The Fintech Blueprint

In this episode, Lex chats with Srini Krish — Co-Head of Financial Solutions at Fiserv, one of the original fintechs, in business for nearly five decades and sitting at the intersection of commerce and banking.

Lex and Srini discuss how Fiserv acts as the technology backbone for 5,000+ US banks and credit unions that lack the wherewithal to match JPMorgan or Wells Fargo on their own, and how the firm is packaging AI into that distribution layer through Agent OS and partnerships with OpenAI and Anthropic. Srini lays out his four-bucket framework for enterprise AI - better client service, internal productivity, AI embedded in products, and a platform banks can use to build their own agents - and explains why money demands deterministic outcomes rather than probabilistic guesses, keeping a human in the middle as commercial loan underwriting compresses from weeks to hours.

They explore the competitive race against challengers like Mercury and Ramp, the mainframe that has outlived thirty years of obituaries, and where power sits between the AI labs and their distribution channels once inference commoditizes.

NOTABLE DISCUSSION POINTS:

MIPS became tokens. Srini frames the whole AI shift through continuity: engineers once measured effectiveness by MIPS consumed and how often they compiled code; today the metric is token consumption. Same discipline of doing more with minimal resource, thirty years apart.

Money forces determinism. Probabilistic outputs are fine for many tasks but unacceptable for balances - a figure 1% or 5% off is a failure, it has to be right every time. So Fiserv’s Agent OS rollout starts with non-real-time, human-in-the-middle use cases and only graduates toward autonomy and eventually customer-built agents. It’s a crawl-walk-run path, and Fiserv says it’s clearly still crawling.

The moat is distribution, not model access. Fiserv’s 5,000+ banks and credit unions can’t engage OpenAI or Anthropic directly at scale, so Fiserv becomes the platform that packages agentic workflows - turning commercial loan decisions from a multi-week process into hours, with the auditability and observability those institutions could never build alone.

TOPICS

Fintech, Fiserv, EmbeddedFinance, AgenticAI, EnterpriseAI, Banking, Payments, DigitalBanking, CommunityBanks, FinancialInfrastructure, AIAgents, OpenAI, Anthropic, ClaudeCode, JPMorganChase, FirstData, Mercury, Ramp, Plaid

 

ABOUT THE FINTECH BLUEPRINT

🔥Subscribe to the Fintech Blueprint newsletter to stay at the forefront of Fintech and DeFi: https://bit.ly/3hyhlC2

🤝 Partner with Fintech Blueprint through sponsorships: https://bit.ly/3UZllsV

👉 Twitter: https://twitter.com/LexSokolin

 

TIMESTAMPS

1’12: Fintech Before It Was Fashionable: Five Decades at the Intersection of Commerce and Banking

6’13: Access, Move, Trust: What Actually Defines a Fintech Across Three Decades

10’28: A Loan at the Mechanic's Shop: How Embedded Finance Widened the Market and the Money Behind It

13’22: Four Buckets for Enterprise AI: Where Agent OS and the OpenAI Partnership Actually Fit

20’47: Not Savviness but Wherewithal: Why 5,000 Institutions Can't Build JPMorgan's Stack Alone

25’39: Mercury, Ramp, and the Mainframe That Never Died: Why the Incumbents Aren't Going Anywhere

29’51: Both Labs, All Three Clouds: Why the Distribution Channel Sits in the Middle

33’16: The Engineer Who Stops Writing Code: Why Replacement and Expansion Can Both Be True

36’50: It Has to Be 100% Correct Every Time: Why Money Demands Deterministic AI

Disclaimer here — this newsletter does not provide investment advice and represents solely the views and opinions of FINTECH BLUEPRINT LTD.
Contributors: Lex, Laurence, Matt, Farhad, Mike, Daniella
Want to discuss? Stop by our Discord and reach out here with questions.

In this episode, Lex chats with Jeff Grimes — who is Head of Live Events Products at Perplexity, the AI company that has evolved from an "answer engine" into an "agent platform" built around Perplexity Computer, its multi-agent digital worker. They discuss how Perplexity has shifted financial research from the how to the what, letting a user describe an outcome in a single sentence while Computer orchestrates 20+ frontier models, direct tool calls to licensed live data, and finance-specific skills to produce the artifact.

