AMD Q3 2026: Data Center Revenue Up 97% to $8.5B as MI450 and Helios Ship

AMD reported its third quarter of 2026 on 3 November, and for the first time the MI450 generation showed up in the revenue line. Total revenue was $13.4 billion, up 45% year on year, and the Data Center segment hit a record $8.5 billion, up 97%, according to the company's 8-K earnings release. AMD guided to about $17 billion for the fourth quarter and told investors it now expects total revenue to grow more than 80% in 2027. It also booked a $1.1 billion non-cash charge tied to the warrants it handed to customers, which is the part of the story worth reading twice.

What happened

In the prepared remarks, Lisa Su said AMD "started shipping MI455X in volume in the quarter", and that lead customers Anthropic, Meta, Microsoft, OpenAI and Oracle brought their first Helios racks online. Helios is AMD's rack-scale system that pairs MI450-series GPUs with 6th Gen EPYC "Venice" CPUs, pitched directly against Nvidia's NVL72 racks. AMD expects Helios deployments to "increase substantially" in the fourth quarter as HPE, Dell, Lenovo, Supermicro and other OEMs start shipping systems.

CFO Jean Hu said sequential Data Center growth of 27% was "primarily driven by initial MI450 Series sales and strong server revenue growth". Server CPUs did a lot of the lifting: AMD says cloud and enterprise server sales both more than doubled, and it attributes part of that to agent workloads, which run continuously on CPUs alongside the accelerators.

The background is the deal AMD announced with OpenAI on 6 October 2025: 6 gigawatts of Instinct GPUs over several generations, with the first gigawatt of MI450 starting in the second half of 2026, and a warrant for up to 160 million AMD shares that vests as deployments and share-price milestones are met. The third quarter is when that machinery started turning.

The numbers

For scale, Nvidia's Data Center segment reported $89.0 billion in its quarter ended 26 July 2026, per its 10-Q. The quarters do not line up exactly, but AMD's whole data center business is still under a tenth of Nvidia's.

Why it matters

For three years, the case for AMD in AI accelerators rested more on roadmaps than on racks. MI450 in Helios is the first time AMD ships a complete rack, with its own scale-up fabric, to five of the most demanding buyers in the industry at the same time. If those racks run production training and inference without drama, the market has a credible second supplier at rack scale, and that changes pricing conversations everywhere.

The guidance is the other signal. A roughly 50% sequential jump in Data Center revenue in the fourth quarter and more than 80% total growth in 2027 are only possible if Helios volume ramps on schedule, and AMD says it has "secured the wafers, substrates, advanced packaging, memory and other key components" to support it. That claim will be tested against TSMC's packaging capacity and an HBM market that every accelerator vendor is fighting over.

What sceptics say

The warrant charge deserves attention. AMD now records a quarterly fair-value adjustment on warrants issued to customers, and Hu said these adjustments "are expected to increase as we and our customers progress under these agreements". In plain terms, AMD's largest new GPU customer is being paid partly in AMD equity to take AMD GPUs. That is the circular-financing pattern bubble sceptics keep pointing at, and it is now visible on the income statement: GAAP EPS was $1.46 against $2.09 non-GAAP.

Cash is the second question. Operating income was strong, but free cash flow was $554 million as inventory rose about $3 billion to $11.5 billion. Building inventory ahead of a ramp is normal; it is also exactly what you would see if deployments slipped.

Third, concentration. The five named Helios customers are also the five companies most likely to build their own silicon or negotiate hard with Nvidia. OpenAI, AMD notes, picked EPYC as the host CPU for its first custom accelerator platforms, which is good for AMD's server business and a reminder that OpenAI is designing its own accelerators too.

Finally, the 2027 numbers are guidance. AMD's own remarks say PC demand is weakening on memory costs and it plans for a double-digit unit decline in the PC market next year, so Data Center has to carry the whole plan.

What it means for engineering and DevOps teams

My view: the practical news is that MI450 capacity will be purchasable through ordinary OEM channels and clouds from the fourth quarter onward, not just through hyperscaler deals. That gives platform teams a stronger hand in 2027 GPU negotiations, provided their software can actually move.

That is the part you control. If your serving stack is vLLM or SGLang with standard model formats, test it on ROCm now (AMD launched ROCm 10 this quarter) rather than when procurement asks. If you depend on custom CUDA kernels or TensorRT-LLM specific features, write down what breaks; my comparison of inference engines covers which engines are portable. On Kubernetes, the device plugin and DRA story is similar across vendors, as explained in GPUs on Kubernetes, but your node images, drivers and monitoring will not be.

Rack-scale systems also change the operational unit. A Helios or NVL72 rack fails, drains and gets upgraded as one thing, with liquid cooling and a power envelope your facility has to support. Capacity planning for 2027 is as much a data-centre conversation as a cloud one.

FAQ

When did AMD start shipping MI450?

AMD says it began recognizing MI450-series revenue in the third quarter of 2026 and started shipping MI455X in volume, with Helios racks online at Anthropic, Meta, Microsoft, OpenAI and Oracle. OEM Helios systems from HPE, Dell, Lenovo and Supermicro begin shipping in the fourth quarter.

What is AMD's warrant charge?

AMD issued warrants to customers under large GPU agreements, including up to 160 million shares to OpenAI. As MI450 ships, AMD remeasures their fair value each quarter; Q3 included about $1.1 billion of non-cash charges, which are excluded from non-GAAP results.

Is AMD catching up with Nvidia in data center revenue?

Not in absolute terms. AMD's Data Center segment was $8.5 billion in Q3 2026, while Nvidia reported $89.0 billion of Data Center revenue in its quarter ended 26 July 2026. AMD's growth rate is high, but from a much smaller base.

Related: LLM Inference Engines 2026: vLLM vs SGLang vs TensorRT-LLM vs Dynamo, GPUs on Kubernetes Explained (2026): Device Plugin, MIG, Time-Slicing, DRA and GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon.