NAVER, NVIDIA, and Brookfield Put $10 Billion Behind Korea's 200 Megawatt Sovereign AI Factory. The Financing Template Is the Story.

NAVER, NVIDIA, and Brookfield put out a joint statement on Friday, July 25, saying they will expand NAVER's DSX AI factory at the GAK Sejong hyperscale campus from 55 megawatts to 200 megawatts by 2028. Total bill: about $10 billion. NVIDIA writes an equity check of $1 billion. Brookfield puts up to $9 billion on a non-binding term sheet. NAVER covers the remainder and operates the site. The compute starts on Blackwell today, moves to Vera Rubin in 2027 and 2028, and the trio's stated long-term ambition is one gigawatt of sovereign AI capacity on the same campus.

The megawatt number is not the story. The financing structure is.

The Numbers

The Vendor Equity Loop Just Recruited a Project Finance Firm

The $1B NVIDIA equity into NAVER continues a pattern we have been tracking all year. Google put roughly $40B of equity into Anthropic and is recovering it through TPU spend in the deal we broke down in May. AMD wrote a $5B check to Anthropic to close 2 gigawatts of MI450, walked through in last Thursday's piece. NVIDIA itself sent $40B toward OpenAI to anchor a Vera Rubin commitment, the closed loop we sketched in April. Every one of those deals closes inside the chip vendor's balance sheet. Customer takes equity, gives commitment, vendor gets a demand floor, and the equity becomes a hedge on the commitment.

Those deals worked because the customer on the other side was a US frontier lab with hyperscaler credit behind it and a revenue line that doubles on a quarterly cadence. NAVER is not that customer. NAVER is a Korean national champion with a real advertising and commerce business but nothing that looks like Anthropic's $30B run-rate or OpenAI's ChatGPT cash flow. NAVER is also building sovereign compute, which by definition cannot lean on AWS, Azure, or Google Cloud to underwrite the ramp. The vendor equity loop stops closing at NVIDIA's balance sheet.

The gap gets filled by Brookfield's non-binding $9B, which is not a chip deal at all. It is project finance against a physical asset, sourced from the $100 billion AI infrastructure fund Brookfield stood up last year with NVIDIA and KIA as anchor investors. That is the new template. NVIDIA writes a token equity check for alignment. Brookfield writes the majority of the bill against the site. The customer signs a compute offtake that services the debt. Sovereign compute gets built without a hyperscaler landlord and without the customer needing to be creditworthy at the scale the compute demands.

Two Sovereign AI Templates Are Now in the Field

Beijing already showed the other version. Z.ai turned on a gigawatt in Hebei last week, the story we wrote up on Tuesday, using nothing but Chinese silicon on state-financed grid rails. No vendor equity from anyone because there is no allied vendor to write it. The Chinese template is: burn political will, absorb the yield hit on inferior silicon, close the sovereignty gap on hardware.

Seoul's template is different. Keep the frontier silicon (Blackwell today, Vera Rubin from 2027). Give up a fraction of upside to NVIDIA via the equity leg. Recruit private capital at project-finance scale. Anchor the whole thing on a national champion. What you get is a sovereign AI factory that actually runs the frontier stack, but you now owe Brookfield and NVIDIA both. Ownership is on paper Korean. Financially, it is a triangle.

Neither template is obviously wrong. Beijing's version buys full technical independence at the cost of a 20 to 40 percent per-chip performance gap and a balance-sheet dependence on state actors who do not price risk the way private markets do. Seoul's version keeps hardware parity and dilutes to a set of allied counterparties who now have a stake in the outcome. If you are a US ally sitting on the Korean side of the buildout, that is a feature, not a bug. If you were hoping for a template you could copy without owing NVIDIA and Brookfield, this week did not deliver one.

What This Does to NVIDIA

Two things.

One, NVIDIA now has a repeatable placement mechanism for gigawatt-class capacity in allied jurisdictions where the customer is neither a hyperscaler nor a frontier lab. That was the growth vector nobody had priced. Every US frontier-lab compute deal has already been written. Sovereign compute in Korea, Japan, Taiwan, Singapore, the Gulf, and probably slices of the EU, that market was open, and until this quarter the underwriting mechanism did not exist. Brookfield just handed NVIDIA one.

