Financial markets are supposed to capture the wisdom of the crowd, but on Monday the crowd ran in all directions waving its hands in the air screaming. Japan’s stock market fell the most in 37 years with a 12% plunge that wiped out all its gains for the year, while in the U.S. the VIX index of implied stock volatility briefly had its biggest rise ever. Panic hit.

The selloff was triggered by Friday’s jobs data prompting a sudden switch in the economic narrative from soft landing to hard landing. Add to the mix a period of deflating hype about artificial intelligence and a Bank of Japan rate rise designed to strengthen the yen. News that Warren Buffett’s Berkshire Hathaway had sold half its Apple shares and boosted its cash pile added to the pain.

But the triggers couldn’t possibly justify the scale of the moves. When a new trigger arrived, in the form of better-than-expected data on the service sector, markets partially rebounded and the Vix fell sharply — again, far more than the data could justify.

The selloff—which at one point had chip maker Nvidia down 15%—was so big because investors had been all-in betting that things would work out well. Now things have calmed a bit, the question is whether the unwind of these bets, and the leverage behind them, is done. If it resumes, will the selloff feed back into higher savings and a weaker economy or, worse, hit the financial system?

The extreme examples of past effects from big market falls are 1987’s crash, 1998’s Long-Term Capital Management blowup and 2008’s global financial crisis. History is never perfect, but so far this looks more like a (much milder) version of 1987 than it does the other two.

In 1987, the stock market had its biggest one-day fall ever, with the S&P 500 down more than 20% on Black Monday in October. Investors had built up excessive leverage after a stunning 39% gain in the year to August’s high, and the crash led both to big margin calls and to badly designed automated trading that exacerbated the selling. But the Federal Reserve poured liquidity into the banks, brokers didn’t default and the market made back all its losses within two years. The economy was fine.

The good news was that 1987 was all about markets: They went up, they went back down, no one else was hurt. The S&P made 36% in the eight months to its August 1987 peak, similar to the 33% it rose in the eight months to the end of June this year. As in 1987, this year’s gains came in spite of tight monetary policy and higher bond yields. Just like today, in 1987 investors were on edge and ready to sell to lock in the unexpected profit. The losses are smaller so far, but lucrative trades have reversed , just as they did for the market as a whole in 1987.

In 1998, the situation was much worse, although stocks recovered more quickly. Highly levered hedge fund LTCM was crushed when Russia’s domestic debt default created a flight to safety. LTCM was big enough that it threatened to bring down Wall Street institutions. The Fed cut rates three times and pulled together a group of banks to rescue the firm and wind down its trades slowly. Stocks took just four months to recover, but the easy money helped stoke the dotcom bubble, which popped two years later and led to a mild recession—and gigantic losses for investors in tech stocks.

We don’t know yet if any hedge funds have been taken out by the big moves in markets, which have brought heavy losses for those engaged in the “ carry trade ” of borrowing cheaply in yen and buying higher-yielding currencies such as the Mexican peso or dollar. Large swings in Treasurys on Monday might also have hurt, given the large positions hedge funds hold. Traders are betting that the Fed will slash rates, with a super-sized cut of 0.5 percentage points priced into futures for the September meeting (and far more earlier in the day).

The really bad outcome would be a repeat of 2008, but it seems highly unlikely. True, some large U.S. banks failed last year, due to bad bets on government bonds. But banks are much less leveraged than they were, and the system is less exposed to a liquidity crisis, as private lenders have taken on much of the risk that used to sit in banks. Big losses are entirely possible, and private funds could hit trouble, but that would take time and wouldn’t create the same system-wide crisis.

The ideal would be that excess in the stock market unwinds as in 1987 without creating wider trouble, hopefully more gradually than in 1987. AI enthusiasm could deflate stock prices much more—even after falling 30% from its June high, Nvidia has still doubled in price this year. But the market is already much closer to normal, with Monday’s falls leaving the Nasdaq 100 index up only 6% this year, and the S&P 7%.

