The 2028 Trading Infrastructure Crisis
This piece is a scenario, not a prediction. It is written as a retrospective from June 2028 to explore what happens when AI-agent adoption in trading follows the trajectory that current trendlines suggest. Every earnings report, data point, and headline below is fictional. The structural logic connecting them is the point.
June 2028.
What if our AI bullishness for retail trading continues to be right — and what if that’s actually bearish for the $47 billion trading technology industry?
We have spent the past three years arguing that AI agents would transform how individuals interact with financial markets. That the terminal would replace the dashboard. That encoded expertise would matter more than information access. That process quality would become the only durable edge.
We were right about all of it. And the consequences have been far more disruptive than we expected — not for the traders who adopted early, but for the platforms, data vendors, and brokerages that assumed their business models would survive the transition.
This memo traces how that happened.
I. The Charting Platform Correction
Q3 2025 – Q4 2026
The first cracks appeared where they were least expected: in the products traders loved most.
TradingView entered 2025 with 60 million registered users, premium subscriber growth of 22% year-over-year, and a valuation north of $3 billion. The product was genuinely excellent. Pine Script had created a powerful ecosystem. The social features drove viral adoption. Management had every reason to believe the growth trajectory was durable.
They were pricing in a world where humans look at charts.
By late 2025, agentic coding tools achieved a step-function improvement in financial data synthesis. An intermediate developer using Claude Code could build a research pipeline — pulling market data from free APIs, computing technical indicators, running scenario analysis, and generating structured trade plans — in an afternoon. Not a product. A workflow. Personal, customizable, and free after the initial subscription.
The early adopters were precisely TradingView’s most profitable users: the power traders paying $60/month for Premium, the ones who used 8+ indicators per chart, the ones who needed multi-timeframe analysis across 30 tickers before market open.
These users didn’t cancel because TradingView got worse. They canceled because they stopped looking at charts.
TRADINGVIEW Q3 2026 EARNINGS: PREMIUM SUBSCRIBER CHURN ACCELERATES TO 8.2% Q/Q; MANAGEMENT CITES ‘WORKFLOW MIGRATION AMONG POWER USERS’
The Q3 2026 call was painful. Revenue growth decelerated to 11% from 22% the prior year. But the composition was worse than the headline: free-tier users were still growing. The churn was concentrated entirely in the high-ARPU cohort. Analysts pressed on what “workflow migration” meant. The CFO’s answer was precise and devastating:
“Our highest-value users are increasingly using AI agent workflows that pull the same market data we provide, perform the same technical analysis we visualize, and generate trade plans in text format. They are not switching to a competitor. They are switching to a different modality.”
Switching to a different modality.
Not a better charting platform. Not a cheaper charting platform. A world where charts themselves are optional.
We had overestimated the value of visual interfaces. A significant portion of what traders called “chart analysis” was pattern recognition and data synthesis that agents could perform directly on numerical data — faster, more consistently, and without the cognitive biases that plague visual interpretation.
This did not happen overnight. It was not universal. Discretionary traders who genuinely think in visual patterns — support/resistance, volume profiles, tape reading — continued to derive real value from charts. But the segment of users who used charts as a data access mechanism rather than a visual thinking tool was far larger than anyone had estimated.
And that segment was walking away.
II. The Data Vendor Unwind
2026 – Mid-2027
The Bloomberg Terminal is perhaps the most successful friction-premium product in the history of financial services.
$24,000 per year. Per seat. For data that is, in many cases, available from free or low-cost sources — packaged in a proprietary interface that only humans needed, distributed through a hardware lock-in model designed in the 1980s, and protected by a switching cost that was primarily social rather than technical.
Bloomberg’s actual moat was never the data. It was the terminal’s role as the shared operating system of institutional finance. The chat function alone — Bloomberg Messaging — created network effects that kept desks loyal even when cheaper alternatives existed. You kept your Bloomberg because your counterparties had Bloomberg.
AI agents do not have counterparties. They do not use chat. They do not need a proprietary interface to access earnings transcripts, economic indicators, or pricing data.
What agents need is structured data access. And by mid-2026, MCP servers provided exactly that — typed, reliable, programmatic access to market data, execution APIs, and analytical tools — at a fraction of the cost or entirely free.
The mid-market data vendors felt it first. Refinitiv (now LSEG Data & Analytics), FactSet, and S&P Capital IQ had spent years building “open API” strategies to compete with Bloomberg’s walled garden. They succeeded — and in doing so, made their data accessible to exactly the kind of automated workflows that would eventually undermine their seat-based pricing.
