Ron Barron: Starlink alone can get SpaceX to $10Trillion valuation. Here’s how: Gwynne Shotwell told Morgan Brennan on CNBC in June that Starlink is constrained by capacity, not demand. More orders than they can fill, in the markets that matter most. Then the print. $4.3B of connectivity revenue in the second quarter, 12 million subscribers, double a year ago. That is what sold out looks like. Elon Musk, on the June 6 roadshow, said what lifts the ceiling. V3 is 10 to 20 times more capable per satellite, running three chips the SpaceX chip team taped out itself, 100 times the network bandwidth at half the latency. Run it per launch: a Falcon 9 carries 27 V2s and adds 2.6 terabits a second. Starship carries 60 V3s and adds 60. David Friedberg on All-In, August 8: that market got decimated. Starlink is adding 2 million consumer subscribers a quarter and he sees 4 to 5 million once handsets count. Handsets are the actual next market. Shotwell says mobile will far exceed home broadband, and more than half the planet carries a phone. SpaceX bought EchoStar's spectrum to reach them. Gene Munster put a date on it for Doug Clinton on Deepwater: a fourth national carrier against T-Mobile, AT&T and Verizon by the end of 2027. Whether it takes meaningful share, he says, remains to be seen. Goldman's ramp runs $11B in 2025 to $246B in 2031. Ron Baron told Baron Capital investors on August 10 that Starlink alone is worth $14 trillion in ten years, against a company priced today under $2 trillion.

Bernstein's Stacy Rasgon ran the semiconductor industry's own numbers at McKinsey decades ago: from 1996 to 2004, it destroyed $5 billion in economic profit. Pull Intel out of the math, and that number is $50 billion destroyed. Then COVID happened. On the same CASPA panel, moderator Austin Lyons asked whether every layer of the chip stack would just keep charging for its value with memory margins already near 90%. Rasgon's answer: pricing power used to be a myth the industry told on itself, since nearly all the value it created flowed straight to the customer. Post-2008 consolidation started to change that, and COVID finished the job. Input costs rose, and for the first time, Rasgon says, every company he covers passed the increase through at its existing margin, not dollar for dollar. The old rule said chip pricing always gives back its gains once the cycle turns. Rasgon's own numbers say that rule expired years before this cycle even started. Our coverage of this panel breaks down which parts of the chip stack have kept that pricing power through a downturn before, and which gave it back. podcastalpha.substack.com Source: CASPA (Chinese American Semiconductor Professional Association) - youtube.com/watch?v=YE6IVKcp…

Every constraint slowing the AI buildout reads like a problem. Bernstein's Stacy Rasgon, on a CASPA panel with SemiAnalysis's Dylan Patel @dylan522p, said one of them might be the thing keeping it from becoming one. Moderator Austin Lyons had just asked about the rising cost of capital. Rasgon's answer: a buildout that can't finance everything at once doesn't necessarily crash, it paces. He pointed back to the telecom bubble, when dark fiber got laid in the ground and never lit. His read on today's financing squeeze: it forces supply and demand to stay matched instead of letting the industry build until it snaps. A bottleneck that slows a buildout isn't automatically a bear signal. It can be the thing stopping the buildout from becoming its own bust. Our coverage of this panel tracks where financing, not demand, is already deciding who gets to build AI compute, and what happens to capacity that doesn't. podcastalpha.substack.com Source: CASPA (Chinese American Semiconductor Professional Association) - youtube.com/watch?v=YE6IVKcp…

NVIDIA's edge here isn't more GPUs. John Linford: it's making the CPU do the part the GPU is worse at. Ben Pouladian @benitoz asked what else compares to cuDSS. Linford's answer: a sparse solve runs in three phases, reordering, factoring, solving, and each wants different memory. "You might want the high capacity, low latency memory of a CPU for one part, but then you want the high bandwidth ... memory of a GPU for another part." cuDSS runs each phase on whichever chip fits and coordinates the handoff, so "you get the best of both worlds." The AI-infrastructure story usually collapses to a GPU count. NVIDIA's own engineer is describing a system that deliberately keeps work on the CPU where the CPU still wins. The moat argument here is the orchestration, not the chip. The newsletter covers the rest of this $NVDA interview: the one customer number that holds up, and the two header multiples that don't. podcastalpha.substack.com Source: BEP Research - youtube.com/watch?v=yunV1tlr…

Where does the time saved by a faster solver actually go? John Linford, NVIDIA: designers spend it on more physics, not on finishing early. Ben Pouladian @benitoz asked what cuDSS 1.0 actually changes for designers. Linford's answer opens with the catch: designers always want to capture more physical phenomena in their simulations, and "sometimes they turn that dial so high that their simulation starts to lag ... it just takes too long to resolve." The fidelity goes up. The idle time never shows up at all. Every performance gain in this stack reads like freed-up capacity until you watch where it actually goes. It gets reinvested in scope, not banked. The bottleneck resets at a higher resolution, not a shorter runtime. Full breakdown in the newsletter: what holds up when Linford gets pressed on this $NVDA product's bigger numbers, and what doesn't. podcastalpha.substack.com Source: BEP Research - youtube.com/watch?v=yunV1tlr…

