Signals Inbox·July 25, 2026·AI Chips

Can AMD get bigger than Nvidia?

AMD can become a much larger AI company and a serious second supplier to hyperscalers. Overtaking Nvidia is another matter: the financial gap is still widening, and closing it would require years of near-perfect execution while Nvidia slows.

Q1What would “bigger than Nvidia” actually mean?

AMD would need to overtake Nvidia in companywide revenue, profit or market value for the claim to mean anything substantial.

Winning one benchmark would not qualify. Neither would selling more server CPUs, offering more GPU memory or taking the lead in a particular inference workload. Those achievements could help AMD build a much larger business, but they would leave the overall ranking unchanged.

Revenue gives us the cleanest measure of business size. Profit shows how much money each company can reinvest or return to shareholders. Market value tells us how investors price their expected future earnings. Looking at all three prevents a narrow product victory from being mistaken for a companywide takeover.

This is a much harder test than asking whether AMD can compete with Nvidia. AMD has already cleared that bar. The open question is whether it can close an enormous financial gap while Nvidia keeps launching products, winning customers and expanding into networking, CPUs, software and complete data-center systems.

Q2How far behind is AMD today?

Revenue, profit and market value all point the same way: Nvidia is far larger today.

AMD reported $10.3 billion of revenue in its latest quarter. Nvidia reported $81.6 billion over a roughly comparable period, nearly eight times as much. Nvidia’s $75.2 billion of data-center revenue was about 13 times AMD’s entire $5.8 billion data-center segment, even though AMD’s figure also includes its successful EPYC server CPUs.

The profit difference is wider still. AMD produced $1.5 billion in GAAP operating income, compared with Nvidia’s $53.5 billion. Nvidia earned more operating profit in roughly ten days than AMD earned during the whole quarter.

Their stock-market values reflect that difference. At the time of this review, AMD was valued at approximately $831 billion and Nvidia at nearly $5 trillion. Investors have already placed AMD among the most valuable companies in the world, yet Nvidia remains about six times larger.

AMD versus Nvidia, latest reported quarter

Q3Is AMD closing the revenue gap right now?

No. AMD is growing quickly, but the distance between the two companies is still getting wider.

AMD’s quarterly revenue rose 38% from the previous year, adding about $2.8 billion. Nvidia grew 85% and added approximately $37.6 billion. Nvidia therefore created more than 13 times as much new quarterly revenue during the same comparison period.

The same pattern appears in their latest full years. AMD grew 34% to $34.6 billion, an increase of nearly $8.9 billion. Nvidia grew 65% to $215.9 billion, adding roughly $85 billion. For every extra dollar AMD generated, Nvidia added close to ten.

Large companies usually find it harder to maintain high percentage growth because they begin from a bigger base. Nvidia is doing the opposite. Its growth rate exceeds AMD’s despite starting with more than six times the annual revenue.

AMD may accelerate once its next wave of AI systems ships in volume. That possibility deserves to be taken seriously. The figures available today, though, show a widening gap rather than the start of a catch-up.

Q4Has AMD become a serious Nvidia rival in AI?

Yes. AMD has earned a place among the small number of companies capable of supplying advanced AI computing at hyperscale.

AMD’s data-center revenue reached $5.8 billion in its latest quarter, rising 57% from the previous year. The segment generated around 57% of AMD’s total revenue, which means data centers have become the centre of the company rather than one promising activity among several.

Customer behaviour supports that conclusion. Meta says it already runs significant MI300 and MI350 deployments alongside millions of EPYC processors. Microsoft has used MI300X accelerators for Azure OpenAI workloads and is now preparing AMD’s complete Helios system for wider Azure deployment. OpenAI has selected AMD as a multigeneration compute partner.

AMD hardware has also appeared in recent MLPerf training and inference rounds through AMD itself and system suppliers such as Dell. These tests do not crown one universal winner, since the results depend on the model, system size and software configuration. They do confirm that AMD accelerators can complete standard industry workloads under the same accuracy rules as Nvidia systems.

That is a major change from the period when customers mainly discussed AMD as a possible future alternative. The company now has real products, production workloads and multiyear commitments. Nvidia still dominates, but AMD’s challenge has moved beyond presentations and isolated trials.

Q5Do AMD’s latest Microsoft, Meta and OpenAI deals change the race?

They change AMD’s prospects more than they change today’s ranking.

The freshest development comes from Microsoft, which says it will deploy AMD Helios systems at scale on Azure for frontier-model inference, Azure AI services and customer applications. The agreement also covers new EPYC virtual machines and wider use of AMD Pensando networking. This is AMD’s broadest Microsoft deployment so far because it reaches across GPUs, CPUs, networking and software.

