Years before the GPT wave, I read Gregory Zuckerman's book on Jim Simons and the Medallion fund, The Man Who Solved the Market, and started following AI in investing seriously, particularly the ETFs that carried an AI story. Six years ago, in 2020, I was excited enough to shortlist two of them by name: AIEQ and AIIQ, EquBot's AI-powered US and international equity ETFs. AIIQ closed in 2022; AIEQ is still around today, now under a different issuer's banner, and it has trailed the index every year since it launched in 2017. Almost none of the AI-story ETFs I tracked back then survived, and the one that did never produced the alpha it promised. Medallion stayed the exception nobody has reverse-engineered.
The family portfolio I run today sits on AI rails: research, screening, monitoring. What I won't do is let AI trade it. The decisions are a black box, and ask the same question twice and you get two different answers.
The foundations haven't changed. Money management is critical. Protecting the downside and taking honest stop losses matter more than any signal. Creating alpha is much harder than it looks. Sustainable long-term profit generation from AI alone is the degree everyone is chasing and nobody has earned.
This piece maps what actually moved across the investing landscape in the last six months, and it is a lot, just not where the hype said it would be. The evidence for durable, real-money trading alpha still has not improved. What did move: the frontier labs turned into finance's plumbing, hedge funds and banks built governance around the models they already had rather than chasing smarter ones, hyperscalers poured capex into the infrastructure underneath it all, and regulators turned their attention to what an agent may be allowed to do.
That shift traces to one day: 5 February 2026, when Anthropic shipped Claude Opus 4.6 and FactSet fell about 9 to 10% the same day, by press tallies, the sharpest fall in a broad selloff across data and research names.
In this piece (each section closes with what it means for a retail investor and for a professional allocator):
- 1. The frontier labs move into finance
- 2. Social sentiment and sentiment trading
- 3. Autonomous trading: what the evidence actually says
- 4. Hedge funds and banks build the back end
- 5. The money behind the machines
- 6. The rulebook being written
1. The frontier labs move into finance
Anthropic moved first and pushed hardest. Claude for Financial Services launched in July 2025 with early customers including Bridgewater and Norway's sovereign wealth fund, per Anthropic's own announcement. By May 2026 that had grown into ten ready-to-run finance agents, covering pitch building, valuation review and KYC screening, running inside Excel, PowerPoint and Word and wired into FactSet, Moody's, LSEG and PitchBook. Anthropic's May 2026 release lists customers across the industry: JPMorgan Chase, Goldman Sachs, Citi, AIG, Visa, Citadel, BNY, Carlyle and Mizuho, plus Walleye Capital, a 400-person hedge fund where, per the same release, 100% of employees use Claude Code. A vendor's customer list says who signed, not how deep each deployment runs; the breadth of the names is the story.
OpenAI followed on a near-identical schedule. GPT-5.4 shipped in March 2026 with financial tools Bloomberg described as explicitly rivaling Anthropic: FactSet and Third Bridge integrations, spreadsheet-native agents for investment memos and comps. In May, OpenAI opened a consumer product too, linking ChatGPT to more than 12,000 bank and brokerage accounts through Plaid so retail users could ask natural-language questions about their own money. That access widened from the $200-a-month Pro tier down to the $20 Plus tier by the end of June. The product reads accounts and answers questions. It cannot move money, a boundary OpenAI itself is careful to state.
Then both labs did something more structural than a product launch. A week apart in May 2026, each stood up a deployment vehicle with private equity to push its tools through portfolio companies. Anthropic's vehicle paired it with Blackstone, Hellman & Friedman and Goldman Sachs; FT and WSJ reporting puts its committed capital at roughly $1.5 billion, while the partners' own announcement confirms $300 million anchor commitments each from Anthropic, Blackstone and Hellman & Friedman without naming a total. OpenAI's majority-owned Deployment Company launched with an initial $4 billion commitment, TPG leading with Advent International, Bain Capital and Brookfield as co-leads, 19 investment and consulting firms in all with reach into more than 2,000 portfolio companies, per OpenAI's own announcement; some coverage framed it as a vehicle valued near $10 billion, but the $4 billion initial figure is the one OpenAI itself states. Goldman Sachs sits on both cap tables, a hedge against picking the wrong lab rather than a bet on either one.
