What is the AI chip market actually worth?

Why do AI chip market estimates differ by hundreds of billions of dollars?

AI chip market estimates differ so wildly because researchers are often measuring completely different things under the same name.

At the low end, some recently updated market reports still put the global AI accelerator market at roughly $40 billion. At the other end, Gartner's latest semiconductor work implies that the AI data-center ecosystem alone is worth more than half a trillion dollars. Both numbers are published as measurements of markets created by AI, yet they differ by more than tenfold. Two published numbers more than tenfold apart; we show how big the market really is today in our report on AI chips.

The gap starts with the product being counted.

A narrow estimate may count dedicated accelerator chips. A wider one can include HBM memory, networking silicon and CPUs. NVIDIA increasingly sells complete computing systems rather than standalone GPUs, making the boundary even harder to draw. Google and Amazon consume large quantities of their own custom chips internally, so much of that computing capacity never appears as a conventional chip sale.

Then there is double counting. TSMC manufactures processors that NVIDIA, AMD, Google and others later use or sell. Adding TSMC's manufacturing revenue on top of the finished semiconductor revenue counts part of the same economic value twice.

We therefore need at least two numbers. The core AI accelerator market is currently around $300 billion to $350 billion. The wider semiconductor market created by AI infrastructure is closer to $400 billion to $570 billion.

What we measure	Current order of magnitude	What it includes
Narrow commercial accelerator reports	~$40B–$45B	Selected accelerator categories
AI processors in 2025	>$200B	AI processing silicon
Data-center AI accelerators today	~$300B–$350B	GPUs and custom AI ASICs
Broader AI semiconductor market	~$400B–$570B	Compute, memory, networking and related silicon
AI server systems	>$460B	Chips plus complete server hardware
Major cloud companies' annual capex	>$880B	Data centers, chips, servers and other infrastructure
Entire semiconductor market	~$1.56T	All semiconductor categories

What should actually count as an AI chip?

The cleanest definition of the AI chip market today is the processors built mainly to run AI workloads, while memory and networking belong in a separate broader estimate.

That first category includes NVIDIA GPUs, AMD Instinct processors, Google TPUs, AWS Trainium, Meta's internal accelerators and similar ASICs. It gives us the closest thing to a true market for AI computing power.

HBM is harder. A modern accelerator depends heavily on high-bandwidth memory, and memory can represent a meaningful part of the package's value. Still, adding all HBM to an accelerator number changes what we are measuring. We have moved from processors to the semiconductor stack around them.

The same applies to networking chips. Large clusters need high-speed switches, NICs and interconnects, but those products do a different job from the accelerator.

CPUs create another problem because AMD EPYC and Intel Xeon processors can support AI servers without being primarily AI chips. Counting every CPU installed beside a GPU would inflate the market.

So we use roughly $300 billion to $350 billion when discussing AI accelerators themselves. When the question is how much semiconductor revenue AI infrastructure generates overall, the higher $400 billion to $570 billion range becomes more useful.

Can the AI accelerator market really be worth only $40 billion?

A global AI accelerator market of roughly $40 billion no longer passes a basic revenue check.

Grand View Research, for example, currently estimates the 2026 AI accelerator market at $42.1 billion. Other commercial market reports land in a similar range.

Now compare that with NVIDIA.

NVIDIA generated $89 billion of Data Center revenue in its latest quarter alone, up 117% from a year earlier. Even allowing for networking, CPUs and complete systems inside that figure, one company's three-month revenue is already more than twice some estimates for the entire annual accelerator market.

Gartner's industry data makes the mismatch even clearer. It measured AI processing semiconductor revenue above $200 billion in 2025, before the latest jump in NVIDIA, AMD and custom-chip spending.

There is clearly a category mismatch behind the $40 billion estimates. They may capture a narrower set of products or rely on definitions that made more sense before rack-scale AI infrastructure exploded.

