{"id":213942,"date":"2026-09-03T22:49:22","date_gmt":"2026-09-03T22:49:22","guid":{"rendered":"https:\/\/www.europesays.com\/people\/213942\/"},"modified":"2026-09-03T22:49:22","modified_gmt":"2026-09-03T22:49:22","slug":"when-i-said-here-a-year-ago-that-the-ai-accelerator-market-would-be-500-billion-in-2028-it-felt-l","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/people\/213942\/","title":{"rendered":"“When I said here a year ago that the AI accelerator market would be $500 billion in 2028, it felt l.."},"content":{"rendered":"

Lisa Su “Inference Has Exceeded Learning”
2030 AI Accelerator Market $1.4 trillion
AMD Next-Generation AI Rack ‘Helios’ Production
Anthropic Contracts to Supply Up to 2GW GPUs <\/p>\n

\"AMD \uc0ac\uc9c4 \ud655\ub300 AMD CEO Lisa Su gives a keynote speech at AMD Advancing 2026 in San Francisco on the 23rd (local time). [Picture = AMD YouTube Capture] <\/p>\n

“When I said here a year ago that the AI accelerator market would be $500 billion in 2028, it felt like a huge number. But if you look at it now, the number is conservative.”<\/p>\n

“The AI accelerator market will reach $1.4 trillion in 2030,” AMD Chairman and CEO Lisa Su said in a keynote speech at the “Advanced AI 2026” held at the Moscone Center in San Francisco on the 23rd (local time). “By the end of the decade, the AI accelerator market could approach the size of the entire semiconductor market.”<\/p>\n

As a basis for raising the forecast, he cited that the center of gravity of the AI operation has shifted from learning to reasoning. CEO Sue predicted, “For the first time in 2026, the world will use more AI computing to execute the model than to train it,” adding, “About 60% of the world’s AI computing capacity will be spent on reasoning this year.” It is explained that the number of tokens consumed every month has increased by about 160 times in two years to 3.5 trillion, and the amount of computation for learning advanced models has increased by about five times every year over the past four to five years.<\/p>\n

It is AI agents that push the demand for reasoning. CEO Sue said, “The initial Large Language Model (LLM) stopped short of answering questions, but the agent continues to work until the problem is solved if you give it a goal. When you ask an agent to work, you repeat the process of inferring, calling tools, and accessing data over dozens of steps.” “We also need a lot of GPUs to handle the operation, but we also need a large amount of CPU to coordinate all steps,” he stressed.<\/p>\n

For this reason, CEO Sue also raised the outlook for the server CPU market. AMD predicted that the server CPU market would grow 18-20% annually until last year to reach $60 billion, but on the same day, the market, which is currently around $25 billion, will exceed $200 billion in 2030. The spread of A.I. was much faster than expected. “In conversations with all customers, there is talk that this investment is just beginning,” he said. AMD hit its own high with a 46% share of sales in the server market last quarter, and more than 60% of Fortune 100 companies use Epic processors.<\/p>\n

The core product unveiled on this day is the rack-level AI system ‘Helios’. CEO Sue said, “It is the best AI rack in the industry,” and “Helios has begun mass production.” Shipments will begin at the end of the third quarter and begin in earnest in the fourth quarter. AMD claimed that it outperforms its competitors by 15 percent in computational performance, 50 percent in HBM4 capacity and memory bandwidth, and 50 percent in scale-out bandwidth. CEO Sue did not mention his competitor by name, but the figure is interpreted as targeting Nvidia’s rack system “NVL72.”<\/p>\n

Helios includes Instinct MI455X GPU, Epic ‘Venice’ CPU, and Pensando networking chip. MI455X has integrated 320 billion transistors, and 12 chiplets and 432 gigabytes of memory made by TSMC 2 nanometers and 3 nanometers process are grouped into 3D stacked packaging. One rack contains more than 18,000 CDNA5 GPU compute units, more than 4,600 Gen6 CPU cores, and 31 terabytes of HBM4. CEO Sue held up GPUs, CPUs, and networking chips one after another on the stage and emphasized, “In the era of agentic AI, the best GPUs and CPUs have to move together, and we are the only companies that make both.”<\/p>\n

AMD has put forward an open ecosystem as the second pillar of its strategy. Helios uses UA links for internal rack connections and open Ultra Ethernet standards for external connections. This contrasts with Nvidia tying its ecosystem with its own standard, NV Link. “In both hardware standards and software, an open ecosystem is essential to the future of AI,” CEO Sue said, stressing investment in its own software stack ROCm.<\/p>\n

Earlier on the 22nd, AMD and Anthropic announced a partnership to deploy MI450 series GPUs up to 2 gigawatts in the form of a Helios rack. The first 1 gigawatt deployment will begin in the first half of 2027. AMD has decided to invest up to $5 billion in Anthropic and will use Anthropic’s Claude to design AMD semiconductors and develop ROCm. AMD shares rose more than 8% on the day of the announcement.<\/p>\n

CEO Sue predicted that AMD’s entire market for high-performance and adaptive computing products will grow by 40% annually, approaching $2 trillion in 2030. “The only way to handle such a market is for the ecosystem to move together,” he said. “No one company can solve this problem alone.”<\/p>\n

The domestic semiconductor industry is also watching the announcement. The MI455X uses 432 gigabytes of HBM4 per GPU, which is about 150 gigabytes more than the 288 gigabytes of Nvidia’s next-generation GPU Vera Rubin. The rack unit is 31 terabytes. While Samsung Electronics is expected to be a major supplier of HBM4 for MI450 series, SK Hynix is also seeking to expand AMD volume. If the deployment of 2 gigawatts becomes a reality, demand for HBM4 from both companies could increase significantly. [Silicon Valley correspondent Wonho-seop]<\/p>\n","protected":false},"excerpt":{"rendered":"Lisa Su “Inference Has Exceeded Learning”2030 AI Accelerator Market $1.4 trillionAMD Next-Generation AI Rack ‘Helios’ ProductionAnthropic Contracts to…\n","protected":false},"author":2,"featured_media":213943,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_share_on_mastodon":"0"},"categories":[159],"tags":[13994,75014,36453,692,575,1003],"class_list":["post-213942","post","type-post","status-publish","format-standard","has-post-thumbnail","category-lisa-su","tag-agent","tag-helios","tag-keynote-speech","tag-lisa-su","tag-nvidia","tag-san-francisco"],"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@people\/117209554674313357","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/posts\/213942","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/comments?post=213942"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/posts\/213942\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/media\/213943"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/media?parent=213942"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/categories?post=213942"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/people\/wp-json\/wp\/v2\/tags?post=213942"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}