AI Has Entered Its Rationing Era

The AI story changed in June. OpenAI said its GPT-5.6 Sol model would initially be available only to a small group of Trump administration-approved customers, while Anthropic had already suspended access to its Fable 5 and Mythos 5 models after a U.S. government directive barred access by foreign nationals, including some people inside the United States, according to AP¹ and Anthropic’s own statement². At roughly the same time, G7 leaders, the club of major industrial democracies, discussed a “trusted partners” scheme for access to advanced U.S. models, according to Reuters³. This is not a side plot. It is the beginning of a new question for the AI industry: not simply what the best model can do, but who is allowed to use it, on what chips, powered by which grid, at what price.

My view is blunt: this is becoming a durable constraint on AI economics, not just a temporary supply-chain hiccup. I do not mean AI growth stops. It will not. I mean the industry’s next phase will be governed by rationing: access rationing by governments, compute rationing by chips and memory, and deployment rationing by power and data-center capacity. That makes the bullish case harder. It also makes the sloppy bubble case too simple.

A few terms matter. A frontier AI model is one of the most capable general-purpose systems at the edge of current performance, especially where cyber, biological, military, or intelligence uses worry governments. Export controls are rules that limit the transfer of sensitive goods, software, technology, or services to foreign users. A GPU, or graphics processing unit, is the accelerator chip used for much AI training and inference. HBM, or high-bandwidth memory, is stacked memory attached to advanced AI chips so they can move huge amounts of data fast. Data-center capex is the capital spending on buildings, servers, chips, cooling, networking, and power systems. An AI bubble means asset prices and investment assumptions have run ahead of plausible profits.

The access controls are real, even if the law is messy. The strongest argument against panic is that the June model restrictions look improvised. President Trump’s June 2 executive order created a voluntary framework for companies to give the government up to 30 days of pre-release access to advanced models, and legal analysts at CSIS⁴ argue the Commerce Department’s authority to restrict hosted model access is uncertain. That matters. If the legal foundation is weak, courts and companies may narrow the regime.

But the direction of travel is unmistakable. Anthropic said the government directive cited national security authorities and required suspension of access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including foreign-national employees, according to Anthropic². OpenAI then limited GPT-5.6 Sol to trusted partners approved by the administration, according to AP¹. The G7 talks suggest allied governments are trying to convert ad hoc blocks into a managed-access framework, not return to the old assumption that a frontier model release is just a product launch, according to Reuters³.

That distinction matters for business models. A “trusted partners” system is not a ban on AI. It is still a constraint. It means identity checks, nationality screening, audit logs, model-tiering, cloud-region controls, and delayed releases for the most powerful systems. These are manageable costs for Microsoft, Amazon, Google, OpenAI, and Anthropic. They are not manageable in the frictionless way investors often price software. Software margins look different when the product behaves less like an app and more like controlled dual-use infrastructure.

The chip side is moving in the same direction. The Commerce Department’s Bureau of Industry and Security said in January that exports of Nvidia H200, AMD MI325X, and similar advanced AI chips to China and Macau would be reviewed case by case if security requirements are met, according to the official BIS announcement⁵. That is less severe than a blanket ban. But it is still a political allocation system for the very hardware needed to run frontier AI. In June, Commerce guidance clarified that licensing requirements for advanced AI chips also apply to companies headquartered in China or with a Chinese parent, even when their subsidiaries operate outside China, according to Al Jazeera’s report on the guidance⁶. The global cloud market is not becoming borderless. It is becoming permissioned.

The physical bottleneck is worse than a normal chip shortage. Semiconductor markets are cyclical, and high prices usually call forth new supply. That is the cleanest counterargument. SK hynix plans to double memory wafer capacity within five years, according to Tom’s Hardware⁷. Nvidia says its Rubin platform can deliver up to 10 times lower cost per token than Blackwell, according to Nvidia⁸. If supply rises and inference gets cheaper, bottlenecks should ease.

