AI Chip Export Controls: Who panics, Who Profits?

In May 2026, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) tightened export controls on high-end AI chips once again. This time, the focus isn't just on GPUs themselves—it extends to AI accelerator cards, HBM high-bandwidth memory, and Chiplet interconnect technologies.

The moment the news broke, half the tech industry started combing through their supply chain inventories.

What's Actually Being Controlled?

Let's break down the key changes in the new rules:

• Compute density threshold lowered: A new "compute density per square millimeter" metric now captures a large number of AI inference chips that previously "slipped through the cracks"
• HBM memory added to entity list: Exporting HBM3 and above bandwidth memory now requires a license
• Chiplet packaging restrictions: The route of bypassing single-chip controls by stitching together small chips via Chiplet has been blocked

In plain language: Previously it was "you can't buy the best chips." Now it's "you can't use any method to piece together the best chips."

Who's Panicking?

Tier 1 anxiety: Cloud service providers. Top-tier cloud vendors' AI training clusters rely on NVIDIA H200/B200; upgraded controls mean longer lead times and price spikes for future deliveries. Rumors suggest a domestic cloud vendor has already activated "Plan B"—reallocating existing GPU priorities, with inference services getting top billing and training projects queuing.

Tier 2 anxiety: AI startups. The A100s that could still be sourced through gray channels last year are now impossible to find, even A800s. Teams doing AI art generation and video generation are being forced to pivot to domestic alternative solutions.

The unique opportunity: local AI hardware. As controls tighten, cloud-side AI compute becomes more expensive and scarce, which in turn makes "local deployment" a required answer rather than an optional one. Hardware like Kaihe, which comes pre-installed with open-source LLMs and works out of the box, precisely fills the gap for "I want to use AI

Who's Profiting?

Domestic GPU vendors. Orders for Huawei Ascend, Cambricon, and Biren AI chips have tripled in the past three months. But a Open-source model ecosystem. Tighter controls make open-source more attractive. Meta's Llama series, Alibaba's Qwen series, DeepSeek series—all run locally without depending on cloud APIs. This directly benefits all hardware vendors doing local AI deployments.

RISC-V architecture. Both x86 and ARM are affected by geopolitical tensions; RISC-V, as an open instruction set, has become a "safe-haven asset." AI chip startups are already working on RISC-V + NPU inference chip solutions.

Direct Impact on Regular Users

Here's what matters to everyone: Will AI tools get more expensive?

In the short term, yes. Cloud LLM API call costs will rise due to increasing compute costs—this is already happening. ChatGPT Plus increasing from $20 to $25 per month is just a signal.

But in the long term, controls are forcing local AI to mature faster, which might actually make AI cheaper—once you own a local AI machine, all model calls have zero marginal cost. No API fees, no token billing, no "peak-hour queuing."

Controls have always been a double-edged sword. They push up short-term costs, but also accelerate long-term alternatives. For anyone hesitating about "whether to get a local AI device," the answer is becoming clearer.

Cover image generated by Seedream 4.5 | Original content by nizwo.com