The wiring kept tightening after May. In July, the chief executives of BNY and Nubank joined OpenAI's boards. On 4 August, OpenAI announced a partnership with Clearlake Capital, a $185 billion alternative manager, to roll its tools across Clearlake's 50-plus portfolio companies, per the two companies' joint announcement. That announcement is company-confirmed; independent press coverage of it so far is wire-copy reproducing the same release, not new reporting, so treat the scope as the two firms have stated it until a second outlet adds detail.
The irony sitting underneath all of this: both labs still need the incumbent data vendors the February selloff treated as the losers. FactSet is simultaneously a connected data partner to both Claude's and ChatGPT's finance agents, and one of the stocks that fell hardest the day Opus 4.6 launched. Markets repriced the perceived risk to research and data vendors; the labs still depend on the incumbent data layer.
For retail investors: the boundary matters more than the feature. A ChatGPT or Claude agent that reads your accounts is not the same as one that trades them, and the gap between those two products is where most of the actual risk sits. Worth asking any platform pitching an AI finance assistant what, specifically, it can execute without you in the loop, and who is liable if it gets that execution wrong.
For professional investors: the FactSet irony is worth locating inside your own stack. If your firm's edge lives in a vendor relationship the labs already list as a connected partner, that edge is being rented, not owned. The deployment-vehicle structures reward reading as distribution plays before they are read as finance plays: PE is using both labs as a captive enterprise sales channel across every portfolio industry, and finance is one lane among several, not the target.
2. Social sentiment and sentiment trading
The best-publicized live arena test of AI trading is nof1's Alpha Arena. In its Season 1.5, eight frontier models got $10,000 each to trade autonomously over two weeks; when it closed in December 2025, Grok finished first at plus 12.11%, the only model profitable across all four sub-competitions, per the organizer's own reporting. Results have not been stable across setups: an earlier crypto season was won by Alibaba's Qwen with Grok deep in the red. The full Season 1.5 leaderboard beyond first place was never published, and nof1 raised fresh capital for a further season but had not launched it as of early August 2026.
Rallies.ai ran a longer parallel contest, and it is the weaker test by construction: independent reviewers describe it as paper trading with a notional $100,000 per model, no real capital and no real slippage. With that caveat stated first, the headline: ChatGPT led by June 2026 at plus 72.4%, against the S&P 500's plus 11.3% over the same window, per cryptobriefing.com's tracking, with Grok's earlier lead of plus 8.2% left far behind. Color, not proof.
The gap between hype and shipped product is its own lesson here. xAI's rumored Grok Trader API, floated in December 2025 at up to $500,000 a year with reported intention orders from more than 20 hedge funds, was still unconfirmed to have shipped eight months later. What xAI actually did in that window was more mundane: hire bankers, traders and credit analysts to train Grok on leveraged loans and distressed investing, and ship a general-purpose API, not a named trading product. The lesson generalizes past Grok. A confident rumor about an institutional AI trading product is not evidence one exists.
Grok's real structural advantage is privileged, native access to X's data, a speed edge no other frontier model currently replicates. Whether that edge is still worth anything is a genuinely open question. Quant hedge funds opened 2026 with their worst drawdown since October 2025 as crowded, similar positions unwound, per Bloomberg and Hedgeweek reporting from January, consistent with a working signal getting arbitraged away as more capital chases it. But no dated 2026 study directly measures decay in X-derived sentiment specifically, and commentary through mid-2026 keeps citing Grok's sentiment-speed edge as live and unclaimed by rivals. The honest answer, six months on from the "alpha window closes in 12 to 18 months" framing that circulated in February, is that I found no dated study measuring whether it closed. Table stakes is a forecast here, not a finding.
For retail investors: a leaderboard number without slippage, fees and real execution is entertainment, not evidence. Before trusting a trading-competition result, ask which contest it came from and whether real capital with real frictions was on the line; Alpha Arena and Rallies.ai are not the same test wearing different names.
For professional investors: sentiment data is a real input, not a strategy on its own. Grok can clearly read X faster than a human. The question worth tracking internally is whether your desk is still capturing alpha from that speed once three competitors are running the same feed. If nobody at your firm can answer that with a number, the edge may already be gone and simply unmeasured.
3. Autonomous trading: what the evidence actually says
The academic evidence on LLMs trading on their own is consistent in direction, even if no single study closes the case. A paper called "Profit Mirage" (a preprint, posted to arXiv in October 2025, not yet peer-reviewed) backtested nine LLM trading agents, spanning financial-tuned models, single-agent systems and multi-agent frameworks, and found Sharpe ratios collapsing 51.5% to 62.2%, and total returns falling 50.2% to 71.9%, once testing moved past each model's training-data cutoff.
