# AMD Catches The Agentic AI Wave And Will Ride It Up Masterfully > Source: > Published: 2026-08-05 13:59:09+00:00 # AMD Catches The Agentic AI Wave And Will Ride It Up Masterfully AMD might not be taking any bites out of Nvidia’s market share when it comes to AI systems, but it is capturing its proportional share as the GenAI market expands with rapid inflation like the early universe did. But, we must remember that after the universe expanded in size, from about the size of a single subatomic particle to about the volume of a grapefruit in that first instance, the universe continued to evolve with bouts of punctuated equilibrium, where new conditions manifest and new elements of physics (literally as well as figuratively) emerged and everything changed several times over before reaching what feels like a steady state. It was hard to play catch up with Intel in CPUs – twice – and it is hard to play catchup with Nvidia – first with GPUs and then with rackscale designs. But AMD did catch and then pass Intel, and there is no law of physics that says AMD cannot catch and beat Nvidia. Incumbents, and especially monopolistic ones, have a tendency to eventually choose a holding maneuver rather than a frontal assault. So, yes, it takes AMD until late 2026 to draw even with Nvidia “Oberon” rackscale systems with the “Helios” rackscale designs that it created in conjunction with Meta Platforms and Microsoft. So what? This is only the second skirmish in a very long war that will eventually descend into a price and price/performance war. One that AMD will be willing and able to win in GPUs as it has in CPUs. Nvidia will move higher up in the stack, selling support for its closed source CUDA-X stack and open source models. And eventually – as hard as this is to believe today – they will share this very lucrative market much as Intel and AMD and the Arm collective (comprised of homegrown CPUs from the hyperscalers and cloud builders) do today with CPUs. About a third each. The point is this: If you were expecting for AMD to do better on the AI front than it did in the second quarter, well bucko, that’s your problem. AMD is doing better than its company’s top brass expected on both the CPU and the GPU fronts, and it is poised to more than double its GPU business next year thanks to Helios systems using its flagship “Altair” MI455X GPUs and its very powerful next generation “Verano” CPUs, a high I/O, mid-core count variant of the “Venice” sixth generation Epyc processors that will also be ramping as 2026 unfolds. Chief executive officer Lisa Su laid the scenario all out in the call with Wall Street going over the Q2 financial results. The datacenter AI accelerator business, which will now include the Helio rack components that are sold to the ODMs and OEMs of the world so they can make systems on behalf of AMD’s MI455X customers, will grow in the second half of 2026 faster than it did in the first half, where it was no slouch, and the plan is for the datacenter AI business to accelerate growth even further in 2027. To be more precise, Su said that the Epyc CPU business will grow by more than 80 percent in the second half of 2026 and by more than 70 percent in all of 2027 against what will be a pretty tough compare, and added that the overall datacenter business will more than double in 2027. She also said that AMD had enough wafer, substrate, interposer, and HBM capacity to not only meet these commitments, but had a little spare capacity on the upside so the company does not get caught flat-footed as it did with the sudden rise of agentic AI sandboxes based on fast CPUs this year. No worries, that demand spike just meant that all CPU makers could sell out, no matter what they had on the truck and at prices they could command. I think it is safe to say that AMD is going to gain server CPU share in 2026 and 2027 and will gain share of rackscale systems by definition in the same timeframe because it doesn’t have any right at the moment. So it is all upside with Helios. Unit shipments and average selling prices for both Epyc CPUs and Instinct GPUs are both on the rise, and that is what is driving this revenue boom for AMD, along with increasing NIC and DPU shipments, which are part of the Helios package. As far as we can tell, AMD is keeping pace, more or less, with the total addressable market for the datacenter overall, which it outlined at its recent Advancing AI 2026 event in Silicon Valley and [which we covered in detail here](https://www.nextplatform.com/compute/2026/07/24/the-money-amd-is-chasing-with-its-rackscale-ai-system-roadmaps/5278510). I had to so some spreadsheet magic to fill in the gaps on the