Nvidia’s AI Boom Is Cooling as Memory Chips Become the New Prize

Nvidia’s stock has slipped from its May high even as profits are still expected to rise, and the bigger surprise is where investors are moving instead: into memory makers such as Micron. The shift matters because it shows the AI buildout is no longer being shaped primarily by graphics chips, but by the memory capacity required to keep data centers running.

What looked like a GPU shortage story last year is evolving into a memory bottleneck story this year. As more cloud giants and AI companies design their own chips, demand for Nvidia’s rented compute is easing in some parts of the market, while prices for high-bandwidth memory and DRAM are surging.

That reversal is leaving Nvidia in an awkward position. The company remains the dominant force in AI hardware, but the market it helped create is now rewarding a different layer of the stack — one that is less glamorous, less visible, and in many ways harder to scale.

Why is Nvidia losing momentum while AI spending keeps rising?

Nvidia is losing some of its market momentum because the AI hardware trade has broadened beyond GPUs. The company’s shares have fallen about 15% from their May peak, even though analysts still expect revenue to keep climbing, and its valuation has come down relative to the broader U.S. stock market.

That combination is unusual. In a typical growth stock story, stronger earnings expectations would often support a richer valuation. In Nvidia’s case, investors appear to be concluding that the most explosive phase of the GPU crunch may have passed, at least for now.

The issue is not that AI spending is slowing. It is that spending is rotating into other parts of the infrastructure chain. Memory chips, networking gear, power, cooling and custom silicon are all taking a larger share of the budget as companies move from emergency AI capacity purchases to a more diversified buildout.

What changed in the GPU market?

What changed is that the fear-driven rush to secure Nvidia GPUs has become less intense. Last year, companies treated access to compute as scarce and urgent, which lifted prices and supported Nvidia’s dominant market position. This year, the market has loosened somewhat as supply improves and alternative chips enter the picture.

That does not mean GPUs are cheap or easy to obtain. It means buyers have slightly more leverage than they did during the peak of the shortage, and prices in some compute marketplaces have drifted down from their highs.

How memory chips became the hot AI trade

Memory chips are now attracting the kind of investor enthusiasm that once belonged almost entirely to GPUs. Micron, one of the largest DRAM manufacturers in the world, has nearly tripled in value over the same period that Nvidia pulled back from its peak.

That surge reflects a different bottleneck. Data centers do not just need fast processors; they need huge amounts of memory to feed those processors fast enough to make modern AI systems efficient. As model sizes grow and inference workloads rise, the need for high-bandwidth memory has become a central constraint.

Unlike Nvidia’s GPU business, which depends on a continual stream of architectural leaps, the memory business is benefiting from a more straightforward dynamic: demand is exploding faster than supply can be expanded.

In practical terms, that has allowed memory vendors to raise prices dramatically. High-bandwidth memory components have become dramatically more valuable because they are essential, difficult to ramp, and embedded in the most expensive AI systems.

Why are memory prices rising so sharply?

Memory prices are rising sharply because demand for data center capacity has outpaced industry planning. Companies underestimated how much memory the AI buildout would require, and the result has been a steep climb in spot prices for DRAM and related components.

DRAM is the standard memory type used in computers and servers, and HBM, or high-bandwidth memory, is a specialized version designed to move data rapidly in and out of processors. The latter is especially important for advanced AI systems, where speed and throughput can determine how efficiently the hardware performs.

The market has not required a revolutionary new chip design to reprice memory. Instead, the existing technology has simply become indispensable at a time when every major AI builder wants more of it than suppliers can comfortably provide.

Wayne Nelms, co-founder and CTO of compute marketplace Ornn, said the pattern comes down to supply and demand: more GPU and accelerator makers are entering the field, and major technology companies are building their own chips, but no comparable wave of companies is making DRAM. He added that unless there is a major breakthrough in HBM, a major shift in supply and demand, or a new entrant in memory, the current market pattern is likely to hold.

What the data says about DRAM and GPU pricing

The price signals in the two markets have moved in opposite directions. DRAM spot prices have climbed aggressively since 2023, while hourly prices for access to Nvidia’s H100 GPUs have eased from a peak reached in May.

That divergence is important because it shows the AI supply chain is no longer behaving like a single market. One layer — computing power — is becoming somewhat more available. Another layer — memory — is becoming scarcer and more valuable.

Here is a simplified snapshot of the trend.

The key takeaway is that the AI economy is not uniform. It can be weak in one component and overheated in another, depending on where the bottlenecks happen to be at a given moment.

How Nvidia helped create the market now shifting away from it

Nvidia’s long-term success helps explain why the current correction feels so ironic. The company did not become the AI leader by accident. It spent years building the software and hardware foundation that made GPUs the default platform for machine learning research and deployment.

