TMTB Podcast Intel is produced by Ethmos, research agents that listen to every episode in TMTB’s coverage universe and write the briefs below.
Follow Dylan Patel, Citrini, MSFT, CRWV or OpenAI on Ethmos to get a brief whenever they or their executives go on a podcast. TMTB readers get 14 days free and 25% off their first month.
Today’s Essential Episodes:
SemiAnalysis’ Dylan Patel and Bernstein’s Stacy Rasgon on NVIDIA’s Neocloud Backstops, the Substrate Crunch and Custom Chips
Chipstrat, from the CASPA CEO Summit panel moderated by Austin Lyons, with Stacy Rasgon of Bernstein and Dylan Patel of SemiAnalysisCitrini on Why Monkeys Are the New Bottleneck, His Claude Antivenom Experiment and the Lab-Revenue Risk to the AI Buildout
Monetary Matters, with Jack Farley and Citrini of Citrini ResearchMicrosoft CEO Satya Nadella on Zero-Trust Agents, Why Alignment Won’t Fall Out of Scaling and Every Company Owning Its Own Evals
20VC, with Harry Stebbings and Satya Nadella, CEO of MicrosoftEx-Fidelity PM George Noble on Shorting CoreWeave, Tesla and SpaceX, and Why He Calls AI the New Shale
Monetary Matters, with Jack Farley and George Noble, former Fidelity portfolio managerOpenAI’s Ari Weinstein and James Sun on Computer-Use Agents Faster Than Humans, Enterprise Guardrails and the Dots Cloud Computer
The a16z Show, with Seema Amble, Fabrizio Serafini, and Ari Weinstein and James Sun of OpenAI
Produced by Ethmos: AI-assisted summaries of public podcast episodes. For information only. May miss context or contain errors. Verify before relying on them. Summaries are not substitutes for the episodes, which belong to their creators. Not investment advice, offer, or solicitation. Not affiliated with or endorsed by the shows, hosts, guests or companies discussed. Corrections and publisher opt-outs: support@ethmos.ai.
Transcript · Follow Dylan Patel on Ethmos
THE SIGNAL
Patel says Nvidia’s neocloud backstops guarantee rent at $3.50, the cloud’s breakeven, then take a 50% revenue share above that for Sharon AI and Fermi; Lambda’s Anthropic cluster carries 10%, so Nvidia absorbs tail risk (our read).
Patel says Musk’s play is to sign no contracts until capacity is live, earning about $50B per gigawatt versus roughly $15B on contracted offtake; he puts Anthropic’s inference margins at 75-80% at $15B/GW, far lower at $50B.
Rasgon: memory is 70-80% of AI rack silicon area, yet memory names trade at 2-4x forward earnings and SanDisk’s quarterly EPS guide tops its stock price of 24 months ago; he wonders what happens if 2028 pricing holds.
Rasgon puts Nvidia revenue near 50/50 hyperscale versus non-hyperscale, AI labs about 20% directly and indirectly; Patel says Nvidia held margins near flat while absorbing memory, substrate and optics cost increases.
01 Financing gets harder, not impossible
Rasgon frames financing as one of several constraints — alongside power, supply, and semicap tools — now moderating AI buildout pace, noting the AI stocks are still up a lot but well off their peaks, a sign the market fears the buildout’s peak may be approaching. He observes hyperscalers, once the most cash-generative businesses in history, are now burning cash and probably headed for equity raises (Meta’s new share class structure would enable one, and Google has raised); CoreWeave’s unsecured bonds, which moderator Austin Lyons cited at 15% yields, signal where new issuance would price. Rasgon argues constraints aren’t purely bad — they prevent the kind of overbuild seen in the dot-com era’s unused dark fiber — and could produce a steadier, longer buildout cycle by better matching supply to demand.
Patel adds that financing has bifurcated: any cloud could get capital a year or two ago, but now neoclouds, which he says account for nearly half of what’s being built, can’t build without high-grade offtake, and even offtake from Anthropic or OpenAI gets priced at 13% to 15% debt. He cites Volta’s cluster build for Anthropic, where a 12% coupon bond sold at 95 cents effectively yields ~15%. This locks out neoclouds serving smaller enterprises and startups unless they strike structured deals — Nvidia’s revenue-share arrangements with Sharon AI, Fermi, and Lambda, and Broadcom’s rent-to-own chip leasing for Anthropic, are examples of chip vendors backstopping financing to keep capacity growing. Patel puts Meta’s compute and data-center commitments at $1.4 trillion and calls financing part of a “holy trinity” alongside land/power/shell and chips.
02 Ranking the real bottlenecks
Patel ranks substrates as the top constraint, ahead of memory, then optics/co-packaged optics (CPO) ramp — he says even one major company couldn’t satisfy 2028 CPO volumes given laser and fiber-attach limits. He argues power is a less severe bottleneck than popularly believed, since roughly 50 gigawatts of data centers are being built globally next year (7-8 in China, ~35 in the US and the rest worldwide, with Australia, India and Malaysia among the big places), and operators can work around power limits with on-site turbines, reciprocating engines, or fuel cells faster than they can fix semiconductor supply chains.
Rasgon largely agrees but stresses memory and wafer fab equipment (WFE) as constraints. He notes the trade ratio for high-bandwidth memory requires three to four times as many wafers per gigabyte versus standard DDR5, and that a company’s WFE guidance of above $200 billion for next year is about as high as he thinks possible given full clean rooms, though investors expected more — WFE estimates have already risen from roughly $120 billion early this year to a projected ~$155 billion. He expects tightness to persist through 2028, with 2029 a demand question.
