2026 were 281.2 billion dollars, +169.6 percent year on year, 40.6 percent of all exports. August alone: 47.5 percent (「Korea's exports pass 709.4 billion dollars (September 5)」).

4. Where the chain breaks

FrictionContent
LagQuarters pass between an announced budget and an actual order (「How long semiconductor capex takes to reach prices」)
TariffsThe route itself bends (「US weighs second-round semiconductor tariffs (August 27)」)
PowerIf electricity does not reach the site, the budget goes unspent
Commitment vs paymentAnnounced amounts are not guaranteed to be spent

The last matters most. Capex is a commitment, not a payment. That 800 billion is expected to become 1.2 trillion is itself evidence of how often this number moves.

5. What is left and what is unverified

  • The company set behind '800 billion to 1.2 trillion' is unspecified. Four firms or seven changes the meaning.
  • No generalised component percentages are offered. The split varies widely by operator and site.
  • HBM shares were not confirmed. The three-supplier structure is known; generation and customer breakdowns are not.
  • Depreciation schedules differ by company. Only the direction is described here.
  • No time series exists for capex execution rates. "What share of announced capex is actually spent" would be the single most useful number for judging this chain, and it is not published.

Sources

  1. CBC News — US big tech investment revised to 1.2 trillion dollars; Goldman says KOSPI 12,000 is not absurd
  2. Hankyung — Goldman Sachs keeps KOSPI target at 12,000, rejecting the peak-out thesis
  3. J.P. Morgan — Financing AI infrastructure and U.S. data centers
  4. NVIDIA — GB200 NVL72 product page
  5. Financial News — Cumulative exports reach 709.4 billion dollars in early September, passing last year's full-year total

Verification

Published
Last modified
Cross-check
Checked against 5 independent sources.
Unverified
  • The 800 billion to 1.2 trillion dollar figure is a Goldman Sachs estimate, and Korean reporting does not specify which set of companies it aggregates.
  • The actual split of AI capex across building, power, cooling and accelerators varies widely by operator and site, so no generalised percentages are given here.
  • HBM supply is generally attributed to SK hynix, Samsung Electronics and Micron, but generation-by-generation and customer-by-customer shares were not confirmed.
  • Depreciation schedules differ by company; this article describes only the direction of the effect, not specific figures.
  • No public time series was found for how much of announced capex is actually spent.
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related