Jeff explains the enterprise strategy behind traceability - the north star that 100% of every quantitative figure traces back to its source filing - alongside bring-your-own-license connections via MCP and the consumer "personal CFO" vision powered by Plaid. They explore what 5x revenue growth on a 34% headcount increase signals for finance jobs, and why the future looks like a 24/7 family office that proactively surfaces and, with permission, executes financial actions for everyone.

NOTABLE DISCUSSION POINTS:

The “how to what” collapse is the real product thesis, not just better models. The shift to zero-shot rests on three stacked unlocks: direct tool calls to licensed live data (Quartr for earnings transcripts, unusual whales for insider and political holdings, SEC filings for historicals) instead of relying on web freshness; a thinking-model router that orchestrates 20+ frontier models in parallel, matching the model to the job (a heavy thinking model for macro analysis, a lighter one for ticker-matching 550 names); and ~20 opinionated finance skills (DCF, three-statement, LBO, comps) tuned through expert-led evals. Together they turn one sentence into a polished equity-research artifact.

Traceability is the enterprise wedge, framed as “don’t trust and verify.” The stated north star is that 100% of every number in any output is hover-traceable back to the source filing - pre-scrolled to the page, highlighted, with the full chain of calculations exposed. The framing inverts the usual “trust but verify”: assume the user won’t trust the model, so trust must be earned per number. Paired with bring-your-own-license via MCP (FactSet, LSEG, Morningstar, CarbonArc, PitchBook), this is the concrete answer to why regulated institutions get comfortable adopting.

The productivity and jobs signal is quantified and lived internally. Perplexity grew annual run rate 5x while increasing headcount only ~34%. Computer began as a company-wide Slack bot where every request was visible to all employees; Jeff now runs 9–10 scheduled cron jobs each morning and says essentially all code is written first by his agents. On the consumer side, the emergent pattern is build-your-own long-tail apps that no roadmap-bound product could serve - a DraftKings-addiction accountability system that emails a user’s spouse on any bet, or a GitHub-style heatmap of daily spending - which is the real substance of the personal-CFO bet.

TOPICS

Perplexity, Perplexity Computer, Perplexity AI, Google, Shadebot, Plaid, Yodlee, Claude, ChatGPT, AI, Artificial Intelligence, LLM, CFO, financial services, AI commerce

 

ABOUT THE FINTECH BLUEPRINT

🔥Subscribe to the Fintech Blueprint newsletter to stay at the forefront of Fintech and DeFi: https://bit.ly/3hyhlC2

🤝 Partner with Fintech Blueprint through sponsorships: https://bit.ly/3UZllsV

👉 Twitter: https://twitter.com/LexSokolin

 

TIMESTAMPS

1’05: A company built on failed founders: Why the startup that didn't work out led here

5’10: A playing card company that became Nintendo: How curiosity carried Perplexity from answers to actions

9’52: We're basically at zero shot now: Why prompt engineering is disappearing from finance work

15’21: A thinking model that routes the job: How Perplexity picks Claude Opus for macro and Grok for tickers

20’12: Two tracks, one engine: Building for enterprise workflows and a personal CFO at once

25’25: Bring your own license, or use ours: How Perplexity gets institutions comfortable enough to adopt

32’54: 5x revenue on 34% more headcount: The productivity gain Perplexity lived firsthand

37’28: Connect everything from your mortgage to the painting on your wall: Building the true personal CFO

45’15: A family office that works 24/7 for everyone: The proactive, automated future of the personal CFO

48’01: The channels used to connect with Jeff & learn more about Perplexity Computer

Disclaimer here — this newsletter does not provide investment advice and represents solely the views and opinions of FINTECH BLUEPRINT LTD.
Contributors: Lex, Laurence, Matt, Farhad, Mike, Daniella
Want to discuss? Stop by our Discord and reach out here with questions.

In this episode, Lex chats with Cactus Raazi — CEO Americas at B2C2, one of the original and largest institutional market makers in digital assets, serving roughly 1,500 institutions and pricing across more than 40 exchanges globally.

They discuss what a market maker actually does, how balance sheet and signal generation underpin roughly $1 billion a day of stablecoin flow at B2C2, and why the two extremes of crypto market making - riskless principal aggregation versus proprietary alpha - produce very different client outcomes that buyers rarely understand.

