Making a graphics card for gamers is one thing, but manufacturing a rackscale supercomputer with over 600,000 components that burns 120 kilowatts of power, that has over 5,000 copper cables for an all-to-all interconnect mesh for 72 dual-chip compute engines, and that weighs over 3,000 pounds is another thing entirely.
So when a trade war is escalated between the United States and China, such a behemoth as the Nvidia DGX GB200 NVL72 supercomputer can’t help but get caught in the crossfire when it is not made in the United States but most customers who buy most of the machines are.
And that is why Nvidia is working to onshore the manufacturing of the chips and other components that go into these systems as well as the manufacturing of those systems to the United States. What Nvidia is not doing is starting up factories to be a chip and system manufacturer itself, regardless of many of the statements in the newspapers and trade press that seem to indicate this.
Nvidia doesn’t make servers and switches – it pays others to do it to its specifications, just like the hyperscalers and cloud builders do. It does have a supply chain that includes hundreds of companies that are involved in assembling the components and systems, and it does put together the initial versions of its systems in its own labs so it understands the process. But once it does that, it releases the components and the designs to others to assemble and sell.
So if Nvidia wants to have its AI supercomputers made in America, it needs to convince its partners to make these components in America and assemble them there, too. And this is what Nvidia, in fact, announced it was doing on Monday. Even though the subhead in that announcement from Nvidia itself says “NVIDIA opens first US factories,” we do not believe this is true.
We do believe that Nvidia is commissioning its partners to build the factories and is committing to using these facilities to make indigenous machinery on behalf of Nvidia. There are other stories that claim “Nvidia to Make AI Supercomputers Entirely in US,” which is also almost certainly not true now and will not be true for the next few years if ever. That is because there will be export controls and tariffs everywhere, and it is far more likely that companies like Nvidia will make stuff have stuff made regionally to supply different regions to dodge tariffs as much as possible.
The manufacturing announcement by Nvidia follows the tariff chaos that has been coming out of the White House for the past two months and relatively hot on the heels of a rumored April 4 dinner meeting between Nvidia co-founder and chief executive officer Jensen Huang and President Trump at the latter’s Mar-a-lago resort. The H20 GPU accelerator, a cut down version of the “Hopper” H100 GPU that was permitted for sale without a license from the US Department of Commerce, was reportedly a big topic of conversation between Trump and Huang. Which stands to reason given that China should be a much bigger part of Nvidia’s business than it is. Moreover, Nvidia can charge full price – and perhaps even a premium – for a crippled H20 compared to an H100 or H200, and thus the H20 should be a more profitable device to sell.
The HGX H20 system board that has been available in China since late 2023 has eight H20 GPUs, which have 96 GB of HBM3 stacked DRAM and 4 TB/sec of bandwidth on that memory. The most capacious H100 had 96GB of HBM3 memory with 4 TB/sec of bandwidth and the H200 has 141 GB of HBM3E memory with 4.8 TB/sec of bandwidth. So the H20 has some memory crimping, but not too much.
On the compute side, the H20 is rated at 1 teraflops at FP64 precision (compared to 33.5 teraflops on the real H200), a huge governor on performance, and 44 teraflops at FP32 precision on its vector-style CUDA cores, which compares better to the 67 teraflops for the H200. On the tensor cores, which are designed for matrix math, the H20 is rated at 74 teraflops on FP32 precision (using TF32 matrices) compared to 989 teraflops on the H200, which is a very big cut indeed. FP16 performance on the H20 is 148 teraflops, a hell of a lot less than the 1,979 teraflops on the H200, and FP8 performance (which is important for both training and inference) is only 296 teraflops on the H20 compared to 3,958 teraflops on the H200.
Again, you see why China’s DeepSeek decided to rewrite the script on AI training and inference. DeepSeek used a few thousand predecessors to the H20, called the H800, to train its eponymous model, instead of tens of thousands of real H100s or H200s. The H800 SXM version, which could be linked by NVSwitch, was capped at 1 teraflops at FP64 precision and nothing much else changed compared to the H100. The PCI-Express version of the H800 had some of its Hopper cores de-activated, so its performance went down, and its memory bandwidth was also crimped by 39 percent. (We did a deep dive on these crimped Hopper GPUs and compared them to Huawei Technology’s HiSilicon division Ascend GPUs back in August 2024.)
There were rumors that the Trump administration might let Nvidia keep selling the H20 into China without a license, but a lot of people didn’t trust that apparently given the rumors that Chinese companies acquired $16 billion in H20 chips in the first quarter of 2025.
That stay of execution for the H20 was very temporary, and so Nvidia wanted to get out in front of that news and on Monday announced out of the blue that it was working with partners to bring AI supercomputer manufacturing to the United States.
Technically, it is already here, since Supermicro, Hewlett Packard Enterprise, and Dell all have indigenous manufacturing of AI supercomputers. The trouble is that Nvidia’s hyperscaler and cloud builder customers all use offshore original design manufacturers like Foxconn, Wistron, Quanta, Inventec, and a handful of others to build their server, storage, and switching systems.
Unless they move some of their operations to the United States, machines imported from China, India, Malaysia, and elsewhere will be subject to various levels of tariffs – with China’s expected to be very high indeed.
During a question and answer session at the GPU Technical Conference in March, Huang hinted that it had onshoring plans in the works.
