In September 2000, Goldman Sachs paid $6.5 billion to acquire Spear, Leeds & Kellogg, the largest specialist firm on the New York Stock Exchange. SLK’s business model was elemental: stand between buyers and sellers, capture the bid-ask spread on every transaction. The firm handled more volume on the NYSE than any other specialist. It was a franchise built on physical presence — a position on the trading floor, a book of orders, a set of relationships with brokers who needed someone to make a market.
Within a decade, the investment was largely worthless. Goldman sold the specialist post to IMC Financial Markets, an electronic market maker, in 2014. The franchise that had been worth $6.5 billion was, by then, a rounding error — a desk overtaken by machines that could do the same job faster, cheaper, and without lunch breaks.
What happened between 2000 and 2014 was high-frequency trading.
The friction hunters
A market maker earns money from the spread — the difference between the price at which someone will buy a security and the price at which someone will sell it. Before 2001, the minimum increment on the NYSE was one-eighth of a dollar — $0.125 per tick. Specialists captured that spread on every transaction, and the spread was, by regulation, wider than it needed to be. This was a structural rent: the price of having a human being stand in the middle of every trade.
High-frequency trading firms are, at their core, friction hunters. They identify any inefficiency in a market — a stale quote, a price discrepancy between exchanges, a spread wider than competition requires — and they compete it away. They do this at speeds that are difficult to comprehend without analogy.
When the NYSE moved to decimal pricing on January 29, 2001, the minimum tick dropped from $0.125 to one cent. Effective spreads for NYSE stocks fell 41%. On NASDAQ, total trading costs fell roughly 50%, from 14.6 cents per share to 7.4 cents. Decimalization was the regulatory precondition. But it was HFT that turned the precondition into a permanent compression. Algorithmic firms could now quote in penny increments, updating prices thousands of times per second, hunting every fraction of a cent that the old market makers had captured through position and patience.
Then came Regulation NMS in 2007, which required that orders be routed to the exchange offering the best price, regardless of venue. The rule fragmented trading across dozens of exchanges and alternative trading systems. For human market makers, fragmentation was a logistical problem. For algorithms, it was an opportunity: more venues meant more price discrepancies, more stale quotes, more friction to hunt.
The speed of money
What followed was an arms race measured in microseconds.
In 2010, Spread Networks spent $300 million to build an 827-mile fiber-optic cable from Chicago to Carteret, New Jersey — from the futures exchanges to the equity exchanges — reducing round-trip latency from 16 milliseconds to 13 milliseconds. Three milliseconds. For $300 million, you could trade three-thousandths of a second faster than your competitors.
Within years, microwave networks rendered the cable obsolete. Microwave signals travel through air at nearly the speed of light, faster than light through glass fiber. Jump Trading paid €5 million at a Belgian government auction for a former US Army microwave tower in Houtem, Belgium — to shave microseconds off the Frankfurt-to-London link. Competing providers built chains of microwave repeaters between Chicago and New Jersey, pushing round-trip times below 9 milliseconds, approaching the theoretical minimum at the speed of light along that path.
Today, the state of the art has moved to field-programmable gate arrays — FPGAs — custom hardware that can execute trading decisions in single-digit nanoseconds. Eurex, the European derivatives exchange, has measured response times of 8 nanoseconds. Not milliseconds. Nanoseconds. A nanosecond is to a second what a second is to roughly 31 years.
The economics of this arms race tell you everything you need to know about where it leads. Virtu Financial, one of the surviving HFT firms, disclosed in its 2014 IPO filing that it had lost money on only one out of 1,238 trading days between January 2009 and December 2013. An updated filing showed 1,277 profitable days out of 1,278 over five years. This was not luck and it was not skill in the traditional sense. When you see a price discrepancy and act on it before anyone can respond, you capture the spread nearly every time. The near-perfection of the record is the point: at sufficient speed, the outcome stops being probabilistic.