If panic continues to abate, the Fed cuts and nothing breaks in the financial system, we should count ourselves lucky. But it would be good if investors could remember the sinking feeling they had on Monday morning, and try to be a bit wiser and less speculative.

The red numbers in your 401(k) today might appear to vindicate warnings about an artificial-intelligence bubble and infirm economy. But don’t tilt your portfolio toward full pessimism just yet.

The S&P 500 was down 3% Monday, with the Nasdaq falling even further. Investors have been selling the year’s best performers, concerned that disappointing second-quarter results from big technology companies such as Alphabet , Tesla and Intel are a sign that the AI frenzy is a fad. Also, consumer discretionary stocks have become the worst-performing sector in the S&P 500, as lacklustre labour-market reports have raised worries that the Federal Reserve made a mistake by waiting until September to cut interest rates.

Overseas, the Stoxx Europe 600 closed almost 5% below where it was a week ago, whereas the Swiss franc, a common haven asset, is up roughly 4%. The most eye-popping moves happened in Asia, though, where the Nikkei 225 plunged 12.4% Monday in the worst trading session since Oct. 20, 1987—the day that followed Wall Street’s infamous Black Monday.

Yet it is precisely the breakneck speed with which Japanese equities tumbled that should give most investors a reason to remain calm.

As a guideline, sudden market selloffs are less dangerous than those that unfold progressively over time. This is because investors who rationally price in bad economic data often do so slowly, as it trickles in. Flash crashes, conversely, are often a sign that some tidbit of bad news made speculative bets go awry, triggering a cascade of trades, many of them automated.

Japan is particularly prone to such reversals because interest rates there are so low that many investors use them to fund higher-yielding investments in other currencies. Whenever markets get jittery, these “carry trades” tend to unravel, pushing up the yen and hitting Japanese stocks, many of which are diversified exporters that do better when global growth accelerates. Amplifying this tendency, Japanese stocks had this year become extremely popular among global investors.

The timing of the rout also points a finger at the Bank of Japan , which last week decided to tighten monetary policy for the first time in 17 years with the explicit goal of boosting the yen. Investors who rushed to cover their bets then triggered the reversal of stretched trades elsewhere, including in the U.S.

One of the most striking features of the S&P 500 for most of this year has been its extremely low volatility. Until July, the Cboe Volatility Index, or VIX, was at 2019 levels, and kept sliding lower even as investors made big changes to their monetary-policy forecasts.

While the VIX is often dubbed Wall Street’s “fear gauge,” the options contracts it is based on often themselves influence volatility. Whenever investors make bets against market swings, as they have recently in the U.S. by buying lots of structured products , the banks that sell those options are forced to take the other side. These hedges then suppress volatility in the stock market.

The flip side is that whenever a panic breaks through this feedback loop, volatility skyrockets. As the stock market opened Monday, the VIX hovered above 50, making it the highest weekly jump since the onset of the pandemic, though it later fell below 40.

This suggests the selloff is disproportionate, especially looking at the historical record: 87% of the time, investors who bought the S&P 500 on days when the VIX closed at 30 or higher ended up making money a year later.

The second-quarter reporting season has brought mostly good news, with 78% of the S&P 500 firms that have reported so far beating analysts’ earnings estimates—compared with a 74% 10-year average. Both AI-related companies and the rest are reporting net income above what was forecast a month ago. Overall, the U.S. economy still looks robust: The unemployment rate has gone up because the labor force has expanded.

Also, looking at S&P 500 returns since 1994 shows that selling based on the previous day’s falls is a bad strategy. Electing to move into cash after large monthly declines fared better, but still less well than sitting tight.

This isn’t to say that concerns about an economic slowdown or high tech valuations aren’t warranted. Investors have reasons to diversify away from the AI trend or swap more cyclically exposed stocks for more “defensive” names. Indeed, selling out of stocks after particularly exuberant days and months has historically tended to be a winning move. But hindsight is a terrible guide to investing your savings.

