Opportunities emerging from rising

demand for AI computing power

Alphinity Investment Management 

May 2024

The AI boom is likely to drive a surge in energy consumption after decades of sluggish demand. Powering energy-hungry data centres has become a significant bottleneck within the AI capex cycle.  We take a closer look at the implications for global energy and gas markets, but also highlight various potential beneficiaries across different industries.

• The proliferation of Generative AI is likely to cause electricity consumption to inflect higher and begin grow once again.
• Chronic underinvestment in power generation creates a bottleneck, making gas a practical short-term solution amid costly and highly intermittent renewables.
• Analysis suggests this could drive incremental growth in US gas demand of ~1-2 bcf/d or 1% - 1.8% of current demand. That is on top of the ~1% of demand from LNG export projects that have already reached FID.
• Alphinity have multiple avenues of exposure to this thematic. ConocoPhillips benefits as a leading global gas producer and trader, along with tech giants Nvidia, Microsoft, Google, ASML, and Cadence Design who supply essential technology to enable AI. Trane Technologies, Ferguson, and Prologis facilitate the physical construction of data centres. Additionally, in the Sustainable fund, we own Schneider Electric and Quanta.

Source: Intel, Nvidia, Morgan Stanley, Company Reports, Alphinity

How much energy do data centres consume today and may consume in the future?

Generative AI and large language models have sparked strong rallies in some stocks, particularly those directly exposed to data centres. These centres, vital for AI processing, are reigniting electricity demand, challenging the adequacy of power supplies and infrastructure. The energy demand from data centres, already high, is predicted to grow, prompting significant investments in energy infrastructure, including natural gas and power distribution systems.

Currently, data centres consume about 2.5% of global electricity, a figure Boston Consulting Group expect to rise to 4% by 2027 and as high as 7.5% by 2030, increasing the U.S. electricity demand at twice the rate of the previous decade. If correct, that will be equivalent to powering ~20% of all homes in the US. This surge underscores the urgent need for expanded power generation to prevent potential blackouts and is driving a meaningful uplift in capital expenditures by utility companies. With the U.S. leading the global data centre market, approximately 40% of global capacity, this trend is likely to be mirrored worldwide, affecting energy consumption and infrastructure development on an international scale.

Did you know?

• Data Centres could consume the same amount of electricity as ~20% of total US homes by 2030
• GenAI consumes 4-5x the power of traditional data centres. By 2027, global demand for GenAI alone could drive the same power consumption as the entirety of Spain.
• Virginia is home to around 35% of the world's hyperscale data centres.
• Within the space of 24mths, PJM, the Eastern Interconnection regional transmission organization (RTO), has revised its 10-year average annual peak load growth forecast from 0.5% to 5.5% per year. Unprecedented in recent history.

Estimating how this increased demand translates into actual electricity demand

Bottom-up analysis: GenAI to add 1% of incremental electricity demand growth per annum

Power consumption in data centres is calculated by adding server and storage power use, multiplied by power usage effectiveness (PUE) and operating hours. By estimating GPU growth, chip utilization and power efficiency, one can project potential power needs. AI racks can need up to five times more power than traditional ones. AI data centres using GPU clusters, consume 30-100kW per rack. Taking this fact further, mature LLMs such as ChatGPT in inference mode, consume 400-1300 kWh daily, requiring significantly more power as user queries escalate post-training.

Source: Boston Consulting Group

Assuming total GPU volumes rise from ~1.8m in 2023 to ~4m in 2026, custom silicon & GPU utilization rates range between 70-90% and data centre PUEs (a measure of a data centres energy efficiency) decline from 1.5 to 1.3,  Morgan Stanley predicts incremental global data centre power usage will rise to 326-460 TWh by 2027, a significant increase from the current 360 TWh, with GenAI contributing about one-third of this demand. Schneider Electric estimates AI will account for a slightly lower 15-20% of the growth. This equates to a 105% compound annual growth rate for GenAI and 20% for overall data centre power, making GenAI's 2027 power demand comparable to Spain's and adding 1ppt of growth to overall electricity demand growth rates.

Top-down analysis: AI energy demands to double in 3yrs

The number of data centres are increasing rapidly, exemplified by Microsoft's plan to build up to one new data centre every three days, at the same time as the power intensity increases significantly. As such, BCG predicts U.S. data centre capacity will rise from 2.5% to up to 7.5% of total energy demand by 2030, increasing electricity demand by 260 TWh. McKinsey comes up with similar estimates, foreseeing power needs for U.S. data centres to double from 17GW to 35GW, equivalent to the current combined global capacity of the big 5 tech giants. Globally, institutions like the IEA foresee data centre, AI, and cryptocurrency electricity consumption doubling in the next three years to 1000 TWh, matching Japan's total power use. This surge suggests electricity demand growth could shift from stagnant for the past two decades, to 0.5-2% annually by 2030, raising questions about future supply sources.

What source of energy is most likely to meet this demand?

