Korea Trades Memory for Power

If You Only Read One Thing

The $500 billion in the Nvidia–SK headlines is not a $500 billion investment; much of it is a promise to buy Korean memory. That distinction links Korea Turns Memory Into Power with Waymo Wants the Rider: both suppliers are using scarcity to reclaim customers and margins from platforms that aggregate demand. One holds HBM. The other has the driverless car.

Korea Turns Memory Into Power

South Korea’s most valuable AI asset is not a model or a cloud. It is the high-bandwidth memory, or HBM, without which Nvidia’s accelerators cannot perform as advertised.

SK Group and Nvidia signed letters of intent for a “$500-billion-plus” initiative that combines long-term SK hynix memory supply with an SK Telecom AI factory of up to two gigawatts, whose first Vera Rubin systems are planned for 2027. Seoul separately described SK’s portion of a wider $950 billion package as $750 billion of semiconductor supply, while Samsung’s $200 billion contribution is supply to Broadcom. In another agreement, Naver, Nvidia, and Brookfield plan to expand a Korean AI factory from 55 megawatts to 200 megawatts by 2028; Nvidia’s proposed $1 billion investment depends on Naver securing the other $9 billion in committed financing.

Why it matters: Korea is using a scarce component as negotiating currency for a larger role in the AI value chain. Selling HBM captures a cyclical manufacturing margin; tying that supply to local compute, financing, and co-development can create customers for Korean clouds and software long after today’s memory shortage eases. That is the strategic move the $500 billion headline obscures: much of the number is prospective revenue, while the valuable concession is Nvidia’s participation in infrastructure built and operated in Korea.

The bargain also answers a national vulnerability: as CSIS notes, Korea’s trade surplus with China fell from $99.7 billion in 2018 to $19.6 billion in 2025 as Chinese manufacturers climbed the electronics stack. Converting memory leadership into domestic computing capacity is an attempt to keep more of the next stack’s recurring value at home.

Room for disagreement: Letters of intent and supply commitments are not financed data centers. Korea’s existing semiconductor buildout already faces serious power and water constraints, and the new headline totals may count spending that companies had already planned. If financing or grid capacity fails, Korea remains an exceptionally profitable component vendor rather than an AI operating center.

What to watch: Whether Brookfield’s nonbinding $9 billion term sheet becomes committed financing for Naver’s 200-megawatt expansion; that conversion is the cleanest test of whether the announced ecosystem exists beyond memoranda.

Waymo Wants the Rider

The robotaxi partnership between Waymo and Uber always contained a basic contradiction: Uber wants autonomous vehicles to become interchangeable supply, while Waymo wants its vehicles to be the product.

Waymo is considering ending the partnership and launching direct service in Austin and Atlanta when its contract permits in January 2028, according to a Financial Times report. Uber says direct booking could run alongside the existing deployment rather than replace it. Uber has complained about poor-weather availability and “unsustainable” financial terms; Waymo has raised concerns about routing and cleanliness. Their intentionally small Phoenix deployment, involving just over a dozen vehicles, ended in June, while hundreds of Waymo cars remain exclusive to Uber in Austin and Atlanta.

Why it matters: Uber’s advantage is demand aggregation: one app balances neighborhoods, peak hours, and a mixed pool of drivers. Waymo’s advantage is differentiated supply: it owns a scarce service that riders may seek by name. If Waymo can acquire customers directly, it keeps the booking margin, trip data, and rider relationship instead of helping Uber turn autonomous fleets into another interchangeable driver pool.

This is the same negotiation occurring in Korea from the opposite end of the stack: a scarce supplier first accepts distribution from an aggregator, then uses the resulting demand proof to renegotiate who owns the customer. Uber is preparing for that outcome by diversifying supply, including a deal targeting 10,000 Rivian robotaxis from 2028, with an option to negotiate for 40,000 more beginning in 2030.

Room for disagreement: Direct distribution makes Waymo responsible for the expensive part Uber already solved: keeping vehicles busy when demand shifts across time and geography. Uber’s hybrid network can absorb peaks with human drivers, and Waymo’s average fare was still above Uber’s in a January price comparison. Scarce supply can win the rider only if it can also match the aggregator’s availability.

What to watch: The price and wait-time gap between Waymo One and Uber in cities where riders can choose either; sustained parity would show that Waymo can replace aggregation with product preference.

The Contrarian Take

Everyone says: These announcements show that the scarce supplier now holds all the cards: Korea can trade HBM for an AI ecosystem, and Waymo can leave Uber once enough riders recognize the brand.

Here's why that's wrong (or at least incomplete): Scarcity creates negotiating power, not operating competence. SK can obtain Nvidia’s signature yet still fail to secure $9 billion or two gigawatts of power; Waymo can own the car yet still pay more than Uber to fill it between peaks. In both cases, the aggregator performs an unglamorous but valuable function—coordinating fragmented demand and infrastructure. Suppliers capture that margin only by rebuilding the coordination layer, not by declaring it unnecessary.

Under the Radar

A duress code becomes a federal test case — The Justice Department is prosecuting an Atlanta police-training-center protester under a rarely used statute after he allegedly entered a GrapheneOS duress passcode that erased his phone at the border. The defense says the search lacked a warrant; if the charge survives, a privacy feature designed for coercive searches can itself become evidence of obstruction.
Domo is selling the company but keeping the stock — Progress will pay $400 million for substantially all of Domo’s operating assets, liabilities, customers, and employees, while the listed corporate shell remains public under a new name. The unusual asset sale is a bleak marker for mature SaaS: the installed base still has value, but the growth-story wrapper no longer does.

Quick Takes

Anthropic compresses its own price umbrella. Claude Opus 5 approaches the performance of the higher-tier Fable 5 at roughly half the cost per task, while retaining Opus 4.8’s list price of $5 per million input tokens and $25 per million output tokens. The business signal is faster tier cannibalization: frontier capability is becoming a temporary premium, not a durable product category. (Source)
Europe’s tech rules become a U.S. trade barrier. President Trump ordered a Section 301 investigation one day after the EU fined Google €890 million under the Digital Markets Act, materially advancing yesterday’s retaliation threat. American platforms have effectively recruited the trade apparatus as an appeals court, raising the economic cost of European enforcement beyond any single fine. (Source)
Open weights find a commercially convenient coalition. Microsoft, Nvidia, Meta, OpenAI, IBM, Hugging Face, Mistral, Palantir, and more urged Washington to avoid premature restrictions on downloadable models. The competition argument is real, but so are the incentives: open weights weaken model-vendor lock-in while increasing demand for chips, clouds, security tools, and applications sold by the signatories. (Source)

The Thread

Today’s stories are about suppliers discovering that scarcity is only the first step toward power. HBM lets Korea demand local infrastructure; autonomous driving lets Waymo challenge the app that delivers riders; open weights let the rest of the AI stack resist dependence on a few closed labs. But each supplier then confronts the same hard problem: somebody still has to coordinate capital, electricity, vehicles, or customers. The durable winner will not merely own the scarce input. It will turn that input into a system others depend on.

Predictions

New predictions:

I predict: By January 31, 2027, Waymo will publicly announce direct Waymo One booking in either Austin or Atlanta, with service scheduled to begin no later than January 2028. No announcement in either market makes this prediction wrong. (Confidence: medium; Check by: 2027-01-31)

Issue date: July 25, 2026 · Generated: 3:24 AM EDT