Signals Inbox·July 20, 2026·AI Chips

Nvidia vs AMD: is AMD finally catching up?

Nvidia is still winning the AI infrastructure race by a wide margin, but AMD has finally caught up in enough areas to become a credible second platform for the world’s largest AI buyers.

Q1Why are Nvidia and AMD compared in AI chips?

Nvidia and AMD are now competing for the same strategic position: becoming the main computing platform behind the world’s largest AI models. Both companies sell data-center accelerators, server processors, networking components, software and increasingly complete rack-scale systems.

The comparison has become much more relevant lately because AMD has moved beyond promising specifications. OpenAI and Meta have each signed agreements covering up to 6 gigawatts of future AMD infrastructure. Oracle plans to launch a publicly available supercluster with 50,000 MI450-series GPUs, while several cloud providers already offer earlier Instinct generations.

These commitments place AMD inside the long-term infrastructure plans of some of the world’s largest AI buyers. Three years ago, the Nvidia-versus-AMD debate was mostly about whether AMD could produce a competitive GPU. The current argument is about whether AMD can become the standard second platform for AI infrastructure.

Nvidia remains the company everyone is reacting to. Its chips power a large share of frontier-model training, CUDA remains the default software environment, and its newer platforms bundle GPUs with CPUs, networking, storage and deployment software. AMD is the direct rival because it is the only other merchant-chip company currently attempting to compete across that entire surface area.

Q2Why is it hard to tell whether AMD is catching Nvidia?

AMD is catching Nvidia in selected product benchmarks and future customer commitments, while Nvidia is still pulling further ahead commercially.

The latest independent benchmark rounds show AMD’s MI355X competing closely with Nvidia’s B200 on several important training and inference workloads. AMD has also secured multi-generation agreements measured in gigawatts rather than hundreds of GPUs. Those developments make AMD look much stronger than it did during the MI250 or early MI300 periods.

The financial results point the other way. Nvidia’s latest Data Center revenue was roughly 13 times AMD’s entire Data Center segment, even though AMD includes server CPUs, FPGAs, networking products and other hardware in that segment. Nvidia also grew faster year over year from its much larger base.

There are really three separate races. AMD is approaching Nvidia in selected chip-level tests. It is beginning to secure a meaningful portion of future infrastructure plans. It remains far behind in revenue, deployed systems, software adoption and ecosystem control.

Calling AMD irrelevant would now be wrong. Calling AMD close to Nvidia would be even more misleading.

Q3How much larger is Nvidia’s AI business than AMD’s today?

Nvidia’s AI infrastructure business is currently more than an order of magnitude larger than AMD’s comparable operation.

Nvidia reported $81.6 billion of total quarterly revenue for the period ending April 26, compared with AMD’s $10.3 billion for its quarter ending March 28. The reporting periods differ by four weeks, but the gap is far too large for that timing difference to change the conclusion.

The comparison becomes more extreme inside the data center. Nvidia generated $75.2 billion from Data Center products, against $5.8 billion for AMD. That puts Nvidia at approximately 13 times AMD’s scale.

Even this comparison is generous to AMD. Its Data Center segment contains EPYC server CPUs, Instinct accelerators, Pensando networking products, FPGAs and adaptive computing products. Nvidia separately disclosed $60.4 billion of Data Center compute revenue. That figure alone was more than ten times AMD’s entire Data Center segment.

AMD does not disclose Instinct accelerator revenue separately, so we cannot calculate a clean GPU-to-GPU ratio. Nvidia’s accelerator business still appears to exceed AMD’s by considerably more than tenfold.

The margin gap also reveals who currently controls the economics. Nvidia recorded a 74.9% company-wide gross margin, compared with 53% for AMD. Product mixes differ, but a gap of almost 22 percentage points shows that Nvidia captures far more value from each dollar of sales.

Nvidia and AMD’s latest reported quarterly results

Q4Is AMD growing fast enough to close Nvidia’s lead?

AMD is growing quickly in data centers, but Nvidia is currently adding revenue so much faster that the dollar gap continues to widen.

AMD’s Data Center revenue rose 57% year over year to $5.8 billion. That would normally qualify as exceptional growth. Nvidia’s Data Center business expanded 92% to $75.2 billion over a comparable period.

The percentage comparison already favors Nvidia. The absolute comparison is brutal.

