GUY: Good morning. It is Wednesday, July twenty-second. This is Morning Signal. Today’s central question is not whether artificial-intelligence demand exists. The evidence says it is still accelerating. The harder question is where the scarce inputs, pricing power, and investment returns migrate as generic model access gets cheaper.

AVA: And the answer cuts across markets, technology, geopolitics, and rates. We have nine podcast episodes inside the verified prior-twenty-four-hour window. Every claim we discuss comes from today’s written PodcastBrief, and we will name the source before using it. The short version is that abundant intelligence can coexist with scarce memory, power, causal data, safety validation, and skilled labor.

GUY: Let’s start with Excess Returns, where guest Azeem Azhar presented a bottom-up estimate of the artificial-intelligence economy. He estimated that, excluding China, generative-AI spending reached roughly one hundred ten billion dollars in the twelve months through June twenty twenty-six, and a one hundred seventy-five billion dollar annualized June run rate.

AVA: Excess Returns also reported Azhar’s view that the growth rate was about as high as it had been in December twenty twenty-five and roughly three times the pace of earlier internet, mobile, app, and advertising waves. His team tried to avoid double counting by separating end-user applications, foundation-model revenue, hosting, and chip capital spending. It is a private estimate, not an audited total, but it challenges the claim that the boom is only circular capital expenditure or pilots.

GUY: The crucial distinction from Excess Returns is demand validation versus return validation. A large end-market number can support the theme while telling us nothing about whether every lab, cloud provider, data center, turbine plant, or application earns an acceptable return. Azhar used a railroad analogy: a technology can create enormous consumer surplus while providers overbuild and destroy capital.

AVA: Excess Returns supplied a useful enterprise-adoption marker too. Among roughly seventy thousand Ramp customers, median AI spending was cited near eleven dollars per employee per month. That suggests a great deal of runway, but also shallow deployment at the median. So the confirmation test is not license count. It is whether organizations redesign workflows and can show measurable productivity, revenue, or labor savings.

GUY: Now move one layer down the stack. Excess Returns cited memory rising from about two percent of data-center cost in twenty twenty-one to eighteen percent today, in a market with three suppliers that is effectively sold about a year forward. Azhar argued that high-bandwidth memory, rather than logic compute alone, may be the tightest hardware constraint over the next five to ten years.

AVA: The Real Eisman Playbook reached the same bottleneck-migration thesis through power. Steve Eisman’s guest, Baird analyst Ben Kallo, said the United States may need roughly thirty gigawatts of new generating capacity each year for the next several years. He described constraints in electricity, electricians, permitting, interconnection, and construction labor.

GUY: The Real Eisman Playbook also highlighted GE Vernova’s gas-turbine backlog extending to twenty thirty or twenty thirty-one, with management framed as having visibility into the middle of the next decade. That visibility is attractive, but it is not the same thing as terminal return. Long order books can support pricing while customers still face fuel, permitting, ratepayer, and utilization risk.

AVA: Put those two episodes together and the first portfolio signal is clear. The AI infrastructure sleeve is broader than graphics processors. High-bandwidth memory, turbines, transformers, grid equipment, power electronics, and qualified construction labor may have more visible scarcity than generic model access. The falsifiers are equally concrete: lead times collapse, cancellations outpace starts, or capacity clears without support for price and utilization.

GUY: That brings us to the inflation paradox. Thoughts on the Market featured Morgan Stanley economists comparing policy across the United States, Europe, Japan, and China. Michael Gapen expects U.S. inflation to fall toward three percent by year-end twenty twenty-six and roughly two and a half percent in twenty twenty-seven as energy, tariff, and shelter effects fade.

AVA: Thoughts on the Market said Morgan Stanley expects no Federal Reserve move this year. But Gapen identified persistent core-goods inflation, renewed Middle East energy disruption, or stronger AI-linked demand as the routes to a twenty twenty-six hike. Direct AI-related consumer prices were estimated at less than one percent of the consumer basket. The bigger macro channel is stronger demand and animal spirits.

