Hey, it's Matt. Here’s what’s up in AI + Wall Street.
Was this email forwarded to you? Sign up below:

IN THIS ISSUE

• Price Drop: AI model costs decline and what it means for adoption

• Quantum Finance: BlackRock explores quantum AI for bond analysis

• Local AI Revolution: On-device AI solutions for record keeping

• Adoption Snapshot: JPM's 100,000 daily AI users and KPMG survey

• Market Pulse: Latest AI-finance deals, moves, and executive shifts

HEDGE FUNDS

Cheaper by the Token: The Declining Price of AI

There's so much happening in AI from the latest models to adoption trends that it’s easy to get lost in the daily news.

A trend that's been somewhat underreported is how much cheaper these models are getting — and not just DeepSeek.

You can see API pricing by looking up their respective pages on OpenAI, Anthropic and Meta. To go back in time, I used the Wayback Machine, which archives websites by date.

But I got busy and meant to pick up this research later.

Luckily, this week, OpenAI released its Deep Research for Plus users, $20/ month (I’m one). As the name implies, Deep Research does more than search for answers for you. It can do this sort of intern-level work when given directions.

It “thought” for about 25 minutes before producing a PDF and an Excel file with the underlying pricing. (I had already screenshotted a few periods so used those to ensure accuracy.)

Here are the very detailed (8,000+ words) results. Model comparisons are not always apples to apples since they can have different context windows, but the trend is the same throughout.

From the AI-Generated report

• OpenAI's GPT-4 saw the steepest price decline, with input token costs dropping from $0.03/1K tokens in Q1 2023 to $0.0025/1K in Q1 2025, a reduction of 91.7%.

• The decline in pricing is correlated with improvements in computational efficiency, economies of scale, and competitive market pressures.

This reminds me of what Tharsis Souza said when we had him on the podcast last month: “The price of intelligence is going to zero.”

Just to be clear: using AI is getting cheaper and cheaper and should only help with enterprise adoption. Producing AI at these prices is a very different question given the billions in capex invested.

Btw, if you haven’t tried OpenAI Plus ($20 a month), give it a try. I’d be curious to hear your experience using Deep Research.

RESEARCH

BlackRock Explores Quantum Science for Bond Analysis

Corporate bonds rarely trade. Some can go days or weeks without any buyers or sellers, making it hard to determine how much they're worth. It's like trying to value a house that hasn't been sold in 10 years – you'd look for recently sold similar houses based on size, number of bedrooms, school district etc. to estimate its value.

A team of researchers at BlackRock compared two different methods to compute similarities of corporate bonds using publicly available datasets: ‘traditional’ machine learning method called Random Forests, and a novel quantum-mechanics-inspired AI method called Quantum Cognitive Machine Learning (QCML).

An intuitive description of Random Forests is that it generates a group of decision trees that ask a series of yes/no questions—sort of like playing “20 Questions.”

• Is the bond rating below investment grade? (Yes/No)

• Is the company’s debt higher than a certain threshold? (Yes/No)

The algorithm repeats this process hundreds or thousands of times with slightly different questions, hoping that the aggregate decisions over all these trees would get closer to the “correct” value of the target variable (in this case, bond yield). Then, the similarity between pairs of bonds from this final group of trees is simply the number of times the pair ends up being in the same end node.

QCML — drawing on quantum theory’s ability to handle multiple possibilities — can outperform traditional bond-valuation methods at least for certain high-yield bonds by letting features interact more subtly, according to a new study by researchers at BlackRock, Qognitive, Wayne State University, and Cornell University.

Instead of forcing bond data into forming groups of decision trees as in Random Forest, in QCML, the researchers use quantum states, which are mathematical representations that measure how similar different bonds are to each other, such as risk, industry, country, yield, and liquidity—at once.

Think of quantum states like chords in music: the same notes can produce different sounds depending on their combination. Similarly, quantum states allow bond features to interact with each other in subtle and sophisticated ways, capturing nuanced relationships that traditional binary methods miss.

QCML creates a more balanced distribution of data points, which is especially helpful for evaluating bonds that rarely trade or have unusual characteristics. While random forests tend to push most bonds to maximum distance from each other (making similarity assessment difficult), QCML creates a more natural clustering where similar bonds remain visibly grouped together, even when handling outliers.

This study is only one of the first steps in applying the quantum mechanics-inspired AI methodology in finance. The preprint of the technical paper is here.

COMPLIANCE

Tracking Client Communications With AI — Locally

Regulatory rules require banks and broker-dealers keep all of their communications tied to business. This was problematic during the pandemic when everything shut down and folks relied on WhatsApp to communicate with their colleagues and clients.

These became known as the WhatsApp penalties, totaling more than $2 billion across Wall Street—a topic I covered a lot at Bloomberg News. The agency's position was that they need complete records to "pull the threads" on investigations, even ones that might initially seem unrelated to the bank.

