In the race to supply the infrastructure of artificial intelligence, two semiconductor giants occupy very different positions — Nvidia as the entrenched sovereign, AMD as the patient challenger. Nvidia built its dominance not merely through hardware, but through nearly two decades of software cultivation that made its chips the default language of an entire engineering generation. Yet as AI computing evolves from the brute labor of model training toward the constant, high-volume work of inference, the terrain is shifting — and the company that wins the next phase may not be the one that won th
AMD Poised to Challenge Nvidia as AI Market Shifts to Inference
Inference could become the much larger market, and AMD is better positioned for it.
Mark
So the story here is that Nvidia has built an unbeatable advantage through software, but AMD might win anyway because the market is changing?
Mimi
Not quite. Nvidia's advantage in software is real and probably durable for training work. But inference—which could become the much larger market—is a different game. It's less about raw power and more about cost and efficiency. That's where AMD can compete.
Luke
How confident are we that inference will actually become that much larger? The article cites Chamath Palihapitiya saying it could be 100 times bigger, but that's one person's opinion, not a market forecast. And "100 times" is a huge range from "2 times."
Mimi
Fair point. The article acknowledges that uncertainty—it says "whether the inference market becomes 2 times bigger or 100 times." The exact multiple matters less than the direction. Inference is growing, and AMD is better positioned for it.
Mark
But Nvidia still leads in inference too, right? It's not like AMD is starting from zero.
Mimi
Correct. Nvidia has the lead in inference, but the gap is narrower than it is in training. That's the key difference. AMD doesn't need to take massive share—just enough to move the needle on its much smaller revenue base.
Luke
The valuations are similar, but Nvidia is growing faster. How do you square that with the idea that AMD will outperform?
Mimi
The argument is that Nvidia's growth will slow as it gets bigger, while AMD could accelerate if it captures inference market share. But that's a forward-looking bet, not something we can verify today.
Mark
What's the biggest risk to this thesis?
Mimi
If AI spending slows unexpectedly, both companies suffer. But AMD suffers more because it's smaller and more dependent on growth to justify its valuation.
Luke
And we should note: the article doesn't have hard data on AMD's inference market share or trajectory. It's making an argument about what *could* happen, not what *is* happening.
Mimi
True. The inference opportunity is real, but whether AMD actually captures it is still an open question.
O Pulso
- Nvidia's 80% GPU market share and $39.1 billion in quarterly data center revenue represent a moat built as much from software loyalty as from silicon superiority — its CUDA platform, seeded into universities since 2006, has made switching costs nearly invisible yet nearly insurmountable.
- AMD trails by a factor of ten in AI data center revenue, and its ROCm software platform still carries the scars of arriving a decade late — thinner documentation, narrower hardware support, and a steeper climb to developer trust.
- The fault line cracking Nvidia's fortress runs between training and inference: training is where CUDA's maturity is irreplaceable, but inference — the constant, everyday act of running AI — rewards cheaper, adequate hardware over premium, optimized silicon.
- Industry voices suggest the inference market could eventually be 2 to 100 times larger than training, meaning AMD's lower-cost GPUs and improving software stack could capture enormous volume even without matching Nvidia's technical depth.
- Both stocks trade at nearly identical valuations, but AMD's smaller revenue base means percentage growth can remain high even as absolute gains stay modest — the asymmetric upside scenario that makes it the more speculative, more interesting bet.
In the race to supply the infrastructure of artificial intelligence, two semiconductor giants occupy very different positions — Nvidia as the entrenched sovereign, AMD as the patient challenger. Nvidia built its dominance not merely through hardware, but through nearly two decades of software cultivation that made its chips the default language of an entire engineering generation. Yet as AI computing evolves from the brute labor of model training toward the constant, high-volume work of inference, the terrain is shifting — and the company that wins the next phase may not be the one that won the last.
The artificial intelligence chip market has narrowed to two serious competitors, and the question investors are asking is not whether both will grow, but which will grow harder as the nature of AI computing itself begins to change.
Nvidia's position is formidable by almost any measure — over 80 percent GPU market share, $39.1 billion in data center revenue last quarter, and 73 percent year-over-year growth. AMD, by comparison, posted $3.7 billion in data center revenue with 57 percent growth. The gap is real. But the gap was built on something deeper than hardware.
In 2006, Nvidia released CUDA, a free programming platform that let developers use GPU hardware for tasks far beyond video game graphics. The company seeded it into universities and research labs, making it the default tool through which a generation of engineers learned to program GPUs. By the time AMD launched its competing platform, ROCm, roughly a decade later, the advantage had hardened into something close to permanent. ROCm arrived with less documentation, narrower hardware support, and a steeper learning curve. Nvidia, meanwhile, kept building — layering specialized AI libraries on top of CUDA until each new tool made its chips harder to leave behind. This is network effect in its most durable form.
