AI Chip ETFs Shift Toward Custom Silicon as TPU Inference Costs Fall

The trade has a new center of gravity

For most of the AI boom, buying an AI infrastructure ETF mostly meant betting on Nvidia. That is starting to change. Coverage from late September and early October 2026 shows custom accelerators gaining ground. These include Google's Tensor Processing Units (TPUs) and the chips Broadcom co-designs for hyperscalers and AI labs. The main reason is cost: these chips appear to be making AI inference cheaper. Semiconductor funds have already had a huge year, so investors now need to know which part of the chip supply chain will capture the next round of spending. The coverage points to three answers: custom-silicon designers, the foundries that manufacture their chips, and the memory and networking suppliers that every AI system depends on, whatever processor it uses.

Semiconductor ETFs had a huge year

The returns are hard to ignore. The iShares Semiconductor ETF (SOXX) was up about 90% year to date by late September. Its four largest holdings were Intel, AMD, Micron and Nvidia, which together made up roughly 35% of the fund.22 Those four stocks have produced a median return of about 1,160% since early 2023, more than ten times the S&P 500's gain over the same stretch.22 Money has followed performance. SOXX and the VanEck Semiconductor ETF (SMH) took in about $11.1 billion and $8.7 billion of net inflows this year, respectively.27 SMH is the largest semiconductor ETF, with roughly $74 billion in assets, compared with about $50 billion for SOXX.21

The two funds are built very differently. Nvidia accounts for about 19% to 20% of SMH.2123 In SOXX, by contrast, Nvidia is only the fourth-largest position at about 7.5%.22 One Seeking Alpha analysis rated both funds strong buys but gave SOXX a slight edge for new money because of its lower Nvidia exposure and lower fees.24 That difference matters more now. If custom silicon takes a growing share of AI spending, a fund concentrated in Nvidia is a narrower bet than its "AI infrastructure" label suggests.

Fund issuers are slicing the theme ever more finely. About 15 AI- and tech-related ETFs launched in September alone. They target memory, power infrastructure, optical components, capacitors and robotics.27 Defiance launched AIFR, which it describes as the first U.S.-listed ETF dedicated to semiconductor foundries. Its pitch is that pricing power is moving to the companies that make chips: foundries have announced wafer price increases for the second half of 2026, and TSMC held about 72% of global foundry revenue in the first quarter.29 It is a new and still tiny product, with net assets of about $263,000 as of late September.29

Why custom chips are gaining

The biggest single number in this story is Broadcom's. Its AI semiconductor revenue rose 221% year over year in its fiscal third quarter, to $16.7 billion. Shipments of custom processors grew about 3.5 times, and custom silicon made up nearly three-quarters of AI revenue.1115 The company guided for $21.7 billion in AI revenue in the current quarter and about $58 billion for fiscal 2026.11 It has also put out longer-range targets: about $115 billion for fiscal 2027 and $230 billion for fiscal 2028.1217 CEO Hock Tan says Broadcom now has six custom-chip ("XPU") customers. Google, Meta, OpenAI and Anthropic are confirmed.1518

The argument for these chips comes down to the economics of inference, meaning the ongoing work of running AI models for users. One estimate says custom accelerators can cut inference costs by 40% to 60% compared with GPUs.11 Another analysis argues that custom silicon needs less upfront capital and uses less power for the same work. That would give hyperscalers a better return when capital, not demand, is the limiting factor.16

Google's seventh-generation TPU, Ironwood, is the clearest test of that argument. SemiAnalysis benchmarks found that, when serving each user 100 tokens per second, Ironwood cost about $0.181 per million tokens. That compares with $0.222 for Nvidia's B200 and $0.276 for the B300.36 Reports of the size of Ironwood's advantage over the B200 range from about 18% to 19%. Against the B300, the advantage is about 34%.36 At a 20-second median response-time target, the gap narrows to about 8% against the B200.3 SemiAnalysis also found that the B200 still wins on a small part of the performance curve, including around a 30-second median response time.8

Do the cost advantages hold up?

The coverage disagrees here, and the disagreement matters.