Cactus explains B2C2's 18-month bet that the Circle-versus-Tether debate would give way to a multi-issuer world, the launch of its PENNY product for instant zero-cost cross-stablecoin swaps, and they explore why programmability is the next frontier for digital dollars, why US capital markets have almost no structure for funding genuine risk-taking businesses, and whether the current combination of scale, speed, and complexity makes this the hardest investing environment Wall Street has ever faced.

NOTABLE DISCUSSION POINTS:

Market makers aren’t a homogeneous category, and clients pay for the difference. At one extreme, a market maker is essentially a riskless agent - aggregating prices across 40+ exchanges and quoting on top with no real view. At the other extreme, a market maker is a proprietary quant shop running alpha signals on horizons from seconds to days, and the price you get is heavily conditioned by where the signal says the asset is going. B2C2 sits in the middle, partly because its public-company parent (SBI) constrains risk appetite. The implication for institutional buyers: who you trade with structurally determines the quality of execution, not just the spread.

Algorithmic fixed income market making didn’t fail on technology, it failed on capital structure. US capital markets are excellent at funding venture, growth equity, private equity, and buyouts, but there is almost no domestic pool of “risk equity” - capital comfortable with the possibility that the machines (or the humans) lose money on a given day. Market makers need exactly that kind of balance sheet, and the mismatch between what the business requires and what the US capital base offers is a structural reason firms like Elefant struggled, regardless of execution quality.

The Circle-vs-Tether framing is already obsolete; the next product wedge is interoperability. B2C2 made an 18-month-old contrarian bet that the duopoly narrative was wrong and that Stripe (via Bridge), Western Union, Revolut, and many other consumer and platform companies would issue their own stablecoins. PENNY - instant, zero-cost, zero-counterparty-risk stablecoin-to-stablecoin swaps - is the product expression of that view. The deeper claim is that stablecoins are software, and the SaaS analogy (a base layer plus an app store of programmable financial logic) is the real reason institutional adoption accelerates from here, not the transfer-of-value benefit on its own.

TOPICS

B2C2, Goldman Sachs, SBI Group, Binance, Coinbase, Circle, Tether, Stripe, Kraken, Credit Suisse, Market making, institutional liquidity, stablecoins, fixed income, risk management, algorithmic trading, crypto exchange infrastructure

 

ABOUT THE FINTECH BLUEPRINT

🔥Subscribe to the Fintech Blueprint newsletter to stay at the forefront of Fintech and DeFi: https://bit.ly/3hyhlC2

🤝 Partner with Fintech Blueprint through sponsorships: https://bit.ly/3UZllsV

👉 Twitter: https://twitter.com/LexSokolin

 

TIMESTAMPS

1’17: Rejected by 30 firms: A cold-call advertising inquiry that became a Goldman career

6’43: "I'll go sell jet engines": Complexity as the through line from credit derivatives to crypto

8’53: Scale, speed, and dimensionality: The hardest investing environment in 28 years

12’33: A terrific idea, a brutal execution: Building an automated market maker in 2015

16’22: The used car dealership of bonds: How over-the-counter fixed income actually works

19’51: Price, time frame, and the art of liquidity: What a market maker actually does

24’52: Riskless principal or proprietary alpha: The two extremes of crypto market making

29’54: Priming the liquidity pump: Why new tokens hire market makers and large ones don't

35’40: $1 billion a day in stablecoins: A contrarian bet against the Circle-versus-Tether frame

39’43: 24/7 money movement: The treasurer wish list stablecoins actually deliver

41’04: The channels used to connect with Cactus & learn more about B2C2

Disclaimer here — this newsletter does not provide investment advice and represents solely the views and opinions of FINTECH BLUEPRINT LTD.
Contributors: Lex, Laurence, Matt, Farhad, Mike, Daniella
Want to discuss? Stop by our Discord and reach out here with questions.

In this episode, Lex chats with Mike Milotich — Chief Executive Officer of Marqeta, the modern card issuing platform that processed nearly $400 billion in payments volume in 2025, and is certified to operate in 40+ countries, growing over 30% for the third straight year. They discuss how Marqeta's separation of bank, processor, and brand armed fintech's largest winners across buy now pay later, on-demand delivery, neo-banking, and expense management with the Lego blocks to build their own card programs. 

Mike explains how the company's growth is shifting from enabling new use cases to displacing volume on legacy bank platforms, and they explore why card issuing is going multinational, what the agentic commerce wave actually requires to clear security and behavioural hurdles, and how Marqeta's continued growth runs through embedded finance, real-time personalisation, and the forced modernisation of the banks themselves.

