Gatekeepers Price the Stack

If You Only Read One Thing

Scarcity is no longer just a supply problem; it is becoming a permission system. Polestar Leaves the Road shows Washington using software provenance to decide market access, while Google Rations the Model shows compute becoming an allocation market between rivals. Start with Polestar's own notice because the most important sentence says Europe is now the plan.

Polestar Leaves the Road

The first real casualty of the connected-car rule is not a Chinese import. It is a Swedish-branded EV whose U.S. flagship is built in South Carolina.

Polestar said the Commerce Department's Bureau of Industry and Security did not grant it authorization under the Connected Vehicle Rule to sell vehicles in the U.S. from model year 2027 onward. The company will keep selling existing Polestar 3 and Polestar 4 inventory and support current U.S. customers, but it is increasing its focus on Europe, which Polestar says already represented close to 80% of retail sales and 94% of first-quarter 2026 retail volume outside the U.S. As The Verge reported, the Polestar 3 is made in South Carolina and the Polestar 4 is assembled in South Korea; the problem is Geely control and connected-vehicle software, not final assembly.

Why it matters: This is the point where industrial policy moves inside the car. The familiar version of auto trade policy is tariffs and local-content rules: where was the vehicle assembled, what percentage of parts were local, and how much duty applies at the border. The connected-car rule adds a different test: who can write, update, inspect, or influence the software and communications systems after the sale. BIS's final rule prohibits covered vehicle connectivity hardware and software with a sufficient China or Russia nexus, and it separately restricts manufacturers owned or controlled by those jurisdictions because they have privileged access to vehicle systems before sale.

That turns cars into a software-supply-chain authorization business. The Commerce Department's own rule summary says the covered systems include telematics, Bluetooth, cellular, satellite, Wi-Fi modules, and automated-driving software, and that the prohibition can apply even if the vehicle was made in the United States. The strategic implication is blunt: local factories are no longer enough to prove domestic trust. The trusted asset is the control plane, meaning the entity that governs updates, data flows, connectivity modules, and automated-driving logic.

That is why the Volvo contrast matters. Volvo, also Geely-owned, reportedly received authorization after talks over governance, technology, and data security, while Polestar did not. The rule is therefore not a simple nationality ban. It is an approval regime with room for architecture, segregation, and corporate control arguments. The winners will be automakers that can produce not just vehicles but evidence: supplier maps, code provenance, update controls, data-routing limits, and corporate-governance firebreaks that satisfy BIS.

Room for disagreement: The U.S. market was not carrying Polestar. Europe was already the core business, tariffs made China-made models difficult, and Polestar had warned that the rule could effectively block it. So one reading is that Commerce accelerated an exit that commercial gravity had already begun. That is true, but too narrow. The precedent is not Polestar's lost volume; it is that a Western-listed, Sweden-based brand with U.S. manufacturing can still fail the software-control test.

What to watch: Watch whether BIS publishes any authorization conditions, redacted reasoning, or compliance template before the model-year 2027 cutoff. If the Volvo-Polestar split stays opaque, the authorization process will look less like a rulebook and more like discretionary market licensing.

Google Rations the Model

Meta can spend hundreds of billions of dollars on AI infrastructure and still end up waiting in a rival's capacity queue.

The Financial Times reported, and Mint summarized, that Google told Meta around March it could not supply all the Gemini capacity Meta wanted to buy, disrupting or delaying some internal AI projects. Mint says Meta had used Google's models for work including advertising, scam detection, customer support, and coding, and urged employees to be more efficient with tokens, the usage units that measure model consumption. Google has also exposed the same constraint at the consumer layer: its Gemini help page says limits can change without notice due to capacity constraints and that users may be shifted to lighter models after reaching limits.

Why it matters: The default story about AI capacity is that the biggest companies will buy their way out of the shortage. That is only partly right. Compute is not a commodity barrel of oil; it is a bundle of chips, power, networking, memory, serving software, model weights, safety systems, and priority rules. If the bundle is scarce, the seller does not simply maximize revenue. It allocates capacity across internal products, strategic customers, cloud backlog, consumer subscriptions, and competitive risk.

Google's own numbers make the constraint visible. Alphabet's Q1 earnings transcript says Cloud revenue rose 63% to $20 billion, Cloud backlog nearly doubled sequentially to $462 billion, and Sundar Pichai told analysts Google was compute constrained in the near term; Cloud revenue would have been higher if the company could meet demand. Meanwhile, BloombergNEF estimated in March that the largest data-center firms' capex was nearing $750 billion in 2026 and that more than 23GW of data-center IT capacity was under construction globally. The capex boom and the rationing can both be true because capacity arrives in lumpy, site-specific, power-dependent chunks while model demand arrives immediately.

