David Haber

Blankfein’s core risk-management call is that investors must take risk while planning and buying cheap mitigants in advance. Forecasting matters less than asking what could happen, what the portfolio would do, and which protections are available before “the hurricane is coming.” Thorough contingency planning lets a firm move so quickly that it appears prescient: “I want everybody to be called for a false start.”

Finance will adopt AI aggressively, but its near-zero error budget produces a different adoption curve from Silicon Valley’s. Goldman kept the trusted system running while testing its replacement, sometimes needing “50 times” and perfection on “the last 49” before switching. Technology therefore increased costs initially; regulated institutions could not rely on apologies after a rollout failed.

The underappreciated AI danger is opaque, massively leveraged execution rather than machine supremacy. Software could execute 70,000 transactions without the human intuition or visible reasoning trail that once stopped a trading room cold. Regulation might need to slow deployment “not because it’s smarter than us and it’s going to turn us into pets,” but because institutions cannot yet test whether outputs are right.

AI is “going to be very, very important,” yet that does not make every model or company a winner. Blankfein sees genuine conviction in founder-dominated hyperscalers risking their own wealth and egos, but conviction is not correctness. He suggests the world might need four large language models rather than 10, with two becoming very large winners and the field possibly reducing to two.

Goldman preserved partnership behavior after its IPO by aligning people to the whole firm, then reshaped its earnings for public-market math. Glass-Steagall’s repeal made a larger balance sheet necessary, but shareholders valued consistency: “In a private company, you care about the E; in a public company, you care about P/E.” Moving principal risk into funds converted “100-cent dollars” into lower-risk “20-cent dollars,” requiring more volume but supporting higher P/E and return on equity.

Goldman’s financial-crisis edge came from forcing marks into reality before reality forced the firm’s hand. A separate valuation group told traders to “go out and sell something—sell a fraction,” exposing vanished bids in purportedly AAA assets and embedding losses before positions had to be sold. Goldman also fully hedged AIG exposure, demanded collateral, and honored commitments such as Chrysler’s—though “not for more” and “not sooner” than agreed.

Systemically important AI companies should establish public legitimacy before backlash arrives. Goldman learned that it had become “too important, too influential, too big to be anonymous,” yet had no consumer relationship anchoring its reputation when the crisis hit. Blankfein’s advice to leaders at OpenAI, Anthropic and similar institutions is to explain their economic function early, because being modest and invisible becomes a liability when the public decides something went wrong.

🔗 Original source & video: Goldman Sachs Chairman on Why Finance Adopts AI Differently | a16z

Jennings says the first week of December broke the decades-old link between headcount and company output. Opus 4.6 and Codex 5.3 suddenly became capable inside complex existing codebases, letting one or two engineers become “10, 20, 100 x more productive.” After a Q1 review, Block cut slightly more than 40% of its workforce.

The cut’s composition is Block’s rebuttal to the claim that this was merely a 2021 overhiring cleanup. Reductions were far larger in development, while outbound sales and account management were barely touched; Block also protected compliance and compliance technology. Jennings’s categorical line: “We’re not writing code by hand anymore. That’s over. That’s done.”

Block replaced feature-team economics with small teams supervising abundant machine labor. Money Bot moved from roughly 15 people to four plus $2,000 of tokens, while Jennings says he now context-switches among as many as 14 agents producing parallel PRs. Meetings fell 70-80%, development layers fell roughly 50-60%, and squads now contain one to six people.

The operating leverage extends beyond software development into deterministic business workflows. Block’s chatbots and AI phone support automate a majority of inquiries. Block keeps humans in the loop for risk and compliance today, but Jennings expects systems eventually to outperform “a thousand humans” processing those queues.

Block’s product bet is that static financial-app interfaces will start disappearing within six months. Goose, its model-agnostic harness with access to probably 120 models, underpins Cash App’s Money Bot and Square’s Manager Bot. Generative UI can build customer-specific charts or even a restaurant scheduling app on demand, creating engagement upside alongside a “potentially a nightmare” QA problem across tens of millions of users.

