Top AI ETFs to Watch & Invest for 2026
Six ASX-listed ETFs give Australian investors direct exposure to the AI megatrend, from broad Nasdaq 100 coverage to pure-play semiconductor bets. We break down fees, 2023-2025 returns, the DeepSeek impact, and what each ETF actually owns so you can build your AI allocation with clear eyes.
Every AI-tilted ETF on the ASX posted negative returns in 2025. DeepSeek rattled semiconductor stocks, Tesla dragged FANG down, and the correction reset valuations sharply. But hyperscaler capex is running at $230B combined and NVIDIA still owns 80% of AI training. The question is not whether AI spending continues: it is which ETF gives you the exposure profile you actually want.
- 01NDQ (0.22% MER) returned +53%, +26%, then -8% across 2023-2025: the steadiest AI proxy with ~100 holdings and meaningful NVDA and MSFT weight.
- 02SEMI (0.57% MER) is still roughly 30% below its September 2024 peak, but sits at the hardware layer where every dollar of AI capex lands first.
- 03FANG (0.35% MER) delivered +72% in 2023 and +56% in 2024 before Tesla's 40% drawdown helped drag it to -18% in 2025.
- 04IVV (0.04% MER) quietly carries about 25% AI mega-cap weight at one-sixth the fee of thematic alternatives, and its 2025 drawdown was just -4%.
who this is for: Australian investors who want to understand exactly what they own when they buy an AI-themed ETF, how much concentration risk they are taking on, and what the 2025 drawdown means for the forward opportunity.
This article is general information and does not constitute financial advice or a recommendation to buy or sell any security. ETFCheck.com.au does not hold an Australian Financial Services Licence. All returns are calendar-year, in AUD, and include distributions unless otherwise noted. Past performance is not a reliable indicator of future performance. Read the Product Disclosure Statement before making any investment decision.
2025 was the year the AI trade got its first serious test. Every AI-tilted ETF on the ASX finished the calendar year in the red. SEMI was down 24%. FANG was down 18%. NDQ was down 8%. IVV, the most diversified and cheapest option on this list, managed a mere -4% and came out looking like a masterstroke of restraint.
The cause was not a collapse in AI spending. It was a recalibration of expectations. DeepSeek's R1 model arrived in January 2025 trained at an astonishing cost of approximately $6 million, compared to the $100 million-plus price tag attached to comparable US models. The market interpreted this as a potential collapse in demand for GPU compute. Semiconductor stocks fell sharply. NVIDIA lost trillions in market cap in a single session. The chips-and-infrastructure thesis had to answer a hard question.
This article breaks down each of the five main ASX AI ETFs, what happened to them in 2025 and why, and what the choices actually mean for a long-term Australian investor building an AI allocation from here.
The AI ETF landscape after DeepSeek
The five main ASX AI ETFs split into two philosophical camps. The first camp is index-based diversification: NDQ and IVV hold dozens to hundreds of companies with AI exposure baked in via the natural weighting of large-cap tech. You own AI without making a specific bet. The second camp is concentration: FANG, SEMI, and RBTZ make an explicit wager that AI-specific companies or sectors will outperform the broader market by enough to justify both the concentration risk and the higher fees.
The 2022-2025 period tested both camps. Concentrated funds peaked higher in 2023-2024 and fell harder in 2022 and 2025. Diversified funds gave up less in the bad years and captured substantial gains in the good ones. Which camp was better overall depends on exactly when you started and whether you stayed invested through the volatility.
NDQ: the balanced AI bet
The BetaShares Nasdaq 100 ETF (ASX: NDQ) tracks the 100 largest non-financial companies listed on the NASDAQ exchange. This gives you Apple (~8% weight), Microsoft (~8%), NVIDIA (~5%), Amazon (~4%), Meta (~4%), and Alphabet (~4%) as your top positions. Together, the top five holdings represent roughly 35% of the fund. You get AI exposure at scale without the stomach-churning concentration of FANG or SEMI.
NDQ's 2025 drawdown of 8% was the mildest among the AI-tilted funds, precisely because that diversification buffered against individual names blowing up. Tesla's 40% collapse barely moved the needle. NVIDIA's post-DeepSeek selloff was absorbed by the rest of the 100-name portfolio. The fund's performance is dominated by the collective trajectory of the AI revolution, not any single company's quarterly result.
