AMD’s Comeback: How a $660 Million Loss Turned Into a Challenge to Nvidia

Shares hit a record; market value passes $1 trillion Results

AMD’s stock closed at a record $633.91 on 2 October 2026 and briefly traded near $645, pushing its market value above $1 trillion, according to market reports. Investors were pricing in the MI450 ramp for OpenAI and Oracle, both due to start in the second half of 2026. The same shares traded below $2 in 2015.

Why it matters: one of OpenAI’s warrant tranches is tied to AMD reaching $600 a share. The stock price is now part of the deal mechanics, not just a scoreboard.

Q2 2026: record $11.5 billion, data centre doubles Results

Revenue rose 50% to $11.5 billion, with GAAP gross margin of 54%, operating income of $2.0 billion and net income of $2.3 billion. Data Center reached $6.7 billion, 58% of sales, driven by EPYC CPUs and MI350 GPUs; it swung from a $155 million operating loss a year earlier (hit by MI308 export charges) to $2.1 billion of profit. Client grew 23% to $3.1 billion, Embedded 19% to $977 million, while Gaming fell 31% to $779 million as console chip sales slowed.

Advancing AI 2026: MI400, Helios and EPYC Venice launch Payoff

AMD launched the Instinct MI400 series and put Helios into full production, with revenue shipments from the end of the third quarter. A Helios rack holds 72 MI455X GPUs, 18 EPYC CPUs and about 31 TB of HBM4; AMD claims 50% more memory capacity than Nvidia’s Vera Rubin NVL72, a vendor figure not yet independently tested. EPYC Venice (Zen 6, up to 256 cores) is the first x86 server CPU on TSMC 2nm. On 22 July Anthropic agreed to deploy up to 2 gigawatts of MI450-series GPUs, with AMD investing up to $5 billion in Anthropic. MI500 is promised for 2027.

Meta signs for 6 gigawatts Deal

Days after expanding its Nvidia order, Meta agreed to deploy 6 gigawatts of AMD GPUs over several generations, starting with a custom MI450-based chip and EPYC Venice CPUs in Helios racks in the second half of 2026. Like OpenAI, Meta received a warrant for up to 160 million AMD shares at one cent each, vesting with deployments and share-price targets.

2025 results: $34.6 billion, a record year Results

Revenue grew 34% to $34.6 billion, almost nine times the 2015 figure. Data Center made $16.6 billion (up 32%), Client and Gaming $14.6 billion (up 51%) and Embedded $3.5 billion. GAAP net income was $4.3 billion. The year absorbed about $440 million of net charges after the US required export licences for the MI308, a reduced chip designed for China, in April 2025.

OpenAI picks AMD for 6 gigawatts Deal

Exactly a year before this update, OpenAI agreed to deploy 6 gigawatts of AMD Instinct GPUs, starting with 1 gigawatt of MI450 chips in the second half of 2026. In return AMD granted a warrant for up to 160 million shares, about a tenth of the company, vesting with deployments and AMD share-price milestones. A week later Oracle said it would offer 50,000 MI450 GPUs from the third quarter of 2026. In November AMD told analysts it targeted revenue growth above 35% a year and data-centre growth above 60% a year over the next three to five years.

Why it matters: this was the first time a leading AI lab had committed to AMD at Nvidia-scale volumes.

MI350 launches; Helios is previewed Bet

At Advancing AI 2025 AMD launched the MI350 series and previewed the MI400 and Helios rack for 2026. The rack strategy relied on ZT Systems, a server builder AMD agreed to buy for $4.9 billion in August 2024 and completed in March 2025; it kept ZT’s design engineers and agreed to sell the factories to Sanmina for $3 billion. The MI350 ramp drove record Instinct revenue in the second half of 2025.

US export rules hit the China chip Setback

The US government required licences to export the MI308 to China, forcing AMD to write down inventory. Full-year 2025 charges came to about $440 million net, after a partial reversal in the fourth quarter, when AMD sold about $390 million of MI308s to China. The episode showed how much of the AI chip business depends on trade policy.

MI300X: AMD’s first serious AI challenger Bet

A year after ChatGPT launched, AMD unveiled the MI300X, a chiplet-based accelerator with 192 GB of HBM3, more memory than Nvidia’s H100. Lisa Su forecast a $400 billion AI-accelerator market by 2027. Instinct revenue went from a small base to more than $5 billion in 2024, AMD’s fastest product ramp, though still a fraction of Nvidia’s sales.

Revenue dips as the PC market slumps Setback

After the pandemic boom, PC demand collapsed. AMD’s client revenue fell sharply in late 2022 and 2023, and total revenue slipped from $23.6 billion to $22.7 billion, the only annual decline since the turnaround. Data centre and embedded sales held up, and the AI story was only just starting.

