Is the AI Chip Market growing now?

Is the AI Chip Market Growing Now?

What actually counts as the AI chip market?

The AI chip market we are measuring here is the hardware directly needed to train and run modern AI models.

That starts with accelerators such as Nvidia GPUs and AMD Instinct chips, but stopping there would miss a growing part of the market. Custom accelerators such as Google TPUs and Amazon Trainium, high-bandwidth memory, networking silicon and optical components increasingly determine how much AI compute a data center can actually deliver.

We do not count every processor that happens to include an AI feature. Smartphone chips, PC NPUs and automotive processors belong to much broader markets where AI is only one function. Including them would make the market look larger while making the growth question less useful.

There is a practical reason to use the wider data-center definition. A recent Semiconductor Industry Association and Deloitte teardown estimated that semiconductors account for roughly 95% of the value of a state-of-the-art AI server rack. Accelerators take the largest share, but memory, CPUs, networking and other chips now represent a meaningful business around them.

Is the AI chip market growing right now?

The AI chip market is growing very fast right now, and the latest company results show something much stronger than a normal semiconductor recovery.

Nvidia's latest Data Center revenue reached $75.2 billion, up 92% from a year earlier. AMD's Data Center business reached $6.7 billion, up 107%. Broadcom generated $10.8 billion from AI semiconductors, up 143%. Marvell's Data Center revenue reached $1.83 billion, up 27%, with the company saying AI demand was driving growth across optics, switching and custom silicon.

If we combine those four latest reported quarters, we get roughly $94.6 billion of relevant revenue compared with about $48.2 billion on their respective year-ago bases. That is almost a doubling. It is not a formal market-size estimate because AMD and Marvell include some non-AI data-center products, and the companies do not close their quarters on exactly the same day. We use it as a breadth check, and the result is hard to dismiss: several large suppliers exposed to different parts of AI computing are expanding at the same time.

The wider semiconductor market is booming too. According to the latest World Semiconductor Trade Statistics data, global chip sales more than doubled year over year during the first half of 2026. Memory grew 305%, while logic grew 45%. So we should not attribute every new semiconductor dollar to AI. What stands out is that the businesses closest to AI infrastructure are repeatedly growing faster than already exceptional industry averages.

Company	Latest relevant quarterly revenue	YoY growth
Nvidia Data Center	$75.2B	92%
AMD Data Center	$6.7B	107%
Broadcom AI semiconductors	$10.8B	143%
Marvell Data Center	$1.83B	27%

Is AI chip growth spreading beyond Nvidia?

AI chip growth is clearly spreading beyond Nvidia now, even though Nvidia still dominates the accelerator market.

One of the clearest changes is where the extra money is going. Amazon says its internally designed chip business has passed a $25 billion annual revenue run rate and is growing at a triple-digit percentage. The number includes Graviton CPUs, Trainium accelerators and Nitro networking rather than pure AI accelerators, but it has grown from more than $10 billion at the beginning of the year to more than $20 billion and then more than $25 billion in successive updates.

Memory gives us another view. Micron's Cloud Memory and Core Data Center businesses generated a combined $25.3 billion in its latest quarter, compared with roughly $4.9 billion a year earlier. Some of that jump comes from sharply higher memory prices, but the scale is still extraordinary. AI infrastructure demand has become large enough to change the economics of the entire memory industry.

Marvell is seeing the expansion from another angle. Data centers already account for 76% of its revenue, and management recently raised its future outlook after what it described as exceptional AI-related bookings across 800G and 1.6T optics, Ethernet switches, custom XPUs and interconnect products.

Nvidia remains the biggest company in AI chips, but the surrounding market is becoming much deeper. An AI cluster now creates major revenue pools in custom accelerators, HBM, networking, optics and advanced packaging before we even reach the servers, buildings and electricity around it.

If you want more recent data on this point, please see our latest AI chip market report.

Is AMD actually catching Nvidia in AI chips?

AMD is becoming a serious AI chip competitor, but it is still nowhere near Nvidia in commercial scale.

AMD's latest Data Center revenue reached $6.7 billion, more than double the year-earlier figure. Nvidia reported $75.2 billion of Data Center revenue in its latest quarter. Nvidia's business is therefore roughly eleven times larger, and AMD's number also includes EPYC server CPUs alongside Instinct accelerators, which makes the clean GPU gap wider than the headline comparison suggests.

