AMD

Speaking in Korea on cooperation with Samsung and SK hynix, Lisa Su said AMD had already begun shipping its next-generation MI455X accelerators and Helios rack-scale systems. It is the first statement that the generation behind the OpenAI, Meta and Anthropic gigawatt agreements has started to leave the factory, rather than being tested or scheduled.

AMD said customers were testing AMD Helios systems at their own sites. It said the testing included workload validation and preparation for deployments at scale.

Benzinga and Newsquawk reported that AMD and Nvidia were lobbying the Trump administration over proposed China-related chip-export restrictions in an annual defense policy bill. The reports attributed the account to Politico and sources.

Lisa Su said she would attend the state dinner scheduled for the following night. She said technology leaders would seek stability across the technology ecosystem.

SemiAnalysis head Dylan Patel alleged that Puzhi Electronic Technology was crowdfunding a system based on AMD's Zynq UltraScale+ XCZU47DR RFSoC. The reports described the device as a U.S. export-controlled dual-use component.

Hewlett Packard Enterprise said it would provide trays to the market. The report described the trays as a white-label supply route for server competitors working with AMD on Helios.

Advanced Micro Devices said it expected its Data Center business to double in 2027 to about $70 billion in sales. The company said it had undertaken supply-chain work to support doubling Data Center revenue the following year. Su added that all three anchor customers for MI450 and Helios had given forecasts above AMD's initial expectations, and that AMD was working with suppliers to support them.

Advanced Micro Devices said Instinct MI450 production shipments were set to begin in the third quarter of 2026. It said revenue would step up in the fourth quarter of 2026 and again in the first quarter of 2027.

AsiaToday reported that Howard Lutnick told CNBC on September 2 that he did not expect export-control easing to be on the summit agenda. NationPress, published September 3, quoted him as saying that Donald Trump had offered China H200 chips and equivalent chips from AMD and other suppliers.

Politico reported that NVIDIA and AMD urged the Trump administration to discourage lawmakers from including China-related chip-export restrictions in the annual defense policy bill. The reported outreach sought to prevent the proposed restrictions from being added to that bill.

AMD said the reported instance was unrelated to any direct AMD sales or shipments. The statement addressed the alleged use of an AMD Zynq UltraScale+ XCZU47DR RFSoC in Puzhi Electronic Technology's project.

AMD's second-quarter 10-Q, filed August 4, repeated the February disclosure word for word: some MI325 export licenses had been granted for certain China-based customers, every unit would first have to be inspected in the United States and pay a 25 percent tariff on the way in, and the company still did not know whether China would allow the imports. No MI325 or MI308 China revenue was called out for the quarter.

Second-quarter 2026 results showed revenue of US$11.5 billion and data center segment revenue of US$6.7 billion, up 107 percent from a year earlier on demand for EPYC CPUs and Instinct GPUs. Lisa Su said Instinct deployments were scaling and Helios was beginning to ramp.

AMD said Instinct MI600-series GPUs were coming in 2028. The statement extended AMD's disclosed accelerator roadmap beyond the MI500 series planned for 2027.

AMD said Meta had begun testing and validating workloads on AMD Helios racks. AMD also said Meta was validating sixth-generation EPYC CPU platforms in its labs.

AMD said OpenAI expected to bring Helios online beginning in the fourth quarter of 2026. AMD said deployments were expected to accelerate throughout 2027.

AMD launched the Instinct MI350P GPU at Advancing AI 2026. AMD said the GPU brought AI acceleration to existing infrastructure.

AMD launched the Instinct MI400 series, with the MI455X in Helios racks for frontier AI and the MI430X for sovereign AI and HPC, and named OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr and Cirrascale among the organisations choosing Helios.

AMD said Anthropic was using AMD Instinct MI355X GPUs before its planned MI450-series deployment. AMD described the planned deployment as building on that existing use.

AMD committed to make a strategic equity investment of up to $5 billion in Anthropic in the future. The commitment was disclosed in AMD's Anthropic partnership announcement.

AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of Instinct MI450 series GPUs in AMD Helios rack-scale systems, with the first gigawatt beginning in the first half of 2027.

AMD said Microsoft would ramp AMD Helios at scale on Azure. The deployment was intended to power frontier-model inference for Microsoft, its AI customers and Azure AI services.

At the Great Hall of the People, He Lifeng invited multinationals including AMD to deepen cooperation, citing the outcomes of the May 13 to 15 Xi–Trump summit, and Su said AMD would expand its operations and investment in China. Su had not been part of the executive delegation that travelled with Trump.

Data center AI revenue dipped sequentially in the first quarter, which AMD attributed to lower China revenue after the $390 million MI308 quarter, and CFO Jean Hu said China revenue in the period was not material. AMD did not forecast further China GPU sales, leaving the MI325 licenses as an option rather than a plan.