The optimistic view relies on internal and negotiated prices that ordinary customers do not pay. SemiAnalysis estimates that Google's own cost of running TPUs is about $1.03 per chip-hour. At that rate, Ironwood's performance-per-dollar lead at high usage rises to 76.7% over the B200 and 130.2% over the B300.8 A separate analysis puts Anthropic's estimated Ironwood rate at about $1.60 per chip-hour, roughly 13% of the public on-demand price.5

The public price list tells a different story. Google lists Ironwood at $12.00 per chip-hour on demand. A three-year commitment costs $5.40, and flexible-start capacity costs $6.00.2 One detailed analysis found that, at list prices, Ironwood costs about the same as a B200 on Google Cloud ($12.00 versus $11.28). It also found that Google's price list charges roughly the same amount per unit of memory bandwidth across TPUs and GPUs alike.5 At moderate batch sizes, that analysis found the cost floor per million output tokens falls in a narrow $0.42 to $0.50 range regardless of chip. Choice of chip only matters much when hardware runs at very high usage.5

Our reading is that both views are correct, and that tells investors something. TPUs are dramatically cheaper for organizations that control the hardware themselves: Google internally, and partners operating at gigawatt scale, such as Anthropic. Ordinary cloud customers get a much smaller saving. Switching from Nvidia's software tools also takes work. One estimate puts it at two to eight weeks of engineering to move a workload to Google's JAX framework.1 Google is trying to remove that obstacle with TorchTPU, which adds native PyTorch support.6 So the cost savings go mainly to the biggest buyers. That is why money is going into multi-gigawatt custom programs rather than into broad enterprise adoption.

Big commitments, and big debts

The commitments are now very large. Broadcom expects Anthropic to deploy one gigawatt of Ironwood TPUs in 2026, five gigawatts of TPU v8i in 2027 and roughly ten gigawatts in 2028.19 Anthropic has reportedly committed $125.2 billion to TPU computing capacity over five years.19 Broadcom expects OpenAI's "Jalapeno" inference chip to ramp through 2027 and 2028, and expects OpenAI to become its second-largest XPU customer, ahead of Google.18 Separately, a $22 billion AI infrastructure financing led by Crux AI is focused on Google's TPUs.24

Much of this depends on borrowed money. Broadcom has agreed to provide up to $42 billion in financing to support Anthropic's expansion, and has a dual role as both supplier and financing partner.1219 One pre-market note describes a $60 billion Anthropic package and reports that Broadcom is in talks to finance OpenAI's custom-chip effort as well.20 Reports differ on the size of these packages. They agree, however, that the custom-silicon boom increasingly relies on credit. That concentrates risk in a few AI labs whose spending has run ahead of their revenue. Some of Anthropic's lease arrangements reportedly cannot be canceled except in the case of a default. That gives Broadcom visibility on future revenue, but it also ties its fortunes to those customers.19

The market has noticed. Broadcom shares slipped after its record quarter, and analysts split on the outlook.15 China has also opened a review of Broadcom hardware in state data centers, adding a geopolitical risk.17

Nvidia is far from finished

None of this means Nvidia is losing its place. Its data center business ended the second quarter at a $356 billion annual revenue rate, more than four times the size of Broadcom's XPU business.16 Even Google still buys GPUs. Its own TPUs make up about 80% of its owned AI computing capacity, yet it still had to expand third-party capacity this year.9 Alphabet raised its 2026 capital spending guidance to $195 billion to $205 billion but has never disclosed how that splits between TPUs and Nvidia hardware.9 The most realistic outcome is that custom chips become a permanent second track running alongside GPU purchases, rather than replacing them.15

What it means for ETF buyers

This coverage points to a different way of thinking about "AI infrastructure" funds. Spending is shifting toward custom chips used for inference, and the companies that benefit regardless of which processor wins are the ones that sell into every data center. Those include Broadcom for its networking chips as well as its custom designs, TSMC for manufacturing, and the high-bandwidth memory (HBM) suppliers Micron and SK Hynix.182229 Broadcom's AI business is in fact dominated by networking, and its Tomahawk 6 Ethernet switch chips go into both custom-chip and GPU clusters.1819

For that reason, a broadly diversified semiconductor fund looks like a safer way to own the shift than an Nvidia-heavy one. It does not require predicting whether GPUs or custom chips win any given workload. Leadership in the AI trade has already moved several times, from hyperscalers to chipmakers to memory stocks.23 Valuations are high, and some analysts warn of short-term volatility and cooling profit margins.2425 Investors coming in after a 90% gain should expect the next phase to be bumpier, even if hardware demand holds up.