The academic evidence puts numbers on the dynamic. Aquilina, Budish, and O’Neill, in a 2022 study published in the Quarterly Journal of Economics, analyzed billions of messages on the London Stock Exchange and found that latency-arbitrage races — moments where multiple firms race to exploit the same price discrepancy — occur roughly once per minute for every FTSE 100 stock. The modal race lasts 5 to 10 microseconds. The top six firms account for more than 80% of all race wins and losses. These races constitute approximately one-third of the effective spread, costing global investors an estimated $5 billion per year.
The defining characteristic
The $300 million cable. The €5 million tower. The 8-nanosecond response time. The one losing day in five years.
Strip away the specifics and the pattern is simple. HFT firms do not create products or build brands or cultivate customer relationships. They find the gap between what a price is and what a price should be, and they close it — faster than anyone else, and then faster than each other.
The old rents are gone. The specialists who once captured an eighth of a dollar on every trade have been replaced by algorithms that compete for pennies, and then fractions of pennies, and then fractions of fractions. The question is what happened to those specialists, to the trading floor they occupied, and to the ecosystem that depended on the rents they extracted.
That is the story of what HFT did to the old world — and the template for what cheap intelligence is doing to every market that depends on cognitive work.
The specialists are gone. The question is what replaced them, what the transition cost, and why the pattern matters far beyond finance.
The emptying floor
In the early 1990s, the NYSE trading floor held more than four thousand people — specialists, floor brokers, clerks, runners — filling rooms that the exchange had expanded twice to accommodate, in 1969 and again in 1988. Every trade passed through someone’s hands. The specialist firms, more than fifty of them in the mid-1980s, many family-owned, controlled order flow in their assigned stocks and earned the spread on every transaction.
That ecosystem did not decline gradually. It collapsed.
By mid-2005, the floor was down to 837 brokers and 2,475 clerks. By year’s end, after another wave of electronic adoption, it was 412 brokers and 966 clerks. The expansion rooms from 1969 and 1988 were shuttered in late 2007. Today, the NYSE’s floor broker directory lists five firms. Three designated market makers — Citadel Securities, GTS Securities, and Virtu Americas — handle the work that more than fifty specialist firms once split among themselves. The floor processes a fraction of NYSE order flow. Its remaining functions are opening and closing auctions, the occasional large block order, and the bell-ringing ceremony that still makes the evening news.
The specialist firms followed their own trajectory: consolidation, acquisition, irrelevance. Goldman Sachs paid $6.5 billion for Spear, Leeds & Kellogg in 2000. Fourteen years later, it sold the DMM post to IMC Financial Markets. The franchise did not shrink. It evaporated.
Eating their own
The destruction of the old guard was the first phase. The second is more interesting: the HFT firms turned on each other.
In 2009, at the peak of algorithmic market making’s gold rush, US equity HFT revenue hit $7.2 billion. That figure reflected a market still adjusting to electronic speed — spreads compressed, but not to their limit. Volatility from the financial crisis created constant dislocations to exploit. Firms with faster infrastructure captured outsized returns.
Then the competition caught up.
By 2012, HFT revenue in US equities had fallen to $1.8 billion. By 2016, $1.1 billion. By 2017, TABB Group estimated it had dropped below $1 billion — an 86% decline from the 2009 peak. The firms that had competed away the specialists’ rents were now competing away each other’s. Each improvement in speed or strategy was matched within months, sometimes weeks. The edge compressed to the width of a nanosecond, and the profit compressed with it.
The $300 million Spread Networks cable, state-of-the-art in 2010, sold to Zayo Group in 2017 for $127 million — a $173 million loss on infrastructure that microwave networks had rendered obsolete within years of construction. The fastest became the second-fastest, and in this market, second-fastest earned nothing.
The catastrophic speed
HFT compressed margins to near-zero. It also compressed the time horizon for catastrophic failure.
On August 1, 2012, Knight Capital deployed a software update to its automated order routing system. A technician failed to copy the new code to one of eight servers. That single unpatched server reactivated an obsolete function called Power Peg, which began sending orders into the market without recording their fulfillment. An infinite loop. In forty-five minutes, Knight executed more than four million unintended trades across 154 stocks, accumulating $7 billion in unwanted positions.
