AMD Is Not Catching Nvidia, but It Is Building a Profitable Business

Competition remains important in technology markets because it encourages economic and technical substitution. AMD’s progress illustrates both the opportunities and limits of competing with Nvidia in accelerated computing.

AMD has become a credible provider of datacenter CPUs for the second time and entered the datacenter GPU accelerator market with the Instinct MI200 series. However, AMD is currently putting more pressure on Intel than on Nvidia. Intel has no comparable datacenter GPU or significant AI accelerator business, while AMD’s ability to challenge Nvidia is constrained by limited access to high-bandwidth memory (HBM) from Samsung, SK Hynix, and Micron.

AMD is also using more HBM per accelerator in the Instinct MI400 series than Nvidia uses in its Blackwell B200 and B300 products or its Rubin R200 products. That strategy provides a more balanced relationship between compute and memory capacity, but it also limits the number of GPUs AMD can produce with its available HBM allocation.

One possible future approach would be for the Instinct MI500 series to use optical interposers and disaggregated, composable HBM installed in trays adjacent to GPU accelerators. Such a design could allow memory to be reused as GPU engines are replaced with higher-capability versions. It could provide an architectural advantage over static configurations with intentionally limited memory, which Nvidia uses to increase accelerator volumes. A similar approach could eventually apply to the “Florence” Zen7 Epyc datacenter CPUs, using trays of disaggregated CPUs and DRAM banks so that memory remains in service as processors are upgraded.

There is no clear indication that AMD is pursuing this type of architecture. Its limited HBM supply is one reason the company is not taking substantial datacenter GPU market share from Nvidia. AMD’s gains against Intel in datacenter CPUs reflect not only its x86 product designs, but also its management of process nodes, packaging capacity, and supply chains. These factors have allowed AMD to increase x86 shipments faster than Intel while achieving a higher average selling price.

Third-quarter market-share figures were not yet available, but AMD held slightly more than one-third of x86 server CPU shipments and slightly less than half of server CPU revenue during the first half of 2026.

The widening revenue gap

The overall revenue gap between Nvidia and AMD is expanding. Comparing total company revenue across all products makes AMD appear stronger relative to Nvidia because AMD benefits from substantial client CPU and GPU businesses, but the comparison still shows the scale of the difference.

In 2021, before the generative AI boom and the emergence of a practical ChatGPT chatbot API, Nvidia reported $26.9 billion in sales, compared with $16.4 billion for AMD. Nvidia’s revenue was 1.64 times higher. GPU computing had already been established for roughly a decade and a half, while AMD was still developing its architectural position.

In 2023, Nvidia’s sales reached nearly $60.9 billion, or 2.69 times AMD’s $22.7 billion. AMD’s results benefited from the PC buying boom during the coronavirus pandemic.

For 2026, Nvidia is expected to generate approximately $354 billion in sales, compared with a possible $52.2 billion for AMD. That would put Nvidia at 6.79 times AMD’s revenue. One Wall Street consensus estimate projects that Nvidia will generate $1,091 billion in calendar 2029, compared with $164 billion for AMD, a 6.65-times gap that would be broadly unchanged from today.

Forecasts are not destiny, but the figures show the effect of AMD’s HBM constraints. AMD’s practical strategy is to generate as much revenue as possible from the products it can manufacture. The AI market remains underserved rather than fully competitive, giving suppliers pricing power. Hyperscalers, cloud providers, and some large AI model developers are responding by designing and manufacturing their own computing engines, which can be cheaper and more specialized.

AMD’s third-quarter 2026 results

Against that background, AMD’s third-quarter 2026 results reflect the production ramp of two “Helios” AI rack-scale systems. One version uses CPUs for agentic AI sandboxes. The other combines Epyc CPUs and Instinct GPUs for AI training and inference.

AMD reported third-quarter revenue of $13.38 billion, up 44.7 percent. Operating income increased 2.2 times to $2.83 billion as some Helios-related costs had already been absorbed in earlier quarters and profits from components used in the systems began to increase. OEMs and ODMs manufacture the racks, while AMD supplies components and designs.