AMD reported $5.8 billion in first-quarter 2026 Data Center segment revenue, up 57% year over year. AMD attributed the result to EPYC processor demand and the continued ramp of Instinct GPU shipments.

In a Digital Foundry interview Cerny said the new PSSR shares its core algorithm with AMD's FSR Redstone upscaler and that a co-developed frame-generation library would reach PlayStation platforms at some point, but would not say which machine or whether it meant the PS6. Asked directly, he said Sony had no more releases planned for the year.

AMD and Meta agreed a multi-year, multi-generation partnership to deploy up to 6 gigawatts of Instinct GPUs. Shipments for the first gigawatt were expected to begin in the second half of 2026 using a custom GPU based on the MI450 architecture, and AMD issued Meta a performance-based warrant for up to 160 million shares.

In its annual report AMD disclosed that it had begun MI308 shipments at the end of fiscal 2025, that net export-control charges for the year came to about $440 million after the reversal, and that US officials expected 15 percent of licensed MI308 revenue although no rule establishing the requirement had been published. It warned that any such demand could subject the company to litigation and favour competitors not bound by the arrangement.

AMD's fourth quarter included roughly $390 million of MI308 sales to China that were not in its guidance, filled from orders placed early in 2025, plus the release of about $360 million of the inventory reserves booked in the second quarter. It guided to about $100 million of MI308 revenue in the first quarter of 2026 and nothing beyond, with Su calling China 'a very dynamic situation' and saying MI325 license applications had been submitted.

AMD reported full-year 2025 Data Center segment revenue of $16.6 billion. It said the result reflected growth in EPYC CPUs and Instinct GPUs.

AMD said it benefited from an approximate $360 million release of previously reserved MI308 inventory and related charges in the fourth quarter. The release concerned AMD Instinct MI308 inventory.

AMD's first-quarter 10-Q disclosed that in February 2026 the US government granted some licenses to ship MI325 products to certain China-based customers under the January policy. Each chip must first pass an inspection in the United States, which makes it subject to a 25 percent tariff on importation for that inspection, and AMD said it did not yet know whether China would allow the imports.

Following Trump's December 8 decision to allow H200-class sales to approved Chinese customers, BIS revised its licensing policy so that applications for the H200, the AMD MI325X and similar chips are reviewed case by case rather than under a presumption of denial. Applicants must show exports will not shrink capacity available to US customers, that the buyer screens its own customers, and that the chips pass independent third-party testing in the United States.

BIS revised its license review policy so that applications to export Nvidia H200, AMD MI325X and similar chips to China are reviewed case by case if the exporter shows no reduction in capacity for U.S. customers, the Chinese buyer has export-compliance screening, and the product passed independent third-party testing in the United States. The rule took effect on publication.

AMD introduced the Instinct MI440X GPU for enterprise deployments. It also previewed the next-generation Instinct MI500 Series GPUs.

Reports citing people familiar with the talks said Alibaba was considering buying 40,000 to 50,000 MI308 accelerators, priced around $12,000 each and roughly 15 percent below Nvidia's H20, in what would be AMD's largest China AI deal since the controls began. AMD had excluded MI308 from its guidance, so any order would be upside.

Over two days in Beijing, Su met MIIT minister Li Lecheng, affirming AMD's commitment to deepening its investment in China, and then Commerce Minister Wang Wentao, whose ministry said the two sides discussed AMD's business development in China and strengthening cooperation. The visit came as MI308 shipments were about to start and as Beijing was steering buyers away from Nvidia's China parts.

Speaking at a Wired conference in San Francisco, Su said AMD had licenses to ship some MI308 chips to China and would pay the 15 percent levy if it shipped them, the company's first direct public acceptance of the arrangement.

AMD set long-term targets of more than 35 percent annual revenue growth, a non-GAAP operating margin above 35 percent and earnings per share above US$20, with data center revenue growing more than 60 percent a year. It said Helios systems with MI450 GPUs would begin delivering in the third quarter of 2026.

Reuters reported that Chinese government guidance required new state-funded data-center projects to use domestically made AI chips. Reuters identified AMD as a foreign supplier of data-center chips and said the issuing regulators and nationwide scope were unclear.

AMD reported record revenue with no MI308 shipments to China in the quarter and none assumed for the fourth quarter. On the call Su said the company had received some MI308 licenses and thanked the administration, but called the situation too dynamic to forecast, with customer demand and China's own import stance still unclear.

AMD said its third-quarter results included no revenue from MI308 GPU shipments to China. The disclosure concerned AMD Instinct MI308 products.