Meta and OpenAI have each agreed to deployments reaching six gigawatts across several product generations. The first gigawatt for each is expected to use MI450-based hardware, with shipments scheduled to begin in the second half of 2026. Oracle has separately announced an initial 50,000-GPU MI450 cluster.

These customers are no longer purchasing a few AMD servers to test bargaining leverage. They are helping AMD shape future chips, racks and software around their own workloads. That gives AMD predictable demand, production feedback and reference customers that other buyers will trust.

Still, the largest figures describe multiyear ambitions. Six gigawatts represents the full potential deployment rather than hardware already installed and paid for. Technical milestones, shipment volumes, power availability and later purchasing decisions will determine how much of the headline capacity becomes revenue.

AMD also offered powerful incentives. OpenAI received a performance-based warrant covering up to 160 million AMD shares, and Meta received another warrant of the same maximum size. The shares vest only when deployment and other milestones are reached, but the potential dilution shows how much AMD was prepared to offer for two anchor customers.

What AMD’s largest announced AI commitments prove today

Q6Can AMD win AI inference even if Nvidia keeps the training lead?

Inference is AMD’s best route into Nvidia’s territory.

Frontier-model training tends to reward a mature, tightly connected platform. Thousands of GPUs must communicate reliably for weeks or months, and delays can waste enormous amounts of money. Nvidia’s hardware, networking and CUDA tools give it a strong position in those demanding projects.

Inference gives AMD more room. Once a model has been trained, large customers can optimize a narrower set of workloads around cost per token, memory capacity, energy use and response speed. Microsoft explicitly plans to use Helios for frontier-model inference. Meta has also described efficient inference as a central aim of its AMD partnership.

The largest hyperscalers are especially well suited to AMD. They control their applications, run enormous volumes and employ engineers who can tune software for a particular chip. Even a small saving per request becomes valuable when a company serves billions of model queries.

Smaller companies face a different calculation. They generally have fewer engineers, a wider mix of applications and a greater need for ready-made software support. Nvidia remains the easier choice because more tools, tutorials, libraries and experienced developers already support CUDA.

AMD can therefore build a substantial inference business without becoming the standard supplier for every AI user. Winning a few of the world’s highest-volume workloads could generate tens of billions of dollars. Nvidia is pursuing those same workloads aggressively, so inference gives AMD an opening, not a private lane.

Q7Is ROCm finally good enough to compete with CUDA?

ROCm has crossed the line from experimental to usable, although CUDA still makes Nvidia easier to buy.

The strongest evidence comes from actual deployments. Microsoft already runs Azure OpenAI workloads on MI300X hardware, while Meta and OpenAI are preparing much larger future systems. Those companies would not base important AI capacity on AMD if its software required constant rescue work.

ROCm now includes the compilers, libraries, drivers and development tools needed for large AI and high-performance-computing workloads. AMD supports major frameworks such as PyTorch and offers migration tools for CUDA applications. The ROCm 7 generation also runs across a wider range of AMD data-center and local hardware.

CUDA retains nearly two decades of accumulated advantages. Developers know it, universities teach it and software companies test against it first. Nvidia has also built specialised libraries for model training, inference, data processing, simulation, robotics, healthcare and industrial computing.

AMD does not need millions of developers to prefer ROCm. It needs the software to become reliable enough that large customers can use AMD without delaying projects or rebuilding every application. Recent customer decisions suggest that threshold has been reached for selected workloads.

The remaining gap concerns convenience and breadth. A hyperscaler can dedicate teams to optimisation. An ordinary enterprise would rather install its software and move on. CUDA still wins that everyday usability test.

Q8Can AMD compete with Nvidia as a full AI system supplier?

AMD can now sell a complete AI rack. Nvidia remains much better at turning a rack design into a worldwide platform.

Helios combines 72 MI455X accelerators with EPYC “Venice” CPUs, Pensando networking and AMD software. Microsoft plans to deploy the system on Azure, while Meta’s first large MI450 installation will also use the Helios architecture. AMD has strengthened its system-design team through the ZT Systems acquisition and is investing more than $10 billion across Taiwan’s manufacturing and packaging ecosystem.

This puts AMD in a stronger position than simply shipping accelerator cards. Customers can evaluate one rack-level design instead of assembling CPUs, GPUs and networking from scratch. AMD can also capture more of each project’s budget.

Nvidia has already repeated this process across several generations. Its latest quarterly networking revenue reached $14.8 billion, up 199% from the previous year. That networking activity alone produced more revenue than AMD’s entire company did during the quarter.

Vera Rubin is now in production, with systems expected from AWS, Google Cloud, Microsoft, Oracle and specialist AI-cloud providers. One recently announced Japanese project alone includes 27,500 Rubin GPUs and 13,750 Vera CPUs. Nvidia is selling an established platform through clouds, server manufacturers, governments and research centres around the world.