“We have a really agile network of suppliers,” Huang explained when asked about the impact of tariffs on the manufacturing of its systems. “They are not just in Taiwan or just in Mexico or just in Vietnam or just in – they are distributed in a lot of places, and it depends on what gets built, what gets purchased, where is the final destination of the product. And so there’s just a whole lot of equations involved. So I think in the near term, based on what we know, we’re not expecting significant impact to our outlook and our financials. Long term, we want to retain the agility, but add a very significant part of our agility, which is onshore manufacturing. So if I could – the simplest way to think about that is our agility is terrific today, missing onshore manufacturing. If we add onshore manufacturing by the end of this year, we should be quite good. We should be pretty good shape.”
None of that, of course, will fix the issue of export controls being put on GPUs and possibly other components from going into China, Russia, and other countries. Nvidia announced late last night, after people had been chewing on its US manufacturing plan for a while, that the Trump administration informed it on April 9 that the H20 GPU or any device that could achieve the same memory or interconnect bandwidth would be subject to export controls to China, including Hong Kong and Macau as well as the so-called D:5 countries designated by the Department of Commerce. These include: Afghanistan, Belarus, Burma, Cambodia, Central African Republic, Cuba, Democratic Republic of the Congo, Eritrea, Haiti, Iran, Iraq, Lebanon, Libya, Nicaragua, North Korea, Russia, Somalia, South Sudan, Sudan, Syria, Venezuela, and Zimbabwe. (China is a D:5 country as well.)
To that end, Nvidia announced it will be taking a $5.5 billion charge associated with H20 inventory and purchase commitments in its first quarter, which ends on April 27. And, having got the same notifications about its Instinct MI308 crimped GPU accelerators, AMD announced that the new export control from the Commerce Department, affecting the same countries, could result in it taking an $800 million charge, presumably spread over its first and second quarters of 2025.
And thus, all of the flag waving about Nvidia’s onshoring, and the title of its announcement seemed to be designed for people who need a point hammered home: “NVIDIA to Manufacture American-Made AI Supercomputers in US for First Time”
Here is what is really happening, as far as we can tell. Taiwan Semiconductor Manufacturing Co, which cut its own $100 billion manufacturing investment deal with the Trump administration back in March, boosting its investments in labs and fabs to $165 billion, has started producing more modern “Blackwell” GPUs in its foundries in Phoenix, Arizona using N4 4 nanometer processes. It is unclear if this was always the plan to do it right now, but someone’s chips need to be printed to ramp up production, so why not Nvidia?
Nvidia is working with Amkor Technology and Silicon Precision Industries to build out chip packaging and testing operations in Phoenix – something they also might have done on their own given that TSMC’s customers are also their biggest customers. Amkor was founded in 1968 outside of Philadelphia, with strong roots also in Korea (hence its name, America-Korea) and moved to Arizona (first in Chandler, near Intel’s fabs, in 2005 and then later moved to Tempe, where Arizona State University is located. Amkor had $6.3 billion in sales in 2023, the last full year for which figures are available. Silicon Precision, also known as SPIL, was founded in Taiwan in 1984, and had revenues of $3.64 billion in 2023.
For all we know, Nvidia is just promising Amkor and SPIL some future packaging and testing business and some Omniverse factory simulation and modeling so they can hit the ground running with packaging operations for Nvidia chippery.
The real news is that Nvidia is working with Foxconn to build a factory in Houston, Texas and Wistron to build a factory in Dallas, Texas, and while the statement says “NVIDIA is building supercomputer manufacturing plants in Texas,” we do not believe that. We believe that Nvidia is encouraging Foxconn and Wistron to do this, and may be providing financial and technical assistance to get these factories built, but we have a hard time that believing Nvidia is actually going to own these factories and lease them out to Foxconn and Wistron. (We reached out to Nvidia for comment, and were told it is not answering any questions beyond its statement.)
But, maybe we are wrong. Maybe Nvidia really is getting in the factory business to build AI factories.
This sentence certainly gives that impression: “The company will utilize its advanced AI, robotics and digital twin technologies to design and operate the facilities, including NVIDIA Omniverse to create digital twins of factories and NVIDIA Isaac GR00T to build robots to automate manufacturing.”
I guess we will know when we see the logos outside of the factories. And we won’t be surprised if the one in Houston says “Nvidia/Foxconn” and the one in Dallas says “Nvidia/Wistron” with equal sized logos – no matter what the financial arrangements are that got these factories built.
Just like TSMC had to sacrifice some profits to move manufacturing to the United States, we believe that Nvidia will do so as well. And as Nvidia’s financials show – the company had $130.5 billion in sales and net income of $72.88 billion, a whopping 55.8 percent of revenues, in fiscal 2025 – it can do anything it damned well wants to, including building its own real factories. Anything, of course, except not obey the laws in the countries in which it does business.
Nvidia says that it will take twelve to fifteen months to get these factories online and that “within the next four years” it will “produce up to half a trillion dollars of AI infrastructure in the United States through partnerships with TSMC, Foxconn, Wistron, Amkor and SPIL.”
In our financial model, we have Nvidia’s Datacenter division growing at a slower and slower pace, but against very large numbers, and pulling in $1,191.9 billion. Heaven only knows what will happen, but that $500 billion is somewhere between 40 percent and 45 percent of Nvidia’s datacenter revenues. If this sounds low, it is because the hyperscalers and cloud builders in the US are all designing their own AI accelerators and there is no reason to believe that in the long run the homegrown AI compute engine share won’t be 50 percent. The question we have is where is Nvidia going to sell the other $700 billion or so in accelerators with export controls on D:5 nations?
Perhaps enterprises in the Americas, Europe, and Japan that are not on the D:5 list are supposed to pick up that slack, and given large enough inference workloads using chain of thought models, this might be a good bet to make the numbers work out.
As we are fond of saying, the only precise way to predict the future is to live it. So we shall see.