Frontier: the first exascale supercomputer Payoff

Frontier, built by HPE with AMD EPYC CPUs and MI250X GPUs, became the first machine to pass one exaflop on the TOP500 list. AMD had won the contract in 2019. In November 2024 El Capitan, using the MI300A chip, took the top spot again for AMD. The supercomputer wins gave Instinct credibility years before generative AI.

Xilinx and Pensando: beyond CPUs and GPUs Bet

AMD completed its all-stock purchase of Xilinx, announced at $35 billion in October 2020 and worth about $49 billion by closing. It added FPGAs and adaptive chips for cars, factories, telecoms and aerospace, and the engineers behind AMD’s later AI engines. In May it bought Pensando, a networking-chip maker, for about $1.9 billion. In June it told analysts its target market was about $300 billion.

Revenue nearly doubles in two years Results

Revenue reached $9.76 billion, up from $6.48 billion in 2018. Zen 3 chips in November 2020 gave AMD the fastest gaming CPUs, the last category Intel had held. The PlayStation 5 and Xbox Series X|S, both using AMD chips, launched the same month. The following year revenue jumped 68% to $16.4 billion.

Zen 2 and chiplets: the design that scaled Payoff

Zen 2 split each processor into 7nm core chiplets made by TSMC and a 12nm input-output die made by GlobalFoundries. The 64-core EPYC Rome had more than twice the cores of Intel’s mainstream 28-core Xeon at the time. The first Zen EPYC in 2017 had already joined four dies in one package, but Rome’s mix-and-match design is what made AMD’s products cheap to scale.

GlobalFoundries quits 7nm; AMD moves to TSMC Bet

GlobalFoundries stopped developing its 7nm process, which pushed all of AMD’s leading-edge chips to TSMC. It turned out to be lucky timing: Intel’s own 10nm process was years late, so AMD’s TSMC-made chips had a manufacturing lead over Intel for the first time. A 2016 amendment to AMD’s wafer agreement, which cost $335 million, had already given it the freedom to use other foundries.

Ryzen and EPYC arrive; AMD returns to profit Payoff

The first Ryzen 7 chips offered eight cores for roughly half the price of Intel’s eight-core desktop parts. Zen delivered about 52% higher instructions per clock than the core it replaced, beating the 40% AMD had promised. EPYC followed in June with up to 32 cores. Revenue rose 25% to $5.33 billion and AMD made a $43 million net profit, its first annual profit since 2011.

Selling assets to fund Zen Low point

AMD sold 85% of its chip assembly and test plants in Penang and Suzhou to Tongfu Microelectronics for about $371 million and licensed server technology to a Chinese joint venture for $293 million. Most of the money was used to keep AMD going until Zen arrived. Graphics chief Raja Koduri’s group launched the Radeon Instinct brand and ROCm software in late 2016, AMD’s first focused push into data-centre AI.

The low point Low point

Revenue fell to $3.99 billion with an operating loss of $481 million and a net loss of $660 million. AMD carried about $2.2 billion of debt, and its server CPU share was close to zero. At its May analyst day AMD described Zen and promised a 40% gain in instructions per clock. Jim Keller, who had returned in 2012 to lead the CPU effort, left in September 2015, with the core design largely finished.

Correction: AMD was never ‘near-bankrupt’ in a legal sense. It was in financial distress, but it never filed for bankruptcy or missed a debt payment.

Lisa Su becomes CEO Bet

Su, an MIT-trained engineer who had run semiconductor research at IBM and networking chips at Freescale, became AMD’s president and CEO. Her plan was simple to state and hard to do: build great products, deepen customer relationships and simplify the company. She focused the company on high-performance computing and stopped trying to compete in every market.

Losses, layoffs and a sold headquarters Low point

Under Rory Read, AMD cut staff, took a large charge to give up its remaining GlobalFoundries stake and lost $1.18 billion in 2012. In 2013 it sold its Austin campus and leased it back for cash. Two console wins helped: custom AMD chips inside the PlayStation 4 and Xbox One (November 2013) brought in steady revenue during the lean years.

The factories go: GlobalFoundries is born Bet

Under CEO Dirk Meyer, AMD moved its fabs into GlobalFoundries, a joint venture with Abu Dhabi’s Advanced Technology Investment Company. The new company assumed about $1.2 billion of AMD debt, and ATIC committed $2.1 billion of investment. AMD became a fabless designer, but a wafer supply agreement tied it to GlobalFoundries for most of the next decade.

Background: AMD had paid $5.4 billion for the graphics company ATI in 2006, which gave it Radeon but stretched its balance sheet just as Intel’s Core chips pulled ahead.