Korea spent last week signing AI deals at a scale that is hard to hold in the head. More than $950 billion of agreements, by the Korean government's own tally, announced across a San Francisco summit built around President Lee Jae Myung's visit, where he shared the floor with the chief executives of Nvidia, OpenAI, Anthropic and Broadcom.

One of those announcements is worth reading slowly, because the press release says something different from the headlines it produced.

What was announced

On 24 July, NVIDIA, NAVER and Brookfield said they plan to expand the NVIDIA DSX AI factory at NAVER's GAK Sejong data centre in South Korea from 55 megawatts to 200 megawatts by 2028. The 55-megawatt buildout is itself recent — announced in June, during Jensen Huang's Seoul deal week. NAVER says it intends to reach a gigawatt eventually.

The factory is expected to feature NVIDIA's Vera Rubin and Blackwell platforms on the DSX stack. NVIDIA's own live blog, opened 23 July, puts 200 megawatts at "roughly 100,000 GPUs" — a figure that appears in the blog, not in the press release.

The money, as reported almost everywhere, is $10 billion: $9 billion from Brookfield, $1 billion from Nvidia, with NAVER funding the rest of the project cost.

What the press release actually says

Brookfield has "entered into a nonbinding term sheet to fund up to $9 billion." Not committed nine billion. A nonbinding term sheet, for up to that amount.

And Nvidia's contribution is, in the release's own words, a case of Nvidia "plans to invest" — a planned investment, subject to customary closing conditions and to "NAVER finalizing at least $9 billion of committed financing for the project, separate from NVIDIA's planned investment." (The warmer phrase "planned strategic investment" appears in the release too, but inside a quote from NAVER's chairman, not in its own descriptive text.)

Read those two sentences together and the sequencing inverts the story. Nvidia's billion is conditional on NAVER securing nine billion of committed financing. The nine billion currently on the table is nonbinding. The largest single number in the deal is the one that has to become real before the second-largest number arrives.

None of this is hidden. It is disclosed plainly in the release, in the ordinary language of infrastructure finance. It is simply not what "Nvidia invests $1 billion in Naver" conveys.

The market read it the other way

NAVER shares closed up 8.43% at 225,000 won, comfortably outpacing a KOSPI that finished up 0.97% after an intraday dip. Korean outlets reported that Nvidia's $1 billion would take roughly a 4.5% stake via new shares issued through a third-party allotment — about 7.24 million shares at 204,500 won, NAVER's first third-party allotment since it listed on the KOSPI in 2008 — with NAVER saying in a regulatory filing that Nvidia will purchase them.

And here the picture cuts the other way, which is worth saying plainly: Korean reporting describes binding terms attached to this leg, with a signed strategic investment agreement. So of the three tranches, Nvidia's $1 billion may be the most firmly committed, not the least — even though NVIDIA's own release is the document hedging it. A company announcing a financing plan to its regulator is not the same as money having moved; but nor is a nonbinding term sheet the same as a signed allotment. It is also worth noting that NVIDIA's release does not characterise the investment as equity at all; the 4.5% figure comes from Korean business reporting.

That distinction matters more than it looks. If the money is new shares, it is dilutive primary capital: it does not buy out an existing holder, it becomes the facility. Nvidia's cash would flow into the company building a data centre that then buys Nvidia's chips.

We have seen this shape before — but not here

That circularity is not new, and we wrote about it four days ago: AMD putting up to $5 billion into Anthropic against a chip order worth far more, the same structure Nvidia had already used with OpenAI. Capital flows toward the customer; a larger sum flows back to the vendor.

What is different here is the counterparty. The best-known instances have run from a chipmaker to a frontier lab — though not all of them do: NVIDIA's own 13F for the quarter ended 31 March 2026 discloses a $3.7 billion stake in the GPU cloud provider CoreWeave, and a much smaller one of about $124 million in Nebius, both of which buy its chips. This one runs to a national internet champion, inside a state-backed industrial programme, with a sovereign-AI framing supplied by all three parties. Brookfield's Sikander Rashid put the rationale this way: "As AI adoption accelerates across the global economy, access to trusted, sovereign and scalable AI infrastructure is becoming an increasingly important strategic priority for companies and countries." Huang's line was blunter: "AI factories are the infrastructure nations need to compete and innovate in the intelligence era."

When the structure moves from a private lab to a national champion, the conditionality stops being a commercial detail. A frontier lab that misses a financing milestone disappoints its investors. A national AI factory that does is a public number that a government has already stood behind.

The honest reading

Nonbinding term sheets convert routinely. Staged, condition-linked financing is how projects of this size are ordinarily built, and there is no reason today to think this one will not proceed — Brookfield is one of the largest infrastructure investors in the world, NAVER is an operator with a hyperscale facility already running at GAK Sejong, and the June 55-megawatt commitment is on the books.

But there is a difference between a $10 billion AI factory and a plan under which $10 billion could be assembled if a nonbinding commitment becomes binding and a set of closing conditions is met by 2028. The first is what most coverage said this week. The second is what the companies said.

It is also worth noting that this news is not as fresh as its circulation suggests: NVIDIA's blog is dated 23 July, the press release 24 July, and the wave of coverage carrying it arrived on 26 and 27 July.

The rest of the summit carries the same caveat at larger scale. NVIDIA describes plans for a "$500-billion-plus comprehensive partnership" with SK Group, 50,000 Blackwell GPUs extending Hyundai Motor Group's physical-AI buildout, and a joint AI lab with KAIST, which the two announced. The $950 billion aggregate is largely a sum of announced intentions rather than contracted spend — the presidential office said the companies had agreed to pursue partnerships worth that much, though some components, including NAVER's own agreements, are firmer than others. It is a distinction that tends to survive precisely as long as it takes to write a headline.