Where AMD has changed its position is with customers. Anthropic has agreed to deploy up to two gigawatts of AMD Instinct MI450 systems. Microsoft plans to deploy AMD Helios systems across Azure, while AMD also names Meta, OpenAI, Oracle and other major AI buyers among customers or deployment partners.

AMD's Data Center operating income reached $2.1 billion in its latest quarter. The company has moved well beyond the stage where its AI effort could be dismissed as an interesting product roadmap with little commercial traction.

Catching Nvidia would require something much bigger: several years of faster accelerator growth, much wider software adoption and large deployments converting into recurring sales. We are seeing the beginning of that test now, not evidence that AMD has already won it.

Are custom AI chips becoming a serious business?

Custom AI chips have already become a serious business, with hyperscalers now spending and earning tens of billions of dollars around their own silicon.

Amazon provides the easiest number to understand. Its custom chip business recently passed a $25 billion annual revenue run rate while growing at a triple-digit percentage. Amazon says Anthropic has committed to as much as five gigawatts of Trainium capacity, while OpenAI has made another multi-year commitment covering roughly two gigawatts.

Broadcom shows how much money sits behind hyperscaler chip design. Its latest quarterly AI semiconductor revenue reached $10.8 billion, 143% above the previous year, driven by custom accelerators and AI networking. Broadcom expects the next quarter to reach around $16 billion if current customer programs ramp as planned.

Google is pushing in the same direction. The company has used TPUs internally for years, but it has now started delivering TPU systems into customers' own data centers. Google says most of the revenue from the agreements already signed will arrive next year.

We are past the point where custom AI chips were mostly an insurance policy against Nvidia prices. For Amazon, Google and other hyperscalers, silicon design has become a way to control cost, power consumption and capacity at enormous scale. Merchant GPUs will continue to sell alongside them, but hyperscaler-designed silicon is now a market of its own.

If you want more recent data on this point, please see our latest AI chip market report.

Is HBM memory growing even faster than AI GPUs?

HBM and data-center memory are currently among the fastest-growing parts of the entire AI chip market.

Micron's latest results are unusually revealing. Cloud Memory revenue rose from $3.4 billion to $13.8 billion in a year, while Core Data Center revenue jumped from $1.5 billion to $11.5 billion. Taken together, the two businesses grew more than fivefold.

SK hynix is seeing the same demand wave. Its latest quarterly revenue increased 257% year over year, with the company pointing directly to AI servers, HBM and high-value DRAM as the main drivers. SK hynix has also started mass shipments of HBM4 while signing longer-term agreements with roughly ten major customers.

Industry data confirms the scale. WSTS measured memory semiconductor sales up roughly 305% in the first half of the year.

Revenue is growing much faster than physical bit shipments because memory prices have surged, so a 305% sales increase does not mean the number of chips being consumed rose 305%. The more interesting conclusion is that AI demand has become strong enough to create scarcity and pricing power far outside the GPU itself. HBM used to sit in the background of the AI infrastructure story. Today, it is one of the industry's central bottlenecks.

Is AI networking becoming a major chip market too?

AI networking is becoming a major chip market because enormous GPU clusters are useless if the chips cannot exchange data quickly enough.

Nvidia's Data Center networking revenue recently reached $14.8 billion for one quarter, up 199% year over year. Its networking business is therefore growing substantially faster than its Data Center compute business.

Marvell gives us a second view of the same shift. The company says current AI bookings are coming from 800G and 1.6T optical products, 51.2T Ethernet switches, chip-to-chip interconnect and custom XPU systems. It bought Celestial AI and XConn earlier this year, giving it additional technology around optical interconnect and connectivity between accelerators.

The economics change as clusters get larger. Buying twice as many accelerators does not simply require twice as much useful connectivity. Thousands of chips need to operate almost like one computer, which pushes customers toward faster switches, optical links and increasingly sophisticated rack-scale architectures.

That has created an AI semiconductor market that barely appeared in the original GPU narrative. Networking chips and optical components now capture billions of dollars that would not exist at anything close to today's scale without large AI clusters.

Is TSMC still seeing AI chip demand accelerate?

TSMC is still seeing extremely strong AI-related chip demand, including fresh growth after an already record first half.

TSMC generated $40.2 billion of revenue in its latest quarter, up 33.7% year over year and 12% from the previous quarter. High-performance computing grew 20% sequentially and reached 66% of company revenue, up from 61% one quarter earlier.