At 0.22% MER, NDQ is the most cost-efficient explicitly AI-focused ETF on the ASX, sitting between IVV's exceptional 0.04% and FANG's 0.35%. For most Australians who want meaningful AI exposure without taking on the volatility of a 10-stock fund or a pure semiconductor play, NDQ is the clearest choice. The fee is justifiable relative to IVV given the higher tech concentration you receive in return.
FANG: concentrated, volatile, and recovering
FANG is the high-conviction play. It tracks the NYSE FANG+ index: a concentrated basket of roughly 10 mega-cap tech names in an equal-weight structure. Microsoft, Meta, Amazon, Apple, Alphabet, NVIDIA, Netflix, Tesla, Snowflake, and one rotating name. Equal weighting means every stock starts each quarterly rebalance at approximately the same allocation. Every single name moves the needle.
That structure delivered extraordinary returns in 2023 and 2024. A 72% calendar year followed by a 56% calendar year. If you held FANG from January 2023 through December 2024, you roughly tripled your money before fees. No broad index on earth matched that.
Then 2025 happened. Tesla, one of the ~10 equal-weight holdings, fell around 40% across the calendar year on a combination of brand damage, slowing EV sales, and Elon Musk's increasingly polarising public profile. In a market-cap-weighted index, Tesla's shrinking valuation would have reduced its drag automatically. In an equal-weight fund, Tesla kept being topped back up to full weight at each quarterly rebalance, repeatedly locking in the pain.
Tesla fell approximately 40% in 2025. In FANG's equal-weight structure, it carried roughly a 10% allocation heading into each quarterly rebalance. That single position contributed an estimated 4 percentage points of FANG's total 18% annual loss. Remove Tesla and FANG's 2025 drawdown would have been closer to -14%, still painful but materially better. The equal-weight rebalancing mechanism, which is a return driver in normal conditions, amplified the damage from one broken name.
The deeper question is whether equal weighting is a feature or a bug. In rising markets, it forces you to trim winners and top up laggards, a contrarian rebalancing bonus that historically adds returns in diversified large-cap portfolios. In falling markets with a single badly-broken name, it forces you to keep buying the disaster. For FANG, with only 10 holdings, a single company in freefall is genuinely consequential. At 0.35% MER, you are paying for active-adjacent concentration, which requires genuine conviction on every single holding.
SEMI: semiconductors after DeepSeek
The BetaShares Global Semiconductors ETF (ASX: SEMI) tracks roughly 30 global semiconductor companies through the Solactive Global Semiconductor index. NVIDIA sits at approximately 20% weight, followed by TSMC (which manufactures chips for Apple, NVIDIA, and AMD), ASML (the Dutch company that makes the extreme ultraviolet lithography machines that the entire chip industry depends on), Broadcom, and AMD. The fund owns the physical hardware layer of the AI economy.
In 2023, SEMI returned 95%. That was the market pricing in the first wave of AI infrastructure demand: every hyperscaler racing to purchase as many GPUs as NVIDIA could manufacture, TSMC running fabs at full capacity, and ASML's lithography machines becoming the most strategically important industrial equipment on the planet. In 2024, SEMI added another 31% as the build-out continued.
DeepSeek changed the narrative. The model's low training cost implied that the 'compute moat' driving semiconductor demand might erode faster than expected. If AI models could be trained efficiently at a fraction of the compute, the insatiable demand for NVIDIA H100s and TSMC wafers was suddenly in question. SEMI fell 24% in 2025 and remains roughly 30% below its September 2024 peak as at March 2026.
Despite DeepSeek's efficiency gains, three structural facts underpin the semiconductor thesis. First, NVIDIA's CUDA software ecosystem accounts for approximately 80% of AI training runs globally and represents a multi-year switching cost that no competitor has yet overcome. Second, Microsoft committed $80B in 2025 capex, Amazon $75B, and Google $75B: combined hyperscaler capex of $230B, virtually all of which flows into servers, chips, and networking equipment. Third, cheaper inference does not kill compute demand: it expands the addressable market. When the cost to run an AI query falls 90%, the number of queries deployed commercially grows by orders of magnitude. This is the Jevons Paradox applied to AI compute.