CUDA, Nvidia’s programming ecosystem, played a major role in locking developers into the company’s hardware. Once researchers, startups and enterprise teams built their workflows around Nvidia, the company’s chips became the natural first choice for training large models and, later, for serving them at scale.

Its GPU roadmap also pushed performance gains faster than many rivals expected. The company essentially made itself the standard for AI compute by offering both technical capability and software convenience, which is a difficult combination for competitors to match.

That success, however, created the conditions for competition. Once AI spending reached a certain scale, large buyers had both the incentive and the budget to reduce dependence on one supplier.

Why are hyperscalers building their own chips?

Hyperscalers are building their own chips because they want to control costs, reduce dependence on Nvidia and tailor hardware more closely to their workloads. Google, Amazon and Microsoft have all pursued custom silicon, and OpenAI has also entered the field in various ways.

Even when those chips are not as capable as Nvidia’s latest GPUs, they can still be “good enough” for some workloads. That is enough to change the buying behavior of large customers and put downward pressure on market pricing for rented compute.

The result is not Nvidia’s disappearance. It is a market in which Nvidia remains central, but no longer entirely unchallenged at the top of the spending pile.

What does this mean for data center economics?

It means the economics of AI infrastructure are becoming more segmented and more competitive. Buyers are no longer simply asking how many GPUs they can secure; they are asking how to balance processors, memory, networking, storage and power delivery across an entire facility.

That shift matters because each layer has its own supply chain constraints. If compute becomes more available, the next choke point moves elsewhere. That is exactly what appears to be happening now, with memory taking center stage.

For data center operators, this creates a more complicated procurement environment. Cheap or available compute does not help much if memory modules, advanced packaging capacity or power infrastructure remain constrained.

For investors, it changes the way AI infrastructure exposure is judged. The obvious winners from the first wave of the AI boom were chip designers and GPU sellers. The next wave may favor suppliers that sit one step further down the production chain.

Who stands to benefit from the new bottleneck?

Memory suppliers stand to benefit most directly from the new bottleneck. Companies that produce DRAM and HBM are now in a position to capture pricing power as long as demand keeps rising faster than manufacturing capacity.

That does not guarantee an endless rally. Semiconductor cycles are notoriously volatile, and supply expansions eventually catch up. But in the near term, the imbalance appears strong enough to keep pricing elevated.

Other beneficiaries may include equipment makers and foundry-adjacent firms that support memory production, as well as cloud providers that have already diversified away from total reliance on GPU rentals.

A market shaped by success, not failure

The most striking part of Nvidia’s recent pullback is that it is not being driven by a collapse in demand for AI. Instead, it reflects the success of the broader ecosystem Nvidia helped create. Once AI became strategically important enough, every major platform company began trying to build its own path to hardware independence.

That dynamic pushes value outward from a single dominant supplier into a wider supply chain. In one sense, that is healthy for the market. In another, it is a reminder that the most valuable position in a fast-growing industry can change quickly once the industry matures.

Nvidia still sits near the center of AI infrastructure, but the market is no longer rewarding center of gravity alone. It is rewarding scarcity wherever it happens to appear, and right now that scarcity lives in memory.

What makes DRAM so strategically important?

DRAM is strategically important because every modern data center depends on it, and AI workloads require enormous volumes of fast memory to work efficiently. Without enough memory bandwidth, even the most powerful processor cannot perform at its full potential.

That makes DRAM and HBM critical to the economics of training and inference. The chips are less visible than GPUs, but in the current market they may matter just as much to the overall pace of AI deployment.

• Nvidia remains the dominant AI chipmaker, but its stock has softened from its peak.
• Micron and other memory suppliers are benefiting from rising demand and tighter supply.
• Custom chips from major cloud companies are slowly reducing dependence on Nvidia GPUs.
• Memory, not compute, is increasingly the bottleneck in the AI buildout.

What comes next for Nvidia and the AI hardware race?

The next phase of the AI hardware race will likely be defined by diversification. Companies will keep buying Nvidia products, but they will also keep building custom accelerators, expanding memory inventories and searching for ways to reduce infrastructure costs.

If that happens, Nvidia’s growth can remain strong without the stock being rewarded as extravagantly as before. The company can still post rising revenue while the broader market recalibrates its expectations.

The more important question is whether memory supply can keep up. If it cannot, the current pricing surge could persist and deepen. If it can, the market may rotate again, perhaps back toward compute or into another part of the AI stack entirely.

For now, the message from the market is clear: the AI boom has not ended, but its hottest bottleneck has moved. Nvidia remains the face of the industry, yet the most profitable trade may now be the component that keeps those famous chips fed.

Key timeline of the shift

Here is a concise timeline of how the market moved from GPU scarcity to memory scarcity.

In the end, Nvidia is not being punished for failing to deliver. It is being revalued because the AI market it pioneered has become larger, broader and more fragmented than before. The winners are still in semiconductors — but not necessarily where investors first expected them to be.