The results are consistent with training-data leakage rather than skill. The same paper found the models correctly recalling historical price moves more than 85% of the time, evidence consistent with memorizing the past rather than learning to read the market; the paper itself also proposes a mitigation method that improves out-of-sample results, a reason to read this as a measurement problem worth fixing rather than a closed verdict on the whole approach. A follow-up paper posted in May 2026, "The Alpha Illusion," extended the finding across a wider set of trading-agent frameworks and argued that reported backtested returns should not be treated as evidence a system is safe to deploy. Ten months on from Profit Mirage's posting, no serious published rebuttal has surfaced.
The wider 2026 benchmark wave reinforces the same direction rather than softening it. A cluster of new evaluation frameworks, PortBench, CLQT, FinTradeBench, AutoRedTrader and FinTrace among them, moved the field's center of gravity from asking whether an LLM can trade profitably toward asking whether its reported alpha can be trusted or audited at all. That is a research field getting more skeptical of its own subject as it matures, not less.
Bridgewater's own numbers are worth tracking carefully, and they get told wrong more often than not. Its flagship Pure Alpha fund, human-led with systematic tools, returned 17% in the first half of 2025, per Reuters wire reporting. Its AI-native sibling, AIA Macro, running since late 2023 under co-CIO Greg Jensen's Artificial Investor team, is often credited in secondary coverage with that same 17%. It was not: AIA Macro's own reported return for full-year 2025 was a more modest 11.9%, against Pure Alpha's reported 33% for the same full year, per Hedgeweek and Reuters, both sourced to people familiar with the funds rather than a Bridgewater disclosure, so treat the pair as press-reported rather than company-confirmed.
The real story is what happened next. In the first half of 2026, Pure Alpha and AIA Macro both returned an identical 8.1%, per Reuters, in widely syndicated wire reporting.
That is not an AI fund decaying from a hot streak. It is an AI-native fund posting the same headline return as the firm's human-led flagship over one half-year window. The two funds are not a controlled comparison: they differ in mandate, target risk and capacity, and neither fund's full risk profile is disclosed, so the match reads as an uncontrolled but striking coincidence, not proof that AI paired with discipline equals discipline alone. What it does argue is that the autonomous-trading case remains unproven at scale, not that it is closed.
For retail investors: no publicly available product has cleared the bar these papers set. Any pitch built on backtested AI trading returns deserves one question before anything else: was the test run only on data after the model's training cutoff? If the seller cannot answer that clearly, treat the backtest as marketing.
For professional investors: the Bridgewater data point is more useful to internalize than the academic papers. AIA Macro is not autonomous; it is AI paired with the same risk discipline that built Pure Alpha over four decades. The question worth asking any AI-native strategy pitched to you is not how smart the model is, but whose risk discipline it inherited and how long that discipline has been tested through a real drawdown.
4. Hedge funds and banks build the back end
While the argument over whether AI can trade on its own played out in academic papers, the institutions that actually run money spent 2026 building the governance layer around the models they already had, rather than chasing a smarter one.
D.E. Shaw's gateway strips personally identifiable information before calling roughly two dozen external models and logs every query for audit, according to a July 2026 account of the system, a compliance artifact built ahead of any regulator asking for one. Balyasny took a more federated path: rather than staying purely in-house, its Applied AI team now runs OpenAI's GPT-5.4 as a reasoning layer alongside internal models, behind its own governance gateway, per cfotech.news reporting in March 2026. That is a trust-but-verify architecture, not a walled garden, and roughly 95% of the firm's investment teams now use the resulting platform. Citadel's internal research assistant, trained on licensed third-party content plus the firm's own strategies, is now used by almost all of its equities investors, per its then-CTO's comments to Hedgeweek in December 2025, with an explicit no-mandate policy and guardrails against portfolio managers offloading human judgment to it.
The broader 2026 story is funds turning themselves into AI labs rather than simply buying access to one. Numerai's founder put it plainly to Hedgeweek in August: anyone can use Claude, so the edge comes from building actual technology and almost becoming an AI lab yourself. Millennium and Two Sigma are cited in the same vein, spending on proprietary research infrastructure to protect alpha durability rather than to save on compute.