Near-term defensibility remains distribution, regulation, network effects, and hardware; long-term defensibility becomes proprietary understanding. Anyone might build peer-to-peer software in a week, but not “vibe code” 50-60 million monthly active users. Block’s intended moat is a rapid feedback loop around its distinctive signal—how buyers and sellers participate in the economy—because companies unable to name what they uniquely understand “maybe could get vibe coded away.”

The AI transformation has not yet resolved Block’s public-market disconnect. Haber notes that the business and gross profit per employee grew while the stock stayed roughly flat for six or seven years; Jennings concedes that the roughly $260 price in 2021 was “a little bit irrational.” His answer is deliberately long-term: markets are voting machines now, weighing machines later, so “just focus on building.”

🔗 Original source & video: How to Reorg After AI Changes Everything | Block’s Owen Jennings on the a16z Show

AI is becoming a full software product cycle, not a standalone model cycle, because it compounds every prior layer—PC, internet, cloud, and mobile—and reaches billions of potential users through smartphones. Rampell says “the vast majority of net new revenue” in software is now coming from AI at both infrastructure and application layers, while capabilities advanced in two years from text, images, and basic reasoning to native audio and real-time interaction. The investor consequence is an application market growing on already-deployed distribution rather than waiting for a new device base.

Adoption evidence is moving from novelty to ROI: Ramp’s customer expense data inflected in January 2025, software companies are reaching $100 million of revenue from zero in one or two years, and roughly 15% of adults globally use ChatGPT weekly. Rampell’s behavioral shorthand is that people want to be “richer and lazier”; the “magic trick has actually gone into the enterprise” because it now saves time, lowers cost, or produces revenue, regardless of whether current valuations are rich or cheap.

AI-native replacements have their best opening at greenfield moments, while installed systems of record make brownfield displacement brutally difficult and let incumbents monetize captive workflows. Rillet can win when a 50-person company with three entities and two currencies must graduate from QuickBooks, but an “AI NetSuite” or Mailchimp clone faces switching friction. Rampell’s deliberately sharp maxim is “the best companies have hostages, not customers,” though he distinguishes durable moats from businesses users hate.

The largest new TAM comes from turning labor into software, but the compelling pitch is often revenue creation rather than headcount reduction. Salient reportedly helps auto lenders collect 50% more, speaks 21 languages, tracks legal requirements across all 50 states and sometimes counties, and automates work for a $50 million call center with 40%-70% annual employee churn. “We are going to make you more money, and it’s going to cost you less” is stronger than a savings-only story.

AI capability is differentiation, not defensibility; the moat is owning the end-to-end workflow and compounding private outcome data. EvenUp routes “literally 100%” of cases through intake, evidence gathering, medical chronologies, demand letters, and complaints, then learns which cases may be worth $50,000 versus $5 million—potentially lowering the viable case floor from $50,000 to $5,000. Haber calls that loop “showing up to a knife fight with a gun.”

Walled-garden data businesses can capture far more value by selling the finished answer instead of licensing raw information. OpenEvidence combines an exclusive medical-journal license with a ChatGPT-like interface reportedly used weekly by two-thirds of U.S. doctors; VLex’s AI layer reportedly quintupled revenue after 26 years of aggregating legal records, while Ask Leo uses otherwise unavailable contract history such as 50 Deloitte agreements. Rampell’s metaphor: own the rare “vegetables,” then sell the finished meal.

The startup opportunity survives strong incumbents, but selection shifts toward model aggregators, proprietary corpora, vertical operating systems, and acquisitions that buy distribution once—not endless services roll-ups. Acharya argues aggregators can offer a “single pane of glass” across specialized models, unlike labs tied to first-party models; Rampell prefers buying one shrinking collector with five blue-chip clients at three times EBITDA over integrating 200 accounting firms. Early enterprise retention is described as strong, with spending tilting toward forward-deployed engineering as customers ask startups where AI should be applied.

