GUY: Good morning, Ava. It is Saturday, July 18, 2026, and today’s Morning Signal has one big idea running through almost everything: artificial intelligence is no longer just a contest over which model is smartest. The harder constraints are power, hardware, capital, trust, distribution, and whether the product works reliably enough to earn a return.

AVA: Exactly. And before we get into it, today’s written brief checked thirty-four podcasts and found seven episodes inside the strict twenty-four-hour window. Four had usable, episode-matched public transcripts. We are only using claims from those verified sources. Two title-matched transcripts were rejected because the publisher or hosts did not match, and the official Goldman Sachs episode had subtitles disabled. So the evidence boundary matters today.

GUY: Let’s start with markets. On The Real Eisman Playbook, Steve Eisman used large-bank non-accrual data to argue that a broad American credit cycle is not already underway. He cited JPMorgan at nine point four billion dollars of non-accrual loans, down five percent year over year and two percent sequentially, and Bank of America at five point eight billion, down four percent year over year and flat sequentially.

AVA: But Eisman’s conclusion was narrower than “nothing can go wrong.” His point was that large-bank consumer and commercial credit are concurrent indicators and still look benign. The harder-to-see risk is private credit, especially software exposure. So the recession watch moves away from the transparent bank books and toward opaque lenders and the economics of AI projects.

GUY: Right. And on that same Real Eisman Playbook episode, he described very strong bank returns. The return on tangible common equity figures he cited were twenty-three percent for JPMorgan, sixteen point five for Bank of America, seventeen point seven for Wells Fargo, thirteen for Citi, twenty-five point five for Goldman Sachs, and twenty-six point six for Morgan Stanley.

AVA: His valuation mechanism matters. Eisman frames price to tangible book as a function of sustainable return on tangible equity. Goldman, JPMorgan, and Morgan Stanley deserve higher multiples in that framework because returns exceed twenty percent. But strong trading, investment banking, mergers, initial public offerings, and AI financing do not prove that the projects being financed will ultimately earn acceptable returns.

GUY: That distinction is the heart of the trade. Banks can get paid for funding the buildout before we know whether the buildout generates enough cash. And The Real Eisman Playbook gave a sharp example of budget reallocation. Eisman said IBM pre-announced earnings per share of two dollars and ninety-three cents against three dollars and one cent expected, and revenue of seventeen point two billion dollars against seventeen point nine billion expected.

AVA: He linked IBM’s twenty-five percent stock decline and the broader software selloff to clients shifting late-quarter capital spending toward servers, storage, and memory. On the same episode, Ericsson’s earnings fell seven percent and revenue fell six percent, with weak guidance tied to component inflation, while ASML beat and raised guidance.

GUY: So the causal chain is simple enough to underwrite: AI infrastructure scarcity pulls customer budgets into hardware, component inflation hurts buyers, software and network vendors miss, and upstream semiconductor equipment benefits. The falsifier, from the written brief, is rapid normalization in component pricing alongside stable software renewals and expansion.

AVA: And The Real Eisman Playbook also showed why a headline beat is not enough when the hurdle is higher. Eisman said Elevance beat but fell because guidance captured less than the quarterly upside and membership declined one percent after pricing increases. UnitedHealth beat and raised guidance as pricing helped margins despite flat revenue. GE Aerospace grew revenue twenty-four percent and raised guidance but still fell because investors wanted more. Netflix met expectations, reduced the frequency of engagement disclosure, and fell eight percent after hours.

GUY: That is a useful portfolio rule: compare the result with the implied bar, not just consensus. And reduced disclosure can itself be a governance signal even when the quarter looks acceptable.

AVA: Now payments. On The Real Eisman Playbook, Eisman treated Circle’s trust-bank approval as strategically positive, but smaller than the competitive threat from an integrated group involving Stripe, Visa, Mastercard, Coinbase, and BlackRock. He described Circle’s economics as interest on Treasury-backed U-S-D-C reserves minus BlackRock’s eighteen-basis-point money-market fee.

GUY: A bank charter might internalize that fee, but it does not create distribution. Eisman’s conclusion was that Circle is too small to fight integrated payment networks alone and should look for a larger partner. On that episode, Circle was discussed near sixty-three dollars, versus a thirty-one-dollar initial public offering price and a prior peak near two hundred seventy dollars.