For anyone trying to understand the economic size of today's data-center AI accelerator business, we would discard the $40 billion figure.

What does NVIDIA tell us about the real AI chip market?

NVIDIA alone tells us that today's AI chip market has already reached the hundreds of billions, even before we add AMD and custom silicon.

NVIDIA's Data Center business went from $47.5 billion in fiscal 2024 to $115.2 billion in fiscal 2025 and $193.7 billion in fiscal 2026.

The newest quarter moved to another level. NVIDIA reported $89 billion of Data Center revenue, compared with $41.1 billion a year earlier. Annualizing one quarter mechanically would produce more than $350 billion, although we would not treat that number as pure GPU revenue.

NVIDIA now sells a growing amount of integrated infrastructure. Blackwell systems combine GPUs, CPUs, NVLink, networking and other components, and complete rack-scale configurations are becoming a larger part of what customers buy.

The older full-year disclosure gives us a cleaner split. NVIDIA reported $162.4 billion of Data Center compute revenue and $31.4 billion of Data Center networking revenue in fiscal 2026.

That $162 billion compute number came before the latest quarterly acceleration. It makes a global accelerator market of several hundred billion dollars much easier to defend than one of several tens of billions.

How big are custom AI chips now?

Custom AI chips have already become a major market, with Broadcom alone heading toward roughly $58 billion of AI semiconductor revenue across its current fiscal year if its latest guidance is met.

Broadcom gives us one of the clearest windows into hyperscaler ASIC spending because it designs custom accelerators and networking chips for large technology companies.

Its AI semiconductor revenue reached $8.4 billion in the first fiscal quarter, $10.8 billion in the second and $16.7 billion in the latest quarter. Broadcom now expects $21.7 billion in the following quarter. Broadcom's AI revenue climbing every quarter is why we list the top companies today in our AI chips market report.

Those four numbers add up to about $57.6 billion.

Some of that is AI networking rather than accelerator silicon, so we should not label all $57.6 billion custom-accelerator revenue. The scale is still striking. Broadcom's latest AI semiconductor quarter grew 221% year over year.

Deployment data point in the same direction. TrendForce expects ASIC-based systems to represent roughly 28% of AI server shipments this year, with Google and AWS ramping their next generation of internally designed processors.

Google is especially important because TPUs already handle a large share of its AI infrastructure. AWS is scaling Trainium, while Meta and Microsoft are also developing more of their own silicon.

Custom AI chips have moved well beyond an experimental side business. They are now one of the biggest pieces missed by any market estimate built mainly from merchant GPU sales.

Is the AI chip market still basically NVIDIA?

NVIDIA still dominates AI chip revenue today, but the rest of the market has become too large to ignore.

AMD's Data Center revenue reached $6.7 billion in its latest quarter, more than double the $3.2 billion recorded a year earlier. The category includes EPYC CPUs as well as Instinct accelerators, so $6.7 billion cannot be treated as pure AI GPU revenue. Still, AMD's acceleration shows that NVIDIA is no longer the only merchant supplier adding billions of dollars of quarterly data-center sales.

Custom chips add a much larger second force. TrendForce currently expects about 27% to 28% of AI server shipments to use ASIC-based architectures, while GPU systems retain roughly 70%.

Revenue share remains more concentrated because NVIDIA sells extremely expensive systems and custom chips can be valued differently when cloud companies deploy them internally.

NVIDIA can therefore hold a dominant share of commercial AI accelerator revenue while representing a smaller share of the actual AI computing capacity being installed.

That gap should widen if Google, Amazon, Meta and Microsoft keep moving stable inference workloads onto their own chips.

How much does HBM add to the AI chip market?

HBM adds tens of billions of dollars to the AI semiconductor market and is currently growing fast enough to distort comparisons with previous years.

Gartner measured HBM revenue above $30 billion in 2025, when it already represented about 23% of global DRAM revenue.

