China Export Controls Whack AMD Datacenter GPU Business

As far as we can tell, the export controls on crippled GPU compute engines announced by the US Department of Commerce back in April have had a disproportionately hard impact on AMD compared to Nvidia, as far as we can tell. These controls did not affect AMD’s first quarter financial results announced this week, but the controls have removed $1.5 billion of revenue from AMD’s 2025 and $700 million of that will come out of the hide of its second quarter.

Ooof!

Let’s do some math. AMD said it would be taking an $800 million writedown for the MI308X inventory for China that cannot be sold into the country, and said that this inventory would have driven $1.5 billion in sales. We had been originally projecting that AMD would have $6.83 billion in sales worldwide for Instinct GPUs in 2025, so that is 22 percent of its sales that disappeared and – more importantly – a much larger share of its operating income from datacenter GPUs that is gone. (Those US-based hyperscalers and cloud builders get steep volume discounts, at least compared to their more desperate counterparts in China.)

The crimped GPUs from Nvidia and AMD sell for relatively high prices compared to the real ones that are not crimped. It is a great business if you can get it, and Chinese customers have to buy 2X or 3X the number of devices and scale out their GPU clusters to get a certain level of calculation or memory bandwidth, as we showed here in this comparison of Nvidia GPUs and “DaVinci” Ascend GPUs from the HiSilicon chip arm of Huawei Technology.

Nvidia said at the time that it would have to take a $5.5 billion writedown of inventory on H20 and maybe some H800 crippled GPUs, and if the ratio of writedown to lost revenue is consistent between Nvidia and AMD devices, then this means Nvidia has lost around $10.3 billion in GPU revenue to China in its fiscal 2026 that ends next January. We were projecting that Nvidia would have $183.6 billion in datacenter sales in fiscal 2026. In fiscal 2025, which ended in January, $102.2 billion of the $115.2 billion of sales – that’s 88.7 percent of the total – was for datacenter compute as opposed to networking and software at Nvidia. Assuming the same ratio for fiscal 2026, that means about $162.9 billion in GPU sales would have come this year. The lost $10.3 billion that would have come from China for Nvidia this fiscal year represents a mere 6.3 percent of sales.

Nvidia can make that up with three big deals. The same cannot be said of AMD. At least not yet.

Which leads us to the next point. If Nvidia and AMD want to play it safe with Uncle Sam and not get burned by export controls, Nvidia could just sell “Ampere” A100 GPU engines in China and AMD could just sell “Antares” Instinct MI250X GPU engines in China and keep it simple. Literally only allow China to buy the past.

And if they do, Chinese researchers will just have Beijing fire up two more coal fired plants for each supercomputer center it has and then build advanced nuclear weapons, airframes, and missiles anyway using hundreds of thousands of ancient GPUs. And if they don’t use Ampere and Antares engines, they will use HiSilicon Ascend GPUs that are roughly equivalent to the A100s and MI250s and it all comes to the same.

Either way, China will advance its HPC and AI initiatives. This is inevitable. The only thing we in the West can do is out-innovate China if this really is a cultural and economic war of East versus West. We can trade, or stick to our own domains, or find a happy medium. Out-innovating China is hard to do, unless you have a lot of Middle East oil money to support the innovation. And in case you hadn’t noticed, this is exactly how a lot of things are being funded: Hedge funds and private equity are teaming up with the United Arab Emirates and Saudi Arabia, who are swimming in trillions of dollars of oil money looking to diversify to fund all kinds of things, like the expansion of Cerebras Systems and the launch of OpenAI’s Stargate Project effort, just to name two.

In this game of Risk, it is important to not have the Middle East team up with China. And it is also important for America and Europe to develop their own energy sources to keep the Middle East honest and to present a united front to Russia and China. It is also important not to antagonize Japan too much, which has over $1.13 trillion in US Treasury bills; the West does not need a Japan that thinks it needs nuclear weapons, and the West needs Japan to be a bulwark of capitalism and democracy in the East. By the way, China has only $784 billion in T-bills – it used to be many trillions of dollars – and has lost some of its leverage over the US because of this. And, ultimately, today America has many more bullets (probably on the order of 20 billion to 50 billion, depending on how much you think is stockpiled) than the population of Earth (call it 8.2 billion people), which is its own kind of dark leverage. Should it come to that. All we know is that we had better work to make World War III, The Non-Nuclear Option, not come to pass. If the Nuclear Option happens, we are all dead anyway, and frankly we have been living under that fear for eight decades now. The robots and cockroaches can have Earth if that happens. And if we get cheap and ubiquitous nuclear fusion power, that counterbalances the disruption that will be caused by AI and we can survive the experience and maybe even thrive.

We said maybe.

Happy Philosophical Phriday, everyone!

With all that out of the way, let’s talk about AMD’s financial results for the first quarter and the forecast for datacenter CPU and GPU spending.

In the March quarter, AMD posted sales of $7.44 billion, up 35.9 percent and down only 2.9 percent sequentially. (Thanks to GPU sales to China before the export controls kicked in, we think.) Net income rose by 5.8X to $709 million, bolstered by product mix and profits from both datacenter and client compute engine revenues. AMD exited the quarter with $7.31 billion in cash and investments, which gives it plenty of maneuvering room as it tries to compete with Nvidia, Intel, and the “hypercloud” collective on various fronts.

AMD booked $3.67 billion in revenues in its Data Center group in Q1, up 57.2 percent year on year but down 4.8 percent sequentially, which is what you expect from a Q4 to Q1 transition.

Operating income for the Data Center group was up 72.3 percent to $932 million, and represented 25.4 percent of revenues as AMD begins another investment cycle in its CPU and GPU businesses. (There might be some pricing pressure here, too, in CPUs, thanks to the Arm chips being developed by the hyperscalers and the cloud builders as well as from custom AI chips these companies are deploying.)

The datacenter and other businesses at AMD are almost perfectly balanced in Q1 2025: $3.67 billion for datacenter stuff and $3.76 billion for everything else.

Within the Data Center group, it is harder to get a sense of what is Epyc and what is Instinct. AMD does not drop precise enough clues to pick it apart, which is why the vertical red line is in the chart below, delineating between when we more confident (to the left of the line) and when we are less so (to the right of the line). We tossed in $45 million in Versal datacenter FPGA sales and $85 million in datacenter NICs and DPUs, and then we pulled out $1.16 billion in Instinct GPU sales, up 90 percent year on year but down 33.6 percent sequentially. That left $2.39 billion in Epyc CPU sales, up 46.6 percent year on year and up 20.3 percent sequentially.

The “Turin” Epyc 9005 processors launched in October 2024 are hitting their stride just as the server recession is over and it is clear that companies have to upgrade their old server fleets to free up power and space for AI systems.

Looking ahead in our model, based on the comments made by AMD chief executive officer Lisa Su and chief executive officer Jean Hu on the call with Wall Street analysts going over the Q1 2025 numbers, we think Epyc CPU sales are going to be larger than Instinct GPU sales for the remaining three quarters of the year. Instinct sales, in our model, will be up a tiny bit sequentially in Q2, and will run at around $1.9 billion a quarter as the MI355 ramp kicks in.

We had been projecting $7.61 billion in Instinct GPU sales in 2025, but now we think it will be more like $6.11 billion. We forecast that Epyc CPU sales will be around $8.52 billion. That will be 21 percent growth for the year for Epyc CPUs and 21.6 percent growth for Instinct GPUs. Without the China export controls, we think Instinct GPU sales would have been up 51.5 percent.