GUY: Goldman Sachs Exchanges reached a similar hold conclusion. Chief U.S. economist David Mericle expects no Fed move in twenty twenty-six and a cut in twenty twenty-seven. He assigned roughly twenty-five percent odds to hikes, versus his reading that markets price about one and a half hikes, or approximately a fifty-fifty chance of two or three.

AVA: Goldman Sachs Exchanges also put numbers around the tolerance band. Mericle said monthly core personal-consumption-expenditure inflation around zero point two percent, or slightly above, could keep the Fed patient. But policymakers are less willing to dismiss repeated supply shocks as one-offs. June consumer inflation may have been better than trend, while tariff, oil, and AI measurement effects should moderate if no new shock arrives.

GUY: Goldman’s growth base case is roughly two percent U.S. GDP growth, a little below potential. Business investment offsets softer consumption, housing, and government spending. Lower immigration reduces the payroll break-even rate to about fifty thousand to sixty thousand jobs per month, so slower job creation does not automatically mean recession.

AVA: Goldman Sachs Exchanges also estimated that equity wealth added roughly zero point three to zero point four percentage point to consumption growth over the past year and could do the same over the next year. That is where AI becomes both growth-positive and duration-negative: capital spending boosts investment, equities support high-income spending, and power demand can add inflation pressure.

GUY: So cheap intelligence can be disinflationary at the task level while the buildout is inflationary in equipment, electricity, and skilled labor. The market implication is uncomfortable but coherent. Software pricing pressure can coexist with an equipment supercycle and a central bank that refuses to ease.

AVA: Thoughts on the Market then widened the map. Morgan Stanley expects the European Central Bank to raise its policy rate from two point two five percent to two point five percent in September. Jens Eisenschmidt argued that two point five percent may still sit inside the ECB’s published neutral range, because fiscal expansion, resilient global demand, and political spreads matter more than an American-style AI boom.

GUY: Thoughts on the Market said Chetan Ahya expects Bank of Japan hikes in December twenty twenty-six and June twenty twenty-seven, fewer than the three hikes embedded in markets and the four some macro investors expect. His reason is weak domestic demand: Japanese consumption is only about one percent above its September twenty nineteen level despite nearly seven years passing.

AVA: On China, Thoughts on the Market reported second-quarter growth slowing to four point three percent despite resilient exports. Ahya expects roughly two trillion renminbi of already-budgeted fiscal capacity to support second-half infrastructure and lift growth toward four point six percent. The TIF interpretation is that actual execution matters more than another headline authorization. Watch construction and broader domestic demand.

GUY: The Indicator from Planet Money added a different macro theme: Dallas and what it called Y’all Street. The episode said Goldman is spending more than five hundred million dollars on a fourteen-story campus expected to house thousands of employees in twenty twenty-eight, while Dallas supplied roughly eighteen million dollars of incentives.

AVA: The Indicator also said Texas now hosts more Fortune Five Hundred headquarters than any other state, and that the Texas Stock Exchange opened this month. But its backers described the buildout as additive rather than zero-sum. JPMorgan’s simultaneous three-billion-dollar Manhattan headquarters is the counterexample. Track finance employment, office absorption, and listing share before declaring New York displaced.

GUY: Now to the model layer. TBPN discussed Moonshot’s Kimi K three and Alibaba’s Qwen three point eight Max as evidence that inexpensive Chinese models can compete on selected benchmarks and enterprise tasks. The episode framed U.S. policy as both a security dispute and a profit-pool dispute.

AVA: TBPN said OpenAI and Anthropic executives emphasize cyber and national-security risks, while critics such as David Sacks and George Hotz view restrictions as potential incumbent protection. The episode also reported that China is considering limits on exporting key training data or downloadable weights, even while the United States debates restricting Chinese models.