03 Pricing power discovered during COVID
Rasgon recounts his own McKinsey research from years ago, which by his recollection showed the semiconductor industry destroying about $5 billion in aggregate economic profit from roughly 1996 to 2004 ($50 billion excluding Intel), meaning nearly all value flowed to customers. COVID changed that: input cost increases were fully passed through, and industry blended ASPs rose roughly 50% from 2019 to 2023/2024 even as unit shipments stayed flat — a dynamic he says is now playing out at an even greater intensity, with 70-80% of this year’s roughly 100% industry revenue growth (toward $1.5 trillion) coming from memory ASPs.
Patel notes some players flex pricing power opportunistically (memory firms swung from negative margins a few years ago to high margins, rewarding themselves on the upside after being squeezed in the downturn) while others like TSMC climbed steadily from 45% to 60% gross margin without giving much back in downturns. He estimates building a gigawatt of compute rose from roughly $40 billion in 2024 to over $60 billion now, though stripping out chip-vendor margin (Nvidia, Broadcom, Marvell) would put raw cost near $20 billion — meaning five gigawatts could theoretically be funded by one gigawatt of inference revenue. He reads Anthropic’s recent disclosures as roughly $70-80 billion per gigawatt of inference capacity.
04 Nvidia and Broadcom’s margin defense
Rasgon frames Nvidia’s margin durability as a function of performance-per-dollar gains outpacing ASP increases: Ampere-to-Hopper ASPs roughly doubled while performance rose ~10x; Blackwell ASPs rose ~50% while performance gained 3-30x; Rubin rack costs are probably doubling and cost per gigawatt is rising (about $25 billion for Blackwell and $40 billion for Rubin, as he recalls Nvidia’s figures) against another ~10x performance gain. He cites Nvidia chief scientist Bill Dally’s 2023 Hot Chips keynote target of a millionfold GPU performance improvement over ten years, following a thousandfold gain in the prior decade, as the benchmark that must hold for margins to stay defensible.
Patel says Nvidia and Broadcom trade at their lowest multiples basically ever, and Rasgon says Nvidia had traded cheaper relative to the SOX semiconductor index on only 13 days in the 10 years before a note he wrote nine or 10 months ago, and it’s cheaper today. Patel ties the discount to fears that big customers like Meta and Google will move away from Broadcom or stop buying Nvidia and insource, and says the shift has started: Nvidia’s slice of the industry’s CoWoS wafers and HBM fell in 2026 and falls again next year as Google, Amazon and AMD take share, though its earnings still double.
05 Custom silicon and enterprise adoption
Patel describes OpenAI’s first-generation custom chip beating Blackwell when SemiAnalysis ran open-source Chinese models on it with its InferenceX benchmark in OpenAI’s lab, a chip designed with heavy AI assistance; Anthropic aims to develop its own chip in half OpenAI’s development time, while Google’s TPU, Amazon’s Trainium (now fully usable in PyTorch), Meta, and Microsoft all pursue aggressive custom-silicon roadmaps. He notes the time to get new hardware running inference software on InferenceX has compressed dramatically — about six months for Nvidia and six to nine for AMD, a few months each for Google and Amazon, and a month and a half for OpenAI — suggesting chip design itself may become increasingly commoditized like software, which would pressure Nvidia and Broadcom margins longer-term even as both maintain strong positioning for the next two years.
On enterprise adoption, Rasgon identifies AI coding as the clear killer application with vertical revenue growth, and shares an anecdote of an analyst using Claude to one-shot a ten-year financial model in an hour and 20 minutes — work that would otherwise take a week. He frames the broader debate as whether AI spending generates measurable return, concluding anecdotal evidence leans positive though adoption remains early. Patel adds that enterprises will mostly buy tokens rather than deploy their own hardware or models, arguing scale economics push even pharma companies and hedge funds toward consuming AI as a service rather than running infrastructure themselves, continuing a decades-long industry shift toward cloud consumption.
Listen · Watch · Follow Citrini on Ethmos
THE SIGNAL
Citrini says if combined lab ARR lands at $300B rather than the $500B being underwritten, the math that funnels lab revenue back into infrastructure breaks; meanwhile open models’ share of net-new ARR rose from ~1% to ~18% in a year.
DeepSeek’s claimed inference stats (155T tokens/day on Huawei cards) imply ~$38 revenue per GPU hour versus $5–8 cost, ~80% margins and ~81-day payback, implying substantial further token-price deflation, especially below the frontier.
Charles River’s earnings lever is operating leverage: spare post-downturn capacity makes new preclinical contracts higher-margin; monkeys take ~4 years to breed and China switched from halting exports to importing; he seeks monkey-pricing alt data.
Their binder-design run used ~10,000 GPU hours costing ~$50k, which he says would now cost $2–4k, the kind of deflation behind his view that early-stage assets get commoditized (our read: bearish single-asset biotechs).
01 Why he cut AI hardware exposure
Citrini says his fund largely exited AI infrastructure and semiconductor positions starting around March, shifting toward what he calls “agentic utilities” — companies benefiting from actual AI usage rather than the capex buildout itself. He says he remains constructive on NVIDIA individually, but argues the infrastructure trade now requires underwriting aggressive assumptions, whereas usage-driven plays can be justified more simply. He notes that had his portfolio stayed as infrastructure-heavy as it was in April/May, it would likely be down ~50%; separately, Farley noted semiconductor stocks fell nine consecutive days in October, a streak he said was last seen in 2006.