NOTABLE DISCUSSION POINTS:

The BNPL business model is flipping from merchant rails to consumer cards. Marqeta originally solved the merchant scale problem for buy now pay later via virtual cards, removing the need for tens of millions of merchants to integrate a new button at checkout. The current shift is more important: BNPL players are now issuing consumers their own physical and virtual cards usable anywhere cards are accepted, turning BNPL from a merchant-acceptance game into a direct consumer value proposition. BNPL volume has grown over 50% year-on-year for Marqeta in recent quarters.

Card issuing is going multinational, and that breaks the legacy bank model. Banks have always been local on the consumer side, with only a handful multinational on the commercial treasury side. The next generation of card issuers, neo-banks like Revolut and Nubank, plus large global platforms embedding financial products into existing user bases, are global by default. A single platform that issues cards, and is certified to operate across 40+ countries, becomes the strategic moat, and legacy processors built to serve domestic bank programs aren’t structured to compete.

The growth story is moving from expanding the pie to displacing the incumbents. To date, Marqeta has mostly powered new card use cases that didn’t exist before — on-demand delivery, BNPL, neo-banking, expense management. Mike’s forward thesis is a phase change: pressure from fintech winners is forcing banks to modernise, and the next leg of growth comes from displacing volume sitting on legacy bank-controlled platforms. Real-time personalised rewards, where the same card delivers different offers to different cardholders based on live data, is the wedge that legacy infrastructure can’t deliver.

TOPICS

Marqeta, Visa, Mastercard, American Express, PayPal, Payments, card issuing, embedded finance, fintech, BNPL, neobank, agentic commerce, e-commerce, crypto, stablecoins, programmable money, machine economy, agentic AI

 

ABOUT THE FINTECH BLUEPRINT

🔥Subscribe to the Fintech Blueprint newsletter to stay at the forefront of Fintech and DeFi: https://bit.ly/3hyhlC2

🤝 Partner with Fintech Blueprint through sponsorships: https://bit.ly/3UZllsV

👉 Twitter: https://twitter.com/LexSokolin

 

TIMESTAMPS

1’04: From Math Brain to Payments Career : Finding the Nuance in How Money Actually Moves

7’05: The Narrative Gets Ahead of Reality : Why Agentic Commerce Will Move Slower Than the Technologists Think

10’08: Global But Local : The Balancing Act That Kept Visa on Top of the Payments Network for Decades

12’58: Carve It Out or Watch It Get Trampled : How Visa Incubates Mobile, Crypto and Agentic Without Killing Them

15’03: $400 Billion in Volume, 30% Growth, Three Years Running : The Numbers Behind Marqeta's Compounding Scale

17’05: The Pandemic Poured Gasoline on Everything : Why DoorDash, BNPL, Expense and Neo-Banking All Exploded at Once

24’24: The Lego Blocks for Payments : How Marqeta Armed the Innovators Who Couldn't Build Through Banks

29’19: Visibility as a Weapon : Why Being Public Helps Marqeta Win Customers Against Private and Embedded Competitors

33’15: Fewer Bets, Higher Probability : How Public Market Discipline Reshaped Marqeta's Risk and Profitability Model

36’35: The Legacy Platforms Were Built for Banks : Why Embedded Finance, Multinational Card Issuing and Personalisation Reshape the Pie

41’50: Prompted, Not Replaced : The Ten-Year View on Whether Volume Comes From People or Robots

43’56: The channels used to connect with Mike & learn more about Marqeta

Disclaimer here — this newsletter does not provide investment advice and represents solely the views and opinions of FINTECH BLUEPRINT LTD.
Contributors: Lex, Laurence, Matt, Farhad, Mike, Daniella
Want to discuss? Stop by our Discord and reach out here with questions.

In this episode, Lex chats with Evan Malanga — Chief Revenue Officer of Yuma, a subsidiary of Digital Currency Group focused on growing the Bittensor ecosystem. They discuss how Bittensor's $6 billion protocol incentivises AI builders worldwide through token emissions across 128 competing subnets, and why the network has produced real commercial outputs — including a 72 billion parameter model trained on-chain and a coding agent rivalling Claude at a fraction of the cost. Evan explains Yuma's role as the institutional gateway to Bittensor through its validator, accelerator, and asset management products, and they explore why the concentration of AI in OpenAI and Anthropic is a systemic risk, and whether Bittensor's future extends beyond AI into a broader coordination engine for decentralised work.

