In just a few weeks, three seemingly unrelated news stories have painted the same picture: the AI economy is no longer explained solely by software, models, and prompts. It’s primarily driven by industrial capacity. With electricity, critical minerals, and real access to GPUs that today function as a currency of technological power.
The first signal comes from China and involves a logistical thriller: tech companies are considering turning to the black market to acquire NVIDIA H200 after units get held up at the border, amid regulatory and political uncertainty.
The second signal is from OpenAI itself. In a post signed by CFO Sarah Friar, the company states there’s a direct relationship between available computing capacity and revenue, providing concrete figures: computational power increased from 0.2 GW in 2023 to approximately 1.9 GW in 2025, while annual recurring revenue (ARR) grew from $2 billion to more than $20 billion in the same period.
The third, less publicized but equally critical, points to the physical “bottleneck”: the United States aims to manufacture domestically an essential input for electric motors, robotics, and defense — rare earth magnets. Texas-based Noveon Magnetics announced a financing round of $215 million to expand operations and strengthen domestic supply, at a time when external dependency is viewed as a strategic risk.
China and the “urgent chip”: when borders become the bottleneck
According to a report cited by international media, some Chinese companies are considering buying H200 outside official channels, due to the difficulty in completing imports, which have become effectively “hyper-sensitive.” The most notable symptom is price markup: a server with 8 H200 is offered around 2.3 million yuan (about $330,403), roughly 50% above the list price.
The core issue is not just price. It’s time. When an AI team can’t wait for the next regulatory window, the temptation to “solve it by any means” grows. But that pathway introduces a broad spectrum of risks: from sanctions and contractual losses to an often unspoken but heavily weighted issue in regulated sectors — traceability and supply chain security. In critical hardware, origin matters immensely.
This episode aligns with another recent report: Reuters indicated that, by late 2025, purchases of advanced chips in China have tightened, requiring full upfront payment amid high friction and uncertainty. In other words, when access to compute is uncertain, the market becomes more expensive, less efficient, and more opaque.
OpenAI: “more compute, more revenue”… and a thesis to calm the market
Meanwhile, OpenAI has worked to establish a narrative of economic stability: if AI creates value and that value is monetized, then revenue growth should “follow” capacity growth. Friar frames it as a cycle: compute → better models → better products → more adoption → more revenue → more compute.
This thesis is powerful because it turns the massive infrastructure expenditure into a “measurable” investment. But it also carries an implicit message: compute is becoming the scarcer resource and, consequently, the factor that differentiates those who can scale from those stuck in demos.
On an industrial level, this type of message aims to normalize that AI no longer only competes for talent or algorithms but for physical capacity and supply contracts. This directly links to what’s happening with GPUs in China and rare earths in the U.S.: the “model economy” depends on an infrastructure that isn’t improvisable.
Rare earth magnets: the invisible material powering robotics and defense
Public discourse often focuses on chips but pays less attention to what comes after. High-performance permanent magnets are essential for compact, efficient motors (electric vehicles, drones, industrial automation) and numerous defense systems. Noveon Magnetics has secured $215 million to expand its capacity in the U.S., seen as a move within a broader industrial resilience strategy.
The uncomfortable truth is: although capital is arriving, scaling production takes years, and external dependence doesn’t vanish “from one quarter to the next.” These investments are not about trying to dominate the global market overnight but about creating redundancy to prevent export restrictions from halting entire sectors.
Asimov as a compass: classic rules for modern AI and robotics dilemmas
In a context where hardware has returned to geopolitics, it’s useful to revisit a cultural framework that surprisingly remains relevant: Asimov’s Laws of Robotics. Applied to physical robotics and, with nuances, to modern AI, they could be reinterpreted as follows:
- “Do no harm to humans”
In robotics: functional safety, redundant sensors, emergency stops.
In AI: preventing harm not only physically but also financial, reputational, and social (fraud, disinformation, opaque automated decisions). - “Obey orders”
In robotics: conditioned obedience for safety.
In AI: aligning with user intent, yes, but subordinate to law, internal policies, and compliance (especially in healthcare, finance, and public administration). - “Protect oneself”
In robotics: self-protection to avoid catastrophic failures.
In AI: cybersecurity, robustness against manipulation, access control, traceability of changes, and the ability to degrade service without collapsing.
The key point is that by 2026, the issue isn’t just “what can AI do,” but who controls the supply enabling it and what economic incentives are in place amid scarcity. If borders delay chips, black markets surge, and compute becomes a strategic asset, then security rules cease to be mere philosophy and become trust infrastructure.
Frequently Asked Questions
What is an NVIDIA H200 and why is it so critical for AI projects?
It’s a high-performance GPU geared toward AI training and inference workloads. Its demand is increasing because it accelerates the massive calculations needed for large models and agent systems.
What risks does purchasing GPUs on the black market pose for a company?
Beyond legal and regulatory risks, there are issues related to traceability, fraud, warranty problems, contractual sanctions, and exposure to supply chain attacks in critical infrastructure.
Why does OpenAI insist that revenue growth is tied to compute capacity?
Because more capacity allows serving more users and workloads, enabling monetization through subscriptions and API usage, following a directly correlated curve.
What’s the connection between rare earth magnets, AI, and robotics?
Robotics and automation depend on compact, efficient electric motors, which often require high-performance permanent magnets. Without that supply, the industrial chain slows down even if chips are available.
via: scmp