GUY: TBPN quoted Treasury Secretary Scott Bessent supporting open source while sanctioning intellectual-property theft. The TIF inference is that both governments may converge on controlled application interfaces and restricted weights for strategic models. That would favor sovereign hosting, model routing, security layers, and locally compliant deployment.

AVA: TBPN also reported that Ramp launched a model router. That fits the broader economics from Excess Returns: sophisticated customers can route tasks across providers by cost, security, and performance. Open weights and switching can compress per-token rents while aggregate compute use still grows. Model commoditization is not the same thing as demand destruction.

GUY: The durable moat may therefore move toward difficult-to-copy context. The a16z Show interviewed Applied Intuition founders Qasar Younis and Peter Ludwig. They said roughly seventy percent of the company’s business is already outside automotive, spanning trucks, defense, mining, agriculture, and industrial machines.

AVA: The a16z Show said Applied Intuition has more than one thousand engineers and hundreds of petabytes of collected data. Its Dana platform packages simulation, synthetic data, reinforcement learning, deployment, and evaluation tooling. The founders described three architectural layers: simulation and reinforcement-learning infrastructure, real operating systems for machines, and autonomy or world models.

GUY: The a16z Show made clear why physical AI is not a chatbot extension. Machines need millisecond latency, power efficiency, hardware-specific deployment, redundant steering and braking, safety cases, and statistical reliability. A better general model helps, but production engineering and verification determine whether it can touch a road, a mine, a farm, or a defense system.

AVA: The a16z Show also presented a cost aspiration. The founders expect advanced driver-assistance hardware to fall from roughly one thousand dollars toward five hundred dollars and eventually become subsidized by automakers, with broad adoption in the late twenty-twenties or early twenty-thirties. Those are management aspirations, not independent forecasts. Safety, unit economics, and production adoption are the confirmation tests.

GUY: Sovereignty matters more in physical systems too. The a16z Show cited South Korea’s mapping restrictions and local hesitation toward foreign robotaxi stacks. Vehicles, maps, defense systems, and industrial machines interact with public infrastructure and real-world safety, so localization and government relationships can matter as much as raw model quality.

AVA: Latent Space took the proprietary-data thesis into biology. Bo Wang and Ci Chu explained why observational single-cell data can saturate. The episode’s technical notes said the CELLxGENE database contains roughly one hundred sixty-eight million cells, each mapped across twenty thousand to thirty thousand genes, producing a matrix of roughly four trillion entries.

GUY: Yet Latent Space said observational RNA expression is highly correlated and weak at predicting what happens after a gene is perturbed. Test loss flattened after roughly one and a half billion parameters even as training loss improved. Increasing information content by about thirty times restored the scaling curve.

AVA: Latent Space described Xaira’s response: use CRISPR experiments to alter genes at scale and generate causal data in X-Atlas, then train the X-Cell diffusion model. The investable claim is not that another model won a benchmark. It is that the data-generation engine may improve predictions on unseen interventions.

GUY: Latent Space’s exact falsifier is prospective validation. Does performance survive unseen perturbations in real human cells? Does it improve target selection, molecule decisions, or experimental hit rates? Today’s input was detailed first-party technical notes rather than a complete audio transcript, so this is the one topic where the source depth is partial.

AVA: The Vergecast offered a useful reminder that regulation can create categories as well as block them. David Pierce and Andy Hawkins examined low-speed electric vehicles from Chip Motors, Amble, and Fiat’s roughly fifteen-thousand-dollar Topolino.

GUY: The Vergecast said vehicles capped at twenty-five miles per hour can operate under a much lighter federal safety regime than conventional cars and are generally restricted to roads posted at thirty-five miles per hour or below. That enables off-the-shelf parts, contract manufacturing, smaller batteries, and lower prices when the average U.S. new-car transaction was cited near fifty thousand dollars.