Yet a crack runs through this fortress, and it follows the boundary between two fundamentally different kinds of AI work. Training a large language model demands raw computational power and the kind of software maturity that only CUDA reliably provides — this is where Nvidia's dominance is nearly absolute. Inference is a different problem. Once a model is trained, running it to answer questions or generate predictions is computationally lighter, and what matters most is cost, latency, and power efficiency. AMD's GPUs tend to be cheaper, and ROCm, while still trailing CUDA, is considered adequate for most inference workloads.
The market dynamics here favor AMD's trajectory. Training is a one-time cost per model; inference runs constantly — every query, every recommendation, every automated response. Some analysts have suggested the inference market could eventually be 100 times larger than training. Even a more conservative multiple of two or three would create enormous room for a lower-cost competitor to capture meaningful share.
Both stocks trade at similar forward valuations — Nvidia just above 32 times earnings estimates, AMD at 28 times. Nvidia is growing faster in absolute terms, but from a base ten times larger, where the mathematics of scale begin to work against it. If inference expands as expected and AMD captures even a fraction of it, the smaller company could sustain higher percentage growth than the giant. That is the scenario in which AMD outperforms.
The shared risk is straightforward: both companies depend on continued heavy AI infrastructure spending, and both would suffer if that spending contracted. Nvidia remains the safer, deeper-moated choice. AMD is the higher-risk, higher-reward position — a company quietly positioned to compete for a market that may ultimately dwarf the one Nvidia currently owns.
The semiconductor market for artificial intelligence has become a two-horse race, and the question that matters to investors right now is not whether these companies will win, but which one will win harder as the nature of AI computing shifts beneath them.
Nvidia sits at the top of the GPU market with commanding dominance—over 80 percent market share, data center revenue that hit $39.1 billion last quarter with 73 percent growth year-over-year. AMD, the challenger, pulled in $3.7 billion in data center revenue with 57 percent growth. The gap is real and substantial. But the gap tells only part of the story.
Nvidia's fortress was built on software. Back in 2006, the company released CUDA, a free programming platform designed to let developers write code for Nvidia GPUs beyond their original purpose of rendering video game graphics. The company seeded universities and research labs with the tool, making it the default language in which a generation of engineers learned to program GPUs. By the time AMD launched ROCm—its answer to CUDA—roughly a decade later, the advantage had calcified. ROCm arrived with less hardware support, thinner documentation, and a steeper learning curve. Nvidia, meanwhile, kept building. It layered CUDA X on top of the original platform: specialized libraries and tools tuned specifically for AI work. Each new tool made Nvidia's chips stickier. Each developer trained on CUDA became a reason for the next company to buy Nvidia. This is network effect in its purest form.
Yet there is a crack in this fortress, and it runs along the fault line between two different kinds of AI work. Training a large language model is brutally hard—it demands raw computational power, sophisticated optimization, and the kind of software maturity that CUDA provides. This is where Nvidia's dominance is nearly absolute. Inference, by contrast, is the easier problem. Once a model is trained, running it to generate predictions or answers is computationally lighter work. What matters in inference is latency, power consumption, and cost. AMD's GPUs tend to be cheaper than Nvidia's. ROCm, while still trailing CUDA, is considered adequate for most inference tasks.
The market dynamics are shifting in AMD's favor. Training AI models will always be necessary, but it is a one-time cost per model. Inference happens constantly—every time someone asks ChatGPT a question, every time a recommendation engine makes a suggestion, inference is running. Some venture capitalists and industry analysts, including Chamath Palihapitiya, have suggested the inference market could eventually be 100 times larger than the training market. Even if that estimate proves too bullish, even if inference becomes only two or three times larger, the sheer volume of that market creates room for AMD to capture meaningful share at lower prices.
Both stocks trade at similar valuations. Nvidia's forward price-to-earnings ratio sits just above 32 times analyst estimates for this year; AMD is at 28 times. Nvidia is growing faster in absolute terms. But Nvidia's AI data center revenue is now ten times AMD's. The law of large numbers begins to apply. A company growing from a much smaller base can sustain higher percentage growth even if its absolute growth is smaller. If the inference market expands as expected, and if AMD can capture a meaningful slice of it with cheaper, adequate-for-the-job hardware, the smaller company could begin growing faster than the giant. That is the scenario in which AMD stock outperforms.
The risk cuts both ways. Both companies benefit from continued heavy spending on AI infrastructure, which still appears to be in early stages. Both would suffer if that spending unexpectedly contracted. For investors, the calculus is straightforward: Nvidia is the safer bet, the proven leader with deeper moats. AMD is the higher-risk, higher-reward play—the company positioned to capture a market that could dwarf the one Nvidia currently dominates. Owning both is probably the wisest move, but if forced to choose based on upside potential, AMD has the more interesting case to make.
Citações Notáveis
The inference market could be up to 100 times larger than the market for training AI models.— Venture capitalist Chamath Palihapitiya (cited in article)
AMD doesn't need to take a lot of share in what could be a rapidly growing market to really make a big difference off its much smaller AI data center revenue base.— Article analysis
Quer a matéria completa?
Leia o original em The Motley Fool Australia ↗
Technology