Artificial intelligence is known for its seemingly insatiable appetite for energy. But some tech leaders and analysts are questioning the extent of AI’s footprint going forward, saying innovations in the sector could help offset rising energy demand.

There is certainly no shortage of forecasts detailing the ominous rise in AI’s energy consumption. One commonly-cited report by Climate Action against Disinformation, an association of environmental groups, suggests that AI could drive up global emissions by 80%.

Another estimate by researcher Alex de Vries—which he described as a worst-case scenario—suggests that Google’s AI alone could eventually rack up as much annual electricity demand as the country of Ireland.

Such predictions present a substantial challenge for the relationship between computing and the climate in the coming years.

However, others see reason for optimism. In a Salesforce survey of about 500 corporate sustainability professionals, published Wednesday, nearly half were concerned about AI’s potential negative impacts on sustainability efforts. Meanwhile, almost 60% thought the benefits of AI would offset its risks in addressing the climate crisis.

Microsoft founder  Bill Gates recently weighed in on the subject, urging governments not to go “overboard” on concerns about AI’s energy footprint, and suggesting that the technology  could actually drive a reduction in global energy demands.

Putting the AI boom in context

The rise of AI should be considered in a broader perspective, according to Charles Boakye, U.S. sustainability analyst at Jefferies. He noted that relative to other industries, the technology still makes up only a fraction of global power demands.

“Data centers—the engines powering everything from AI to traditional computing to cryptocurrency—currently account for about 2% of global electricity consumption. Of this, AI accounts for roughly 0.5%,” Boakye said.

In terms of emissions, statistics last year from the International Energy Agency showed in total, data centers and transmission networks are responsible for 1% of energy-related greenhouse gases. For comparison, the oil-and-gas industry contributed just under 15% of emissions in 2022.

Looking ahead, the intergovernmental organisation said that by 2026, demand for AI is expected to increase ten fold compared with 2023. Even so, IEA data showed that it will still only account for roughly an eighth of total data-center electricity consumption.

“So in terms of AI demand, we’re talking about a small piece of a small piece,” Boakye said, noting that other areas of electrification, such as electric mobility, traditional data center growth and the industrial transition, will ask much greater questions of the power sector.

Charting efficiency trends 

Demand is difficult to predict. But recent trends show that in practice, a rise in demand for computing power rarely correlates one for one with a rise in energy consumption, according to climate researcher Jonathan Koomey.

“Between 2010 and 2018, global data centres saw a 550% increase in compute instances and a 2,500% increase in storage capacity. This compares with just a 6% rise in electricity use,” he said.

Koomey, previously a visiting professor at Stanford, Yale and Berkeley, is best known for his work studying long-term trends in the energy-efficiency—now known as Koomey’s Law—highlighting the propensity of computing technologies to become more efficient over time.

AI may be a different animal, with some estimates suggesting large models such as ChatGPT use 10 times more energy than a Google search. But this is unsurprising for a relatively new technology, Koomey noted. In most areas of computing, energy demand tends to spike before levelling off as efficiencies gather pace, he said, noting that forecasts based on this inflection point will often be misguided.

Koomey cited a brief moment in the mid 1990s when internet data flows doubled every hundred days—a statistic that led to overinvestment in networks in the following years and 97% of fibre capacity sitting unused.

Similar efficiency trends can be seen in the development of AI, Jefferies analyst Boakye said. Google’s new TPU processors, for example, are more than 67% more efficient than in 2022, and the energy used to train OpenAI’s ChatGPT models has gone down by 350 times since 2016, he noted.

“It’s in their best interest, and in the interest of their business models, to increase that efficiency,” the analyst said.

Going for gold in the AI Olympics 

If any company will have a say in the future of AI, it’s Nvidia . The U.S. technology company designs roughly 80% of the world’s specialized AI chips. Its next-generation GPUs, known as Blackwell, are touted to be 25 times more energy efficient than current iteration Hopper, while offering 30 times more computing power. Blackwell chips are slated for release later this year.