The pre-tax loss was $440 million. The stock collapsed more than 70%. Knight was rescued four days later by a consortium that injected $400 million in emergency capital, acquired by Getco in December 2012 for roughly $1.4 billion, and absorbed into KCG Holdings. Virtu acquired KCG in 2017. The Knight Capital name ceased to exist — one unpatched server, forty-five minutes, a company gone.
Two years earlier, the Flash Crash of May 6, 2010, had demonstrated the systemic version of the same fragility. A single algorithmic sell program — Waddell & Reed offloading 75,000 E-Mini S&P 500 futures contracts as a portfolio hedge — interacted with thinning liquidity and triggered a cascade. The Dow dropped 998 points in thirty-six minutes. Shares of Accenture traded at one cent. Procter & Gamble fell 37% in minutes before rebounding. Nearly a trillion dollars in market value vanished and reappeared within the hour.
The speed that created the advantage also created the fragility. One server. One algorithm. The margin for error is measured in the same units as the margin for profit: effectively zero.
The new equilibrium
What emerged from this cycle — specialists destroyed, HFT firms cannibalizing each other, consolidation through acquisition — was not chaos. It was a new equilibrium, and it looks nothing like the old one.
Three designated market makers now cover the NYSE. Citadel Securities alone executes roughly 35% of all US retail equity volume and about 25% of total US equity trading — more volume than the entire Nasdaq exchange. Citadel and Virtu together handle approximately 40% of daily US trading flow. Off-exchange trading crossed 50% of all US equity volume in late 2024. More than half of all stock trading in America now occurs away from the exchanges that were, a generation ago, the only game in town.
The surviving firms are not struggling. In 2024, Jane Street generated $20.5 billion in net trading revenue. Citadel Securities posted $9.7 billion. Hudson River Trading earned roughly $8 billion. These are not the compressed margins of a dying industry. These are the returns of an oligopoly operating at the frontier of infrastructure, scale, and computational speed — the only dimensions that still matter once the old rents have been competed away.
The barriers to entry are now different in kind from the old specialist barriers, and arguably higher. Colocation fees run to a million dollars per year. Competitive latency requires custom FPGA hardware. The top six firms account for more than 80% of all latency-arbitrage race outcomes, per the Aquilina, Budish, and O’Neill study. No significant new entrant has broken into the top tier in over a decade.
This is what Richard D’Aveni’s hypercompetition framework predicts as the endgame: not permanent chaos but a new steady state. An oligopoly of the most capable firms, operating where margins are thin, competition is continuous, and the barriers to entry are no longer knowledge or relationships but capital and computation. The old rents are gone. The new rents belong to whoever can afford the infrastructure to compete at the frontier.
The floor traders lost to the algorithms. The algorithms competed each other to near-zero. The survivors consolidated. And the market kept functioning — faster, cheaper for end users, brutally concentrated at the top.
The HFT parallel is not a metaphor. It is a structural description of what happens when the cost of identifying and exploiting inefficiency drops toward zero.
In financial markets, HFT firms scan for any gap between what a price is and what a price should be — a stale quote, a spread wider than competition requires, a momentary dislocation between correlated securities — and they close it, in microseconds, for fractions of a penny. The specialists who once earned the spread are gone. The rents built on the cost of having a human being stand in the middle of every transaction were competed away by machines that could do the same work faster, cheaper, and continuously.
Now apply the same lens to every market that depends on cognitive work.
The friction map
A customer support call handled by a human agent costs $2.70 to $5.60. An AI voice agent handles the same call for roughly $0.20. Klarna deployed an AI chatbot in early 2024 that handled two-thirds of all customer chats within its first month, doing the work of 700 full-time agents, cutting resolution time from eleven minutes to under two. The company’s cost per transaction fell from $0.32 to $0.19 over two years.
A junior associate at a major law firm bills five hours to research a legal question. Thomson Reuters’ CoCounsel performs the same research in five minutes. Harvey AI, valued at $8 billion, is now used by most of the largest US law firms. Law firm AI adoption surged from 19 percent to 79 percent in a single year. And yet only a handful of firms pass the savings to clients — the rest pocket the difference or charge a premium for “AI-enhanced” work. The spread is being captured, not by the firms hunting the friction, but by the incumbents who adopted the tools first. For now.