The structural surprise is Meta's dependence. Meta has its own Llama history, its Muse Spark push, and one of the largest capex programs in the industry. Yet it reportedly still used Gemini where Gemini performed better or was faster to deploy internally. That makes frontier models less like ordinary SaaS and more like strategic inputs. If a rival controls the best input for ads, moderation, support, or coding, then the customer is not only buying performance; it is exposing product timing to another platform's queue.

Room for disagreement: Capacity rationing can be temporary. Google is building, Meta is building, and model efficiency can improve faster than raw demand. It is also possible that the Gemini restriction nudges Meta into a healthier internal substitution path. But the counterargument has to explain why Google, with surging Cloud revenue and a massive buildout, still had to ration a customer as large as Meta. Scarcity that reaches Meta is not a startup inconvenience.

What to watch: Watch Alphabet's next earnings call for a change in the language around backlog and fulfilled demand. If Google starts separating AI infrastructure backlog from ordinary cloud commitments, capacity allocation will have become a financial-reporting category.

The Contrarian Take

Everyone says: The lesson from Polestar and Meta is that governments and hyperscalers are getting more powerful.

Here's why that's wrong (or at least incomplete): Power is not just accumulating at the top; it is moving to whoever controls verification. BIS now decides whether a connected-car stack is trusted enough for the U.S. market. Google decides which model customers receive scarce compute and at what priority. UK courts may decide worldwide patent prices for mobile standards. The scarce asset is no longer only chips, factories, or models. It is the ability to make access auditable.

Under the Radar

• London is pricing global wireless patents. Apple's UK Supreme Court appeal against Optis is not just a $502 million bill. The court's case summary says UK courts can set global fair, reasonable, and non-discriminatory licensing terms for standard-essential patent portfolios, and the dispute moved from a $56.43 million first-instance award to $502 million plus interest at the Court of Appeal. That makes London a rate-setting forum for the connectivity stack.
• China's IPO pipeline is doing industrial policy work. LSEG data cited by Reuters through Investing.com showed Chinese AI and chip companies raised $3.1 billion in domestic listings this year through June 18, more than five times the year-earlier volume. The hidden point is that capital-market reopening is part of self-reliance policy: Beijing is using public listings to finance the suppliers export controls are supposed to slow.

Quick Takes

• Qualcomm is turning export controls into product segmentation. Qualcomm plans China-specific Dragonfly data-center products, including export-compliant AI accelerators that avoid restricted thresholds. The move looks like Nvidia's old China playbook, but Qualcomm's phone and auto relationships give it a different entry point: the server market may be built through existing Chinese device and vehicle customers. (Source)
• SpaceX's compute pivot is starting to look like a financing model. MarketWatch says SpaceX has signed large AI compute deals with Alphabet, Anthropic, and Reflection that could exceed $11 billion by year-end 2026, while allowing termination after early ramp periods. That makes compute leasing a bridge between the IPO story and the AI-capex story, but also exposes SpaceX to the same utilization risk facing every neocloud. (Source)
• Consumer Gemini limits preview enterprise priority tiers. Google's support page says Gemini limits can change without notice due to capacity constraints, and users may continue on Flash-Lite after hitting limits. Consumer limits are not the story; the story is productized scarcity. The same logic becomes enterprise contracting once model access depends on priority, workload type, and fallback rights. (Source)

The Thread

Today's stories are all access stories. Polestar lost access to a market because software provenance beat factory location. Meta ran into a capacity ceiling because even the largest AI customers are not entitled to unlimited model supply. Apple is asking UK judges to narrow a global patent-price precedent. China is pushing AI and chip firms into domestic listings so strategic suppliers can keep raising capital. Qualcomm is carving chips around export thresholds. The old platform question was who controlled distribution. The new question is who controls the permission system that distribution now depends on.

Predictions

New predictions:

• I predict: By September 30, 2026, at least one China-linked automaker or major supplier will announce a U.S.-specific connected-vehicle software stack, data-governance structure, or supplier replacement plan explicitly designed to satisfy BIS's connected-vehicle authorization process. (Confidence: medium; Check by: 2026-09-30)

Generated: 2026-06-28 03:25 EDT