John Linford, NVIDIA: simulating a car used to mean submitting a job and waiting for the airflow solve to finish. Now the designer just changes a number and watches the air move. Ben Pouladian @benitoz asked for one real example of the bottleneck cuDSS removes. Linford's answer: once the sparse solve runs fast enough, "it's like I have a sort of virtual wind tunnel where I can play around with the simulation and change my design." The car that looks good and gets the best mileage comes out of that loop, not one final run. The upgrade here isn't a bigger speedup number. It's the difference between one shot at a design and as many as the engineer wants to try before the deadline. That's what actually decides how good the finished version is. The newsletter has the rest of this $NVDA conversation, including the two headline multiples NVIDIA's own spokesperson couldn't back up on tape. podcastalpha.substack.com Source: BEP Research - youtube.com/watch?v=yunV1tlr…

John Linford, NVIDIA's Director of Industrial Engineering Products, says the point of an engineering agent isn't speed. It's never having to wait. Ben Pouladian @benitoz asked him where agent orchestration actually changes engineering work. Linford's answer: a designer can simply say, "go generate a swath of designs for me to evaluate," then step in as the expert to judge the results. The agent runs the exploration. The human keeps the call on what's actually good. That split holds past this one interview. Every "AI agent" pitch in engineering ends up drawing the same line: the software absorbs the waiting, the judgment stays human, and the judgment is still the expensive part. The newsletter walks through the rest of this $NVDA interview: the 32x customer case study NVIDIA leans on, and the 245x and 460x multiples Linford couldn't source when pressed. podcastalpha.substack.com Source: BEP Research - youtube.com/watch?v=yunV1tlr…

Technical debt isn't an accident. It's a budget line nobody gets credit for filling. Nikesh Arora @nikesharora and Chamath Palihapitiya @chamath explained why it never gets fixed until it breaks. On All-In @theallinpod, host Jason Calacanis @Jason asked why technical debt keeps resurfacing across every enterprise category. Chamath's answer: once software ships, nobody wants to own maintenance, so every client drifts onto its own unsynced version, multiplied across millions of companies and trillions of lines of legacy code. Arora's answer: "CIOs are not incented to maintain the most modern infrastructure. They maintain to keep costs low." Every upgrade risks an outage, with no budget line that rewards fixing it anyway. Arora's next move: that same incentive gap is why he thinks the CIO's job is about to get bigger, since every corporate function now wants AI headcount and the CIO already controls the deployment. The executive everyone blames for old infrastructure is the one about to run point on buying the new infrastructure too. Our newsletter ties that incentive gap to the forced security re-rate the rest of this episode predicts. podcastalpha.substack.com Source: All-In Podcast - youtube.com/watch?v=6czo-C3F…

Nikesh Arora @nikesharora shut down every Palo Alto Networks data center eight years ago to move to the cloud. Now he calls going back "kind of an empty threat." Chamath Palihapitiya @chamath asked Arora, on All-In @theallinpod, whether more enterprises will follow the one major US bank that just built six independent data centers instead of scaling further on AWS and GCP. Arora's answer: it sounds good on a slide, but Nvidia can't ship you chips until 2028, and almost nobody can run 50 data centers worldwide the way a hyperscaler does. The real move is hybrid: lease bare metal from a hyperscaler or "a cousin of a hyperscaler," and run open-source models you can actually inspect, while global compute stays this scarce. The cloud-versus-own-infrastructure debate didn't resolve toward either side. It resolved toward whoever controls the compute, since that's the one input a slide deck can't substitute for. Our newsletter tracks which enterprises are actually building their own infrastructure versus bluffing, and what tight compute means for the hyperscalers. podcastalpha.substack.com Source: All-In Podcast - youtube.com/watch?v=6czo-C3F…

Who's liable when your AI agent drains $50,000 from your account? Nikesh Arora @nikesharora gave the group a one-word test: agency. Chamath Palihapitiya @chamath posed the hypothetical to Arora and hosts Jason Calacanis @Jason and David Friedberg @friedberg on All-In @theallinpod: a shopping or banking agent goes rogue, money disappears, who pays. Arora's framing: "There's a word in agent, it's called agency. Somebody has to give the agent agency." Whoever grants that access carries the liability. Chamath agreed it's the only rule that scales, with one carve-out: an agent that breaks its own stated limits becomes the platform's product-liability problem, not the user's. Every AI product this year ships with an "allow access" prompt. That click is now the liability contract, whether or not anyone reads it that way. Our newsletter maps exactly where Amazon, the banks and the hosts draw that line today, and where the rogue-agent exception could move it. podcastalpha.substack.com Source: All-In Podcast - youtube.com/watch?v=6czo-C3F…

