Meta Supplies, TikTok Returns

If You Only Read One Thing

A government ban and a $10 billion rental deal now turn on the same question: who controls the asset, not who built it. Meta Makes Compute Wholesale turns Anthropic into a customer without making it an ally; TikTok Passes the Control Test makes a sale enough to change federal security policy. The DOJ opinion makes ownership the switch for permission.

Meta Makes Compute Wholesale

Meta's first reported anchor customer would be a model rival, evidence its infrastructure can become a business separate from its AI ambitions.

Meta is negotiating to supply Anthropic with as much as $10 billion of computing capacity over two years, according to a New York Times report summarized by Reuters. July 2's briefing said a named customer would separate a product line from valuation management. Anthropic would supply that evidence, though talks remain early.

Meta can sell excess capacity without becoming Amazon Web Services. This month it expanded its planned Louisiana campus to five gigawatts and more than $50 billion, supported by seven new gas plants, three grid batteries and nuclear uprates. Anthropic, meanwhile, says its contracted supply already includes 300 megawatts at SpaceX, up to five gigawatts from Amazon and another five from Google. A Meta contract would make a nominal rival one more supplier in a deliberately diversified portfolio.

Why it matters: AI infrastructure is separating into ownership and use. Training clusters have enormous fixed costs, long power commitments and chips that lose economic value quickly. Selling a block to one sophisticated tenant fills capacity without the software catalog, sales force and support burden of a general-purpose cloud. Meta needs a small number of buyers, not millions of developers.

That makes hyperscale computing resemble wholesale energy or telecom capacity. The owner captures a return on the physical network; the tenant keeps control of its workload and can move future demand among suppliers. Meta gains utilization and purchasing scale without requiring Anthropic to adopt Meta’s models. Anthropic gains bargaining power against Amazon, Google and Microsoft without building power plants itself.

The second-order effect is a different kind of concentration. More suppliers can reduce cloud lock-in, but the viable owners remain companies able to finance gigawatt campuses and negotiate generation. Competition moves from the API layer toward balance sheets, power contracts and chip allocation.

Room for disagreement: No contract has been signed. The talks could be evidence that Meta overbuilt because its own models cannot consume the capacity profitably, while Anthropic may only be shopping bids. Wholesale supply also carries commodity risk: Meta would absorb hardware depreciation and power obligations without the margins or customer stickiness of a mature public cloud.

What to watch: Watch the final contract, if signed, for a take-or-pay capacity commitment. That clause would show whether Anthropic is merely renting optional overflow or underwriting Meta’s buildout.

TikTok Passes the Control Test

Meta’s talks separate infrastructure ownership from workload control. TikTok shows the same distinction turning government permission back on.

Federal employees may now install TikTok on government devices, subject to their agencies’ discretion. The app did not become more secure overnight. The Justice Department concluded that its legal identity changed.

The 2022 No TikTok on Government Devices Act applies to an app developed or provided by ByteDance or an entity it controls. In a July 16 opinion, DOJ’s Office of Legal Counsel said TikTok USDS falls outside that definition. ByteDance owns 19.9% of the new joint venture; Oracle, Silver Lake and MGX collectively hold 45%. The venture has its own board, security program and a licensed recommendation algorithm that is being revised and retrained on U.S. data.

Why it matters: Washington has converted a technical national-security dispute into a corporate-control test. Congress prohibited a supplier because ByteDance could direct it. Once a transaction moved formal control to American-majority owners, the service crossed from forbidden to permissible even though ByteDance remains an investor and licenses the central algorithm.

That distinction creates a reusable policy template. Governments do not need to prove that code is harmless if they can dictate who appoints the board, holds data, audits updates and can remove personnel. Corporate governance becomes a security product: a company can sell control rights while preserving a minority economic interest and commercial license.

The template also changes acquisition economics. The U.S. venture was valued at roughly $14 billion, and the ownership structure gives three managing investors a decisive bloc. The scarce asset was not just TikTok’s audience. It was a governance arrangement capable of turning regulatory permission back on.