Advanced manufacturing has also taken over the mix. Chips produced at 7 nanometers or below accounted for 77% of TSMC's wafer revenue, including 30% at 3 nanometers and the first 3% contribution from 2 nanometers. Those are the nodes used for the most demanding accelerators, CPUs and related processors.

The freshest monthly numbers remain strong. TSMC's latest reported monthly revenue was 44.7% above the previous year, while cumulative revenue for the first seven months was up 37%.

TSMC's HPC category includes some computing products outside AI, so we cannot label two-thirds of its revenue "AI." Still, the combination of a rapidly rising HPC mix, advanced-node demand and continued monthly growth gives us manufacturing-level confirmation that the AI chip cycle remains very strong today.

Is inference creating a second AI chip boom?

AI inference is starting to create a second wave of chip demand, and lately we have begun seeing real revenue and customer deployments rather than just ambitious startup forecasts.

Cerebras gives us one of the cleanest examples. Its latest quarterly cloud and services revenue reached $126 million, up 281% year over year, while total GAAP revenue grew 74% to $180 million. Hardware revenue itself was lower than a year earlier, so almost all of the company's incremental growth came from selling access to inference computing.

Cerebras now has more than 600 megawatts of data-center capacity live or contracted for delivery and reported $25.4 billion of remaining performance obligations. Those commitments will take years to convert into revenue, but they show that specialized inference infrastructure can attract very large commercial contracts.

Etched is much earlier, yet its recent progress is worth watching. Jane Street tested Etched's hardware, received the startup's first production rack and is now deploying it on its own workloads. Etched says it has signed more than $1 billion of customer contracts. Investors subsequently valued the company at $21 billion after a $700 million financing round.

The funding valuation itself tells us very little about the eventual size of Etched's business. The useful part is that a specialized inference chip has moved from prototype silicon into an actual customer's data center.

Inference could eventually consume more aggregate compute than model training because successful models are trained occasionally but queried constantly. We are still early in that transition, but the first commercial numbers are now large enough to take seriously.

If you want more recent data on this point, please see our latest AI chip market report.

Is China building its own AI chip boom?

China is building a fast-growing domestic AI chip market, helped rather than stopped by restrictions on access to the most advanced U.S. accelerators.

Cambricon's first-half revenue reached roughly 6 billion yuan, or about $890 million, up 108% year over year. The company also reported 2.3 billion yuan of net profit. Growth slowed sequentially in the second quarter, so Cambricon is no longer accelerating at the almost absurd rates it posted from a tiny base, but it is now a profitable chip company with substantial sales.

Huawei appears to be operating on a much larger scale. Reuters reported earlier this year that ByteDance, Tencent and Alibaba were seeking additional Ascend 950 chips after the release of DeepSeek V4 increased demand for Huawei-compatible compute. Huawei was preparing roughly 750,000 Ascend 950PR units for the year, according to people familiar with the plans, while production remained constrained by limited access to advanced manufacturing equipment.

China's foundries are benefiting too. SMIC's latest quarterly revenue rose 36% to more than $3 billion, while profit attributable to shareholders more than tripled. The company specifically pointed to strong AI-related chip demand and said it was accelerating new production capacity.

Export controls have changed the competitive structure of the AI chip market without removing Chinese demand. China is increasingly building a parallel ecosystem around Huawei, Cambricon, SMIC and other domestic suppliers. Manufacturing technology remains the biggest constraint, but customer demand is already there.

Are AI chip buyers still running out of capacity?

Major AI chip buyers are still short of computing capacity today, even after adding huge amounts of hardware.

Microsoft provides unusually concrete evidence. The company said roughly two-thirds of one recent quarter's capital expenditure went to short-lived assets, primarily GPUs and CPUs. Microsoft added another gigawatt of capacity during that quarter and says it still expects to remain constrained through the end of the year.

Google is dealing with a similar problem. After reporting a sharp acceleration in Google Cloud, management said it would use additional third-party capacity while internal infrastructure catches up with demand. Google also raised its capital-spending plan because capacity was arriving too slowly relative to customer demand.

Amazon's custom-chip commitments point in the same direction. OpenAI and Anthropic have together committed to several gigawatts of future Trainium capacity, while Amazon says its chip business continues to grow at a triple-digit rate.

This is a useful stress test for the current AI chip boom. Customers are spending unprecedented amounts while simultaneously saying they still need more compute. We would become much more cautious if cloud providers started talking about idle accelerators, falling utilization or excess inventories. Their current language and spending behavior show the opposite.