Demand has moved higher since then. TrendForce's latest AI server work says HBM demand is more than doubling this year as Blackwell systems, newer NVIDIA platforms and hyperscaler ASICs consume larger memory configurations.

Price changes make the effect even bigger.

Gartner's newest semiconductor forecast expects total memory revenue to reach roughly $837 billion this year, up from about $220 billion in 2025. That extraordinary increase covers much more than HBM and reflects a much wider memory-price cycle, but AI data centers are one of the major sources of pressure on high-end capacity.

Broad AI semiconductor revenue can therefore rise much faster than accelerator shipments. Buyers may purchase more chips, use more HBM per accelerator and pay more for each unit of memory at the same time.

For that reason, HBM belongs in our roughly half-trillion-dollar AI semiconductor estimate. We keep it outside the narrower $300 billion to $350 billion accelerator figure so the two numbers stay understandable.

How much of the AI chip boom is actually networking?

AI networking has become a multibillion-dollar market of its own because enormous GPU clusters are useless when processors cannot exchange data quickly enough.

NVIDIA's Data Center networking revenue reached $31.4 billion in fiscal 2026, up 142% from the previous year. That business includes NVLink infrastructure, Ethernet and InfiniBand.

Broadcom adds another large pool of AI networking revenue, although the company reports its custom accelerators and networking products together. Its latest AI semiconductor quarter reached $16.7 billion. Networking sits beside accelerators as a separate pool of revenue, and we look at which segments and regions lead in our AI chips market report.

Those numbers explain why broad AI semiconductor estimates rise so far above GPU-only estimates.

A cluster with tens of thousands of accelerators needs switches, network interface chips and high-speed interconnects capable of keeping those processors busy. As clusters get larger, networking becomes a bigger part of the cost.

We therefore include AI networking in the wider semiconductor market while keeping it out of the core accelerator calculation.

Should TSMC's AI revenue be added to NVIDIA and AMD?

TSMC's AI-related manufacturing revenue should not be added directly to NVIDIA, AMD or Broadcom revenue because much of it sits inside the value of chips those companies eventually sell.

TSMC is still extremely useful for checking whether the AI boom is showing up deeper in the supply chain.

High-performance computing represented 66% of TSMC's $40.2 billion revenue in its latest reported quarter, which works out to roughly $26.5 billion. HPC covers more than AI, but TSMC has repeatedly described AI accelerator demand as a major driver of advanced-node and advanced-packaging capacity.

The problem comes when we turn that supply-chain evidence into a market total.

Suppose NVIDIA pays TSMC to manufacture a processor and later sells a system containing that processor. Adding both companies' full revenue would count the manufacturing stage and the finished product as separate pieces of final chip demand.

Advanced packaging creates the same trap. So do substrates, boards and complete server assembly.

TSMC confirms the scale and durability of AI semiconductor demand. We use it as supporting evidence rather than another pile of revenue to add to the market.

Is the AI server market much bigger than the AI chip market?

The AI server market is already above $460 billion in annual system value, but only part of that money ends up with chip suppliers.

TrendForce's latest industry analysis expects total server output value to exceed $624 billion this year, with AI servers accounting for more than 74%. That implies more than $460 billion of AI server systems.

Those systems include GPUs or ASICs, CPUs, HBM, conventional memory, storage, networking hardware, motherboards, power equipment, cooling and the manufacturer's own margin.

The spending pattern is becoming more extreme because buyers increasingly want expensive rack-scale machines. IDC has also found server spending rising far faster than server unit shipments, reflecting the much higher value of GPU-heavy systems.

Buyer budgets are large enough to support this scale. TrendForce now estimates that nine major cloud companies will spend more than $886 billion in capital expenditure this year. Google, Amazon, Meta, Microsoft and Oracle represent close to 90% of that total.

We cannot treat $886 billion as AI chip spending because those budgets also cover buildings, power infrastructure, storage and many other assets. But they remove one obvious objection to a $300 billion-plus accelerator market: the customers clearly have enough money committed to support it.