Owners of data centres, most of whom have committed to net-zero targets, face a dilemma with surging electricity needs. Renewable energy sources offer intermittent supply with 25-40% capacity factors and when paired with energy storage, higher costs. With the rapid growth in AI demand necessitating immediate and reliable power, gas generation emerges as a logical short-term solution, providing consistent electricity around the clock. Renewables will still grow in importance, especially as the cost of energy storage falls. Nuclear energy & small modular reactors are potential alternatives (see Amazon's DC acquisition), but challenged due to the long lead times to deploy and cost challenges.

Recent discussions among CEOs and experts highlight the shifting demand landscape and gas's role in meeting these urgent energy needs.

"The energy needs over the course of the next couple of decades (driven by AI) is going to be enormous and over the next 5-10yrs, the amount of gas fired power generation that will be required to deliver that electricity, to fire all the data warehouses, that are going to be required is going to be enormous" - Ryan Lance, CEO of ConocoPhillips

"Electricity demand is experiencing three times faster growth per year this decade than what we've seen in previous decades, driven by the increase in electric vehicles and emergence of new, large-load data centres..... This is a major shift for our country, and it's a major shift for the natural gas market to be able to keep up with this" - Alan Armstrong, CEO of Williams Company

This could drive an incremental 1% to 1.7% pa of unexpected demand for natural gas

Bernstein projects that with ~25% of the U.S. incremental demand met by gas-fired generation, an additional ~230 TWh would necessitate ~15GW new gas capacity, increasing gas demand by ~1.1 bcf/d or 1% of the total market. Under Morgan Stanley's scenario--70% renewables and 30% fossil fuels with 70% server utilization--gas demand would rise by ~13GW over 5 years. Jefferies, accounting for upcoming renewable installations and existing energy plans, forecasts an annual increase of 1.6-1.8 bcf/d in gas demand. Consequently, AI could boost U.S. gas demand by 1-1.7% annually in the late 2020s, potentially significant against an already planned 10 Bcf/d of incremental LNG exports (or ~1%).

Additionally, this growth in data centre energy use could induce market volatility and regional imbalances. You are already seeing examples of crypto miners building server farms in the Permian where they can take advantage of cheap associate gas. Companies that can take advantage of the volatility through globally balanced gas portfolios and/or large trading capabilities will likely benefit.

Addressing one key assumption - advancements in chip power efficiency gains

This analysis contains multiple complex variables, such as diminishing chip power efficiency, the shifting power requirements as AI models transition from training to inference, and regulatory influences on renewable energy development and new gas infrastructure permitting.

A critical factor in this context is compute intensity. Notably, while advancements in chip technology, exemplified by Nvidia's B200 Blackwell, yield substantial efficiency improvements per compute unit, the absolute power consumption continues to escalate, with the Blackwell chip (announced a fortnight ago) experiencing a 43% increase in maximum theoretical load compared to its predecessor, the H100.

Source: Schneider Electric White Paper: Challenges and Guidance for Data Centre Design

Conclusion

The tech sector's rapid evolution often outpaces slower-moving industries and governments, potentially causing bottlenecks. While technological advancements may offer some relief, they may not suffice to meet the surging demand for power. Consequently, this growing need is expected to spark a renewed interest in gas-fired generation. Companies such as ConocoPhillips are well positioned to benefit from fulfilling this demand and capitalizing on the ensuing market volatility.

Additionally, Alphinity has significant positions in the technology companies that are crucial to the rapid growth of AI, as well as exposure to the "picks and shovels" companies further up the value chain. These businesses are actively addressing emerging challenges, including managing heat generation, enhancing, and securing the electrical grid, and overcoming the scarcity of strategically located land.

These include technology companies like Nvidia, Cadence Designs, Alphabet and Microsoft, alongside companies such as Trane Technologies, who are the leaders in data centre cooling solutions. Prologis, who are spending ~US$7-8bn over the next 5yrs to build 20 new DCs. ASML who manufacture advanced AI chip equipment, and Ferguson who supply infrastructure essentials such as pipes and fire suppression for DCs. Furthermore, our sustainable strategy owns Schneider Electric with 20% of sales to DCs and Quanta, the top U.S. grid EPC.

Author: Chris Willcocks, Portfolio Manager

Funds operated by this manager:

Alphinity Australian Share Fund, Alphinity Concentrated Australian Share Fund, Alphinity Global Equity Fund, Alphinity Global Sustainable Equity Fund, Alphinity Sustainable Share Fund

Hedge Clippings | 24 May 2024

News & Insights

Market Commentary | Glenmore Asset Management

Investment Perspectives: 14 of the most important charts for data centre investors right now | Quay Global Investors

April 2024 Performance News

Argonaut Natural Resources Fund

Bennelong Emerging Companies Fund

PURE Income & Growth Fund

Delft Partners Global High Conviction Strategy

Bennelong Twenty20 Australian Equities Fund

Equitable Investors Dragonfly Fund

If you'd like to receive Hedge Clippings direct to your inbox each Friday

JOIN OUR MAILING LIST

Tax Loss Selling

Marcus Today

May 2024

We are approaching June, and it's a good time to start talking about tax loss selling in stocks.