🔗 Original source & video: The AI Opportunity that goes beyond Models

Fintech has moved from a startup category into the operating fabric of financial services, but its capital cycle remains brutally seasonal. About 25% of venture dollars flowed into fintech during the mid-2020-to-early-2022 “big EDM pumping summer,” followed by “basically 0%” from the second half of 2022. David Haber places the current cycle in “early to mid-spring”: stronger survivors, new startups, and lingering snow in lending and the broader economy.

The first fintech wave solved access; the next must improve the underlying financial product. Digitization can now produce 30 mortgage offers in an hour or a Rocket application in five minutes, but, as Perret put it, “We’ve made it digital. We haven’t necessarily made it excellent.” Cash-flow-informed underwriting, better fraud controls, embedded finance, and full-stack product bundles are the next value pools.

AI’s clearest near-term institutional role may be automating labor rather than launching autonomous consumer finance. Haber sees financial institutions moving from defensive, in-house development toward outside software that “can actually do the work” across compliance, risk, servicing, treasury, and trading. Voice agents operating in 50 languages illustrate the new economics: “The TAM is largely labor.”

Consumer financial agents remain a trust problem disguised as a technology opportunity. Perret wants an app that automatically routes his paycheck among expenses, high-yield savings, and investments, yet doubts mainstream users will accept unexplained money movement: “Where’s my money? What’s going on?” Plaid’s strategy is infrastructural—let users safely connect data and authorize actions, then observe emerging behavior while watching for new risks.

AI is already scaling the adversary faster than the defense. Financial fraud is growing 18% to 20% annually, and Perret’s bleak 2026 call is that “the mouse is winning right now,” even if the cat eventually prevails. AI has replaced the human factories described in some pig-butchering operations, while Plaid’s Protect combines bank, device, and cross-network behavior to score users, accounts, and actions.

Crypto’s mainstream path runs through familiar financial behavior, not necessarily a wholly separate system. Perret framed speculation, prediction markets, saving, investing, and spending as durable consumer desires whose form factors change. Haber expects some convergence between core financial services and stablecoins such as USDC, while Perret was not sure crypto would merge with banks; Haber left room for its more decentralized frontier.

Plaid’s next leg rests on products that became possible only after its network reached sufficient scale. Protect and the income-and-expense-based Lens Score are positioned as major 2026 drivers, following an 11-to-13-year journey that included paperwork to sell to Visa, the deal’s reversal, fintech winter, and several “refounding” moments. Perret argues the downturn—not the boom—created the discipline and data foundation for faster product development.

🔗 Original source & video: How AI Will Transform Fintech In 2026

AI turns software from a claim on IT budgets into a claim on labor spend because the product can now perform the work. David Haber’s defining example is software that speaks 50 languages, compliantly, 24/7; Alex Rampell’s is even broader: “I’ve never been able to hire somebody for a dollar. Now I can hire software for a dollar.” That should create new consumption rather than simply eliminate jobs.

The AI capability differentiates a product, but it does not by itself defend the company. Haber’s durable moats remain familiar: owning the end workflow, owning the context in which it is applied, becoming the system of record, generating network effects, and embedding deeply enough that the customer depends on the product. “AI is an incredible tool for differentiation,” but its ubiquity makes it weak as a standalone moat.

AI lowers software-production costs while making the race to defensible scale more brutal. Rampell’s anti-fraud analogy: four customers versus three proves little, but four billion observed customers versus one billion can produce a real data advantage. With “nine million ankle biters” competing around obvious ideas, momentum matters because it offers the best path to “gravitational scale.”

Per-seat SaaS faces a pricing-model problem, not necessarily an extinction event. Adobe or Zendesk may sell fewer seats when AI reduces the labor attached to them, yet they could potentially quadruple revenue by charging for outcomes. The more credible disruption is concentrated where wall-to-wall licenses are expensive and underused; software whose payment is tied directly to actual usage, like payroll, is much harder to rationalize away.