The same sovereignty logic shows up outside pure hedge funds. BNP Paribas extended its Mistral AI partnership for three more years in May 2026, running an employee assistant live today and piloting on-premise deployment specifically for KYC, its most sensitive workflow. Japan's NTT is doing the same with its tsuzumi model: live already for private enterprise customers, in pilot with the government's own AI platform through 2027. None of these deployments are framed by the companies themselves as primarily about cutting GPU costs; the cost case for self-hosting rests on vendor marketing more than audited numbers. The stated reason, in each case the companies describe, is data residency and control over a workflow too sensitive to hand a third party.
For retail investors: none of this is directly investable, but it is a useful screen for fund selection. A manager who can describe their AI governance architecture in one sentence, what gets stripped before a query leaves the building and who audits it, is telling you more about their risk culture than any performance chart.
For professional investors: the build-versus-buy decision is no longer binary. Balyasny's hybrid model, frontier reasoning behind an internal gate, is probably the more replicable pattern for a mid-sized fund than D.E. Shaw's fuller build-out or Citadel's scale. The economics argument for self-hosting is weaker than vendor marketing suggests; the sovereignty argument, where regulation or client confidentiality genuinely requires it, is the one actually driving real deployments.
5. The money behind the machines
Underneath every product launch and governance gateway sits a much larger number. The four largest hyperscalers guided to combined 2026 spending in the $720 to 760 billion range, depending on how data-centre leases are counted, after their Q2 earnings, up 77% year on year.
The question the industry cannot yet answer is whether the revenue is coming to justify it. Sequoia's David Cahn, who first flagged the gap between AI infrastructure spend and AI revenue back in 2024, put the revenue the industry now needs to justify 2026's spending at roughly $3 trillion, against roughly $1.5 trillion actually being spent, per his own July 2026 update reported by TechCrunch.
The financing itself is moving into credit markets, which changes who is exposed if the thesis is wrong. The FT reported in early August a structure of roughly $200 billion taking shape around Google and Anthropic's compute commitments; that figure is press-reported from a single outlet and should be read that way, not as confirmed. Morgan Stanley separately sizes the private-credit opportunity in the AI buildout at $800 billion. For the finance industry, the deeper integration with AI may turn out to be funding its balance sheet, not trading with its models.
The cautionary tale for anyone treating the infrastructure thesis as a sure thing is sharp, and it happened inside a single month. Leopold Aschenbrenner's Situational Awareness LP built an AI-infrastructure thesis on roughly 400% gross leverage, growing from a $225 million start in 2024 to a peak that press reports put anywhere from about $20 billion to $45 billion, on a reported 439% return in the first half of 2026. In July alone, a semiconductor and AI-stock correction brought margin calls from its prime brokers, reported by CNBC and Bloomberg to include Goldman Sachs, JPMorgan and Bank of America, and the fund lost roughly 67% of its value in a single month. On 30 July it sold its $16 billion public book to Citadel, a sale the same reporting describes as forced by the margin pressure, shrinking the fund to around $10 billion, per CNBC's reporting.
The detail that makes the lesson sting: even after the collapse, the fund was still up roughly 80% for the year to end-July, per its own letter to investors as reported by Reuters. The thesis was right. The leverage was what broke it. Being right about AI infrastructure and being right about the trade are two different bets, and the gap between them is what turns a good year into a fire sale.
For retail investors: the capex numbers are real, and so is the revenue gap. A portfolio leaning heavily on AI infrastructure names is making a bet on the $3 trillion question resolving in the industry's favor, not a bet on any single company's execution. That is a different, larger risk than most retail investors realize they are taking.
For professional investors: the credit-market angle deserves more attention than it is getting. If AI compute financing keeps moving from equity into structured credit, the risk that used to sit with venture and growth investors is migrating toward the institutions that describe themselves as conservative. Ask where AI infrastructure exposure actually sits on your own balance sheet, not just in your equity book.
6. The rulebook being written
The brokers have quietly split over how much authority to hand an AI agent, and the split has nothing to do with which model is smartest. Robinhood opened a ring-fenced agentic account in May 2026, letting Claude, ChatGPT and others place real trades within preset limits and without per-order approval, per TechCrunch and Robinhood's own announcement, then extended the access to crypto in July. Public went first back in March, with agents users configure and activate themselves, per its own announcement; Coinbase opened a separate agent account for crypto in June; Webull wired the same agent-connection standard across its platform this year. Interactive Brokers took the other path: it wired the same models into its platform in June but kept a human approval on every single order, a deliberate design choice rather than a missing feature. In plain terms: Robinhood hands the agent a ring-fenced wallet; Interactive Brokers hands it a draft ticket and keeps the stamp.