AVA: Both The Real Eisman Playbook and the All-In Podcast discussed unconfirmed reports of a PayPal bid. Eisman cited sixty dollars and fifty cents per share and roughly eleven times estimated 2026 earnings. The All-In panel discussed a roughly fifty-three-billion-dollar offer and more than four hundred million PayPal consumer accounts.

GUY: Emphasis on unconfirmed. The strategic logic is clearer than the transaction status. In the All-In discussion, Stripe’s merchant base, PayPal and Venmo consumers, Braintree, a possible Block point-of-sale contribution, and stablecoin rails could create more transactions inside one network and reduce reliance on Visa and Mastercard rails. But without a named buyer group, filed terms, financing, and an antitrust path, it is optionality, not an investable floor.

AVA: Let’s connect that to wealth infrastructure. On Masters in Business, Barry Ritholtz interviewed Altruist founder Jason Wenk. Wenk described the independent-advisor market as roughly ten trillion dollars across about thirty-five thousand firms, with Schwab and Fidelity holding about eighty-five percent of standalone registered investment advisor assets and Schwab above fifty percent.

GUY: And he explained how the plumbing makes money. According to Wenk on Masters in Business, incumbent custodians monetize idle cash spreads, fund distribution, and payment for order flow, while advisors often buy separate accounting, billing, reporting, and tax systems. Altruist tries to integrate those functions, support fractional shares, reduce idle cash, and earn across custody, software, asset management, and automation.

AVA: Wenk also said Altruist raised more than six hundred million dollars over seven years, and estimated that a credible new custodian requires at least five years and two hundred fifty million dollars. That tells you why this apparently boring infrastructure is hard to disrupt. The moat is not a flashy interface. It is engineering, capital, licensing, security, and trust.

GUY: So our markets setup is unusually coherent. The Real Eisman Playbook says broad bank credit is benign and financing activity is strong. It also says AI spending is crowding out parts of software. Masters in Business says modern financial infrastructure can create measurable operating leverage, but only after years of investment. The common question is where capital earns a return rather than merely where it is being spent.

AVA: Which takes us to technology and AI. On the All-In Podcast, David Sacks, Chamath Palihapitiya, David Friedberg, and Jason Calacanis discussed Demis Hassabis’s proposal for a federally overseen, industry-funded self-regulatory organization modeled on FINRA.

GUY: The All-In panel described a thirty-day pre-release submission period for frontier models, quarterly updated risk benchmarks, and tests for cyber, national-security, and biological risks. Sacks supported the concept only with conditions: broad representation including startups and open source, review limited to genuinely frontier models, scope confined to catastrophic risks, voluntary operation before it becomes mandatory, and replacement of a new agency rather than adding another layer.

AVA: Palihapitiya’s concern on All-In was speed, because well-capitalized firms can shape regulation into a barrier to entry. The investment issue is not simply whether self-regulation happens. Certification cost, release delay, board composition, and the definition of “frontier” can become moats. Capability-based scope and startup representation reduce capture. Fixed compute thresholds or a state-by-state patchwork favor incumbents.

GUY: Trust is the second layer. On All-In, the panel covered an acknowledged Grok Build data-handling failure in which codebases were reportedly sent to remote servers despite privacy expectations. Elon Musk said previously uploaded data had been deleted, and the harness was open-sourced.

AVA: Palihapitiya’s point on that All-In discussion was that a zero-data-retention promise is not enough if non-obvious leak paths remain. Sacks cited private evaluations, tenant-level trust boundaries, decoupled orchestration, and customer fine-tuning rights as ways enterprises can retain control. So data sovereignty becomes architecture, not just contract language.

GUY: The All-In Podcast also cited Ramp chief executive Eric Glyman saying AI token spend among Ramp customers rose twenty-one-fold over the prior year. The panel’s implication was that uncontrolled model selection can become a real expense and margin risk. They discussed higher-memory Macs as one possible route to more local workloads, lower token cost, and less leakage, but the written brief correctly labels that as a scenario rather than a verified purchasing forecast.

