GPUs, Passports, and the Price of Progress: Why 2025 Export Controls Feel Like 1973 All Over Again

They called oil the lifeblood of the twentieth‑century economy. In 2025 the fluid that keeps things moving is electricity coursing through a slab of silicon the size of a hard‑cover novel. Whoever controls the flow of H100s and their even‑hungrier cousins controls who gets to build the next foundation model, and by extension, who gets to write the safety rules.

As a political science student with dual U.S. and Canadian citizenship, I have long followed the intersection of international policy and tech infrastructure. So when the outgoing Biden Administration’s U.S. Commerce Department dropped its Interim Final Rule in January, I paid close attention. The document expanded advanced-GPU restrictions from a China-only leash to a 120-jurisdiction no-fly list; it felt less like a surprise and more like a plot point I had been anticipating for months. Within hours Nvidia warned investors it might take a $5.5 billion hit because chips already on freighters could no longer be delivered.

That same week my go‑to cloud spot instance in Montréal jumped from about C$3.80 to C$10.70 per H100 hour. No new hardware had vanished from the data‑center rack; only the legal latitude to turn it on had changed. If you are a grad student budgeting compute like a household ration book, those few extra dollars per hour mean delaying a paper, shrinking a dataset, or begging another lab for time.

We Have Been Here Before

Scroll back five decades. On 6 October 1973, during the Yom Kippur War, a coalition of Arab nations led by Egypt and Syria launched, a surprise attack on Israel on Judaism’s holiest day. As the fighting intensified, the United States organised an emergency “Operation Nickel Grass” air‑bridge to resupply Israel with ammunition and hardware, a mission documented by the U.S. Air Force Historical Support Division.

On 17 October 1973, the Organization of Arab Petroleum Exporting Countries (OAPEC), Saudi Arabia, Kuwait, Libya, and others, announced that they would cut production and embargo oil shipments to the United States and the Netherlands in retaliation for Western support of Israel, according to the U.S. State Department Office of the Historian. Crude prices leapt from roughly $3 to more than $11 a barrel within three months, a shock traced by the Federal Reserve History, ushering in a wave of inflation, unemployment, and economic stagnation across the West.

The 1973 OPEC embargo was not only about energy, it was a calculated use of economic power to influence Western foreign policy, particularly in relation to the Middle East. By targeting countries that supported Israel during the Yom Kippur War, the embargo exposed how geopolitical alliances could be punished through trade controls. It revealed the vulnerabilities of an interconnected global economy where energy supply became a proxy for political leverage. The decision by oil-producing nations to tie resource access to diplomatic alignment set a precedent for the kind of strategic decoupling we are now seeing in the AI and compute space. As historian Avraham Sela has noted, the war and its aftermath reshaped Israeli defense doctrine and pushed American foreign policy toward deeper strategic entanglement with Israel.

Fast forward to today. The tools of economic leverage are different, but the logic is hauntingly familiar. We no longer ration oil and LNG, we gatekeep GPUs. This embargo does not involve tankers but export licences. The choke-point is not the Strait of Hormuz, but a box on a customs form that determines whether an H100 can leave San Jose without an escort of lawyers. The battlefield is no longer the Sinai, but the server racks of data centers. And once again, scarcity is not natural, it is imposed.

• January 2025: Washington widens the net beyond China, formally placing dozens of Middle‑East and African nations on the naughty list, and EU lawmakers protest a “two‑tier Europe” that pits license‑free members against quota‑bound neighbours.
• April 2025: The newer H20 chip is added to the restricted roster. Companies scramble for work‑around SKUs.
• Days later Jensen Huang hops a flight to Beijing, attempting diplomacy with customers who now see every Nvidia roadmap slide as a maybe.

The embargo’s defenders frame it as a necessary brake on autonomous weapons and influence campaigns. Fair. But brakes work best when the whole car shares one pedal. The moment Washington squeezed, Huawei doubled production of its Ascend 910C accelerator and announced plans to mass‑produce an even beefier variant by year’s end.

Domestic startups felt the tail‑winds too. A SemiAnalysis deep‑dive dubbed it “Tariff Armageddon,” with venture capital pouring into any company that could promise a licence‑free tensor core. Meanwhile, chip‑broker headlines got weirder: Malaysia recorded a 366 percent spike in high‑end GPU imports, a surge linked to grey‑market resellers funneling boards onward to embargoed labs.