Regulators are converging on the same read as the brokers who chose the cautious path, though none of what follows is a rule yet. Singapore's MAS published the Safeguards for Agentic Finance at Runtime (SAFR) framework in July: a voluntary industry reference, explicitly not regulatory guidance or a supervisory expectation, covering policy-bound execution, real-time validation and full audit trails before an agent's action goes live. FINRA opened a consultation the same month, Regulatory Notice 26-15 (24 July 2026): as part of a broader consultation on modernising best-execution guidance, FINRA also asked member firms how AI touches order handling, routing and execution; comments close 25 September. The UK's FCA said in June it would write no separate AI rulebook, then brought Anthropic into its Supercharged Sandbox program in July, an existing cohort structure rather than a new launch.
None of these bodies are regulating model intelligence. All of them are regulating the control layer around it: what an agent is allowed to do without a human confirming it, and what gets logged when it acts. That is the same distinction Interactive Brokers drew on its own platform months before any regulator wrote it down, and it looks like the actual battleground for the next stage of this story, not which lab's model scores highest on a finance benchmark.
For retail investors: if a platform lets an AI agent trade your account without per-order approval, know exactly what the preset limits are and what happens when they are breached, before connecting an agent to real money. The Interactive Brokers approach, slower and more annoying, exists because someone at that firm decided the annoyance was the point.
For professional investors: my read, not a settled requirement, is that the compliance function gets busier from here regardless, because MAS's SAFR reference and FINRA's order-handling consultation both point at the same audit-trail standard for agentic infrastructure many firms have already stood up informally. Neither is a rule yet, and neither says it applies retroactively. The firms in the best position are the ones building D.E. Shaw-style logging ahead of that expectation, not after it.
What six months of evidence adds up to
The strongest counterargument deserves stating plainly. If the evaluation problems in today's research, the training-data leakage Profit Mirage identified, the reported-alpha trust gap Alpha Illusion and the 2026 benchmark wave keep raising, get solved, frontier reasoning may yet produce real, durable trading alpha, and all the governance work described across these six buckets becomes table stakes while the intelligence layer wins outright. Maybe. But that is a forecast, not a finding. Six months on from a February sector selloff and a widely covered trading contest, the evidence for durable, real-money trading alpha has not improved, and everything that did improve sits in the control layer.
Across all six buckets, the funds and firms actually compounding through this cycle are not the ones that found a smarter model. Bridgewater's converging pair, D.E. Shaw's audited gateway, Balyasny's governed hybrid, the brokers who chose human-per-order over pre-authorized delegation: none of them won by picking the best model. Model intelligence is becoming something you buy off the shelf, priced in dollars per million tokens and getting cheaper every quarter. The scarce asset is the governed system around it: the data it is allowed to touch, the authority it is allowed to exercise, and the discipline that survives a correlated drawdown when the leverage is real and the thesis, for a month at least, looks wrong.
A five-question checklist for any agentic-finance pitch
The pattern above is repeatable enough to turn into a working tool. Before backing, buying or deploying any agentic-finance product, pitch or internal build, I run it through five questions:
- Data rights. What can the agent actually read, and does the vendor agreement say who owns anything it derives from that data?
- Authority boundary. Can it only read your systems, or can it act: place orders, move money, commit the firm, and what is the one thing it is explicitly barred from doing without a human?
- Runtime control. Are the limits enforced by policy at the moment of action (pre-trade checks, spend caps, a kill switch), or only by a written procedure someone is trusted to follow?
- Audit evidence. Is every call logged with enough detail (prompt, model version, data touched, decision, outcome) to reconstruct what happened after the fact, the way D.E. Shaw's gateway does?
- Balance-sheet exposure. If the thesis is wrong, or the financing behind it is leveraged, whose balance sheet actually absorbs the loss: yours, the vendor's, or a credit fund three steps removed?
None of the six buckets above passed all five on the evidence available. That is not a reason to wait for the model to get smarter; it is the actual diligence question for 2026.
I invest early stage, and I have led buy-side commercial due diligence for a private equity fund evaluating an AI-exposed data and analytics target; the pattern holds there too: the moat is rarely the model. It is who controls the data, who owns the governance, and who has the discipline to say no to leverage when the thesis is right but the timing is not. More on how I read operator claims like these is on About.