AVA: The Vergecast approached trust through Apple’s trade-secret suit against OpenAI. Nilay Patel and David Pierce emphasized that Apple’s complaint is one-sided and that the merits still need discovery and adjudication. Their legal framing was that Apple is using trade-secret law because there is no shipped OpenAI device for a patent claim and no clearly copied code or expression for copyright.

GUY: On The Vergecast, Patel and Pierce expect the case to test the boundary between manufacturing knowledge employees can carry in their heads and protected materials they cannot take. The commercial question is whether this becomes a years-long constraint on OpenAI’s hardware ambitions, but the next pleadings and any request for injunctive relief matter more than speculation.

AVA: The Vergecast also gave us the cleanest consumer-product test. Patel and Pierce found Apple’s beta Siri more capable at some multi-step tasks but less predictable on simple commands such as reminders. Their car example showed how replacing a known command path with a cloud-mediated model adds routing and security ambiguity, even if a local button remains.

GUY: That is the ceiling-versus-floor problem. A fantastic demo raises the capability ceiling. But consumers stop using an assistant when the reliability floor collapses. Nobody wants to debug a timer, a light, or a reminder every morning.

AVA: And on The Vergecast, they applied that test to reports that OpenAI’s first device could be a screenless, battery-powered smart speaker with cameras and sensors. A mostly home-based device may ease some connectivity problems, but it inherits music services, smart-home standards, microphones, local-versus-cloud processing, privacy, and predictable-action requirements.

GUY: Their skepticism was not that conversation cannot improve. It was that conversation alone may not overcome phones, HomePods, Google Home, Alexa, and the existing ecosystem. The measurable catalyst is repeat use beyond timers, music, and retrieval, not launch-day attention.

AVA: Masters in Business gave the enterprise counterexample. Jason Wenk said Altruist can validate accounts, transfers, and bank links in under two minutes, with more than ninety-eight percent of workflows avoiding human intervention. He said Hazel, Altruist’s AI product, targets tax planning, financial plans, data gathering, and client responses, with some complex tasks reaching a claimed unit cost of three to five dollars.

GUY: The critical difference is structure. In Wenk’s Masters in Business account, enterprise custody workflows have defined inputs, auditable outputs, transaction rails, and a financial incentive to fix failures. That makes operating leverage measurable. His argument that cloud-native microservices are easier to defend than old monolithic mainframes still needs controls testing, but strong authentication, security keys, smaller services, and restrictions on high-risk phone workflows have a clear mechanism.

AVA: So the investable AI question is becoming: who owns the scarce complement? Power, components, proprietary workflow data, security, distribution, and verified reliability can all retain value even as model prices compress.

GUY: Let’s add geopolitics and policy. On All-In, the panel argued that electricity and permission to build are the next scarce inputs. Chamath Palihapitiya said near-term energizable data-center capacity commands an extreme premium because many projects are delayed or abandoned. David Sacks argued that New York’s announced hyperscale data-center moratorium would push construction toward places that allow behind-the-meter generation.

AVA: David Friedberg’s point on All-In was brutally physical: graphics processors have little economic value until they can be powered. The panel’s precise estimates for the American energy deficit and project cancellations were not independently verified in the written brief, but their directional mechanism aligns with Eisman’s budget-reallocation evidence. Scarcity is moving upstream into chips, generation, and interconnection.

GUY: On the same All-In episode, Sacks argued that restricting domestic data-center construction while limiting advanced-chip exports to allies could reduce U.S.-aligned compute capacity in both places. That is an attributed policy view, not a settled fact. The falsifier is visible: allied capacity expands under the controls and American projects reach operation without longer queues.

AVA: The Vergecast discussed another regulatory lever: Federal Communications Commission control over broadcast ownership and merger review. Patel and Pierce argued that removing the thirty-nine-percent broadcast-reach cap would consolidate ownership without solving competition from YouTube, TikTok, and Instagram, where user-generated content has a different cost base.

GUY: They also criticized reported hospitality provided to F-C-C officials by companies with transactions under review. Those are the hosts’ interpretations of reported events, not a legal finding. The market takeaway is narrower: regulatory permission can move media asset values, but consolidation cannot repair a structurally inferior content-cost model.