Financial firms have to maintain these records, and they have to do so securely, i.e., mitigate the risk of having their data hacked. I say mitigate because there’s an old trope in cybersecurity that there are two kinds of companies: 1) they know they’ve been hacked and 2) the ones that don’t know they’ve been hacked.

Last week, I got a quick demo of Knapsack from Mark Heynen, a former Google and Meta exec who co-founded the company with Cooper Lindsay. This AI tool for financial advisors saves data locally rather than in the cloud and integrates with email providers to sync meetings and conference call summaries in one centralized place. They also have other automations like drafting follow-up emails and social media posts.

Saving data locally reminded me of the conversation I had last summer with Professor Alejandro Lopez-Lira on where he sees AI in five years:

I expect that local models are going to become easier to host and smarter. You’d take an open-source model, fine tune it without it ever touching the internet. A lot of tasks that are sensitive will be done with AI in five years for sure. Whereas now there may be some concern that “hey, I don't know if I can send my client’s information” [to an AI provider]. In five years, you will not need to send anything to anyone.

HEADLINES

OpenAI has 400 Million Weekly Active Users

The company has added 100 million users since December, when it had 300 million. It now has 2 million paying enterprise users. (CNBC)

• Also: BNY, America’s Oldest Bank, Signs Multiyear Deal With OpenAI (WSJ)

CredCore Secures $16M to Bring AI to Debt Markets

CredCore uses AI to streamline debt market transactions, accelerating deal processing, risk assessment, and document analysis for lenders and asset managers. (Press Release)

17-Month-Old AI Startup Sells for $220 Million

MongoDB acquired Voyage AI for $220 million after just 17 months in business. The startup’s anti-hallucination AI technology improves accuracy for enterprise applications. (Inc.)

BlackRock Aladdin Tech Head Joins Optiver as CTO

Electronic trading giant Optiver has hired a new CTO, one of the most influential technologists in financial services. (eFinancialCareers)

DBS to Cut 4,000 Jobs as AI Expands

Singapore’s DBS Group expects to cut 4,000 contract and temporary employees over the next three years as AI expands. (Finextra)

JPM Leads $20M Investment in AI Startup Albert Invent

The growth-equity group’s latest deal more than doubles the value of the AI-driven engine for research chemists. (WSJ)

ADOPTION

JPM: Half of Employees with AI Access Use It Daily

About 100,000 JPMorgan employees actively use the bank’s AI tools daily out of about 200,000 who have access, according to Teresa Heitsenrether, JPMorgan’s Chief Data and Analytics Officer, in a recent WSJ interview.

• Use Cases: “Think about anyplace in the bank where people are preparing to go and talk to their clients… pulling briefing memos together… lots of information to sift through… it’s a real time saver.”

• Frontier Models: “Building our own model isn’t how we differentiate ourselves. The secret sauce for us is how you use them with our information.”

• Proprietary Data: “What the models don’t have access to is the data that’s specific to the enterprise. That then becomes the differentiator in terms of how much value JPMorgan gets versus somebody else in financial services.”

Last month, I wrote about how megabanks have a competitive AI advantage thanks to their massive proprietary datasets. It’s always more interesting to see how traditional firms like JPMorgan implement AI rather than hearing AI companies talk about it.

AI Adoption Surges in Finance: KPMG Survey

AI at financial firms has surged, with 71% now using the technology and 41% doing so extensively, according to KPMG's global survey of 2,900 companies across 23 countries. Notably, 57% of "AI leaders" (the top 24% of adopters) report that their investments have exceeded return expectations.

• Generative AI is gaining traction quickly: 38% of AI leaders already use it for financial reporting, compared to just 3% of other organizations

• Geographic divide emerging: North American financial firms lead with 41% wide adoption, significantly ahead of Asia-Pacific (27%) and Europe (22%)

AM I MEMING RIGHT?

WHAT ELSE I’M READING

Four Ways AI and Tech Are Transforming Finance

Generative AI and financial data are opening up the digital economy to more consumers and small businesses, but crypto concerns remain. (MIT)

Microsoft Sees OpenAI as an Existential Threat

• OpenAI’s ChatGPT Enterprise (2m paying subs) is about to eclipse Copilot for Microsoft business.

• OpenAI can potentially rug pull on Microsoft by invoking its AGI clause

But obviously Satya won’t tolerate this insolence. (Enterprise AI Trends)

Rare Interview with Paul Singer of Elliott Management

At 34:40, Singer says, “AI is way over its skis in terms of practical value being brought to users.” Unfortunately, he doesn’t elaborate on why.

BNP Uses AI to Build Thematic Investment Funds

BNP Paribas’ Theam Quant funds use AI to identify stocks aligned with themes such as healthy living and energy transition, aiming for diversified and scalable thematic investing. (HedgeWeek)