Nvidia added $36.1 billion of quarterly Data Center revenue in one year. AMD added approximately $2.1 billion. Nvidia therefore created about 17 times more new quarterly Data Center revenue than AMD during the same broad period.

Nvidia’s annual increase alone was more than six times the size of AMD’s current Data Center segment. AMD could have duplicated its entire Data Center business several times and Nvidia would still have added more revenue.

The most recent sequential results tell the same story. Nvidia’s Data Center revenue grew about 21% quarter over quarter, while AMD’s rose approximately 7%. Nvidia then guided total revenue to around $91 billion for the following quarter, another increase of roughly 12% from an already enormous base.

AMD is building a much larger AI business. It is not yet growing fast enough to reduce Nvidia’s commercial lead.

Q5Are major AI companies really using AMD GPUs at scale?

Major AI companies are now using AMD GPUs in production, although AMD deployments remain smaller and less widespread than Nvidia deployments.

Microsoft’s Azure platform offers MI300X virtual machines that can scale to thousands of GPUs. Microsoft has also said AMD accelerators power proprietary and open-source models in production on Azure, including workloads related to Azure OpenAI services.

Oracle has moved beyond a limited preview. OCI made MI355X instances generally available and designed its Zettascale Supercluster to scale as high as 131,072 GPUs. DigitalOcean currently sells MI300X and MI325X capacity, while Vultr offers MI300X, MI325X and MI355X systems.

Cohere has deployed its Command models on MI300X infrastructure. Several specialist AI clouds and system manufacturers have also submitted independently validated results using AMD hardware. AMD performance is no longer limited to systems operated by AMD’s own engineers.

We are now seeing commercially available cloud instances, production inference and multi-node training rather than a collection of evaluation agreements.

Nvidia still occupies far more of the critical infrastructure inside frontier laboratories. OpenAI, Anthropic, Meta, xAI and other major developers continue to plan large deployments around Blackwell and Rubin. AMD has established that customers can depend on its hardware. It has not established that they are ready to make AMD their primary platform.

Q6Do AMD’s OpenAI and Meta agreements change the Nvidia race?

AMD’s agreements with OpenAI and Meta create a credible path to much greater scale, but most of that demand has not reached production yet.

OpenAI agreed to deploy 6 gigawatts of AMD GPUs across several generations, beginning with one gigawatt of MI450 infrastructure. Meta later announced an agreement covering up to another 6 gigawatts, also beginning with an initial one-gigawatt MI450 deployment.

A combined headline figure of 12 gigawatts is too large to dismiss as experimentation. These companies are coordinating roadmaps with AMD across accelerators, server CPUs, rack systems and software. Meta’s first deployment will even use a custom MI450-based GPU optimized for its own workloads.

Oracle adds a clearer near-term distribution channel. Its first public MI450 supercluster is expected to contain 50,000 GPUs, with further expansion planned afterward. Customers will therefore be able to rent the architecture rather than building their own gigawatt-scale data centers.

The timing still limits what these agreements prove. Initial OpenAI and Meta shipments begin during the second half of 2026, while Oracle’s MI450 service is planned for the third quarter. AMD’s latest reported revenue contains little or none of this volume.

Gigawatt agreements are infrastructure frameworks rather than revenue delivered immediately. Rollouts depend on data-center construction, electricity, cooling, networking and software readiness.

AMD has secured its route into the next generation of AI infrastructure. The difficult part now is delivering the systems on time.

AMD’s largest announced AI infrastructure commitments

Q7Are AMD’s AI chips now as fast as Nvidia’s?

AMD’s latest AI chips can match or beat Nvidia on selected workloads, although Nvidia still demonstrates stronger performance across a broader range of systems.

The MLCommons Training 6.0 results gave AMD its clearest training evidence so far. MI355X systems finished within roughly 5% of Nvidia’s B200 on Llama 2 70B fine-tuning and within about 6% on Llama 3.1 8B pre-training.

AMD also improved its Llama 2 performance by 3.5 times between its first MI300X submission and the newer MI355X result. Hardware explains part of that jump, while ROCm optimizations, lower-precision formats and improvements to AMD’s training framework contributed substantially.