The best entry markets combine greenfield customer creation with patient founders. ADP and Paychex inhabit a “Goldilocks zone of irrelevance”: payroll fees are too small relative to payroll itself to justify switching, making entry difficult and retention excellent. A new EHR faces the opposite problem—almost no new hospital systems are created—so even superior software has no clean beachhead.

AI features can reach meaningful revenue unusually fast, but they must still backfill into products and companies. An orthodontic receptionist may look like a feature layered on existing software yet command $20,000 annually because it replaces labor; Rampell’s warning is that “the feature has to backfill product, backfill company as quickly as possible.” Haber’s “messy inbox” wedge shows the path from ingesting email, fax, and phone data into owning downstream scheduling, benefits, and eventually the system of record.

Platforms and incumbents remain advantaged, but the vertical opportunity is too broad for one provider to absorb. OpenAI can pursue five billion ChatGPT users, the developer back end, coding, and large-enterprise deployments without building every obscure vertical workflow. Consolidation should still punish undifferentiated number-three-through-number-100 players, while incumbents that preserve distribution and adopt AI may turn labor replacement into higher margins rather than disruption.

🔗 Original source & video: Why AI Moats Still Matter (And How They’ve Changed)

AI turns software from a passive filing cabinet into an active labor substitute, opening a market potentially far larger than software itself. Alex Rampell traces a 65-year progression from on-premise databases to cloud systems of record and financial-services-enabled vertical SaaS; agents can now perform the work those systems merely recorded. The new formula is “Input, Coffee, Output, Code.”

The addressable budget shifts from software spend toward wages: US registered nurses alone represent more than $600 billion annually, versus under $600 billion for the entire worldwide software market. AI cannot perform CPR, but it can call patients before a colonoscopy, converse in 45 languages, and absorb work hospitals cannot staff. The operative question is how far customers let “their software budget bleed into their labor budget.”

Per-seat incumbents face a brutal cannibalization choice: lose most revenue as AI reduces seats, or reprice around outcomes and potentially grow 10×. Rampell’s Zendesk example pairs roughly $1.4 million of annual software spend with $50 million of support labor; copilots could cut 1,000 seats to 100, while autopilot could eliminate them. For Salesforce and Zendesk, AI is “both defense and offense.”

The strongest startup wedge is the “messy inbox problem”: automating judgment-intensive work between unstructured inputs and legacy systems of record. David Haber’s healthcare example, Tennr, trained against, he thinks, roughly 4 million documents and cut patient-intake administration costs about 90%, then began eating into scheduling, eligibility, and benefits. The AI capability may commoditize, so durability still comes from owning workflows, integrations, network effects, and the eventual system of record — “moats still matter.”

Previously uninvestable niches become venture-scale when software captures labor budgets or bundles labor with a 10×-better replacement system. Compliance is the model: it is reportedly America’s fourth-fastest-growing job, often runs on Excel, and remains chronically understaffed. David Haber describes AI agents clearing tens of thousands of alerts while helping introduce a better transaction-monitoring system, in the context of TD Bank’s $4 billion fine related to transaction monitoring.

AI may automate routine white-collar tasks while making scarce human interaction more valuable. The panel expects at least every white-collar job to gain a copilot, with some roles fully agentic; Rampell’s extreme formulation is that people may either “tell a computer what to do” or be “told by a computer what to do.” Yet once automated outreach becomes ubiquitous, relationships built face-to-face — even “over golf” — may command a premium.

Business fundamentals do not change, but falling costs expand both market size and competitive risk. Investors still need retention, gross profit, overhead discipline, and the “present value of future profits”; meanwhile, AI makes software easier to build and pushes prices inexorably downward. The most attractive hunting grounds are obscure industries where domain experts understand a workflow, the technology is already good enough, and 30-year-old systems can become “10× better.”

🔗 Original source & video: How AI is Reshaping Labor Markets: A $Trillion-Dollar Opportunity Explained