In drug discovery, the traditional path from hypothesis to clinical candidate takes 2.5 to 4 years and costs hundreds of millions of dollars. Insilico Medicine’s AI platform nominates drug candidates in an average of 13 months. Its lead compound, rentosertib — where both the target and the molecule were identified by AI — completed a Phase IIa clinical trial with positive results published in Nature Medicine in June 2025. Only 60 to 200 molecules were synthesized per project, compared to thousands in conventional screening. The preclinical cost: roughly $2.6 million versus an industry average of $430 million.
The pattern repeats across every sector where cognitive labor sets the price floor. Lemonade’s AI cut insurance claims processing from $44 to $14 per claim. GitHub Copilot users complete coding tasks 55 percent faster, with the tool now generating 46 percent of all code written by its users. AI-generated marketing content costs an average of $131 versus $611 for human-written equivalents. Ramp’s analysis found that more than half of companies using freelancers in 2022 had stopped entirely by 2025, substituting $1 of reduced freelance spending for $0.03 in AI costs — a 33-to-1 replacement ratio.
Every one of these numbers is a spread. Every spread is a margin that someone built a business on. And every one is being compressed by competitors who can identify the friction and undercut it at machine speed and machine cost.
The speed arms race
In HFT, the modal latency-arbitrage race lasts 5 to 10 microseconds. The top six firms account for more than 80 percent of all race outcomes. The races constitute roughly one-third of the effective spread. The competition is not over whether someone will capture the friction. It is over who captures it first.
In AI markets, the units are different but the dynamic is identical.
OpenAI launched Deep Research on February 2, 2025. Within twenty-four hours, Hugging Face had an open-source clone scoring 55 percent on the GAIA benchmark versus OpenAI’s 67 percent. Five more open-source alternatives appeared within the same week. Jina AI’s CEO had a functional clone merged into his codebase in twelve hours.
Google’s NotebookLM went viral in September 2024. Meta released an open-source equivalent, NotebookLlama, within weeks. Claude Code launched in February 2025; OpenAI’s Codex CLI followed fifty-one days later; Google’s Gemini CLI arrived four months after that. The terminal coding agent category went from one to three major players in under six months.
In November 2025, four frontier AI labs released flagship models within twenty-five days of each other: Grok 4.1, Gemini 3, Claude Opus 4.5, GPT-5.2. Google’s Gemini 3 reportedly triggered a “code red” at OpenAI, accelerating the release of GPT-5.2. By February 2026, Anthropic and OpenAI were releasing competing models on the same day — Claude Opus 4.6 and GPT-5.3-Codex launched within minutes of each other.
This is not the diffusion curve of previous technologies, where a dominant design emerged and competitors iterated over years. This is latency arbitrage applied to product development. Every advance by one firm forces an immediate response from every other. The window between launch and replication has compressed from years to months to weeks to, in some cases, hours.
The margin compression
HFT compressed market-maker spreads by more than 40 percent within years of decimalization. Then it compressed its own margins: US equity HFT revenue fell 86 percent from peak as firms competed with each other toward zero.
AI is on a parallel trajectory.
Traditional SaaS companies operated at 80 to 90 percent gross margins — software, once written, costs almost nothing to distribute. Companies integrating AI into their products report gross margins of 60 to 70 percent. AI-centric companies run at 50 to 60 percent. Early-stage AI startups, the ones burning through inference costs to acquire customers, operate at roughly 25 percent. Eighty-four percent of companies in a 2025 survey reported margin erosion of 6 percent or more from AI infrastructure costs alone.
The cost of the intelligence itself is falling faster than HFT ever compressed spreads. Sam Altman claims the cost to use a given level of AI falls roughly 10x every twelve months. Epoch AI’s independent analysis supports this: the price per token at equivalent model performance has declined by a median of 50x per year since early 2024, with some benchmarks showing 200x declines. OpenAI’s own pricing fell 150x from GPT-4 to GPT-4o in roughly fourteen months. When DeepSeek V3 demonstrated frontier-class performance at roughly one-tenth the cost per token, Anthropic responded by cutting Claude’s pricing 67 percent.