The MLPerf Inference 6.0 round was arguably more encouraging. On one single-node Llama 2 70B configuration, AMD reported 111% of B200 offline throughput and 115% of B200 server throughput. Against the newer B300, MI355X reached 91% of offline performance and 82% in the server test.

AMD also exceeded one million tokens per second in several multi-node inference submissions. It maintained strong throughput after moving beyond a single eight-GPU machine.

These results do not establish universal AMD leadership. MLPerf contains specific models, precisions and implementation rules. Performance changes with model architecture, context length, batch size, latency requirements and software maturity. Nvidia also appears across more submissions and deployment scales.

The assumption that an AMD accelerator automatically requires accepting dramatically weaker performance has become outdated.

AMD MI355X in recent independent benchmarks

Q8Does AMD still have a memory advantage over Nvidia?

AMD currently offers more memory than Nvidia in several GPU comparisons, although Nvidia has already removed part of that advantage with Blackwell Ultra.

MI355X includes 288 GB of HBM3E memory and 8 terabytes per second of bandwidth. Nvidia’s original B200 carries 180 GB, giving AMD 60% more memory capacity per accelerator in that comparison.

More memory can allow customers to keep a larger model on fewer GPUs, reducing model partitioning and communication between accelerators.

Nvidia’s B300 Blackwell Ultra also provides 288 GB per GPU, eliminating AMD’s capacity lead at that tier. AMD plans to move ahead again with MI455X, which is designed with 432 GB of HBM4 and up to 19.6 terabytes per second of bandwidth.

Memory remains one of AMD’s clearest product strengths, but Nvidia has shown that it can answer within the following generation.

Q9Has AMD’s ROCm become good enough to replace Nvidia CUDA?

AMD’s ROCm software has become good enough for serious production deployments, but Nvidia CUDA remains the easier and safer choice for most customers today.

AMD now supports the main frameworks used for modern AI work, including PyTorch, JAX, vLLM, SGLang and Megatron-LM. Prebuilt containers and optimized libraries have reduced the amount of manual porting required to run common models on Instinct GPUs.

A particularly fresh development came from the PyTorch team, which recently demonstrated its Monarch distributed-training system on AMD GPUs. The implementation included fault recovery across ROCm clusters, allowing healthy nodes to continue working while failed nodes recover. That addresses a real operational problem in large training jobs.

PyTorch 2.11 also introduced additional ROCm debugging and performance improvements. AMD’s current MLPerf results were reproduced across systems from several manufacturers, showing that its software can operate outside AMD’s internal test environment.

CUDA still benefits from almost two decades of accumulated libraries, documentation, developer familiarity and production debugging. New models and optimization techniques usually arrive on Nvidia first. Companies with large CUDA codebases must also account for migration work and engineer retraining.

Nvidia keeps extending the software layer above CUDA. Dynamo coordinates distributed inference, while TensorRT, NCCL and NeMo cover deployment, communication and model development.

ROCm now clears the threshold required for AMD to win large contracts. Choosing AMD still carries more operational risk than choosing Nvidia.

Q10Is Nvidia vs AMD now a rack-scale systems race?

Nvidia and AMD are increasingly competing through complete AI racks rather than individual GPUs, and Nvidia currently leads that transition.

A frontier AI system depends on far more than accelerator speed. Thousands of processors must exchange data, remain synchronized, recover from failures and stay busy despite changing workloads. CPUs, switches, network cards, storage, cooling and orchestration software all affect the useful output of the cluster.

Nvidia’s Vera Rubin NVL72 rack combines 72 Rubin GPUs with 36 Vera CPUs, NVLink 6 switches, ConnectX network interfaces and BlueField data-processing units. The broader platform also includes dedicated storage and Ethernet racks.

Nvidia said seven Rubin components had entered full production, with systems becoming available through major cloud and hardware partners during the second half of 2026. More recently, Japan’s Noetra announced a national AI facility containing 27,500 Rubin GPUs, 13,750 Vera CPUs and 140 megawatts of data-center capacity.

AMD’s Helios platform follows a comparable 72-GPU structure. It brings together MI455X accelerators, Venice EPYC CPUs, Vulcano network interfaces, UALink connectivity and ROCm. AMD’s acquisition of ZT Systems added rack-design and customer-enablement teams that previously built systems for hyperscale buyers.

Nvidia controls most of its critical components and optimizes them as one proprietary platform. AMD uses more open standards and gives customers greater freedom to modify the design.