Market layer	Approximate current scale	Why the numbers differ
AI accelerators	~$300B–$350B	GPUs and custom AI processors
Broad AI semiconductor stack	~$400B–$570B	Adds memory, networking and related silicon
AI server systems	>$460B	Adds complete server hardware
Nine major cloud companies' capex	>$886B	Adds buildings, power and other infrastructure
Total semiconductor industry	~$1.56T	Includes every chip market, AI and non-AI

Are custom AI chips actually taking business away from GPUs?

Custom AI chips are gaining share rapidly, but GPU demand is still growing because the whole AI infrastructure market is expanding faster than the substitution.

TrendForce initially expected AI server shipments to rise more than 28% this year. Its latest forecast has increased that to nearly 31% as cloud companies accelerate infrastructure spending.

Within those shipments, ASIC systems are approaching 28% of the market.

That gives custom processors plenty of room to gain share without requiring NVIDIA GPU shipments to fall. A growing market can support both at once.

The supplier numbers show exactly that pattern. Broadcom's latest AI semiconductor revenue grew 221% year over year, while NVIDIA's latest Data Center revenue grew 117%. AMD's Data Center business grew 107%.

All three are expanding at the same time.

The pressure becomes more interesting over several years. TrendForce expects ASIC-based AI servers to approach 40% of shipments by 2030 as hyperscalers push more inference onto chips designed around their own workloads.

For now, custom silicon is taking a larger piece of a rapidly expanding pie rather than shrinking the GPU business.

Is AI inference making the chip market even bigger?

AI inference is widening the chip market because companies need computing power every time users, agents and applications run models, not only when developers train them.

Training drove the first wave of enormous GPU clusters. Inference creates a different demand pattern because the workload continues as long as people keep using the product.

TrendForce estimates that the five largest North American cloud providers will increase their combined AI inference computing power by about 122% this year. Their training computing power is also expected to increase, but at a much slower rate of roughly 56%. Inference compute is growing about 122% in a year; we look at how fast the market grows, and why, in our AI chips market report.

That difference helps explain the custom-chip boom.

A cloud company may tolerate the flexibility and high cost of top-end GPUs when training a frontier model. Repetitive inference workloads give it much more incentive to design a specialized processor that lowers cost per token.

Google TPUs and AWS Trainium are already doing exactly that. Broadcom's rapidly growing custom-AI business provides another financial trace of the same shift.

Inference expands both the amount of compute being consumed and the number of chip architectures that can compete for that spending.

How fast is the AI chip market growing right now?

The core AI chip market appears to be growing by roughly 50% to 70% in value this year, although individual suppliers are moving much faster.

A useful supply-chain model from GF Securities estimated the global accelerator market at about $207 billion in 2025 and roughly $349 billion this year. That works out to growth of around 69%.

The company-level data are even stronger. NVIDIA's latest Data Center quarter grew 117% year over year. AMD's Data Center revenue rose 107%. Broadcom's latest AI semiconductor revenue climbed 221%.

We should not average those percentages because the categories overlap and the companies start from very different revenue bases.

Physical deployment is growing more slowly than revenue. TrendForce expects AI server shipments to rise nearly 31%, well below the growth rates reported by the major chip suppliers.

The gap makes sense. New rack-scale systems contain more expensive processors, more HBM and more networking per deployment. Memory prices have also risen dramatically.

So the market is growing through both volume and semiconductor content. Roughly 50% to 70% growth in core accelerator value looks much more believable today than either a normal semiconductor growth rate or the triple-digit percentages reported by individual challengers.

Why has the broader AI chip estimate jumped so much lately?

The broad AI semiconductor market has become much larger because memory prices and memory content have surged on top of already strong accelerator demand.

Earlier this year, Gartner expected total semiconductor revenue of about $1.32 trillion and said AI semiconductors would account for roughly 30%. That implied a market near $396 billion.

