The Trump White House published its highly-anticipated AI Action Plan this morning. It is a 20-plus page document that anchors America’s national point of view on AI along three pillars – accelerating innovation, building more infrastructure, pushing AI diplomacy and security around the world – with recommendations on how each pillar ought to be effectuated.
It is quite a dense document and has clearly gone through a very intense interagency process to produce it. Most of the recommendations name check several departments and relevant bureaus within those departments for coordination and execution. It is a notable departure, and a welcome one, from most of the executive orders that have come out of the White House so far, many of which are quite off-the-cuff.
Depending on where you are on the AI debate spectrum, there is a lot to like or dislike about this action plan. Having thoroughly read this document once through (I will likely read it again after more reflection), here are my own off-the-cuff share of the top five most interesting things that stood out to me. (I’m using the word "interesting" neutrally here to highlight sections where when I read them, I literally went “Huh! Interesting!")
Open Source AI with GPU as Commodity
A prominent section of the plan is encouraging more open source and open weight AI coming out of the US. This “open source AI deficit”, especially vis-a-vis China, has become very pronounced over the last few months. According to Artificial Analysis, the top four open source AI models are all from Chinese labs. (It was three just a few weeks ago. Then Kimi K2 was released and is gaining traction.) This plan is putting forward recommendations to deal with this gap in open source AI.
The one recommendation that I’m most intrigued by is solving the lack of GPU availability for university labs that are both capable of doing frontier model development and willing to open source that research by treating it like a commodities marketplace.
The lack of large GPU clusters available for any organization other than the well-funded labs and large tech companies is a major problem. It is also a tough nut to crack because the builders of these GPU infrastructures, e.g. the neoclouds like CoreWeave or Crusoe, would only build out large clusters if they have at least a three-year commitment from one of these large labs, at which point they take out a mountain of debt from the capital market to fund the purchase of GPUs and energy, as well as constructing the site. When the site is done and online, there is almost never any capacity left for other organizations that are without the deep pocket or the long-term contractual commitment ability, i.e. university labs, who are struggling with dwindling funding from the government already.
To the extent that GPU capacity can be treated as a commodity, how “spot and forward markets” mechanism works in this context to "accelerate the maturation of a healthy financial market for compute" is a black box but an intriguing one. I can see a scenario where the various government agencies work with, for example, Nvidia's newly-launched DGX Cloud Lepton GPU marketplace to pool GPU capacity from various neocloud providers to make a meaningfully large cluster, then offer that at a preferential spot price for leading university AI labs, who are doing leading open source AI research. We shall see if that, or something similar, is what will happen to make this recommendation real.
American AI in a Box
The “International AI Diplomacy and Security” pillar led with aggressively exporting American AI to allied countries by having the government and the private sector work together to come up with “full-stack AI export packages” to push abroad. This recommendation basically amounts to an “American AI in a box” sales strategy for Nvidia, AMD, OpenAI, Oracle, etc.
Reading this section made me chuckle as this is more or less Huawei’s “AI in a box” go-to-market strategy, by packaging its hardware with leading Chinese models, like Zhipu AI, to present an all-in-one solution. While Huawei’s push may or may not receive explicit endorsement and support from the Chinese government, the Trump administration is clearly more than happy to take up the “pre-sales” role for the American AI tech stack around the world. This imitation is not so much problematic as it is a logical extension of America’s global technology strategy – if the overarching goal is to counter the Huawei stack, the only way would be to compete head-on.
When I was looking at the evolving “cloud war” between American hyperscalers and Chinese hyperscalers a couple of years ago during the Biden administration, I mused about whether foreign service officers from the State Department would start doing pre-sales for the American cloud companies. Looks like that is indeed about to happen during Trump’s reign.
More importantly, this concerted effort to export American AI will likely be factored into all the ongoing bilateral trade and tariff negotiations. Products we will push other countries to buy to reduce trade deficits will now include not just soybeans, oil, SUVs, and Boeing airplanes, but also American AI in a box!
GPU Tracking to Reduce Smuggling
One of the most frustrating policy debates in the last couple of years is whether and how many GPUs have gone into China via smuggling to circumvent export control. While there are some good reports that indicate some scale of smuggling is occurring, Nvidia is still pushing back with full-throated denials, while people who claim to know the scale of the smuggling won’t show their work.
Meanwhile, there are tracking and monitoring systems that can be placed to bring more technology and rigor to this debate. Looks like the administration is finally taking this route to close this loophole (or at least know with more precision when and how much the loophole is being exploited).
Using location verification features seems like the obvious approach to strengthening chip export control, which will continue on advanced AI systems for as far as the eye can see. Interestingly, this does not address the cloud service rental loophole, where a large and perfectly legal cluster of Blackwell systems can be installed in, say, Malaysia and be rented by any number of companies or organizations that may be doing things that the US government deems objectionable. Restrictions on cloud services do not appear anywhere in the action plan based on my reading.
Nevertheless, leveraging these location verification capabilities will at least bring more clarity, integrity, and less he-said-she-said to the GPU smuggling debate.
Sub-Systems Export Control on Semiconductor Manufacturing
Continuing on the export control thread, another recommendation that caught my eyes is the AI action plan potentially going deeper on restrictions of semiconductor manufacturing equipment to include not the equipment themselves, but the equipment and components that can be used to make those equipment, or “sub-systems”.
This is an aggressive extension of the export control regime. Currently, almost all the advanced semiconductor manufacturing equipment is already barred. This recommendation is looking to further slow down China’s indigenous semiconductor manufacturing efforts, by expanding the scope and making the acquisition of the parts and components of building one’s own lithography machine, for example, harder. If implemented, it will be quite an escalation of the tech cold war vis-a-vis China and goes counter to the current wind of increasingly constructive stance between the two countries on trade, with Sweden hosting round 3 of the negotiation next week.
This is, after all, only a “recommended policy action”, not policy, so we shall see what comes of it next.
“Try-First” Culture
The last interesting element I want to highlight is in the “Encouraging AI Adoption” section, where the Federal government may take on a more proactive role to push companies and enterprises across industries, but especially in regulated industries, to “try” AI more willingly.
The subsequent recommendations outline multiple scenarios, where agencies from the FDA and SEC to the Department of Defense and the Intelligence Community will lead on encouraging industries that are under its regulatory jurisdiction to test, deploy, and experiment with AI services more aggressively.
If this happens – a significant shift from “regulate first” to “try first” – the obvious beneficiaries would be all the companies, big and small, that are building AI tools and services for different verticals. But the more notable part is the tone of this section, promoting a “try-first” culture led by the Federal government, which is quintessential Silicon Valley. “Try-first”, or variations of this ethos (e.g. “move fast, break things”), runs through the spirit of almost everything that happens in Silicon Valley, from the way startups get built and funded, to how people live their normal lives.
It looks like the Silicon Valley ethos is steadily taking over the White House on all things AI and technology.