• The Midas Report
  • Posts
  • Design for augmentation, not automation with these lessons from Nvidia’s China compromise

Design for augmentation, not automation with these lessons from Nvidia’s China compromise

3 min read.

Author

Midas AI
August 12, 2025

Export controls, degraded chips, and a 15% revenue skim point to a future where capability is tiered and oversight is built in

The United States has begun issuing licenses for Nvidia to sell its H20 chips in China, and Nvidia and AMD agreed to send 15% of their China AI chip revenues to the U.S. government as part of the new arrangement, according to officials and people familiar with the deal. A US official described the payments as voluntary and structured to avoid being treated as an export tax, while reporting by the Financial Times framed the payment as a prerequisite for the licenses.

The legal and policy pushback was swift. Critics argued the Constitution forbids taxes on exports and said export controls should protect national security rather than raise revenue, calling the arrangement bribery or blackmail. The episode underscores a simple reality for AI builders. Performance is not the only constraint that matters. Policy, oversight, and trust set the ceiling on what you can ship and where.

Licenses and a 15 percent skim reshape incentives

Nvidia has already been designing to constraints. The H20 was introduced with China’s export market in mind after prior restrictions, and it lacks the power of Nvidia’s most advanced chips like the H100 and H200 that are better suited for training large language models. Yet the H20 remains state of the art in important ways, with memory capabilities that make it sophisticated and valuable.

A Chinese model, DeepSeek, was reported to have used H20 chips. China still accounted for 13% of Nvidia sales in 2024, and CFRA Research estimates Nvidia and AMD together could generate up to $35 billion annually from H20 and MI380 sales in China under the new regime, which would send roughly $5 billion to the U.S. government at a 15% rate. Even with downgraded capability, there is a large market for products that meet the letter and the spirit of the rules.

Capability tiering is now policy and product strategy

Washington is signaling that capability tiering is a feature of the landscape, not an exception. President Donald Trump said he might allow an unenhanced version of Nvidia’s next generation Blackwell platform into China and described the idea as taking 30% to 50% off. He called Blackwell super high end and said he would not make a deal for it today, while describing the H20 as an old chip that still has a market and does not advance China’s capabilities.

At the same time, independent experts argued H20 class chips still carry significant capability, reinforcing the point that regulatory line drawing is not the same as harmless. Nvidia’s public stance has been consistent. It will follow the rules the U.S. government sets and hopes export controls will let America compete in China and worldwide. The company warned that America cannot repeat 5G and lose telecommunication leadership, arguing the U.S. AI tech stack can become the global standard if it races.

The policy process itself is now a core part of product strategy. In April the administration halted H20 sales to China citing security risks, then announced a reversal in July. Licenses were granted on a Friday, and no shipments had been made as of reporting. A U.S. official said the payment structure helps maintain control over the export process while generating revenue. Others warned that easing restrictions and improvising payment mechanisms could erode trust in U.S. export controls, and one source suggested companies can pay for the right to undermine national security.

Treasury leadership even described Nvidia export controls as a negotiating chip in broader trade talks. For founders, the takeaway is not about geopolitics alone. It is that systems succeed when they respect constraints and make room for human and institutional oversight. If regulators can halt or reshape your go to market in a single quarter, you should assume your customers and operators will do the same when faced with black box automation that bypasses their judgment.

There is also a simple business lesson in the numbers. Nvidia reported billions in charges and lost revenue from export controls in the first quarter and projected a similar outcome in the second, even as its valuation touched roughly $4.5 trillion and shares rose modestly on the licensing news. The company has been preparing China specific variants at lower cost, has not shipped H20 to China for months, and is navigating scrutiny on every lever from model capability to revenue sharing. The profitable path runs through products that deliver value within hard boundaries. That means tiered offerings, explicit controls on capability, and designs that center the human and institutional operators who bear the risk.

Augmentation beats automation in environments defined by trust and accountability. The U.S. is allowing sales of certain chips again, but only with licensing, a revenue cut, and active debate over national security. Nvidia is positioning versions of its hardware that are useful without being unbounded. AI builders should do the same with software. Treat human cognition and institutional oversight as first class constraints. Build copilots that keep people in the loop, make capability legible, and degrade gracefully for sensitive contexts. The market is telling you there is ample demand for powerful but governed tools. The policy environment is telling you those are the only tools that scale.