Gartner's newest forecast looks very different. Total semiconductor revenue is now expected to reach about $1.56 trillion, with the AI data-center ecosystem representing 36.5%.

Multiplying those two figures gives roughly $568 billion.

The wording and market definitions are not perfectly identical between the two Gartner forecasts, so the figures should not be read as a clean like-for-like revision. But the direction is clear: a lot of the extra value is coming from memory.

Memory revenue is now forecast at roughly $837 billion, compared with $220 billion in 2025. AI infrastructure is consuming huge amounts of high-end memory at the same time that prices are rising sharply.

The $568 billion figure tells us how economically important AI has become to the whole semiconductor industry. It should not be mistaken for $568 billion of GPUs.

Could AI chips really become a $1 trillion market?

A $1 trillion AI semiconductor market is now plausible later this decade, but the accelerator market itself has not reached that level today.

AMD has previously forecast a data-center AI accelerator opportunity above $500 billion by 2028 and now talks about a data-center silicon market exceeding $1 trillion by 2030. A trillion-dollar forecast only helps if you know where the openings are, and we look at the biggest opportunities in our AI chips report.

Those are vendor forecasts, so we would never use them as proof of current market size.

Gartner provides a more useful independent benchmark. Its latest forecast puts the AI data-center ecosystem at 36.5% of semiconductor revenue today and above 53% by 2030.

The underlying semiconductor industry is also becoming much larger. Gartner currently expects total global semiconductor revenue of roughly $1.56 trillion this year and around $1.94 trillion next year.

If the industry keeps expanding and AI eventually represents more than half of it, a trillion-dollar AI semiconductor ecosystem stops looking mathematically extreme.

The harder question is whether today's exceptional capital spending continues long enough to get there. Cloud companies are currently investing at an extraordinary pace, but semiconductor cycles have repeatedly shown that supply can eventually catch demand.

So $1 trillion belongs in the plausible future range for the broad AI semiconductor market. It would be misleading as a description of the current accelerator business.

So what is the AI chip market actually worth today?

The best current answer is about $300 billion to $350 billion for AI accelerators themselves and roughly $400 billion to $570 billion for the wider AI semiconductor market.

The lower range is the one we would use when someone simply asks, “How big is the AI chip market?”

It captures the economic core of the boom: NVIDIA GPUs, AMD Instinct processors and the growing fleet of custom accelerators from Google, Amazon and other hyperscalers.

Several independent checks make that range credible. Gartner had already measured AI processors above $200 billion in 2025. NVIDIA has since reached $89 billion of Data Center revenue in a single quarter. Broadcom's AI semiconductor business is approaching a $22 billion quarterly run rate under its latest guidance. TrendForce has raised its AI server shipment forecast to nearly 31% growth.

The higher range answers a broader question. Once we include HBM, AI networking and related data-center silicon, Gartner's newest forecast implies an AI data-center semiconductor ecosystem of roughly $568 billion.

The smaller $40 billion estimates now look detached from observable industry revenue. The trillion-dollar figure is getting easier to imagine, but it still describes where the broad market could go rather than where accelerator sales stand today.

For now, roughly $350 billion is the clearest number for the core AI chip market. Around half a trillion dollars is the better number for AI's wider semiconductor footprint.

Definition	Best current estimate	How we would use it
Narrow legacy accelerator estimates	~$40B–$45B	Too small for today's data-center market
AI processors in 2025	>$200B	Useful historical floor
Current AI accelerator market	~$300B–$350B	Best answer to “How big is the AI chip market?”
Broad AI semiconductor market	~$400B–$570B	Best measure of AI's semiconductor footprint
AI server market	>$460B	Useful system-level comparison
Current cloud capex	>$880B	Checks whether buyer budgets support the market
Future broad AI silicon market	Potentially $1T+	Plausible later-decade scenario, not current revenue