CAPTCHA was built to prove a human was behind the keyboard. Nikesh Arora @nikesharora just told All-In that AI clears it better than most humans do. David Friedberg @friedberg asked the Palo Alto Networks CEO how banks are supposed to tell a user's own browser login from a personal AI agent logging in with the same credentials. Arora's answer: they can't. Both show up as one login. Then he added the harder part: there are already ways to solve CAPTCHA with AI, so you don't even need a human on the other end anymore. Host Jason Calacanis @Jason, on All-In @theallinpod, piled on that AI now reads the "click all the bicycles" puzzle faster than people do. CAPTCHA was the internet's cheap proxy for "a human is here." That proxy just failed, quietly, before anyone updated what replaces it. Our newsletter covers what actually replaces CAPTCHA as a line of defense, and who eats the liability when it doesn't hold. podcastalpha.substack.com Source: All-In Podcast - youtube.com/watch?v=6czo-C3F…

The same crowd that said all of crypto's value would accrue to Bitcoin and Ethereum said all of AI's value would accrue to Anthropic and OpenAI. Pompliano has watched that argument lose twice. On the Pomp Podcast, Anthony Pompliano @APompliano pointed to crypto's old fat protocol thesis, the idea that app and infrastructure layers would capture nothing, which proved wrong once value spread across the stack. He argued AI's fat model thesis is making the identical mistake. Jordi Visser @jvisserlabs agreed, citing Jensen Huang's framing of a 90 trillion dollar compute buildout large enough for nearly every layer to survive, while still expecting Anthropic's own returns to slow from here. The thesis that one layer eats the whole value chain keeps getting proposed and keeps losing to the same math. A market big enough to support a stack supports more than one winner in it. Podcast Alpha's coverage of this episode breaks down where the value actually sits across models, infrastructure, and the application layer. podcastalpha.substack.com Source: Pomp Podcast - youtube.com/watch?v=5nUxci_0…

Jordi Visser @jvisserlabs: Stripe is worth about as much as Citigroup ($C). Stripe runs that business with 10,000 people. Citigroup needs 200,000. On the Pomp Podcast, Visser pointed to Stripe's purchase of OpenRouter as the reason why. Stripe can pay customers yield on their balances instead of taking the deposit spread to zero, the same spread JPMorgan ($JPM) collects $107 billion a year from while spending roughly $16 billion a year on technology trying to defend it. Twenty times the headcount for the same valuation is the gap AI agents are being pointed at next. The deposit spread banks live on is the line item closest to actually disappearing. Podcast Alpha's coverage of this episode runs the numbers on where that spread goes next and who it hits first. podcastalpha.substack.com Source: Pomp Podcast - youtube.com/watch?v=5nUxci_0…

Anthony Pompliano @APompliano: Sylvia's AI compute bill hit $550,000 a month within five months of launch. The fix wasn't switching providers. It was deleting a bug. On the Pomp Podcast, Pompliano walked Jordi Visser @jvisserlabs through what actually brought the number down. Before touching which model Sylvia ran on, the team found pages that re-queried the model on every refresh and stopped doing that. Only after that fix did they move 500 million tokens off Anthropic onto open source models and owned hardware, cutting costs 97% while latency and accuracy both improved. The model swap gets the headline number. The real saving came from killing waste that had nothing to do with which lab's model was running underneath. Podcast Alpha's coverage of this episode weighs that evidence against Visser's case for why JPMorgan-scale enterprises won't follow the same path. podcastalpha.substack.com Source: Pomp Podcast - youtube.com/watch?v=5nUxci_0…

Jordi Visser @jvisserlabs: core CPI came in hot this week. He didn't read the report. Anthony Pompliano @APompliano asked him directly on the Pomp Podcast why neither of them had CPI on their list of topics, even as the rest of the market called it the week's biggest story. Visser's answer: core CPI sits at 2.6%, and whether it prints 2.2% or 3.2% has never decided a year's return. The S&P has compounded more than 10% a year across every trailing window from 10 years back to 90. The only thing that has ever broken that run is job losses, and weekly jobless claims, adjusted for population, sit near an all time low. CPI reads like the story because it lands every month on a schedule. Jobless claims are the one number that has actually forecast a downturn, and right now they are not saying anything. Podcast Alpha's coverage of this conversation lays out the two catalyst windows Visser is watching next and what would prove the framework wrong. podcastalpha.substack.com Source: Pomp Podcast - youtube.com/watch?v=5nUxci_0…