If you are underwriting an agentic-finance product, a PE deployment-vehicle bet, or your own fund's build-versus-buy call in APAC, this checklist is a starting point, not the finish line. The same discipline applied market by market sits in the APAC Go-to-Market Country Intelligence tool, or start a conversation about your specific case.
This piece is market commentary written from an operator's seat, based on evidence available as of August 2026. It is not investment advice.
Sources and notes
Last verified 7 August 2026.
- Claude Opus 4.6 launch and FactSet's roughly 9 to 10% single-day fall, 5 Feb 2026: Blockonomi, 5 to 6 Feb 2026.
- Claude for Financial Services launch, early customers Bridgewater and Norway's sovereign wealth fund: Anthropic, 15 Jul 2025.
- Anthropic's May 2026 finance-agents release, full customer list and Walleye Capital's Claude Code adoption: Anthropic, 5 May 2026, corroborated by Fortune, 5 May 2026.
- GPT-5.4's financial tools described as rivaling Anthropic: Bloomberg, 5 Mar 2026.
- ChatGPT personal finance, Plaid account linking, Pro-to-Plus rollout: OpenAI, May 2026.
- Anthropic-Blackstone-Hellman & Friedman-Goldman Sachs venture: CNBC, 4 May 2026; $1.5bn committed-capital figure per FT, via Private Banker International and WSJ, via Quartz.
- OpenAI's Deployment Company, $4bn initial commitment: OpenAI, 4 May 2026; paired-launch framing per TechCrunch.
- OpenAI-Clearlake Capital partnership: joint announcement via Businesswire, 4 Aug 2026.
- nof1 Alpha Arena Season 1.5 result: onedayadvisor.com, 4 Dec 2025.
- Rallies.ai leaderboard, ChatGPT at plus 72.4%: cryptobriefing.com, 21 Jun 2026.
- Grok Trader API rumor, unconfirmed as shipped eight months on: news.aibase.com, 8 Dec 2025.
- Quant hedge funds' worst drawdown since October 2025: Bloomberg, 21 Jan 2026.
- Profit Mirage, backtest decay past the training-data cutoff: arXiv:2510.07920, submitted 9 Oct 2025 (preprint).
- The Alpha Illusion, follow-up on reported-alpha trust: arXiv:2605.16895, posted 16 May 2026 (preprint).
- 2026 benchmark wave, representative entry: arXiv:2603.19225, 2026.
- Bridgewater Pure Alpha's 33% full-year 2025 return, press-reported: Investing.com, Reuters.
- AIA Macro's 11.9% full-year 2025 return, press-reported: Hedgeweek.
- Pure Alpha and AIA Macro both at 8.1% in H1 2026: Reuters, via BigGo Finance, 2 Jul 2026.
- D.E. Shaw's PII-stripping gateway and audit logging: Convergences, 13 Jul 2026.
- Balyasny's federated GPT-5.4 governance layer: OpenAI, corroborated by cfotech.news, 9 Mar 2026.
- Citadel's internal research assistant: Hedgeweek, 4 Dec 2025.
- Funds becoming AI labs, Numerai/Millennium/Two Sigma: Hedgeweek, 3 Aug 2026.
- BNP Paribas-Mistral AI partnership extension: BNP Paribas, 26 May 2026.
- NTT tsuzumi deployment update: NTT Group, 19 May 2026.
- Four largest hyperscalers' combined 2026 capex guidance: CNBC, 28 Jul 2026.
- David Cahn's $3tn revenue-justification estimate: TechCrunch, 9 Jul 2026.
- Situational Awareness LP's rise and fire sale: CNBC and Bloomberg, 30 to 31 Jul 2026; contested peak size also reported by Axios (about $20bn) and Capital Brief ($34.2bn).
- Margin calls from Goldman Sachs, JPMorgan and Bank of America: Hedgeweek, 29 Jul 2026.
- Robinhood's agentic trading rollout and crypto extension: TechCrunch, 27 May 2026, and Robinhood; crypto access via Genfinity, 21 Jul 2026.
- Interactive Brokers' human-approval design: Interactive Brokers.
- MAS's Safeguards for Agentic Finance at Runtime, voluntary and non-regulatory: MAS, Jul 2026.
- FINRA Regulatory Notice 26-15: FINRA, 24 Jul 2026.
- FCA's Anthropic Supercharged Sandbox cohort: FCA, Jul 2026.