FACTSET Q1 2027: ENTERPRISE SEAT COUNT DECLINES 6% Y/Y; FIRST NET CONTRACTION IN COMPANY HISTORY
The mechanism was straightforward. A mid-tier hedge fund paying for 40 FactSet seats discovered that a team of three analysts using agent workflows could cover the same analytical surface area that previously required twelve. The data wasn’t less valuable. The human interface to the data was less necessary.
Multiply this across thousands of firms. Seat-based pricing assumes a stable ratio between data consumption and human headcount. When agents break that ratio, the pricing model collapses even if data demand increases.
BLOOMBERG LP FY2027 REVENUE GROWTH SLOWS TO 3.1%; FIRST TIME BELOW 5% SINCE 2009. TERMINAL UNIT SHIPMENTS FLAT Y/Y.
Bloomberg’s diversification into data licensing, indices, and media provided a cushion. But the terminal business — still the profit engine — was showing deceleration that management attributed to “cyclical headwinds in financial services hiring.” Analysts were less generous. A widely-circulated sell-side note titled The Last Terminal argued that Bloomberg’s addressable market was structurally shrinking:
“The total number of humans who need a dedicated financial data workstation is declining for the first time in four decades. Not because financial data is less important, but because the primary consumers of that data are increasingly non-human.”
III. The Brokerage Execution Revolution
Mid-2027
Retail brokerage in the United States was built on a simple insight: most individual traders don’t optimize execution quality. They optimize convenience.
This is why Robinhood could offer “free” trades while routing orders to Citadel Securities and Virtu Financial, who paid for the privilege of filling those orders at slightly worse prices than the best available. Payment for order flow generated $1.8 billion in revenue across the industry in 2024. The individual trader lost perhaps 1-3 basis points per trade. On a $5,000 position, that’s $0.50-$1.50. Invisible. Irrelevant, for a human.
Agents are not humans.
Agent-routed orders do not default to a single broker. They do not have a “home screen.” They do not have brand loyalty. When a trading agent is instructed to execute a buy order, it queries available execution venues, compares bid-ask spreads, evaluates fill probability, factors in any explicit routing costs, and routes to the best available option.
Every time. For every order. In milliseconds.
ROBINHOOD Q2 2027 EARNINGS: PFOF REVENUE DECLINES 34% Y/Y AS ‘AGENT-MEDIATED ORDER FLOW’ REACHES 28% OF TOTAL EQUITY VOLUME
The category “agent-mediated order flow” didn’t exist in brokerage reporting before Q4 2026. By Q2 2027 it had its own line item. Robinhood’s 10-Q disclosure was blunt:
“Agent-mediated orders exhibit significantly different routing characteristics than traditional retail flow. These orders are more likely to be routed to exchanges rather than wholesale market makers, are more price-sensitive, and show lower effective spreads. Our payment for order flow revenue per share for agent-mediated orders is approximately 64% lower than for traditional retail orders.”
The mechanism was the same one Citrini Research identified in consumer commerce: agents eliminate habitual intermediation. A human trader opens Robinhood because it’s on their home screen. An agent opens whatever execution venue offers the best fill for this specific order.
VIRTU FINANCIAL Q3 2027: NET TRADING INCOME FALLS 19% Y/Y; CEO CITES ‘STRUCTURAL SHIFT IN RETAIL ORDER FLOW QUALITY’
The wholesale market makers had built their businesses on the predictable behavioral patterns of retail traders — patterns that agents systematically eliminated. Not through sophistication, but through indifference. Agents don’t have behavioral biases. They don’t panic-sell at the bottom or FOMO-buy at the top. They execute the plan they were given, at the best price available.
The irony was sharp: the same AI technology that the major market makers were deploying to improve their own execution was simultaneously making their most profitable customers — uninformed retail flow — less profitable to serve.
IV. The Platform Consolidation Spiral
Late 2027
With charting commoditized, data access democratized, and execution comparison-shopped by agents in real-time, the traditional competitive moats in retail trading technology dissolved in sequence.
The question became: what’s left to differentiate on?
The answer, for platforms that survived, was brutally simple: API quality, MCP server reliability, and latency.
Not the user interface. Not the mobile app. Not the social features or the gamification or the confetti animation on your first trade. The only thing that mattered was how well your infrastructure served agents.
INTERACTIVE BROKERS ACQUIRES THREE NEOBROKERS IN SINGLE QUARTER; COMBINED ENTERPRISE VALUE 62% BELOW 2025 PEAKS
Interactive Brokers had built for this world accidentally. Their API had been comprehensive and reliable for over a decade — originally for institutional clients and algorithmic traders, now repurposed for agent workflows. While competitors scrambled to build MCP servers and typed API interfaces, IBKR’s infrastructure was already agent-compatible.