Nvidia currently leads in execution. Rubin components are in production, while Helios is only beginning its defining deployment phase.

Q11Can AMD catch Nvidia in AI networking?

AMD remains far behind Nvidia in AI networking, one of the most important parts of large-scale AI infrastructure.

Nvidia generated $14.8 billion of Data Center networking revenue in its latest quarter, up 199% year over year. That single business produced about 44% more revenue than AMD generated across its entire company.

The networking operation is also growing faster than Nvidia’s compute business. Its 35% sequential expansion shows that customers increasingly purchase networking alongside GPUs.

Large AI clusters lose enormous amounts of potential output when accelerators wait for data or for other processors to complete their work. Nvidia controls NVLink inside its racks and sells both InfiniBand and Spectrum-X Ethernet for larger clusters.

AMD has assembled the necessary pieces. Pensando provides network interfaces and data-processing units, while Helios uses Vulcano NICs and UALink for scale-up communication. AMD has also partnered with Celestica and HPE on switches for MI450 systems.

The architecture looks credible. Its commercial footprint remains tiny beside Nvidia’s networking business.

Q12Can customers access AMD GPUs as easily as Nvidia GPUs?

Customers can access AMD GPUs through several major clouds today, but Nvidia still offers broader and more standardized distribution.

AMD hardware is available through Azure, Oracle, Vultr and DigitalOcean, among others. Oracle’s MI355X infrastructure can scale into very large clusters, while smaller platforms give developers access to individual or eight-GPU configurations.

The AMD offer remains inconsistent across providers. One cloud may sell MI300X, another MI325X and another MI355X. Regional availability, cluster sizes and software images vary.

Nvidia’s distribution is broader. AWS, Google Cloud, Microsoft Azure and Oracle all plan to offer Rubin infrastructure. CoreWeave, Lambda, Nebius, Nscale and several other AI-cloud providers are also preparing the platform.

Nvidia additionally works with Dell, HPE, Lenovo, Supermicro, Cisco and a large network of regional manufacturers. Customers can rent Nvidia capacity, install systems themselves or use a managed AI provider.

AMD has secured enough distribution to compete for serious workloads. Nvidia remains easier to find and easier to scale in a familiar configuration.

Q13Does AMD’s open ecosystem give it an advantage over Nvidia?

AMD’s open ecosystem gives hyperscalers more control than Nvidia’s tightly integrated platform, which helps AMD enter strategically important accounts.

Helios supports UALink, Ultra Ethernet and Open Compute Project rack specifications. Customers can combine AMD components with technology from other suppliers instead of adopting one vendor’s complete architecture.

Meta’s agreement provides the clearest proof. Its initial deployment uses a custom MI450-based accelerator, Helios racks and AMD server CPUs. Meta is buying an architecture it can shape around its internal workloads.

Nvidia has responded by supporting standard Ethernet, contributing designs to the Open Compute Project and offering NVLink Fusion for customers developing their own processors.

AMD’s openness creates a credible reason for hyperscalers to maintain a second platform. Nvidia’s tighter integration remains more attractive when rapid deployment and operational certainty are the priority.

Q14Is AMD releasing AI products as fast as Nvidia?

AMD has accelerated its AI product cadence, while Nvidia currently turns new architectures into deployable systems faster.

AMD has moved from MI300X to MI325X, MI350-series products and the coming MI450 family on a roughly annual cadence. MI500 is already on its roadmap, suggesting that AMD intends to maintain that rhythm beyond Helios.

The progress extends beyond silicon. ROCm releases now arrive alongside hardware improvements, AMD is developing full rack designs, and the retained ZT Systems teams help customers qualify and deploy those systems.

Nvidia follows a similarly aggressive cadence through Hopper, Blackwell, Blackwell Ultra and Rubin. Its larger advantage appears during the transition from announcement to production.

Rubin’s principal components entered production well before partner availability. Nvidia lined up the four largest public-cloud platforms, numerous specialist AI clouds, server manufacturers and frontier laboratories around the architecture before the broad customer rollout.

AMD has collected impressive commitments for MI450 and Helios, but those commitments place more weight on upcoming execution. The next test involves delivering custom accelerators, CPUs, switches and complete racks together at gigawatt scale.