This is the AI equivalent of Spread Networks spending $300 million on a cable that microwave networks rendered obsolete within years. Except in AI, the obsolescence cycle is measured in months.
The February 2026 SaaS crash made the dynamic visible to everyone at once. More than a trillion dollars in enterprise software market capitalization erased in weeks. Salesforce down 38 percent. ServiceNow down 50 percent over twelve months. The market was repricing the entire sector around a single realization: if ten AI agents can do the work of a hundred employees, companies do not need a hundred SaaS seats. The per-seat licensing model that enterprise software was built on is a spread, and the spread is being hunted.
Every margin built on the cost of human cognitive labor is a bid-ask spread waiting to be compressed. The firms doing the compressing are not, in most cases, the incumbents. They are the algorithmic equivalent of HFT market makers: lean, fast, infrastructure-native, scanning continuously for the next friction to exploit. And like HFT, the competition among the hunters is itself compressive — each price cut forcing the next, each temporary advantage eroding before it can be consolidated.
In 1994, Richard D’Aveni published Hypercompetition. His framework identified four arenas in which firms compete: cost and quality, timing and know-how, stronghold creation and invasion, and deep pockets. In each arena, he described escalation ladders — sequences where every competitive move provokes a response, and every response intensifies the competition rather than resolving it. A year later, he put the thesis bluntly in Fortune: “This is not an age of castles, moats, and armor, where people can sustain a competitive advantage for very long. This is an age that calls for cunning, speed, and enterprise.”
The framework was controversial. Michael Porter’s Five Forces model, which had dominated strategic thinking since 1980, held that firms could achieve sustainable competitive advantage through structural positioning — choosing the right industry, defending a profitable position within it, building moats. D’Aveni argued the opposite: the only sustainable advantage was the ability to string together temporary advantages faster than competitors could match them. Porter saw defensible castles. D’Aveni saw an open field where every castle was under siege.
For two decades, the debate was unresolved. Then the data started to arrive. Robert Wiggins and Timothy Ruefli, in a 2005 study in the Strategic Management Journal, analyzed thousands of firms over decades and found that fewer than 5% sustained superior economic performance for ten years or longer. Their title posed the question plainly: “Schumpeter’s Ghost: Is Hypercompetition Making the Best of Times Shorter?”
The answer, increasingly, is yes — and AI is the accelerant.
The four arenas, mapped
D’Aveni’s four arenas map onto the AI economy with a precision he did not anticipate.
Start with cost and quality — the arena where HFT manifested as spread compression, market-maker margins falling from an eighth of a dollar to fractions of a penny. In AI markets, the equivalent is the collapse in the cost of cognitive output: inference prices falling 50x per year according to Epoch AI’s analysis, customer support dropping from $5.60 per interaction to $0.20, SaaS gross margins compressing from 80-90% toward 50-60%. Each price cut forces the next. Each efficiency gain becomes the new baseline.
Then timing and know-how — the arena that in HFT produced the $300 million cable, the €5 million tower, the 8-nanosecond response time. In AI, it shows up as time-to-clone: Deep Research reproduced in 24 hours, four frontier models released in 25 days, competing products launched within minutes of each other. The window between innovation and imitation has compressed from years to months to, in some cases, hours. Know-how that once took years to accumulate can now be distilled from a foundation model in days.
Stronghold creation and invasion is where the HFT parallel gets structurally interesting. In financial markets, the shift ran from NYSE floor presence to colocation at data centers, from human relationships to direct market access. In AI, the Bruegel research group has documented a two-tier structure: a foundation layer trending toward oligopoly — Meta’s AI infrastructure alone represents more than $10 billion in GPUs — while the application layer is radically contestable, with thousands of startups competing atop the same models, any stronghold lasting months before it is invaded.