Winners, Losers, and the Rest of Us

If you already sit on thousands of H100s (hello, Anthropic, Meta AI, OpenAI)  you are laughing. You can train the next frontier model while charging rivals a premium for inference. If you are MILA or EleutherAI, you are suddenly priced out of experiments that felt routine a year ago. Even the big‑cloud bargain hunters are hurting; analysts warn of a “tumultuous era of GPU cost volatility” likely to last through the election cycle.

China is not exactly standing still. Two days after the H20 notice, Huawei unveiled CloudMatrix 384, an Ascend‑powered rack that slots into existing data centers and, in theory, sidesteps U.S. rules entirely. If compute really is the new oil, Beijing just started drilling its own wells.

The Ethical Knot

Export controls are a legitimate instrument of national policy, especially when adversarial regimes seek to develop military-grade autonomous systems. The United States has valid reasons to limit who gets access to advanced U.S.-designed hardware. But the current structure of enforcement involving broad and blunt policy risks collateral damage to the very institutions driving responsible AI development. Three questions keep me up at night:

1. Does scarcity push research underground? Grey-market GPUs are already swapping hands at prices that would make a Scalper-era PS5 blush. When access becomes privilege, transparency suffers. History says science conducted in black boxes rarely serves the public good.
2. Who gets to audit frontier systems when only cash-flush labs can run them? Open-source safety research flourishes under diverse scrutiny. When only a handful of well-funded players can run meaningful experiments, we get less oversight, not more.
3. Will parallel tech stacks splinter global safety norms? A fractured compute landscape leads to competing AI architectures, separate optimization paths, and ultimately inconsistent safety standards. If open tooling cannot cross borders, neither can the governance frameworks we need to keep frontier systems in check.

The West has a chance to lead not just with control, but with coordination. But that means recognizing that open research is a strategic asset, and not a liability.

A Better Way, Maybe

What would a 2025 “energy treaty” for compute look like? My sketch is simple, think cap‑and‑trade for FLOPs:

• Transparent ledgers. Every advanced‑node chip ships with a public identifier tied to a compute registry. Regulators see usage, researchers see availability.
• FLOP‑tier licences. Less paperwork for projects under, say, 100 PFLOPs. Bigger jobs require disclosure, not outright bans.
• Efficiency incentives. Grants and tax credits for work that delivers safety or performance gains per watt.

It is imperfect, but so was OPEC. Over time the cartel model gave way to futures markets, strategic reserves, and eventually the green‑energy transition. Compute can follow a similar arc if we start replacing fear with data.

Why Hugging Face Readers Should Care

I am not a machine learning engineer by trade. My background is in political science, and my daily tools are more likely to be policy briefs than PyTorch notebooks. But that’s exactly why this conversation matters. Export controls on GPUs are not just a supply-chain hiccup, they are a form of international regulation that will define who gets to participate in the future of AI.

Hugging Face readers sit at a critical junction. You build the infrastructure, run the benchmarks, and set the tone for what responsible AI development looks like in practice. The embargoes, licensing rules, and geopolitical tensions shaping access to compute will ripple directly into your workflows, your collaborators, and your ability to publish reproducible results.

Here is what we can do right now:

• Contribute price data to a public Compute Ledger so that legislators and trade officials can see the real-world impact of their decisions.
• Document and share parameter-efficient techniques: **QLoRA, Mix-of-Experts, and beyond that help level the playing field for researchers outside the hyperscaler class.
• Foster cross-sector dialogue. If we want good regulation, policymakers need your insights. And if you want good policy, they need your participation.

Our uncertain moment in AI is not just technical, it is structural. And what we build together, across roles and disciplines, will shape AI’s trajectory for years to come.

Closing Thought

My Zaide still talks about lining up for expensive gas in the 1970s. He was living in Montreal during the OPEC embargo and remembers how fuel was rationed, tempers flared at gas stations, and even city transit was disrupted. As a young man working in a white collar job, he had to adjust his hours around fuel availability, and he said it felt like the world was being punished for standing by its allies.

That personal story stuck with me. It was the first time I understood that geopolitics wasn’t just a game played in Washington or Riyadh; it shaped the choices made at corner stores, classrooms, and community centers.

I never thought I would feel a digital version of that scarcity while queueing a training job, but here we are, refreshing the cloud dashboard, praying the price bar slides back under five bucks. If we do not address the ethics of compute control now, the next generation of researchers will inherit a world where curiosity itself requires an export licence.

And that would be a tragedy no amount of silicon can solve.