datacenter compute TAM for 2026 through 2029, but what AMD said this week matches with what I hunched out two weeks ago. Now, for the datacenter AI accelerator market, AMD is growing faster than the TAM it did outline for those years as well as for the 2025 and the 2023 endpoints. The datacenter AI GPU/XPU TAM is expected grow by 75.6 percent to $360 billion and by another 75 percent in 2027 to $630 billion, and then cool off after that, according to AMD’s own market analysis. The datacenter CPU TAM will double in 2026 to $53 billion and will grow again by 81.1 percent in 2027 to $96 billion, according to AMD’s own prognostications, as I outlined in my detailed monster table in the story mentioned above. The CPU TAM starts slowing down, too, after 2027, but there is no decline in either the CPU or GPU markets out through 2030. This will normalize, and then the fight really begins. In the meantime, let’s drill down into AMD’s second quarter. It was another record setting quarter for the company, with revenues up 50.1 percent to $11.54 billion across all product groups. Operating income was $1.99 billion, a sharp reversal from the $134 million operating loss the company had last year. Net income was up by a very good 2.6X to just a tad under $3 billion, helped by AMD’s interest on its cash pile and its investments in other companies, which are rising in value – a value it must put on its books even if it is just a paper profit. Net income, thus boosted, represented 19.9 percent of revenues, a high level that AMD has only just barely beaten twice in its history thus far. AMD exited the quarter with $13.11 billion in cash and investments, which is more than a quarter of the annual run rate for the company and which is a healthy amount of simoleons for any tech company gearing up for battle and having to prepay for a lot of components for its hardware to ensure supply. AMD still has healthy if under pressure PC and gaming chip businesses, which are feeling the profit pinch as component prices keep going up and customers are getting fussy about paying more for machines with less DRAM and flash storage. The embedded business – meaning mostly FPGAs from the Xilinx line – is still the most profitable business that AMD has, with an average of around 38 percent of the money dropping to the middle line over the past two years. So that helps counterbalance the issues in PC CPUs and GPUs and gaming GPUs. But the datacenter business, which includes selling CPUs, GPUs, DPUs, and other IP inside the Helios racks, is the main driver for the company and will be increasingly so as we drive towards the magical land of 2030 out there on the horizon. The Data Center group at AMD had sales of $6.72 billion, up 107.3 percent year on year and up 16.3 percent sequentially. Operating income was $2.1 billion, 31.3 percent of revenues and a hell of a lot better than the $155 million operating loss this time last year. The thing to remember here about the datacenter business at AMD is that the company is not just optimizing to boost revenue, but learning how to optimize its HBM and wafer allocations to be competitive with Nvidia, Intel, and the Arm collective across all manner of compute and I/O engines for profitability, customer happiness, and competitiveness. For instance, AMD is 50 percent more generous with the HBM memory capacity in the Helios racks than Nvidia will be with its Vera-Rubin Oberon racks, while giving customers 15 percent more AI flops and 30 percent better cost per token per watt according to early benchmarks. AMD does not tell us Epyc CPU and Instinct GPU revenues explicitly, but gives us enough hints that we can suss it out to within maybe five points or so here or there. My best guess is that Pensando DPUs and AI NICs drove maybe $275 million of revenue, nearly triple from a year ago, and datacenter-class FPGAs accounted for maybe $95 million, up 1.7X year in year. If you back those out if the Data Center group revenue, and then take a stab at what “more than 70 percent” growth means for CPUs and GPUs in Q2 2026, as Su said, you might conclude (as I did) that Epyc CPUs accounted for $3.35 billion, up 74 percent but down 8.3 percent sequentially, while Instinct GPUs rose by 2.6X to just a tad over $3 billion. It is a wonder why Su did not call out this GPU growth explicitly. It is possible that I am underestimating DPU and FPGA sales. My best informed guess is that the hyperscalers and cloud builders accounted for 73 percent of Epyc CPU sales in the second quarter, up 75 percent year on year, while enterprises, telcos, academia, and sovereigns accounted for $905 million in Epyc CPU sales, up 71.4 percent.