Trump's trade war with China to cost AMD $1.5B in lost rev

But a multi-billion dollar contract with Oracle for a pile of Instincts and Epycs should take the edge off

AMD expects the Trump administration's newly implemented export controls on GPUs and AI accelerator sales to China to take a $1.5 billion byte out its 2025 revenues, executives revealed on a Tuesday earnings call with Wall Street.

Despite the setback, AMD CEO Lisa Su remains confident in the House of Zen's ability to grow Instinct GPU revenues by "strong double digits" in the full 2025 fiscal year as it looks to steal yet more share from rival Nvidia in the AI arena.

AMD learned sales of its China-spec MI308-series accelerators would be subject to new licensing agreements restricting their sale to China and other nations of concern late last month.

"While we face some headwinds from the dynamic macro and regulatory environment, including the recently announced export controls for instinct, MI308 shipments to China, we believe they are more than offset by the powerful tailwinds from our leadership product portfolio," Su said on the call while going over AMD's latest financial figures, the ones covering its first quarter of 2025.

Had it not been for the export controls, AMD's forecast for its second quarter of the year might have looked quite a bit rosier, climbing 9.4 percent sequentially and 47 percent year-over-year to an estimated $8.1 billion. Alas, the latest salvo in the US-China trade war will shave about $700 million off those anticipated Q2 revenues, which it now expects to come in at $7.4 billion give or take $300 million or so.

"Sequentially, we expect the client and gaining segment revenue to increase by double digit percentage; embedded segment revenue to be flat-ish; and we expect datacenter segment revenue to decrease due to the exclusion of MI308 revenue," CFO Jean Hu told analysts on the call.

While less than ideal, Hu emphasized that at the midpoint of the forecast, Q2 revenue growth would still top 27 percent compared to the same time last year.

Most of the remaining $800 million in revenues lost to unrealized MI308 GPU sales in China are expected to come out in Q3 with "very little in Q4," Su explained.

In addition to the expected total loss of $1.5 billion in revenues this fiscal year, the x86 giant had already announced in an SEC filing that it would take a charge of up to $800 million for "inventory, purchase commitments, and related reserves" - this most likely refers to stockpiled chips that it will not be able to sell. In its Q1 earnings statement [PDF] Tuesday, AMD reiterated this, saying it expected to take "approximately $800 million in charges for inventory and related reserves due to the new export controls" in Q2.

The situation could be worse. AMD's much-larger rival Nvidia recently revealed it'll book a $5.5 billion charge on account of H20 GPUs it can no longer sell to its Chinese partners.

AMD preps next-gen Instinct accelerators for early arrival

AMD expects its prospects to improve significantly in the second half as the US corp begins shipping its next generation of AI accelerators designed to compete directly with Nvidia's Blackwell generation of GPUs.

Initially expected late in 2025, AMD moved up the launch, presumably to get out ahead of Nvidia's newly announced Blackwell Ultra accelerators. As we previously reported, the MI355X promises to deliver peak floating point performance roughly equivalent to that of Nvidia's B200 while also delivering 50 percent more memory capacity at 288GB of HBM3e memory.

Oracle, Su said, will be among the first cloud providers to deploy a "multi-billion" cluster of the accelerators alongside its 5th-gen Epyc datacenter processors, and Pollara 400 smartNICs.

This higher capacity has allowed AMD to serve larger gen AI models like Meta's Llama 3.1 405B and OpenAI's GPT family of models using fewer resources than comparable Nvidia systems. However, Blackwell Ultra slated to ship later this year, Nvidia will close the gap with 288GB of memory and 50 percent higher dense 4-bit floating point performance.

Su also teased its upcoming MI400-based rack-scale systems, which Su suggested would "deliver leadership performance in both inferencing and training," when they arrive in 2026. Up to this point, AMD has lacked a rack-scale architecture to compete with Nvidia's GB200 and GB300 NVL-72 platforms.

We expect to get more information on both the MI355X and MI400-series accelerators at AMD's Advancing AI event next month.

Q1 Performance

As for the company's Q1 performance, AMD banked $709 million in profits – up significantly from the year-ago quarter's $123 million – on revenues which grew 36 percent year-on-year to $7.4 billion. This growth, Su said, was once again driven by strong demand across its Epyc, Instinct, and Ryzon product lines.

Datacenter sales accounted for the lion's share of AMD's revenue gains during the quarter, with the division generating $3.7 billion, up 57 percent from a year ago, as the biz ramped production of its MI325X accelerators and 5th-gen Epyc CPUs during the quarter.

According to Su, AMD will begin production of the latter in the United States later this year at TSMC's Arizona fab site.

AMD's Client and Gaming segment revenues, meanwhile, jumped 28 percent to $2.9 billion in the first quarter. Of that, Ryzen processor and SoC sales accounted for the majority of that at $2.3 billion. Gaming revenues, on the other hand, slid a precipitous 30% from last year to $647 million on weak demand for custom SoCs used in Sony, Microsoft, and Valve game consoles.

Finally, the embedded compute segment, which includes FPGAs and adaptive-SoCs, pulled in $823 million during the quarter, down 3 percent from this time last year. ®