And deep pockets. In HFT, colocation fees, custom FPGA hardware, and microwave infrastructure meant that the cost to compete at the frontier rose even as margins fell. The same dynamic is emerging in AI: OpenAI projects $14 billion in losses for 2026 while seeking $100 billion in additional funding. At the foundation layer, the capital requirements are an entry barrier as formidable as any in economic history. At the application layer, the capital requirements are approaching zero — which is exactly why competition there is so intense.
Running to stand still
The endgame of D’Aveni’s escalation ladders is not a final resting point. It is what the evolutionary biologist Leigh Van Valen described in 1973 as the Red Queen effect, named for Lewis Carroll’s character who tells Alice: “It takes all the running you can do, to keep in the same place.”
William Barnett, in The Red Queen among Organizations, applied the concept to business competition. Organizations facing competitive pressure search for improvement. Successful search increases competitive strength — but it triggers learning in rivals, which triggers further learning in the first, which triggers further response. The cycle never terminates. “Competition,” Barnett wrote, “concerns relative performance, not absolute performance.” It does not matter how good you are. It matters how good you are relative to the firm that improved last week.
This is what the HFT equilibrium actually looks like from the inside. Jane Street generated $20.5 billion in trading revenue in 2024, and is spending a significant fraction of it on AI infrastructure. Citadel Securities handles 35% of all US retail equity volume. These are not firms resting on competitive advantage. They are firms running as hard as they can to maintain the positions they have — because the moment one slows down, a rival at the same speed takes the spread.
The AI economy has entered the same dynamic. OpenAI held 50% of enterprise LLM spending in 2023; by 2025 that had fallen to 27%, with Anthropic rising from 12% to 40%. ChatGPT’s consumer market share dropped from 87% to 65% in a single year. Leaderboard positions change monthly. The Red Queen is running.
The restructured economy
The HFT endgame is the template.
Financial markets were not destroyed by high-frequency trading. They were restructured. The old specialist system — fifty firms, rents built on physical presence and human judgment — gave way to an oligopoly of three designated market makers operating at compressed margins, with barriers made of infrastructure rather than knowledge. The markets kept functioning. Spreads narrowed. Costs fell for end users. And the competition never stopped.
The broader economy is heading somewhere similar. Ronald Coase argued in 1937 that firms exist because using the price mechanism has costs — discovering prices, negotiating contracts. A firm expands until its internal coordination costs equal the cost of transacting in the open market. Shahidi, Rusak, Manning, Fradkin, and Horton, in an NBER analysis they titled “The Coasean Singularity,” argued that AI agents are collapsing transaction costs toward zero — not quite reaching it, but approaching asymptotically — which reshapes the fundamental logic of why firms exist and how large they need to be.
This maps directly onto the two-tier structure already visible. At the foundation layer: concentration. Three providers account for 88% of enterprise LLM spending. Training costs are doubling annually. The barriers are capital, compute, and data at a scale that only a handful of organizations can sustain.
At the application layer: hypercompetition. Thousands of startups built on the same foundation models, competing on speed, design, distribution, and price. Ninety percent failing within eighteen months. The survivors building temporary advantages that erode before they consolidate. This is the rest of the market — firms hunting every friction, finding it, compressing it, and then watching as a dozen competitors do the same.
Jay Barney, who for decades championed the resource-based view of sustainable competitive advantage — the theoretical counterpart to Porter’s Five Forces — conceded the point in MIT Sloan Management Review in May 2025. AI, he wrote, will be “a source of homogenization, not differentiation.” The tools are available to everyone. The capabilities converge. What remains is what Barney called “residual heterogeneity” — the ability to go beyond what is accessible to all. That is D’Aveni’s thesis, stated by a former opponent: sustainable advantage is the exception, not the rule.
The economy is becoming a trading floor. Not in the sense that everything is financialized, but in the sense that every market built on cognitive work is being scanned continuously for friction, and every friction is being competed away at the speed that current technology allows. The specialists are being replaced. The algorithms are competing with each other. The margins are compressing.
And the competition — continuous, overlapping, relentless — is not a transition to something else. It is the destination. Hypercompetition is the new steady state.