AVA: Now the cross-currents. First, The Real Eisman Playbook and All-In together describe a strange split. AI financing supports bank fees and returns. Scarce hardware and power pull capital spending away from some software. Private-credit lenders may hold the opaque exposure. So financials and semiconductor equipment can work at the same time that application vendors and software lenders struggle.

GUY: Second, All-In, The Vergecast, and Masters in Business all point toward verifiable control. Open systems do not mean weak boundaries. The Grok Build incident shows the cost of leakage. Apple’s lawsuit shows provenance risk in employee knowledge and supply chains. Wenk’s custody model shows the regulated answer: integrate data and execution, authenticate strongly, restrict access, and keep an audit trail.

AVA: Third, All-In and Masters in Business show two modernization paths. Acquiring PayPal could buy consumer accounts and distribution, but it also inherits architecture and culture. Altruist’s clean-sheet path can create operating leverage, but it consumes time, capital, licensing effort, and trust. The integration outcome matters more than the deal announcement.

GUY: Fourth, The Vergecast, Masters in Business, and The Real Eisman Playbook put reliability between capability and free cash flow. Siri shows higher capability can reduce usefulness if simple tasks fail. Altruist shows narrow automation can save labor and errors. IBM shows enterprise budgets are moving before the return is proven.

AVA: And fifth, distribution remains harder than creating the product primitive. The Real Eisman Playbook says Circle can improve reserve economics but still lacks network scale. All-In says PayPal has hundreds of millions of relationships but an aging interaction model. Masters in Business says Altruist distributes through advisors whose clients add assets over time. The Vergecast says a new OpenAI device still has to overcome established ecosystems.

GUY: The absence signal is as important as any headline. Today’s usable episodes did not provide a company-level bridge from incremental AI capital spending to incremental free cash flow for hyperscalers or model labs. They were much more precise about shortages, power, token cost, financing, and regulation than realized return on compute. That is still the central underwriting gap.

AVA: Let’s finish with what we are watching. The Real Eisman Playbook plans a July twentieth interview with Baird sustainable-energy and mobility analyst Ben Kallo on how AI data-center construction is changing energy markets. That directly tests the power-bottleneck thesis.

GUY: And The Real Eisman Playbook has a July twenty-second interview scheduled with Autonomous payments analyst Ken Suchoski. We want evidence on PayPal, stablecoins, and agentic-payments consolidation, while keeping unconfirmed transaction claims out of the base case.

AVA: In the next bank reporting cycle, watch sequential consumer and commercial non-accruals at JPMorgan, Bank of America, Wells Fargo, and Citi. Broad deterioration would falsify Eisman’s claim that the visible bank books remain benign and the harder risk sits elsewhere.

GUY: In the next IBM and software updates, watch whether server, storage, and memory purchases remain a substitute for software spending or whether late June was only timing. If component pricing falls while software renewals and expansion stabilize, the crowd-out thesis weakens.

AVA: For Apple and OpenAI, watch the next pleading and any request for injunctive relief. For AI governance, watch who selects the board, which federal body oversees it, whether state laws are pre-empted, and whether “frontier” is defined by capability rather than spending.

GUY: For New York data centers, track whether projects migrate, secure exemptions, or arrange behind-the-meter power elsewhere. For PayPal, require a named buyer group, filed terms, financing structure, and antitrust route before treating the reported price as a floor.

AVA: And for consumer AI, measure repeat behavior. Does beta Siri complete multi-step work while preserving reminders, smart-home controls, and app intents? If users do not return after novelty fades, capability has not crossed the reliability bridge.

GUY: The portfolio message is selective, not universally bullish or bearish. Prefer transparent balance sheets over opaque credit. Prefer scarce complements and audited workflows over vague AI exposure. Treat rumored deals as optionality. And demand evidence that spending produces retention, task completion, unit-cost savings, or free cash flow.

AVA: That is Morning Signal for Saturday, July 18. The models are improving, but today’s evidence says the scarce assets increasingly sit around the model: electricity, components, trust boundaries, regulated infrastructure, and distribution.

GUY: Have a great Saturday. We will be watching the bottlenecks, the earnings hurdles, and the evidence that turns capability into cash. See you next time.