I buy part of that. I just do not think it solves the investment problem fast enough. AI demand has pulled HBM and server memory into the center of the semiconductor cycle, and the squeeze is spilling into ordinary DRAM, the memory used in servers, PCs, and devices. S&P Global reported that as capacity tilts toward HBM, traditional DRAM is facing shortages and higher prices, with 2026 HBM average selling prices expected to rise by about 8% at Samsung, 1% at SK hynix, and 22% at Micron, according to S&P Global Market Intelligence⁹. Tom’s Hardware reported that SK hynix shares surged on HBM demand, that the company held 61% of the global HBM market in 2025, and that memory shortages could run past 2027 as capacity shifts toward AI memory rather than commodity chips, according to Tom’s Hardware¹⁰.

This is why the “temporary cycle” story feels too tidy. HBM is not just more memory. It is a specialized, high-margin input for AI accelerators, and shifting wafer capacity toward it can worsen shortages in conventional memory. TrendForce data cited by Tom’s Hardware showed DDR2 contract prices rising 55% to 60% in the second quarter of 2026 as major memory makers prioritized newer, higher-margin AI-related products, according to Tom’s Hardware¹¹. That is the strange shape of the AI boom: the newest models are bidding away capacity from old memory standards that still sit inside industrial and consumer hardware.

Then comes power. This is the bottleneck the software world is least prepared to think about. Data centers do not run on venture capital. They run on electricity, transformers, switchgear, substations, water, land, and local permission. Sightline Climate found that power constraints and grid equipment shortages could delay 30% to 50% of data-center projects in 2026, according to Semafor¹². TechSpot, also citing Sightline, reported that between 30% and 50% of U.S. AI data centers planned for deployment this year could be delayed or canceled because of power availability and electrical equipment shortages, according to TechSpot¹³. That is not a bug fix. It is civil infrastructure.

The International Energy Agency’s analysis is more balanced, and it should be taken seriously. The IEA expects data-center electricity demand growth to be met by a mix that includes expanding renewables, according to its report on energy supply for AI¹⁴. But “met eventually” is not the same as “available when hyperscalers want to deploy.” If an AI lab can improve model efficiency in 12 months but a utility interconnection, transformer order, or transmission upgrade takes years, the scarce input decides the schedule.

This is where valuations become fragile. Goldman Sachs says large technology companies leading the AI buildout could spend a combined $5.3 trillion from 2025 through 2030 on technology and data centers, according to Goldman Sachs¹⁵. S&P Global estimated that Alphabet, Amazon, and Microsoft alone projected about $495 billion of 2026 capex, much of it for AI-heavy technical infrastructure, up 61% from 2025 and six times the 2020 level, according to S&P Global¹⁶. Microsoft told investors it expected to remain capacity-constrained at least through 2026 even while bringing GPU, CPU, and storage capacity online faster, according to its FY2026 third-quarter earnings call¹⁷.

The market has noticed. Axios reported in February that the four major hyperscalers lost a combined $1 trillion in market value after their latest earnings releases as investors worried about AI overbuild, according to Axios¹⁸. Axios also reported in June that Alphabet, Amazon, Meta, Microsoft, and Oracle had raised $255.34 billion through equity and debt in 2026, more than twice as much as in all of the prior year, according to Axios¹⁹. That is not proof of a bubble. It is proof that the AI trade has become capital intensive in a way the first ChatGPT-era narrative obscured.

The best bullish reply is that demand is real. Enterprises are adopting AI, cloud revenue is growing, and inference, the day-to-day running of models after training, should become cheaper as chips improve and smaller models handle more tasks. S&P Global Ratings said the AI spending glut is beginning to deliver through strong first-quarter 2026 results at Alphabet, Meta, Microsoft, and Amazon, and projected free operating cash flow for most hyperscalers to recover beyond 2027 and 2028 as investments mature, according to S&P Global Ratings²⁰. That is the strongest case against doom.

But it still concedes my main point. The question is not whether AI has demand. It does. The question is whether the market priced an almost software-like expansion path for something that now looks like a regulated, energy-hungry industrial buildout. I think it did. Frontier AI is starting to resemble telecom, cloud, defense technology, and power infrastructure all at once: high demand, high strategic value, high fixed costs, and heavy political oversight.

So no, I do not think the AI boom is fake. I think it has become harder. The winners will not simply be the companies with the cleverest models. They will be the companies with approved users, reliable memory supply, enough GPUs, cheap power, patient capital, and compliance systems that can survive a government deciding that a model is not merely software but a national-security asset. The bottleneck has moved from intelligence to access. That is a very different industry from the one investors thought they were buying two years ago.