Room for disagreement: This is not arbitrary relabeling. A qualified divestiture was the remedy Congress demanded, and refusing to recognize one would eliminate the incentive to comply. But DOJ relied heavily on representations from TikTok USDS’s own general counsel, while the current federal procurement rule still states the broad prohibition. Formal independence is evidence about control, not proof that every data or algorithmic risk has disappeared.

What to watch: Watch whether the government rewrites Federal Acquisition Regulation 4.2202. A formal revision would make the new control test operational for agencies and contractors; silence would leave procurement officers reconciling a live rule with DOJ’s opinion.

The Contrarian Take

Everyone says: Meta’s Anthropic talks admit that the company built too much infrastructure for models that disappointed.

Here's why that's wrong (or at least incomplete): Capacity is lumpy: a five-gigawatt campus cannot be resized each quarter, and the newest chips are most valuable before the next generation arrives. Selling blocks of compute can therefore be rational even when Meta’s model plans are on track. Airlines sell seats to partners, telecom networks sell wavelengths and energy producers sign offtake agreements; none of those transactions proves the owner no longer needs the asset. The harder question is whether Meta can make wholesale demand recurring without accepting commodity returns. One Anthropic agreement would improve utilization, but only a portfolio of multi-year anchor tenants would establish a new business; until then, “failed AI bet” and “new cloud rival” both exceed the evidence.

Under the Radar

Korea’s AI boom is becoming a macro event. South Korea’s ICT exports reached $253.9 billion in the first half, up 120.5% year over year. Shipments to the U.S. rose 215.6%, versus 141% to China. The base effect matters, but so does the direction: AI infrastructure demand is no longer just lifting chip-company earnings. It is reshaping national trade balances and external exposure.
Wildfire smoke is entering the sanctions toolkit. Senator Bernie Moreno says he will introduce legislation sanctioning Canada and officials for failing to contain wildfires whose smoke crosses the border. The bill is a proposal, not policy. Its novelty is the instrument: a transboundary environmental externality is being framed as enforceable foreign misconduct, creating a route from disaster management to trade retaliation.

Quick Takes

Apple’s settlement problem has 20-plus principals. Apple and the Justice Department have reportedly begun early talks over the 2024 iPhone monopolization case, but a bilateral handshake may not end litigation joined by numerous states. The durable question is remedy design: rules governing defaults, messaging, wallets and developer access could alter platform economics more than a fine, while state plaintiffs can resist a narrow federal deal. (Source)
New York is pausing data centers to set a price. The state’s one-year hyperscale moratorium is less an anti-development ban than a negotiation window. New York plans environmental standards, a community-payment formula within 60 days, grid contributions and possible repeal of sales-tax exemptions. The state is turning access to power and permits into a standardized bargain rather than negotiating every campus from scratch. (Source)
An AI regulator funded by AI companies has a capture problem by design. The administration is considering an SEC-supervised body resembling Demis Hassabis’s proposal for an industry-funded, FINRA-style model evaluator (first reported by Bloomberg [paywalled]). Technical staffing and faster review are genuine benefits. But when incumbents fund the standards body and help define acceptable testing, compliance can become an entry fee that entrenches the firms best able to pay it. (Source)

The Thread

Contracts are becoming products in their own right. Meta can sell a long-term block of capacity without building a public cloud. TikTok's investors bought a governance package that switches federal permission back on. New York wants permits to carry a standard community tariff, while an industry-funded AI evaluator would sell regulatory legitimacy.

This is the financialization of compliance and capacity: institutions divide a system into rights, attach a price or permission to each, and assign them to whoever will fund the bargain. The weak point is incomplete contracts. Power reliability, algorithm updates and model behavior are hardest to specify in advance, so risk accumulates in the dependencies the agreement does not price.

Predictions

New predictions:

I predict: By September 14, New York will publish a Community Investment Framework that includes a project-size or power-capacity formula for data-center payments rather than leaving compensation entirely to local negotiation. (Confidence: medium; Check by: 2026-09-14)

Issue date: July 18, 2026 · Generated: 3:39 AM ET