Days after the Amethyst video, the AMD-focused leaker KeplerL2 said a 2027 PS6 launch was the plan unless unexpected delays happened, and Moore's Law Is Dead said the next Xbox, on an AMD chip he called Magnus, was targeting 2027 as well. A 2027 date would have matched the seven-year gap between the PS4 and PS5; within four months the memory market had put it in doubt.

AMD said Oracle would be the first hyperscaler to offer a publicly available AI supercluster powered by 50,000 MI450 Series GPUs. The offering was scheduled to begin in the third quarter of 2026.

In a nine-minute video with AMD's Jack Huynh, Cerny unveiled the technologies the partnership had produced: Neural Arrays that let compute units work as one AI engine, Radiance Cores, a dedicated hardware block for ray and path tracing, and Universal Compression to cut GPU memory bandwidth. All exist only in simulation, he said, but he was excited to bring them to a future console in a few years' time, the closest Sony had come to confirming the PS6. Cerny later told Digital Foundry the shift to co-engineering AMD's own roadmap was a step change from building custom PlayStation-only tech.

AMD and OpenAI announced a multi-year agreement under which OpenAI will deploy 6 gigawatts of AMD Instinct GPUs across generations, starting with 1 gigawatt of MI450 GPUs in the second half of 2026. AMD issued OpenAI a warrant for up to 160 million shares at US$0.01 that vests with deployment milestones and AMD share-price targets.

Six U.S. senators wrote to President Donald Trump asking him to reverse the arrangement allowing AMD and Nvidia to sell advanced AI chips to the People's Republic of China for a fee. The letter said the sales conflicted with U.S. national-security interests and questioned the fee's legal basis.

After the Financial Times reported the arrangement, a White House official confirmed to CBS News that Nvidia and AMD had agreed to pay the US government 15 percent of revenue from H20 and MI308 sales in China as a condition of the export licenses granted the previous week. No regulation was published, a point AMD would later flag to investors.

Reports, first from the Financial Times, said Nvidia and AMD had agreed to hand the U.S. government 15 percent of revenue from H20 and MI308 sales in China as the condition for export licenses. The Register described the arrangement as revealed that week; President Trump confirmed the terms publicly.

AMD booked roughly $800 million of inventory and related charges for the MI308, cutting non-GAAP gross margin to 43 percent from about 54 percent otherwise. Data center revenue still grew 14 percent on EPYC demand, and the company excluded any MI308 revenue from its third-quarter guidance while the licenses were under review.

Moore's Law Is Dead reported that Sony was developing two AMD APUs under star-themed codenames: Orion for the PlayStation 6 itself and Canis for a handheld able to run PS4, PS5 and PS6 games at 720p or 1080p, sharing a CPU and GPU design to simplify development. Sony has never confirmed the codenames, but the handheld idea resurfaced in executives' 2026 comments.

On the second-quarter 2025 call, CEO C.C. Wei said construction of the second Arizona fab, for 3nm, was complete and its volume production schedule was being sped up by several quarters. Work on the third fab, for 2nm and A16, had begun. Wei said about 30 percent of TSMC's 2nm-and-smaller capacity would eventually be in Arizona.

The day after Nvidia said it had been assured H20 licenses would be granted, AMD said the Commerce Department would move forward with its MI308 license applications and that it planned to resume shipments as approvals came, praising the administration's trade negotiations. The stock rose more than 6 percent.

Cerny told Tom's Guide that the first fruit of the AMD collaboration, an advanced upscaling algorithm co-developed with AMD, would ship to PS5 Pro in 2026 as a drop-in replacement for PSSR, work that he said would otherwise have taken years. He stressed the technology was not proprietary and was meant to move the whole industry, a preview of how the next console's features would be built.

Xbox president Sarah Bond announced a strategic multi-year partnership with AMD to co-engineer silicon for the next generation of Xbox consoles, handhelds, PC and cloud, with backward compatibility promised. Both console makers were now publicly building their next machines on AMD, setting up the parallel 2027 timelines that leakers would report in October.

AMD reported that Microsoft said Instinct MI300X was powering proprietary and open-source models in production on Azure. The statement concerned Azure production workloads.

AMD reported that Meta detailed broad MI300X deployment for Llama 3 and Llama 4 inference. The statement concerned AMD Instinct MI300X deployments.

At Advancing AI 2025 AMD said MI350 series accelerators were already deploying at Oracle Cloud Infrastructure with broad availability in the second half of 2025, and previewed Helios, a rack-scale system built on the coming MI400 series GPUs, Zen 6 EPYC "Venice" CPUs and Pensando "Vulcano" NICs.

Reporting first-quarter results, CFO Jean Hu said the export license requirement removed an estimated $700 million from second-quarter revenue and about $1.5 billion for the year, mostly in the second and third quarters. The next morning Su told CNBC that China was a 'large opportunity' and argued for a balance between national-security controls and the widest possible adoption of US technology.