AMD is moving faster than its historical reputation suggests. Nvidia still coordinates its supply chain and partner ecosystem more effectively.

Q15Can AMD beat Nvidia by offering cheaper AI compute?

AMD can undercut Nvidia on some cloud workloads, but cheaper GPU rental does not automatically produce lower AI costs.

DigitalOcean lists 12-month reserved MI325X capacity at $2.10 per GPU-hour and Nvidia B300 capacity at $5.65. The products differ in generation and performance, so the 63% price difference is not a direct performance-per-dollar comparison.

AMD’s larger memory configurations and competitive inference benchmarks can make the economics attractive for suitable workloads.

Customers must also account for networking, power, software engineering, utilization and downtime. Transparent contract pricing is too sparse to calculate a reliable market-wide cost per token.

AMD clearly has room to compete on price. The overall advantage depends on the workload and the customer’s ability to optimize ROCm.

Q16Does AMD have enough money to compete with Nvidia?

AMD has enough cash and operating income to remain a serious Nvidia competitor, although Nvidia can invest at a completely different scale.

AMD ended its latest quarter with $12.3 billion in cash and short-term investments. It generated almost $3 billion of operating cash flow and spent $2.4 billion on research and development, equal to roughly 23% of revenue.

Nvidia spent $6.3 billion on research and development and produced $50.3 billion of operating cash flow during one quarter. Its quarterly cash generation was around 17 times AMD’s.

AMD is financially secure enough to fund new GPUs, server CPUs, networking products and software. Nvidia can finance all those layers simultaneously while also investing in suppliers, cloud partners and adjacent markets.

Q17Are hyperscaler AI chips a bigger threat to Nvidia than AMD?

Google, Amazon, Microsoft and Meta may place a larger long-term limit on Nvidia than AMD does, because each company is building chips for its own enormous workloads.

Google’s Ironwood TPU is generally available in pods containing as many as 9,216 chips. AWS offers Trainium3 UltraServers with up to 144 chips per system. Microsoft has introduced Maia 200, while Meta plans several new generations of MTIA accelerators.

These companies can design hardware around their own models, data centers and software. They do not need to support every workload served by Nvidia or AMD.

Every workload transferred to a TPU, Trainium, Maia or MTIA chip is unavailable to both merchant suppliers. AMD cannot assume that companies seeking alternatives to Nvidia will automatically buy Instinct GPUs.

AMD’s strongest position may be as the external alternative for customers that want more choice but cannot justify creating an internal chip program.

Q18Who is winning Nvidia versus AMD right now?

Nvidia is clearly winning the AI infrastructure race today, while AMD has finally become credible enough to challenge individual parts of Nvidia’s platform.

Commercial scale carries the greatest weight. Nvidia’s Data Center operation is approximately 13 times larger than AMD’s, and Nvidia is growing faster from that much larger base. AMD’s new contracts could change the trajectory later, but future gigawatts do not cancel current revenue.

Full-system execution comes next. Nvidia already sells a mature combination of accelerators, CPUs, networking, storage and software. As discussed above, Rubin components are in production and large deployments are beginning to appear. AMD’s Helios strategy is convincing, though its most important production test still lies ahead.

Ecosystem control completes the lead. CUDA, Nvidia’s libraries, its networking products and its distribution network reinforce one another. AMD has greatly improved ROCm and can now compete on important benchmarks, but customers still accept more engineering risk when moving away from Nvidia.

AMD deserves credit for changing the nature of the debate. Its accelerators sometimes approach or exceed Nvidia performance. Production customers are using them. OpenAI and Meta have made multi-gigawatt commitments. Oracle is preparing a 50,000-GPU MI450 cloud cluster.

These achievements give AMD a realistic opportunity to become the default second supplier. They do not put AMD near Nvidia’s overall market position yet.

Over the next several quarters, AMD needs to convert announced capacity into delivered systems, demonstrate reliable Helios clusters at tens of thousands of GPUs and grow accelerator revenue faster than Nvidia. Nvidia needs to prove that Rubin’s integrated architecture continues to produce better customer economics despite pressure from AMD and custom silicon.

The answer remains decisive: Nvidia is winning by a wide margin. AMD is catching up in enough areas to become strategically important, but it has not begun closing the overall commercial gap.

Nvidia versus AMD across the AI infrastructure stack