The Future Growth of Semiconductor Manufacturing

For most of its history the chip industry rose and fell with demand for PCs and phones. Today its largest customers are the companies building AI data centers, and they are buying on a much larger scale.

A trillion-dollar year

Global semiconductor sales reached $791.7 billion in 2025, up 25.6% on the prior year and the highest total in the industry’s history.¹ The Semiconductor Industry Association expects sales to pass $1 trillion in 2026, a threshold that until recently looked about a decade away. The first quarter alone brought in $298.5 billion, 25% more than the fourth quarter of 2025, normally the slowest quarter of the year.²

AI accounts for most of that gain. Logic chips from companies such as NVIDIA, AMD and Intel grew 39.9% to $301.9 billion in 2025, and memory grew 34.8% to $223.1 billion, both pulled by data center construction.¹ IDC expects data center chips alone to generate $477.1 billion in 2026 and $843.2 billion by 2030, close to half of the entire semiconductor market.³

Capacity sits in very few places

The growth is concentrated in a small number of manufacturers and countries. The foundry market reached a record $320 billion in 2025, and TSMC captured most of the revenue at the advanced end.⁴ Nodes of 7nm and below made up about 74% of TSMC’s wafer revenue by the third quarter of 2025, up from roughly half a few years earlier.^4,5

Governments and companies are spending heavily to move fabrication beyond Taiwan and South Korea into the United States, Europe, Japan and India. Under the CHIPS and Science Act, TSMC was awarded $6.57 billion for three Arizona fabs, Intel $7.87 billion and Micron $6.14 billion.⁶ The SIA projects U.S. fab capacity will grow 203% between 2022 and 2032, against 108% worldwide.⁷ Most of those fabs will not be producing for several years, so the reliance on Taiwan will not change much in the near term.

Where middle-market deals are forming

Advanced packaging, specialty fabrication and custom chip design are all short of capacity, and buyers are acquiring capabilities they cannot build or buy on the open market quickly enough. McKinsey expects semiconductor consolidation to stay active through 2026, with acquirers paying for control of data, AI models and computing efficiency more than for revenue or market share.⁸

Advanced packaging is the clearest example. Capacity for chip-on-wafer-on-substrate (CoWoS) packaging is now nearly as hard to secure as the wafers themselves. ASE Technology, the world’s largest outsourced assembly and test provider, bought a plant from WIN Semiconductors in August 2025 for NT$6.5 billion, about US$217 million, to expand advanced packaging for AI and high-performance computing chips.^8,9

That is relevant for smaller companies in packaging, specialty fabrication and chip design. Strategic buyers cannot add this capacity fast enough on their own, and they have shown they will pay to own it.

Custom chips and the challenge to NVIDIA

For more than a decade, NVIDIA’s graphics processors were the standard engine for AI, and its CUDA software made switching away costly. Application-specific integrated circuits (ASICs) and tensor processing units (TPUs) designed in-house by the largest technology companies now challenge that position. On large workloads they can match or beat GPU performance at a much lower total cost.^10,11

Google’s seventh-generation TPU, Ironwood, delivers 4,614 FP8 teraflops per chip with 192GB of memory. SemiAnalysis estimates an Ironwood server costs about 44% less to own than a comparable NVIDIA GB200 Blackwell configuration for Google’s own systems, and roughly 30% less for outside customers.¹⁰ For a company running data centers at Google’s scale, a cost gap that size is worth billions of dollars a year.

A challenger outside the hyperscalers

SambaNova Systems, founded in Palo Alto in 2017, takes a different approach with its Reconfigurable Dataflow Unit. A GPU executes instructions across thousands of cores. The RDU maps a model’s full computational graph directly onto the hardware, which cuts the repeated memory calls that make inference expensive and delivers more tokens per watt.¹² Its fifth-generation SN50, built for agentic inference, claims five times the compute and four times the bandwidth of its predecessor.¹³

In February 2026 SambaNova closed a $350 million Series E led by Vista Equity Partners, with Intel, BlackRock, T. Rowe Price and Google Ventures participating. Intel had earlier explored acquiring the company for about $1.6 billion.¹⁴ The size of the round, and Intel’s earlier interest, suggest investors see this kind of chip as a real alternative to GPUs for inference.

Custom chips still go through TSMC

Amazon’s Trainium3, its first chip on a 3nm process, has passed one million deployed processors, and AWS’s custom chip business is running at a $20 billion annual revenue rate.¹⁵ Bloomberg Intelligence expects the ASIC segment to grow 27% a year through 2033, against 16% for GPUs.¹¹

Custom silicon does not spread manufacturing out. Google’s TPU, Amazon’s Trainium, Microsoft’s Maia and Meta’s MTIA are all made at TSMC.^10,16 In May 2026 the Center for a New American Security concluded that chip production has become a binding constraint on the AI build-out, with demand outrunning manufacturers’ forecasts at every step of the supply chain.¹⁷ The shift to custom chips puts even more of the world’s advanced production in TSMC’s hands.

How NVIDIA and Google are responding

At Computex 2026, NVIDIA introduced the RTX Spark superchip, built with Microsoft to run AI agents on consumer devices. NVIDIA has sold mainly to enterprises and data center operators. Consumer products give it a broader customer base and make it less dependent on a small number of very large buyers.^18,19

NVIDIA has also committed at least $6.5 billion in 2026 to companies developing photonics, which moves data between processors with light instead of over copper and lowers the energy cost of doing so. AMD bought the photonics startup Enosemi in 2025.²⁰ Power use is becoming a real cost constraint in AI, and chipmakers are investing to bring it down.

Google, meanwhile, is offering TPUs to outside customers and has adapted them to PyTorch, the framework most AI developers already use. TPUs have trained frontier models including Google’s Gemini 3 and Anthropic’s Claude Opus 4.5.¹⁰ Much of NVIDIA’s advantage came from its software rather than its chips. If customers can use the same tools on Google hardware, switching becomes easier.

GPUs are still expected to dominate. Bloomberg Intelligence expects the AI accelerator market to exceed $600 billion by 2033, with GPUs holding about 81% and ASICs more than doubling their share.¹¹ If ASICs keep growing faster than GPUs after that, their share could end up larger than current forecasts show.

Risks to watch

Chip supply chains take years to expand, and demand keeps beating forecasts. GPU rental prices have risen despite steady technical progress, and AI is expected to take almost all of TSMC’s 3nm capacity, most of it allocated to NVIDIA.¹⁷ TPU and ASIC makers are waiting on the same TSMC capacity as the GPU makers they compete with, so building an alternative chip does not get them around the shortage.