NOTABLE DISCUSSION POINTS:

Bittensor has crossed from experimentation into shipping benchmark-competitive work at a fraction of centralized cost. Three recent proof points: Templar (subnet 3) completed the largest decentralized pre-training run of a 72B parameter model using only the network’s token incentives. Ridges, an AI agent platform, is hitting 88–90% on software engineering benchmarks, on par with Claude-class agents at ~5x cheaper, built by a 3-to-5-person team under $10M of token emissions. Score (subnet 44) is doing computer vision 200x faster than centralized counterparts. Small distributed teams are producing outputs competitive with frontier labs without raising venture capital or hiring staff.

Dynamic TAO restructured emissions from validator-curated to market-curated, making each subnet its own tradeable asset. Previously, dominant validators assigned weights that determined how the 7,200 daily TAO emission flowed across subnets. Under Dynamic TAO, each of the 128 subnets has its own token denominated in TAO, and any holder can buy or sell into specific subnets, pricing them like a market rather than a committee vote. Subnet owners, miners, and validators earn fees in the respective subnet token. Distribution has settled into a power law: the top ten subnets hold ~80% of market cap. This is the move that turned Bittensor from “decentralized AI protocol” into a financial hyperstructure with hundreds of tokenized work markets layered on top.

The economics for subnet owners are genuinely unusual — hundreds of millions in annual incentives, fully subsidized labor, no fundraising. A subnet owner gets access to up to ~256 miners globally competing to satisfy their problem statement, with miner compensation paid by protocol emissions rather than the subnet owner. At current TAO prices, annual incentives across the network run into hundreds of millions; at higher prices, this approaches $1B/year up for grabs. No hiring, no benefits, no recruiting, the network runs as a continuous adversarial competition where validators rank miner outputs. This is the mechanical answer to “why would an AI researcher choose Bittensor over Silicon Valley”, and explains why researchers at Meta and Google reportedly mine Bittensor on nights and weekends, with top miners on subnets like Ridges earning ~$30,000/day.

TOPICS

Yuma, Bittensor, Digital Currency Group, DCG, OpenAI, Anthropic, Foundry, Templar, Ridges, Bitcoin, Meta, Google, BlackRock, JPMorgan, Decentralized AI, Crypto, Blockchain, AI, Tokenomics, Decentralized Science, DeSci, AI Agents, Computer Vision, Proof of Work, Tokenization, Real World Assets, RWA, Machine Economy

 

ABOUT THE FINTECH BLUEPRINT

🔥Subscribe to the Fintech Blueprint newsletter to stay at the forefront of Fintech and DeFi: https://bit.ly/3hyhlC2

🤝 Partner with Fintech Blueprint through sponsorships: https://bit.ly/3UZllsV

👉 Twitter: https://twitter.com/LexSokolin

 

TIMESTAMPS

1’09: The World Wide Web of Intelligence : How Bittensor Turns AI Into Open Competition

9’48: Decentralized AI or Financial Hyperstructure : Unpacking Bittensor's Tokenomics and the Shift to Dynamic TAO

15’04: 256 Miners, Zero Payroll : How Bittensor Subsidizes the Labor Behind Every Subnet

18’03: The Olympics of AI : How Subnet Competitions Replace Bitcoin's Proof of Work

20’09: The Grayscale Playbook for Bittensor : How Yuma Is Building the Institutional On-Ramp

23’19: AI Is the Wedge, Not the Ceiling : Bittensor's 3-to-5-Year Path to Coordinating All Work

28’03: Right but Early : Why the Vision for Decentralized AI May Take 15 Years to Realize

30’52: Decentralized Science as the Next Wedge : Why DeSci Could Be Bittensor's Most Underrated Use Case

34’10: $30,000/Day Mining on Nights and Weekends : Why Meta and Google Researchers Are Quietly on Bittensor

35’56: The channels used to connect with Evan & learn more about Yuma and Bittensor

Disclaimer here — this newsletter does not provide investment advice and represents solely the views and opinions of FINTECH BLUEPRINT LTD.
Contributors: Lex, Laurence, Matt, Farhad, Mike, Daniella
Want to discuss? Stop by our Discord and reach out here with questions.