So far, the progress appears consistent, with Hopper 25 to 30 times more efficient than Nvidia’s previous generation of chips. Overall, the company said it has experienced a 45,000 times improvement in GPU energy efficiency over the last eight years.

Nvidia added that software optimisation also plays a big role in increasing the energy efficiency of its products.

“Once a platform has been launched, we will make it more efficient in a single year,” the company said. In the year following its launch in 2022, Hopper became two times more efficient after taking part in MLPerf, otherwise known as the Olympics of AI, in which tech companies compete and collaborate to drive improvements in the speed and efficiency of their models.

Part of this optimization involves taking large, energy-intensive AI models—such as ChatGPT—and refining them to perform more specific tasks. On July 18, for example, OpenAI launched a  smaller, smarter and more energy efficient  version of its previous GPT model, known as GPT-4o mini. Koomey sees these more “lightweight” models  driving demand in the future .

“I’m not convinced large-scale AI has a good business model at this point, despite them driving all the investment. The slam dunk machine learning usually involves smaller, more efficient models, focused on a specific task. For me, this is where most of the business value lies,” he said.

Training and education will be essential for getting a more targeted and efficient experience with AI, helping to narrow the gap between businesses and their sustainability goals, said Suzanne DiBianca, Salesforce’s chief impact officer.

According to Nvidia, another way of alleviating AI’s impact is directing workloads—particularly the more energy-intensive jobs of training AI models—to regions with more abundant resources, ideally renewables which are set to keep a lid on electricity emissions in the coming years.

One company making strides in this space is NexGen Cloud, which builds renewable-powered data centers in areas with untapped energy resources.

According to co-founder and CSO Youlian Tzanev, a large portion of AI workloads don’t need to be performed close to traditional logistical hubs like London or New York, and can be powered from more remote areas in countries like Canada and Norway with excess hydropower.

“We have significantly more power than people believe. The power just isn’t reaching the grid in many cases and is going to waste, and so that is where we focus our efforts,” Tzanev said.

In the U.S., Crusoe Energy Systems offers another example of startups finding innovative ways to power the AI boom. Crusoe’s modular data centers are designed to run on excess natural gas produced at oil wells, achieving a 99.9% methane reduction in the process.

Common forecasting pitfalls 

Projecting an outlook for AI’s energy demands is far from straightforward. In its midyear electricity update, the IEA noted that estimates exhibit a wide range of uncertainty, with some analyses following overly “simplistic” extrapolations.

For instance, the organization said certain studies make the mistake of assuming data center operators build all the facilities for which they apply to utilities. Given that several applications can be made for each new data center, this can lead to a multiplication of estimates.

Within data centers themselves, it is tempting for forecasters to imagine computers working flat out around the clock, Koomey said. In practice, GPUs usually operate on much less than their full power capacity, he added.

Based on modeling carried out by Nvidia, GPUs on average tend to run on less than 70% of their potential power. One particular function, known as Multi-Instance GPU, enables workloads—and therefore energy consumption—to be split into seven distinct components, with each able to function independently.

In addition, the company noted the role of substitution effects, in which traditional computing workloads are transferred onto AI platforms and subsequently performed more efficiently—an aspect that can be easily overlooked.

Forecasts can also  conflate local data-centre developments with broader energy demands , Koomey added, noting that the most common estimates for electricity demand come from local utility companies.

In the U.S. as a whole, electricity use actually fell in 2023 compared with the previous year, according to the U.S. Energy Information Administration. In March 2024—the most recent month of available data—total demand reached 306 billion kilowatt-hours, down from 317 billion kWh in the year-prior period.

“I worry that people are jumping on the explosive demand-growth train before really understanding what’s going on,” Koomey said. “If you cluster data centers in certain places you’re going to see some local power constraints, but that doesn’t necessarily mean that AI will be a key driver of electricity use more broadly.”

Corrections & Amplifications undefined Nvidia said it has experienced a 45,000 times improvement in GPU energy efficiency over the last eight years. An earlier version of this article incorrectly said the company developed its first GPU eight years ago. (Corrected on July 24)