AVA: The Vergecast’s sober conclusion is that this is a second-car niche in warm, lower-speed communities, not a broad replacement for pickups or highway-capable compact cars. Confirmation means repeat sales outside golf-cart communities and road networks that connect homes, schools, and retail. Falsification means weak safety acceptance, no service or resale network, or roads households cannot actually use.

GUY: Across all of these episodes, regulation is not simply pro-innovation or anti-innovation. It decides which bottleneck earns a scarcity rent. Air permits and interconnection can delay data centers. A lighter vehicle category can admit new manufacturers. Mapping and defense rules can force physical-AI localization. Financial governance and taxes can move jobs without moving the whole exchange network.

AVA: The most important cross-current is value capture. Excess Returns validates a powerful demand direction, but it also warns that model switching and efficiency can compress rents. The Real Eisman Playbook validates long equipment backlogs, but permitting and project economics still matter. The a16z Show and Latent Space point toward proprietary physical or causal data, but only real-world validation can convert technical promise into return.

GUY: Here is the practical watch list from today’s PodcastBrief. For the week of July twenty-seventh, watch the next Fed meeting and communication under Chair Kevin Warsh. Goldman expects a hold but sees more market-volatility risk if the Fed provides less guidance.

AVA: From Thoughts on the Market, watch September for Morgan Stanley’s expected ECB hike to two point five percent, then December twenty twenty-six and June twenty twenty-seven for its expected Bank of Japan moves. The key tests are fiscal impulse and spreads in Europe, and wages plus consumption in Japan.

GUY: From Thoughts on the Market, watch second-half China infrastructure execution against the roughly two-trillion-renminbi budget capacity. The thesis is not confirmed by an announcement. It needs actual spending and broader domestic-demand improvement.

AVA: From Excess Returns, watch whether the one-hundred-seventy-five-billion-dollar annualized AI-spending estimate persists and whether Ramp’s eleven-dollar-per-employee median rises. Median penetration catching the frontier cohort would be stronger evidence than another headline partnership.

GUY: From Excess Returns and The Real Eisman Playbook, watch high-bandwidth-memory lead times, power-equipment prices, order cancellations, and visibility into twenty thirty through twenty thirty-five. These are the fastest tests of the bottleneck thesis.

AVA: From Latent Space, watch X-Cell on unseen perturbations in real human cells and whether predictions improve experimental decisions. From The Indicator, watch Goldman’s Dallas campus target for twenty twenty-eight alongside finance payrolls, office absorption, and Texas exchange listings.

GUY: The absence signal matters. None of today’s episodes supplied audited hyperscaler AI revenue, project-level data-center returns, primary high-bandwidth-memory contracts, independently replicated Kimi benchmarks, or prospective clinical validation for X-Cell.

AVA: So today’s evidence is strongest on demand direction, bottlenecks, and competitive mechanisms. It is weaker on final capital returns. That is exactly why the best formulation is not, quote, AI demand is fake, or AI demand guarantees profits. It is: demand looks real, while value capture is migrating and must be underwritten layer by layer.

GUY: The portfolio takeaway is to separate abundant from scarce. Generic model access is getting cheaper and more interchangeable. Energized land, high-bandwidth memory, qualified power equipment, real-world data rights, causal experiments, safety validation, and organizational redesign are harder to copy.

AVA: The risk takeaway is to separate a full order book from a good investment. A supplier can have visible demand while customers overbuild. A model can win benchmarks while pricing collapses. A platform can have proprietary data while real-world validation fails. Each thesis needs its own confirmation and falsification condition.

GUY: And the macro takeaway is that AI can lower the cost of cognition while raising the cost of its physical buildout. That supports productivity in the long run, but it can also keep equipment, power, and labor tight enough to make central banks cautious now.

AVA: That is Morning Signal for Wednesday, July twenty-second. Today’s nine-episode brief used eight full transcripts and one detailed first-party partial. The complete written briefing and source table are in the TIF PodcastBrief.

GUY: We will be back tomorrow. Until then, watch the bottlenecks, watch the rate path, and keep demand validation separate from return validation.