SEMI at 0.57% MER is the most expensive fund in this comparison by a wide margin relative to its breadth. If you believe the semiconductor thesis and want concentrated hardware exposure, SEMI delivers it, but you need to accept a fund that can peak 95% in a year and then lose 24% the following one. The standard deviation of returns is genuinely extreme. This is not a core holding for most portfolios: it is a high-conviction satellite position sized accordingly.
RBTZ: industrial AI and robotics
The BetaShares Global Robotics and Artificial Intelligence ETF (ASX: RBTZ) tracks the ROBO Global Robotics and Automation index, holding approximately 80 companies selected for their exposure to robotics, automation, and AI-adjacent manufacturing technology. Holdings include NVIDIA (~5%), Intuitive Surgical (surgical robots), Keyence (industrial sensors), Fanuc (factory robots), and various mid-cap specialists in automation hardware.
RBTZ is structurally different from the others: it targets the physical deployment of AI rather than the software or chip layer. Industrial robots, automated warehousing systems, autonomous machinery, and precision sensors. This makes RBTZ's performance more correlated with global manufacturing investment cycles than with pure software AI trends. When software AI stocks surged in 2023, RBTZ's 41% return was strong but well below SEMI's 95% and FANG's 72%, because factory automation spending cycles more slowly than software hype.
The same dynamic provided modest insulation in 2025: RBTZ fell 12% while SEMI fell 24%, because DeepSeek's impact on inference efficiency was largely irrelevant to a Fanuc factory robot or a Keyence vision sensor. If your AI thesis centres on long-term physical automation adoption rather than software AI spending, RBTZ gives you a differentiated exposure that does not simply replicate NVIDIA's trajectory. The caveat is the 0.57% MER, identical to SEMI, for a fund with a narrower global investment universe and more limited liquidity than the semiconductor alternatives.
IVV: the AI ETF nobody talks about
Most investors hunting for AI exposure overlook the most obvious option. IVV tracks the S&P 500, charges 0.04% per year, and gives you roughly 25% of the portfolio in the five largest AI companies on the planet: Apple, Microsoft, NVIDIA, Alphabet, and Amazon. You get NVIDIA at ~6%, Microsoft at ~6.7%, and Alphabet at ~4%, all inside a 500-company portfolio that also includes healthcare, financials, energy, and consumer companies as ballast.
In 2025, that ballast mattered. While SEMI was down 24% and FANG was down 18%, IVV fell just 4%. The non-tech holdings provided genuine downside protection when AI sentiment reversed. Over the full 4-year period from January 2022 through December 2025, IVV's lower drawdown in both 2022 and 2025 means it actually competes surprisingly well with the AI-concentrated alternatives on a cumulative basis, at one-sixth the fee of FANG and one-fourteenth the fee of SEMI.
Over 20 years on $100k at 8% gross return, the compounded fee drag is approximately $8k for IVV versus $118k for SEMI and RBTZ. That $110k difference is the direct cost of the AI concentration premium. Whether the thematic funds deliver enough return premium to justify that fee gap is the core question every SEMI or RBTZ investor needs to answer honestly. On the 2022-2025 evidence, they did not.
3-year performance: what the returns actually look like
The chart below shows calendar-year returns for all five funds across 2023, 2024, and 2025. The variation is substantial. 2023 and 2024 were exceptional for concentrated AI funds: SEMI nearly doubled in a single year, FANG tripled across the two years. 2025 erased a significant portion of those gains across the board.
Every AI ETF posted negative returns in 2025, but the starting points matter. FANG entering 2025 was up approximately 170% from January 2023. Even after the -18% year, it had roughly doubled from the 2023 start. SEMI, starting from a much lower recovery base (it bottomed hard in 2022), still held significant 2023-2024 gains despite the DeepSeek selloff. Investors who held through 2022's brutal drawdown and through 2025's correction are still meaningfully ahead on most funds. Investors who bought in late 2024 near the peak are not.