Two, the $100B AI infrastructure fund is the external mirror of the customer-investor loop we have been describing all year. Google recycles Anthropic equity into TPU revenue. NVIDIA does the same trick but funds the deal externally through Brookfield, which is significantly more capital efficient at the vendor level. NVIDIA takes multi year offtake without carrying multi year balance sheet risk against a customer that is not one of its natural credits. It is the cleanest version of the compute buildout that has been assembled anywhere in the industry.

The catch is concentration. Brookfield's fund is $100B and has NVIDIA anchoring the anchor tenants list. If Brookfield ends up on every meaningful sovereign AI deal outside China, and NVIDIA sits on the same side of every one, the CFIUS-adjacent question becomes real quickly, and so does the antitrust one. Kimi K3's open weights drop today (July 27), and if every sovereign AI factory on the allied side of the wall ends up running the same Brookfield-financed NVIDIA rack architecture, the word sovereign starts to do a lot of work in the marketing copy.

The 2028 Delivery Window

The Korea site does not fill up in 2026. The 200 MW target is 2028. Vera Rubin arrives in 2027 and 2028. That puts NAVER's build alongside essentially every other frontier compute commitment we have been tracking. OpenAI's 10 gigawatt Vera Rubin ramp with NVIDIA, Anthropic's $200B on Google TPU, Meta's Anthropic-adjacent buildout, and now this. The 2027 to 2028 window is the physical delivery ceiling on close to a trillion dollars of announced compute demand. Fab allocations, substation permits, water rights, and grid interconnects all converge in the same 18 month band.

That matters because the offtake math on a sovereign AI factory is nothing like the offtake math on a frontier lab. NAVER cannot double revenue every six months. It runs a business inside Korea's advertising, commerce, and mapping economy, and its top line is roughly $10B. The compute at Sejong has to service a longer, flatter demand curve than what Anthropic or OpenAI runs against, and Brookfield knows this. That is precisely why the term sheet is non-binding today and why the guarantee, when it hardens, will be a much stronger take-or-pay than any hyperscaler has ever asked a customer to sign.

That guarantee lives on NAVER's balance sheet, which is why the stated 1 GW long-term target is the number the deal actually depends on. The 200 MW at Sejong is only the first tranche. The financing model only works if Korean demand keeps climbing through the decade, which is a bet on NAVER pulling US-based AI customers into the site through the trio's stated intent to serve Korean and US innovators from the same racks.

Our Take

The pattern we saw with Anthropic-Google (customer pre-finances vendor buildout via offtake), Anthropic-AMD (vendor equity check locks in commitment), and NVIDIA-OpenAI (larger vendor check locks in larger commitment) just added a fourth variant. Project finance replaces the vendor's balance sheet. The vendor still gets the customer, and now also gets to preserve capital for the deals that need equity to close.

That template is the exportable version of the Google-Anthropic loop. It is why NAVER got Vera Rubin capacity, and why every sovereign AI buildout on the allied side of the wall for the next 18 months is going to look like this one on the org chart. The customer sits on the sovereignty. NVIDIA sits on the silicon. Brookfield sits on the beam.

Practical implication for builders. If your AI product is going to run inference in Korea, the price floor by 2028 is going to look a lot like the inference price floor everywhere else, because the underlying hardware and vendor economics are the same. The compute floor we described in our pricing-floor analysis is now genuinely global at the frontier tier. What differs is who owns the copy of the model and what jurisdiction its training data sat in. That is a policy floor, not a price floor.

Three signposts to watch:

One, whether the Brookfield term sheet hardens into binding debt before year end. That is what turns the $9B soft commitment into the actual financing engine of the site, and the terms of the offtake will tell you how tight Brookfield is squeezing NAVER on take-or-pay.

Two, whether Japan (SoftBank plus Rapidus) or Taiwan (Foxconn plus TSMC) announces a comparable NVIDIA plus project-finance sovereign AI deal before Q1 2027. If Brookfield shows up on either, this becomes a template. If it does not, this stays a one-off.

Three, whether the AI FINRA gate we covered in June ends up applying to sovereign AI factories that host US-designed model weights, not just to the labs that train them. If the answer is yes, the offshore version of the American frontier stack picks up an additional compliance surface, and Brookfield's legal team gets very busy.