AMD’s open design may appeal to customers that want more control and less dependence on one supplier. Its immediate challenge is simpler: deliver Helios on time, make it reliable and repeat the process with the next generation. Until those deployments are running in volume, Nvidia holds the clear advantage in complete systems.

Q9Does AMD’s broader chip portfolio help it catch Nvidia?

AMD’s wider portfolio makes the company stronger, but the extra businesses are too small to erase Nvidia’s AI lead.

EPYC is the most useful part of that portfolio. Server CPUs handle data preparation, storage, orchestration and general computing around AI accelerators. AMD can sell EPYC inside Nvidia-based clusters, then use those customer relationships to introduce Instinct GPUs and Pensando networking.

Meta says it has deployed millions of EPYC processors and will become a lead customer for the next “Venice” and “Verano” generations. Microsoft is also expanding its EPYC virtual machines. AMD recently began ramping Venice on TSMC’s 2-nanometre process, giving it another chance to strengthen its server position.

PCs and gaming provide useful revenue, but their scale is modest beside AI infrastructure. AMD’s combined client and gaming segment generated $14.6 billion during 2025. Nvidia’s data-center operation generated more than that amount in each month of its latest quarter.

Embedded chips add diversification and steady customer relationships in industrial, communications and automotive markets. They can soften the damage from a weak PC or gaming cycle. They will not drive a companywide overtake unless the economics of the semiconductor market change dramatically.

AMD’s broad product range gives it more ways to enter a customer account. Nvidia’s narrower focus has produced far more revenue because AI infrastructure has become the most valuable computing market. Breadth helps AMD compete. AI execution decides whether it ever catches up.

Q10Can AMD afford to keep up with Nvidia’s product pace?

AMD can fund a serious challenge. Nvidia can fund several challenges at once.

AMD spent $8.1 billion on research and development in its latest full year, up 25% from the previous year. That represented almost one-quarter of its revenue, showing how aggressively it is funding new CPUs, GPUs, networking products and software.

Nvidia spent $18.5 billion in its latest full year, more than twice AMD’s amount. Its R&D spending rose 43%, including a 79% increase in computing and infrastructure costs. Nvidia ended the period with approximately 31,000 employees working in research and development.

Cash generation creates an even larger difference. AMD’s continuing operations produced $6.5 billion in operating cash during its latest full year. Nvidia produced $102.7 billion, almost 16 times as much.

That money lets Nvidia develop several chips at the same time, reserve advanced manufacturing capacity, build internal computing systems, support software projects and invest in other companies. Nvidia recently committed $27 billion to multiyear cloud services that it expects to use largely for research and development.

AMD has enough resources to remain competitive, especially because it outsources chip production and works closely with large customers. Its choices still carry a higher opportunity cost. Money spent improving ROCm cannot also fund a new consumer GPU, networking chip or acquisition.

A product delay would hurt AMD more severely because each new Instinct generation carries so much of its expected growth. Nvidia has more products, more cash and a larger installed base to absorb mistakes. That is a huge advantage.

Q11Could better margins help AMD close the gap faster?

Higher margins would make AMD much stronger. Margin improvement alone cannot close a gap of this size.

AMD’s latest GAAP gross margin was 53%, against 74.9% for Nvidia. Its operating margin was 14%, while Nvidia kept around 66 cents of GAAP operating profit from every dollar of revenue.

AMD’s reported margin includes costs linked to acquisitions and other items that make the comparison somewhat harsher. Its non-GAAP operating margin reached 25%, which better reflects the performance management uses to judge the business. Nvidia’s non-GAAP operating margin still exceeded 65%.

The difference comes partly from product mix. AMD continues to sell console chips, PC processors and embedded products with lower margins than high-end AI systems. Nvidia now earns most of its revenue from expensive data-center platforms sold into a market where demand remains extremely strong.

A growing share of Instinct revenue should lift AMD’s profitability. The company has set a long-term target above a 35% non-GAAP operating margin. Reaching that level would produce much more cash for software, system development and future chips.

Nvidia could also face pricing pressure as AMD and custom accelerators become more credible. Even then, AMD must first ship enough AI systems for the richer product mix to transform its companywide margins. The process has started, but it remains far from complete.

Q12Where is Nvidia genuinely vulnerable?

Nvidia’s weak spots are real: customer concentration, custom chips, export controls and dependence on enormous AI spending.

Two direct customers generated 22% and 14% of Nvidia’s latest annual revenue. The identities were not disclosed in the filing, and direct customers can sometimes be manufacturers or distributors serving several end users. Still, more than one-third of revenue passing through two buyers creates obvious negotiating pressure.