The cumulative 4-year picture (from January 2022, including the 2022 tech crash) tells an important story about what broad-market diversification is worth:
IVV's resilience in 2022 (-11% compared to NDQ's -27% and FANG's approximate -42%) provided a base that compounded well through the 2023-2024 surge. Ending December 2025, IVV competes with NDQ and FANG on a total-return basis despite never having a single spectacular year. SEMI peaked highest in mid-2024 but gave back substantial gains. RBTZ consistently lagged the software-AI plays while providing somewhat lower volatility.
The DeepSeek question: does cheaper AI hurt AI ETFs?
On January 20, 2025, the Chinese AI lab DeepSeek released R1, a reasoning model that matched or exceeded GPT-4 on key benchmarks. The stated training cost was approximately $6 million. By comparison, the widely-cited estimates for training GPT-4 class models in the US ran to $100 million or more. The market reaction was immediate and severe: NVIDIA fell 17% in a single session, its worst day in market-cap terms in US stock market history, wiping approximately $600 billion from its valuation.
The market's logic was straightforward: if AI models can be trained efficiently for a fraction of the assumed cost, the demand for high-end NVIDIA GPUs, TSMC wafer capacity, and data centre infrastructure would be lower than projected. SEMI, with its 20% NVIDIA weight, was the most direct casualty. The fund fell sharply through January 2025 and never fully recovered within the calendar year.
Despite the DeepSeek shock, the three largest hyperscalers each announced record capital expenditure plans for 2025: Microsoft committed approximately $80 billion, Amazon approximately $75 billion, and Google approximately $75 billion. Combined hyperscaler AI capex of $230 billion was not the behaviour of companies who believed compute demand had collapsed. Their explanation: cheaper inference expands the addressable market dramatically. When the cost to run an AI query falls 90%, businesses deploy AI across 10 times as many use cases. The Jevons Paradox applies: more efficient resource use drives higher total demand, not lower.
For ETF investors, the practical answer is nuanced. DeepSeek does not break the AI thesis, but it does change which layer benefits most. Inference efficiency gains (cheaper to run models) benefit the companies deploying AI (Microsoft, Google, Meta) more than the companies selling the hardware to train it (NVIDIA, TSMC). This is a relative argument for broader AI ETFs like NDQ or IVV over semiconductor-specific plays like SEMI, at least while inference efficiency is the dominant narrative. When the next generation of training runs begins (GPT-5 class models, Gemini Ultra successors), the semiconductor argument reasserts itself.
The fee cost of AI concentration
The chart below shows the compounded fee drag over 20 years on a $100,000 investment at an assumed 8% gross annual return. The difference between the cheapest and most expensive options in this comparison is not trivial: it is the equivalent of a new car, annually, for two decades.
SEMI and RBTZ each carry a 0.57% MER, which compounds to approximately $118,000 in foregone wealth over 20 years on a $100,000 starting investment. IVV at 0.04% costs approximately $8,000 over the same period. That $110,000 gap is the fee you are paying for the thematic concentration. To justify it purely on a fee-adjusted basis, SEMI or RBTZ need to outperform IVV by enough to overcome the annual 0.53% headwind: every year, in every market environment.
The 2022-2025 data does not support that case. Over the full 4-year period, SEMI's higher-volatility returns netted out to a result broadly comparable with IVV's steady compounding, before fees. After the 0.53% annual fee headwind, IVV has the upper hand. The strongest argument for paying the SEMI premium is conviction about a specific 3-5 year window when semiconductor demand will surge beyond what the broad index captures. That is a timing call, not a structural advantage.
A 80% IVV + 20% NDQ allocation captures broad market diversification at low cost (blended MER of approximately 0.07%) while overweighting the Nasdaq 100's AI-heavy holdings. You get meaningful NVIDIA, Microsoft, and Meta exposure without betting the portfolio on a 10-name fund or a 30-stock semiconductor index. Rebalance annually. This is not exciting, but the cumulative return data suggests it produces comparable or better results to the high-fee thematic alternatives over market cycles.