The neobrokers — the generation of platforms built on beautiful mobile apps, social trading features, and frictionless onboarding — found themselves structurally disadvantaged. Their entire product thesis was the quality of the human interface. When the human stepped back and the agent stepped forward, the beautiful interface became irrelevant.
We watched three distinct platform failure modes play out:
The GUI Trap. Platforms that had invested heavily in proprietary charting, custom indicators, and visual analytics found that these features — their primary differentiators — were the first things agents made unnecessary. Years of UI/UX investment became stranded assets.
The Data Silo Trap. Platforms that had locked proprietary data behind closed APIs found that agents simply routed around them. If your options chain data requires a proprietary desktop client, agents will pull it from the broker that exposes it via MCP. Data silos became a competitive disadvantage, not a moat.
The Social Trap. Copy-trading platforms, social sentiment feeds, and community-driven features assumed that traders value other traders’ opinions. Agents don’t follow influencers. They don’t copy trades. They execute systematic analysis. The entire social trading category contracted by an estimated 40% in user engagement between Q1 2026 and Q4 2027.
V. The Skills Economy
2027 – 2028
Something unexpected happened as the old infrastructure collapsed: a new value layer emerged.
Not in tools. Not in data. Not in execution. In encoded expertise.
The early CLI trading community had been developing what they called “Skills” — reusable, domain-specific reasoning workflows that agents could execute. A Skill wasn’t a trading algorithm. It wasn’t a signal generator. It was a structured way of thinking about a specific problem — IV surface analysis for options, liquidity assessment for DeFi positions, event probability modeling for prediction markets — encoded in a format that agents could reliably execute.
By mid-2027, the best Skills were demonstrably superior to generic agent analysis for their specific domains. Not because the underlying LLM was smarter, but because the Skill encoded the domain knowledge, decision frameworks, and risk heuristics that took experienced traders years to develop.
“THE INTELLIGENCE PREMIUM DIDN’T DISAPPEAR FROM TRADING. IT MIGRATED. FROM THE FIRMS THAT SOLD YOU TOOLS TO WATCH THE MARKET, TO THE INDIVIDUALS WHO ENCODED HOW TO THINK ABOUT IT.”
This was the structural insight that most industry observers missed. They saw the platform collapse and concluded that AI was commoditizing trading entirely. But commoditization only applied to the infrastructure layer — the data access, the charting, the execution routing. The expertise layer was actually becoming more valuable, not less.
A trader who had spent fifteen years developing intuition about options market microstructure could now encode that intuition into a Skill that worked 24/7, applied consistently to every opportunity, and improved iteratively. Their edge didn’t disappear. It scaled.
The analogy to software development was precise: open-source frameworks commoditized the infrastructure of web applications, but the value of domain-specific software increased. Nobody pays for a web server anymore. Everyone pays for the application running on it. Similarly: nobody would pay for data access or charting. But encoded trading expertise — the application layer — commanded a premium.
A cottage industry emerged. Independent traders and small firms began publishing Skills on open repositories. Some were free. Some were licensed. The best ones — the ones built by traders with genuine domain expertise and rigorous backtesting — developed reputations that attracted serious capital.
VI. The Individual Trader Ascendancy
Q1 2028
By early 2028, the operational capability gap between a disciplined individual trader and a mid-tier institutional desk had narrowed to a degree that would have been inconceivable three years earlier.
Not in every dimension. High-frequency execution remained the domain of co-located infrastructure and billion-dollar budgets. Prime brokerage relationships, regulatory licenses, and access to private capital markets were still walled off. The structural advantages of institutional scale in certain strategies were real and durable.
But for the strategies where the edge comes from research quality, analytical rigor, and process discipline — fundamental analysis, event-driven trading, cross-asset macro, options relative value, DeFi yield optimization — the playing field had functionally leveled.
Consider what an individual trader with a terminal-first workflow had access to by Q1 2028:
- Data: The same earnings transcripts, economic releases, options chains, on-chain data, and alternative datasets available to institutional desks — via MCP servers that cost nothing or near-nothing to run
- Analysis: Agent-assisted research pipelines that could synthesize 50 data sources in minutes, apply consistent analytical frameworks, and generate structured trade plans
- Execution: Agent-routed orders that achieved execution quality competitive with institutional VWAP algorithms for standard order sizes
- Risk management: Encoded guardrails — position limits, correlation checks, drawdown controls, exposure caps — that ran automatically on every trade
- Review: Automated trade journals with full decision audit trails, performance attribution, and process quality metrics
The $180,000-per-year junior analyst at a multi-strategy fund and the independent trader working from a terminal in their apartment now used functionally equivalent data, equivalent analytical frameworks, and equivalent execution infrastructure. The remaining differences were risk capital, regulatory access, and — critically — the quality of their encoded expertise.
