In 2010, a company called Spread Networks opened a fiber-optic line between a trading data center in Chicago’s South Loop and a building across the street from Nasdaq’s servers in Carteret, New Jersey. Forbes estimated it cost about $300 million. The line cut the one-way trip for a trade signal from about 8 milliseconds to 6.65, according to a history of market technology published by the Cato Institute.
Rishabh Singh, co-founder and chief technology officer of JoinMeNow Inc., says he worked on the software side of that race, spending years “shaving nanoseconds off trading systems at global banks,” as he put it in response to written questions. He says that work meant writing ultra-low-latency trading software in C++.
His argument concerns where the competitive edge moved next. Singh says the edge in finance is shifting from how fast a firm can act to how well its software can interpret what is happening. “The story of how finance’s edge shifted from raw speed to intelligence is one I never get tired of telling,” he wrote.
In his written answers, he describes the kind of systems he is after as ones that “don’t just predict, but reason, adapt, and learn.”
The market record supports part of that argument and complicates the rest.
Two different problems
Singh’s career, as he describes it, covers both sides of the question. After the low-latency work, he says, he built machine learning models for credit risk and pricing, fine-tuned language models on financial text, and developed AI systems for financial forecasting and optimization.
Those systems solve different problems. Low-latency engineering shortens the time between a signal and a trade. Forecasting, pricing, and risk models try to improve the decision itself. Neither is simply more advanced than the other. A firm can have the fastest connection in the market and still act on a bad prediction, or have the best model and lose the trade to a quicker rival.
Singh argues that more of the competitive differentiation now comes from the second problem. He describes his own path the same way, moving from low-latency systems to AI forecasting work.
What speed still buys
Fast execution still matters. A 2020 study by economists at the UK Financial Conduct Authority and the University of Chicago looked at “latency arbitrage” races, in which the fastest firms trade against prices that have not yet incorporated the newest market information. The researchers found about one latency-arbitrage race per minute per FTSE 100 symbol, and the typical race lasted 5 to 10 millionths of a second.
The amounts involved were substantial. Races accounted for about 20% of trading volume in the London data. Extrapolating from that UK data, the researchers estimated roughly $5 billion a year was at stake across global equity markets, according to the FCA paper. Six firms accounted for more than 80% of race wins and losses.
At the same time, the economics of high-frequency trading became less lucrative. Tabb Group estimated that U.S. high-frequency trading revenue from stocks fell from $7.2 billion in 2009 to $1.1 billion in 2016, Markets Media reported. The causes cited included calm markets, thinner margins, and competitors closing the technology gap.
One reasonable reading of that last point is that once many firms reached similar speeds, speed alone did less to separate them. That reading fits Singh’s argument. The data does not establish the cause.
Where AI fits
AI adoption rose sharply among the surveyed UK financial services firms. Among firms responding to a survey by the Bank of England and the Financial Conduct Authority, the share using AI rose from 58% in 2022 to 75% in 2024. Another 10% of the 118 firms that responded in 2024 said they planned to adopt it within three years.
The International Monetary Fund, in its October 2024 Global Financial Stability Report, said AI adoption in capital markets “is likely to increase significantly in the near future.” It also warned of a downside. More AI in trading could mean “increased market speed and volatility under stress, especially if trading strategies of AI models all respond to a shock in a similar manner.”
That warning undercuts any clean before-and-after story. AI does not take speed out of markets. It can add to it. Taken together, the evidence points less to fast execution being replaced than to speed and increasingly sophisticated analysis operating together.
Singh’s bet is that the next competitive advantage will depend less on shaving another fraction of a millisecond and more on what software can infer from the information that arrives in that time. By Forbes’s count, Spread Networks’ line cut roughly three milliseconds from a round trip. By 2016, microwave networks had reduced the one-way time further, to about 3.98 milliseconds. Whether an advantage built on better interpretation and prediction proves more durable than one built on faster infrastructure remains an open question.