LNAS: do not mistake leverage for an AI strategy
BetaShares LNAS is a 2x leveraged Nasdaq 100 ETF. It is designed for short-term tactical use by sophisticated traders and is explicitly not appropriate for buy-and-hold investors. The daily reset mechanism in leveraged ETFs causes 'volatility decay': in a volatile but sideways market, a 2x leveraged fund will lose money even if the underlying index is flat. In a 30% Nasdaq drawdown (as occurred in 2022), LNAS would approximate a 60% loss. At 0.35% MER plus daily compounding costs, the long-term fee drag is brutal.
LNAS appears in searches for 'AI ETF Australia' because it tracks the Nasdaq 100 and AI is prominent in that index. But buying LNAS as an AI investment is a category error. It is not a higher-conviction version of NDQ. The leverage introduces path dependency that makes it permanently inferior to unleveraged NDQ for periods longer than a few weeks. The product disclosure statement warns explicitly that LNAS is not for long-term investment.
If you believe AI will drive Nasdaq 100 outperformance over the next 5 years, buying NDQ directly and reinvesting distributions captures that thesis without the volatility decay, the daily reset risk, or the compounding fee structure that works against you in every sideways or declining period. LNAS has its place in the toolkit of sophisticated short-term traders. That place is not in a long-term AI portfolio.
Building your AI allocation
The right AI ETF allocation depends on your conviction level, your volatility tolerance, and how long you intend to hold. Here is a practical framework by investor type:
- ✓Core-only investor (low complexity, long horizon): 100% IVV, or 70% IVV + 30% NDQ. You get 25-35% AI mega-cap exposure without thematic concentration. Blended MER stays under 0.10%. No active monitoring required. Rebalance annually.
- ✓Moderate AI tilt (some conviction, 10+ year horizon): 60% IVV or VGS + 40% NDQ. You are meaningfully overweighting the Nasdaq 100 AI leaders while maintaining a diversified base. Blended MER approximately 0.12%. This is the most defensible AI allocation for most retail investors.
- ✓High-conviction AI thematic (3-5 year view, accepts volatility): 50% NDQ + 30% FANG + 20% SEMI. All three funds have different AI exposures: NDQ is software-plus-hardware broadly, FANG is concentrated mega-cap, SEMI is pure-play hardware. Combined MER approximately 0.37%. Expect meaningful 20-30%+ annual swings. Size this as a satellite, not a core.
- ✓Robotics-specific thesis (physical AI automation, 7+ years): 50% NDQ + 30% RBTZ + 20% IVV. RBTZ gives you industrial automation exposure that does not simply replicate the software AI cycle. Suitable if your thesis is on factory automation and physical AI deployment rather than chatbot adoption.
- ✓All-in semiconductor (maximum conviction, maximum volatility): SEMI as a satellite at 10-20% of total portfolio alongside a diversified core. Never as a standalone holding. SEMI's annual return range of -24% to +95% across the past 3 years means you must be prepared to hold through severe drawdowns without selling.
Whatever AI ETF allocation you choose, size it so that a 40% drawdown in that portion does not change your behaviour. If you own $50,000 in FANG and it falls to $30,000, will you hold? If the honest answer is no, you are oversized. The investors who lost money in AI ETFs in 2022 and 2025 were mostly the ones who bought at the peak and sold at the bottom. The investors who are ahead bought diversified, sized conservatively, and held. The ETF selection matters less than those two decisions.
Methodology
MER and holdings data is sourced from each fund provider's current Product Disclosure Statement and monthly fund update reports: BetaShares (NDQ, FANG, SEMI, RBTZ, LNAS) and BlackRock iShares (IVV), as at March 2026. Top holdings weights reflect the most recent monthly holdings publication and shift between rebalances.
Calendar-year returns are approximate and expressed in Australian dollars (AUD), incorporating the impact of AUD/USD exchange rate movements. They include distributions where applicable. Cumulative growth figures in the chart are illustrative and should not be treated as precisely audited returns. For current and historical performance data, refer directly to fund providers: BetaShares, BlackRock iShares. Fee drag calculations assume an 8% p.a. gross annual return on $100,000 over 20 years. Hyperscaler capex figures are sourced from company earnings calls and investor relations disclosures. DeepSeek training cost figures are sourced from the DeepSeek-V3 technical report (January 2025). This article is general information only and does not constitute financial advice. All investments carry risk including possible loss of capital.