CITADEL SECURITIES Q1 2028 INVESTOR LETTER: “THE RETAIL PARTICIPANT HAS EVOLVED FROM A LIQUIDITY PROVIDER TO A LIQUIDITY COMPETITOR IN SEVERAL STRATEGY CATEGORIES”
That sentence landed like a grenade in institutional allocator circles. Ken Griffin’s firm — the single largest beneficiary of unsophisticated retail order flow for over a decade — was now acknowledging that the flow had changed character. Not all of it. Not most of it. But enough to compress margins in strategies where institutional and retail participants overlapped.
The mechanism was not that individual traders became smarter. It was that agent-assisted workflows eliminated the process errors that had historically separated retail from institutional performance.
Retail traders didn’t lose money because they lacked information. They lost money because they:
- Skipped pre-trade checklists under time pressure
- Sized positions inconsistently based on emotional state
- Failed to log and review losing trades
- Abandoned systematic analysis when markets moved fast
- Let winners ride too long and cut losers too late
Every single one of these failure modes is a process problem. And process problems are exactly what agents solve.
When your risk controls are encoded rather than remembered, when your analysis pipeline runs the same way every time regardless of your emotional state, when your trade journal is generated automatically rather than voluntarily, when your position sizing is calculated by formula rather than feeling — the behavioral gap between retail and institutional narrows dramatically.
VII. The Uncomfortable Question
June 2028
The trading technology industry is undergoing a repricing that mirrors what happened to enterprise SaaS: not because the products were bad, but because the interface assumptions they were built on turned out to be temporary.
The total market capitalization of publicly traded retail trading platforms has declined 44% from 2025 peaks. Data vendor revenue growth has turned negative across the mid-market. Brokerage PFOF revenue has fallen below the level that sustains commission-free trading for some smaller firms. The social trading category has effectively ceased to exist as a standalone business model.
At the same time, individual traders executing disciplined, agent-assisted terminal workflows are reporting the best risk-adjusted returns in the short history of the practice. Not because markets are easy — 2027 was one of the most volatile years on record — but because process consistency shines brightest when conditions are hardest.
The uncomfortable question is this: was the entire retail trading technology stack — the $47 billion ecosystem of charting platforms, data terminals, social trading apps, and gamified brokerages — an artifact of a temporary constraint?
Was it all built to compensate for the fact that humans couldn’t efficiently process market data in text form? That we needed visual interfaces because our cognitive architecture demanded them? And now that agents can mediate between raw data and human decision-making, does the interface layer retain any structural value?
We do not have a definitive answer. The transition is still early. Many traders still prefer visual workflows. Many strategies genuinely require visual pattern recognition. The charting platforms that survive will serve these users well.
But the direction is clear. The value is migrating from the tool layer to the expertise layer. From the platform to the process. From the interface to the intelligence.
You’re Not Reading This in June 2028.
You’re reading this in February 2026.
TradingView is still growing. Bloomberg is still charging $24,000 per terminal. Robinhood is still collecting PFOF on 100% of its order flow. The neobrokers are still raising venture capital. The social trading platforms are still signing up new users.
The canary is still alive.
But the MCP servers are proliferating. The agent coding tools are improving every quarter. The early adopters are building terminal-first workflows and discovering that their process quality has jumped measurably. The Skills repositories are growing. The first traders who encoded their expertise are reporting that their agents are generating better research than they could produce manually.
None of the individual signals look alarming. A few power users churning from a charting platform. A handful of institutional desks experimenting with agent-assisted research. Some retail orders routing differently than they used to.
Each signal, in isolation, is noise.
Together, they describe a trajectory.
The question is not whether AI agents will restructure the trading technology stack. The infrastructure assumptions that underpin today’s $47 billion industry — that humans need visual interfaces to process market data, that data access requires expensive terminals, that execution quality is invisible to retail traders, that behavioral consistency is a function of willpower rather than workflow design — every one of these assumptions is being tested by technology that improves monthly.
The question is how quickly. And whether you’re building your workflow for the world you’re in, or the world you’re heading toward.
The terminal is waiting.