Net income rose 95.3 percent to $2.43 billion, equal to 18.2 percent of revenue. AMD remains substantially less profitable than Nvidia, but this was nevertheless a strong net margin for AMD.

The company’s cash balance more than doubled year over year to $15.78 billion. AMD now holds more than one quarter’s worth of revenue in cash, providing a level of financial security that it has rarely had before. The result is a stronger position in markets characterized by high production costs and intense competition.

Approximately two-thirds of AMD’s quarterly revenue came from datacenter products, including Epyc CPUs, Instinct GPUs, Pensando DPUs, Helios rack designs, ROCm software, and related products.

Segment performance

Data Center group revenue was $8.55 billion, an increase of 96.9 percent and above the company’s earlier target of more than 80 percent growth. Data Center group operating income increased 2.7 times to $2.88 billion, representing 33.7 percent of company revenue.

AMD’s Client group, which sells PC CPUs and GPUs, generated $2.95 billion in sales, up 7.2 percent. In the underlying model, operating income declined 18.8 percent to $401 million.

Gaming revenue, which primarily represents discrete Radeon GPUs for gamers, was $773 million, down 40.4 percent. The decline would have been a major concern several years ago, but it now represents a relatively small part of AMD’s overall business. Estimated Gaming group operating income was $172 million, down 53.9 percent.

Embedded revenue, primarily from Xilinx FPGAs, reached $1.11 billion, up 29.4 percent. Operating income increased 60.1 percent to $453 million, roughly three times the profitability of AMD’s PC chip business.

Epyc and Instinct revenue

CPUs again generated most of AMD’s Data Center revenue, marking the third consecutive quarter in which Epyc processors led the segment.

AMD chief executive officer Lisa Su said sales of Epyc server CPUs to hyperscalers and cloud providers more than doubled. Sales of CPUs for enterprise, telecommunications, service provider, government, and academic deployments also more than doubled.

Based on those comments, the estimates in the underlying model put hyperscaler and cloud-provider CPU sales at $3.53 billion, up 2.2 times. The enterprise and related segment reached an estimated $1.15 billion, a 105 percent increase. Together, those figures imply total Epyc CPU sales of $4.68 billion, up 116 percent year over year.

After estimating revenue from Pensando DPUs, other network interface cards, datacenter-class FPGAs, and components for CPU and GPU system boards, the model places Instinct GPU sales at $3.43 billion, up 72.3 percent. Greater HBM allocations and additional packaging capacity could have allowed AMD to sell substantially more GPUs. The difference between potential demand and reported revenue illustrates the effect of HBM availability.

The forthcoming “Venice” Xen6 Epyc processors began shipping for revenue in the third quarter of 2026. Su said the first OEM systems would be available before the end of the year. OEMs and ODMs are developing 240 Venice platforms, 20 percent more than the design wins for the preceding “Turin” Xen6 Epyc processors.

Su also said that Helios racks equipped with “Altair” MI455X GPUs had begun shipping to Meta Platforms, Microsoft, OpenAI, and Oracle.

Outlook for 2026 and 2027

AMD raised its datacenter guidance and said 2026 datacenter revenue was expected to more than double compared with 2025. The AI portion of the business, including CPUs, GPUs, and DPUs, is expected to grow 50 percent sequentially in the fourth quarter of 2026. GPU sales are expected to exceed CPU sales during that quarter, supported by additional Helios systems for production workloads.

AMD also reported improving visibility into 2027. The company previously expected annual server revenue to grow by more than 70 percent and datacenter segment revenue to more than double. Since then, customers have expanded their Epyc and Instinct deployment plans, while AMD has secured additional supply.

The updated expectations are for server revenue to double year over year in 2027 and datacenter AI revenue to grow faster than previously projected. AMD expects datacenter products to account for more than 80 percent of total company revenue in 2027 and to produce most of the company’s revenue and earnings growth.