Large cloud companies want alternatives. An AMD deployment gives them another source of supply and strengthens their position during price negotiations. They are also designing internal accelerators, including Google’s TPUs, Amazon’s Trainium and Microsoft’s Maia chips. Nvidia can lose spending to AMD, but much of it may instead move to chips built by the customer.

Export restrictions create another risk. Nvidia excluded China data-center compute sales from its latest quarterly outlook. AMD has also faced charges and sales restrictions related to its MI308 accelerators, so geopolitical pressure can hurt both companies rather than handing the market directly to AMD.

The AI investment cycle itself may eventually slow. Nvidia’s current revenue assumes that cloud companies, model developers, governments and enterprises will continue building infrastructure at historic speed. If those customers struggle to earn a return, orders could fall quickly.

A broad slowdown would probably delay AMD’s catch-up. AMD needs the market to remain large enough for customers to fund second suppliers and new platforms. Its best scenario is continued AI growth combined with gradual diversification away from Nvidia.

Q13Could AMD become more valuable than Nvidia before earning more revenue?

A temporary market-value overtake could happen during extreme volatility. A lasting lead would require AMD’s earnings to move much closer to Nvidia’s.

AMD is currently valued at around one-sixth of Nvidia. That gap is already smaller than the operating-profit difference, showing that investors expect AMD to grow much faster from here.

AMD also trades at a much higher multiple of its trailing reported earnings. Investors are paying in advance for the expected impact of MI450, Helios and the large customer agreements. Strong expectations help AMD’s valuation today, but they leave less room for delays or disappointing margins.

For AMD to pass Nvidia without first approaching its revenue, investors would need to believe that AMD’s future growth and profitability had become clearly superior. That could follow several successful Instinct generations, major market-share gains and a sustained weakening of Nvidia’s pricing power.

Nvidia’s own value would keep changing during that period. AMD is chasing a moving valuation supported by extraordinary cash flow, a large software ecosystem and a leading position in several emerging AI markets.

Stock prices can briefly produce almost any ranking. Holding the lead would require AMD to earn far more money than it does now.

Q14What would have to happen for AMD to overtake Nvidia?

AMD would need nearly everything to go right for years while Nvidia slows sharply.

AMD’s long-term plan calls for companywide revenue growth above 35% a year over a three-to-five-year period. The target is unusually ambitious for a company already producing more than $34 billion in annual revenue.

Even that growth would take time to close the existing gap. Starting with the companies’ latest full-year revenues, AMD growing at exactly 35% annually would need just over six years to catch a completely stagnant Nvidia. If Nvidia kept growing at 10%, the process would take almost nine years.

During those years, AMD would need to turn the Meta, OpenAI, Microsoft and Oracle commitments into repeated orders. Helios would have to ship reliably. ROCm would need to become easier for ordinary companies. AMD would also need enough manufacturing capacity and advanced memory to support rapid growth without damaging margins.

Nvidia would simultaneously need to lose momentum. Custom chips could take some hyperscale demand, AMD could pressure prices and slower returns on AI investment could reduce infrastructure spending. A normal cyclical slowdown would probably be insufficient; Nvidia’s growth would have to stay far below AMD’s for a long period.

This calculation assumes smooth annual growth, which semiconductor companies rarely achieve. Product delays, supply shortages and investment cycles would make the real path far less orderly.

How long AMD would need to catch Nvidia under different growth assumptions

Q15Can AMD get bigger than Nvidia?

No, the evidence available today does not make AMD overtaking Nvidia a likely outcome.

AMD has already achieved something important. It has become the only major merchant-chip company with a credible chance of building a large business beside Nvidia in advanced AI accelerators. Its products now run real workloads, major customers are planning gigawatt-scale installations and its latest Microsoft expansion covers the whole AI system rather than one chip.

That progress can make AMD enormous. A durable place as the preferred second supplier could produce tens of billions of dollars in annual AI revenue. AMD could also benefit from EPYC growth, networking sales and customers seeking more control over their infrastructure.

Nvidia currently operates on another financial level. Its quarterly revenue is almost eight times larger, its operating profit exceeds AMD’s by more than 30 times and its annual cash generation gives it far greater freedom to invest. Its next platform is already entering production while AMD prepares the first large Helios shipments.

The gap may narrow once AMD’s new deployments start contributing revenue. AMD should take some AI share and become a tougher competitor from here. Nothing in the latest financial figures shows it moving towards an overall lead yet.

Our judgment is direct: AMD can become much bigger, more profitable and more important than it is now. Becoming bigger than Nvidia remains possible only under an unusually favourable sequence of events, and it is too remote to serve as the base-case outcome.