JEDEC released JESD270-4, the HBM4 standard, which doubles independent channels per stack from 16 to 32 with two pseudo-channels each, keeps backward compatibility with HBM3 controllers and adds directed refresh management. SK hynix, AMD, Synopsys and Cadence gave statements.

A day after Nvidia disclosed the H20 license requirement and a $5.5 billion charge, AMD filed an 8-K saying the same new rule covered exports of its MI308 to China, Hong Kong, Macau and other D:5 countries. It expected up to $800 million of inventory, purchase-commitment and reserve charges, said it would apply for licenses with no assurance of approval, and its shares fell more than 7 percent.

AMD announced successful bring-up and validation of its fifth-generation EPYC CPU products at TSMC's new Arizona fabrication facility. AMD said the result underscored its commitment to US manufacturing.

AMD said it delivered more than $5 billion of AMD Instinct accelerator revenue during 2024. It said Data Center annual revenue nearly doubled as EPYC adoption accelerated.

In a 37-minute technical seminar on the PS5 Pro, lead architect Mark Cerny closed by announcing Amethyst, a joint program with AMD to design a more ideal architecture for machine learning and a shared set of neural networks for game graphics, which both companies could use on their own hardware. Cerny said not to expect a hardware announcement soon; it was the first public step toward the next PlayStation's silicon.

AMD launched the Instinct MI325X, shared details of the MI350 series expected in the second half of 2025, and said the MI400 series based on CDNA Next remained planned for 2026.

AMD said OCI bare-metal instances using MI300X accelerators had been adopted by Fireworks AI. The instances were part of Oracle Cloud Infrastructure's MI300X offering.

AMD reported that revenue increased 21% sequentially. It attributed the increase primarily to the ramp of AMD Instinct GPU shipments.

At Computex in Taipei AMD laid out a yearly roadmap for its AI accelerators: the MI325X with 288 GB of HBM3E in the fourth quarter of 2024, the CDNA 4-based MI350 series in 2025, and the MI400 series on a CDNA "Next" architecture in 2026.

AMD said Azure ND MI300X V5 instances were generally available. Hugging Face was identified as the first customer.

Lisa Su met Chinese Commerce Minister Wang Wentao in Beijing to discuss China-U.S. trade relations and AMD's development in China. Su said China was one of AMD's global strategic priorities and that AMD would increase investment and work with local partners.

Bloomberg reported that the Commerce Department had declined to clear AMD's cut-down AI chip for sale in China because it remained too capable, meaning AMD would have to apply for a BIS license rather than ship under the notification route. The report was the first public sign that the MI308 would live or die by license decisions.

At Advancing AI 2023 AMD announced availability of the MI300X accelerator and the MI300A APU. Microsoft announced Azure virtual machines built on MI300X, Oracle Cloud planned MI300X offerings, and Dell, HPE, Lenovo and Supermicro showed systems using the parts.

The October 2023 update replaced the interconnect-bandwidth parameter with a total processing performance and performance density test, closing the loophole that Nvidia's A800 and H800 had used, and added a notification requirement for chips just under the line. AMD's MI300 family fell above the new threshold, so any China part would have to be cut down much further.

On the second-quarter 2023 call, as AMD prepared to ramp the MI300, Su said the company would be fully compliant with US export controls but saw an opportunity to develop an accelerator specifically for Chinese customers, following Nvidia's A800 and H800 and Intel's Gaudi variant. That product became the MI308.

At its Data Center and AI Technology Premiere AMD introduced the MI300X, a GPU-only version of the MI300 family built on the CDNA 3 architecture with up to 192 GB of HBM3 memory, aimed at large language model training and inference, and said it would sample to key customers from the third quarter.

The Commerce Department turned the August letters into a published rule: advanced computing chips above a performance and interconnect threshold, the supercomputers they feed and the equipment to make them all needed licenses for China. The thresholds caught the MI250 class and set the ceiling under which AMD, Nvidia and Intel would design China-specific parts.

Days after Nvidia disclosed a license requirement for the A100 and H100, AMD confirmed it had received a similar notice covering the Instinct MI250 and MI250X, while the older MI100 was unaffected. AMD said the impact on its data center business would be marginal, since most of its compute GPUs went into US and European supercomputers, and disclosed the requirement in its third-quarter 10-Q.

The Bureau of Industry and Security added Sugon, the Wuxi Jiangnan Institute of Computing Technology, Higon and the two Chengdu Haiguang companies that AMD had formed with Sugon's THATIC consortium in 2016 to license Zen server-CPU technology. The listing cut the venture off from US technology a week before Trump's G20 meeting with Xi Jinping, and it marked the first time Washington's China controls landed directly on AMD's business.