Can hyperscalers really keep spending this much on AI chips?

For now, hyperscalers can keep funding the AI chip boom because their AI and cloud businesses are growing fast enough to make another year of heavy spending rational.

Google Cloud recently reached $24.8 billion of quarterly revenue, up 82% year over year. AWS reached $42.2 billion, up 37%, its fastest growth in 18 quarters. Microsoft has said its AI business surpassed a $37 billion annual revenue run rate while growing 123%.

The investment needed to keep that growth going is enormous. Microsoft expects around $190 billion of capital expenditure this calendar year. Alphabet recently increased its range to $195 billion to $205 billion. Meta expects $130 billion to $145 billion. Those three companies alone are therefore planning roughly $515 billion to $540 billion of annual capital expenditure.

Only part of that money buys AI chips. Data centers, land, power, cooling and network infrastructure absorb large amounts too. Microsoft, however, has given us a useful clue about the mix by saying that roughly two-thirds of one recent quarter's capex was going into short-lived assets led by GPUs and CPUs.

The uncomfortable part comes later. Higher depreciation, electricity costs and financing needs are already hitting cash flow. If AI revenue stops scaling quickly enough, these companies can cut infrastructure budgets much faster than semiconductor suppliers can adjust factories and inventories. We do not see that break today. The current cloud revenue numbers still support aggressive spending.

Company	Current annual capex indication	What is driving it
Microsoft	~ $190B	Cloud and AI capacity, with GPUs and CPUs representing a large share of short-lived assets
Alphabet	$195B–$205B	AI compute, servers, networking and data centers
Meta	$130B–$145B	AI infrastructure and future data-center capacity

If you want more recent data on this point, please see our latest AI chip market report.

What could actually stop AI chip growth?

The biggest threat to AI chip growth is a slowdown in spending from a very small number of enormous buyers.

Nvidia's own filings show how concentrated the market has become. Three direct customers represented 21%, 17% and 16% of its latest quarterly revenue. Together, three customers accounted for 54% of the company.

That concentration runs through much of the supply chain. Custom accelerators are built around a handful of hyperscaler programs. HBM suppliers depend heavily on a small group of high-end accelerator customers. Advanced packaging capacity is being expanded around projected AI volumes years into the future.

The second risk is economic rather than technical. AI chips keep getting better and cheaper per unit of useful computation. That has been excellent for the market while usage grows even faster, but eventually demand for tokens and AI workloads has to keep outrunning improvements in price-performance. If companies can suddenly handle twice as much AI work with half as many dollars of hardware and usage fails to expand enough, semiconductor revenue growth would slow quickly.

The third risk is that hyperscaler returns disappoint. Hundreds of billions of dollars are currently being committed to infrastructure before the full depreciation cost has flowed through income statements. Strong cloud growth makes that bet look defensible today. A couple of quarters of weaker AI monetization would make boards and investors ask much harder questions about the next data-center build.

We would therefore watch customer capex, accelerator utilization and AI revenue growth before worrying about startup funding rounds or short-term chip prices. Those three variables can actually change the direction of the market.

So, is the AI chip market growing now?

Yes. The AI chip market is growing extremely fast right now, and the expansion has become broader than the original Nvidia GPU boom.

We now see strong growth in accelerators, custom silicon, HBM, networking, optics and advanced foundry production at the same time. Inference is creating another source of demand, while China is building a separate domestic supply chain rather than dropping out of the market.

The strongest part of the case comes from what buyers are doing. Microsoft and Google still describe themselves as capacity constrained after enormous infrastructure investments. Amazon's internal chip business has become a multibillion-dollar operation. Large AI labs are making multi-gigawatt hardware commitments years ahead.

There is one important weakness: a surprisingly small number of companies fund much of this expansion. That makes the AI chip market vulnerable to a sharp capex reset if AI revenue eventually fails to justify the infrastructure being built.

As of now, that reset has not started. Demand is still outrunning available capacity, suppliers across the stack are posting extraordinary growth, and buyers are raising rather than cutting infrastructure plans.

Our judgment is clear: the AI chip market is growing now, and it is still in a powerful expansion phase. The next question is no longer whether the boom exists. It is whether AI usage can keep growing fast enough to support the extraordinary amount of chip capacity now being built.

If you want more recent data on this point, please see our latest AI chip market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.