It is something that we should all think about in May rather than June. It works like this:Assuming you pay capital gains tax - many of you don't - on June 30th the tax man will freeze your portfolio, add up all the capital gains you've made this year, all the capital losses, offset the losses against the gains, and come after you for CGT on any net profit.If your accountant hasn't reminded you, then this is a reminder that you can minimise ('defer' really) your capital gains tax (CGT) liability this year by realising some losses as well as gains before the year-end, rather than just holding on. Basically, if you have a net capital gain from stocks sold, or any other capital gain, you should now be looking through your collection of current holdings for any stocks with losses attached (that you could sell now), and in so doing crystallise a loss that can be used to offset any gains made during the year. In so doing you can minimise your need to pay tax this year. Ultimately, it is not about how much tax you pay, that will not change. It is about when you pay it. It's obvious stuff.

Capital losses can offset capital gains.

• Stocks that have not performed well over the year, stocks in which a lot of shareholders are holding at a loss coming into the financial year end, tend to get sold down again towards the end of June. Thanks to tax loss selling, bad performers perform badly in June.
• Small illiquid stocks, ones that have not performed well where again, a lot of shareholders are likely holding at a loss, can get absolutely massacred by tax loss selling in June because the selling is forced through when the liquidity is not there. Net result, if you hold an illiquid stock that is running at a loss, take the loss early, because if you wait until the end of June when every other dodo wakes up to it, you'll be selling into a last-minute frenzy.

If you do want to take a loss before the end of June but want to continue holding the stock long-term, it is a good idea to take the loss early (now for instance), and not buy it back immediately, but wait until the other tax loss sellers destroy it. If all goes well, you can buy it back lower down as the liquidity issue bites the share price closer to the financial year-end.Even if you don't have tax issues, thanks to tax loss selling, trading opportunities can arise, especially in the small illiquid stocks that get massacred. Once the tax loss selling fades away, sold-down stocks can bounce significantly as small buy orders rebound the prices. So traders should be looking for the lows in small illiquid sell-offs over June, hoping for a rally into July. But another word to the wise, whilst you might think you should wait until July 1st to buy, experience suggests that many of the stocks impacted by tax loss selling tend to bounce before the year-end, a week or so before. So for a trader, the game is to identify the stocks getting destroyed and watch for the first rally, rather than July 1st, to get stuck in. The rebounds can be just as sharp as the last-minute drops. Hesitate and you'll miss it.

Which Stocks Do You Sell for Tax Reasons?

Legal Note: ATO Wash Sales Provisions

Taking Advantage of the Selling

Hints for Taking a Loss

• If a stock is going down, it is far more likely to continue going down than it is to turn on a sixpence to suit you.
• An early loss is the smallest loss - The further a stock falls, the more intense the selling becomes as higher losses cause more selling decisions, so sell early.
• The odds are in your favour - If you sell 10 falling stocks, it will be the right thing to do in nine cases, but you will only remember the other one.
• Clear the mind - If you sell now, you are no longer exposed, and all you have to do is come to terms with the loss.
• You can buy it back - If you sell now, you can always buy it back - you might even buy it back lower than you sold it, and in the meantime, you have a period of clarity.
• The eye of the hurricane - If you sell now, you enter the eye of the storm and all becomes calm. You have a moment to think and can watch from a distance. You can always choose to enter the storm again, and you will be thinking more clearly and be armed with a plan if you do.
• Would you buy this stock now? - If you are making a loss on a stock, think to yourself... "If I had cash, would I buy the stock now at this price?" If the answer is 'no', why are you holding it? Sell it. Most people begin to 'hate' the stocks they lose money in... so this argument always works.
• Your state of mind has a value - What would your spouse pay (or you pay) to have you carefree at the weekend, instead of ripping the heads off the kids? Look after yourself. There are not that many weekends in the year (or your life). Don't ruin too many of them by keeping risky loss-making positions until Monday because you didn't have the guts to sell them on Friday.
• Averaging down is a mug's game - If you have money to invest, you should be putting it in the best investment in the whole world. Do you really think that will be the very same stock you have already bought at a higher price and that is falling at the moment? Very unlikely. You already have an exposure... why do you need more of something that has already proved itself to be a dog?
• There is no logic in being emotional about losses - Do what most brokers do with their own shareholdings. They have an Excel spreadsheet linked to live prices monitoring all their holdings and what they are worth. At the bottom of the page is a total of what all the holdings are worth now. That clicks over every second, all day. Up $500, down $500. This figure is the only truth. This is what the shares are worth.
• The price you paid is irrelevant - If it's gone down, it's gone down. XYZ stock doesn't 'owe you'... what you paid for it is irrelevant to where it goes next.
• Opportunity cost - Not selling a dog ties up capital that could be employed somewhere more gainfully than in a dog stock.
• If in doubt, sell it - It crystallises a capital loss for this tax year. Why wait until the end of the year to take your losses? Taking losses today could set you up for making and taking gains this year. You can always buy it back once you've made the sale.

Hopefully, you hold good long-term stocks and won't have to take a loss, but when you do, read this again and see if you can get to the bottom of the list before you put the sell order on.

Author: Marcus Padley

Funds operated by this manager:

Marcus Today Equity Income SMA, Marcus Today Growth SMA