2026.10.07 | Substack | AI Demand | TMTB Podcast Intel: AI Payback & ASICs

TMTB Podcast Intel is produced by Ethmos, research agents that listen to every episode in TMTB’s coverage universe and write the briefs below.

Today’s Essential Episodes:

1. SemiAnalysis’s Dylan Patel and Jordan Nanos on Whether AI Spending Can Keep Growing Exponentially, NVIDIA as

Financier and the Anthropic IPO

Big Technology Podcast, with Alex Kantrowitz, Dylan Patel (founder and CEO of SemiAnalysis) and Jordan Nanos of SemiAnalysis

1. Marvell CEO Matt Murphy on the Long-Term Revenue Guide, Custom Silicon and Breaking Through the Copper Wall

CNBC’s Squawk on the Street, with Carl Quintanilla, Jim Cramer and David Faber, and Matt Murphy, chairman and CEO of Marvell Technology (the Murphy segment)

1. Figma CEO Dylan Field on Figma’s Design Agent, AI Lowering the Floor for Creation and Code as a Canvas Material

MTS, with Dylan Field, co-founder and CEO of Figma

1. Ben Bajarin and Jay Goldberg on Whether the Memory Cycle Has Broken and NVIDIA and Broadcom’s $40B+ Financing

Playbook

The Circuit, with Ben Bajarin and Jay Goldberg

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.

Listen

THE SIGNAL

• Patel says Anthropic already covers all compute (training plus inference) from revenue, OpenAI follows within two to three quarters per SemiAnalysis modeling, and Amazon’s AI infrastructure is profitable, with Bedrock lifting gross margin.
• Patel’s IPO math: Anthropic used to raise at about 30x revenue and may list near 20x on about $100B of revenue; at a $10 trillion valuation, he says, it would need something like a trillion dollars of revenue.
• Oracle’s OpenAI deals are effectively fixed-margin, missing rising GPU rental prices; Patel says New Mexico could slip multiple years and Wisconsin also has delays.
• The guests expect OpenAI and Anthropic each to add 5-10 GW next year versus a few GW for Google, and Patel argues Google’s TPU sales to Anthropic directly undercut DeepMind.

01 Scale and trajectory of the buildout

Patel opens by correcting a widely-circulated chart comparing AI CapEx to railroads as a share of GDP: he argues the real figure is higher than reported (5-6%, not ~3%) because US-based CapEx serves inference demand for OpenAI and Anthropic users in Europe and Asia, unlike railroads or fiber, where CapEx and usage were geographically matched. Host Alex Kantrowitz notes hyperscaler CapEx growth hit 116% year-over-year in Q3 but Wall Street expects deceleration toward 70% by Q1 next year; both guests agree growth must decelerate mathematically (you can’t repeat 10x growth off an ever-larger base) but push back on the idea this signals trouble, framing it as definitionally inevitable rather than demand weakening. Kantrowitz reads out Bain’s estimate that hyperscalers need $6 trillion a year in revenue by 2031 to break even and a Columbia professor’s estimate of $3.5 trillion a year in AI spending by 2032; Patel works the payback math himself, putting the revenue needed against a trillion dollars of capex at roughly $1.5 trillion over six years, or about $250 billion a year.

02 Where the revenue actually comes from

Patel argues AI revenue won’t just come from existing software budgets but will accelerate because labor’s share of GDP — already declining since the 1970s — will fall faster as AI tokens substitute for labor, while the overall economy grows fast enough that labor’s dollar value may stay flat or grow slowly even as its share shrinks. He points to chip design as a precedent: the number of US R&D chip engineers has been flat for 20 years while the value created by firms like NVIDIA and Broadcom exploded, arguing this pattern — same headcount, AI-assisted tooling, exploding output value — will replicate across pharma, trading, robotics, and self-driving. Nanos adds that investors lack imagination about non-token revenue: if Anthropic or OpenAI apply frontier models to bring products like a cancer drug or self-driving vehicle to market, returns could dwarf per-token API pricing, citing GLP-1 drugs’ financial impact on Eli Lilly as an analogy.

03 Anthropic’s economics and IPO stakes

Patel calls his own viral tweet about the Anthropic IPO a joke but defends its underlying logic: at a $2 trillion valuation Anthropic could go to zero, but at a $20 trillion valuation the implication is that AI has become so dominant and wealth so concentrated among equity holders that social unrest becomes a real risk. He separately argues Anthropic’s unit economics already beat classic SaaS — Salesforce-level 75% gross margins achieved much faster, near-zero customer acquisition cost, and sticky revenue that doesn’t shift even when competing Chinese models emerge. On “pacing the frontier” (deliberately holding back capability), Patel calls competitive dynamics between Anthropic’s Dario and OpenAI’s Sam largely self-reinforcing — neither can slow down much given competition from each other and from China.

04 NVIDIA’s financing role and neocloud contract risk

Nanos details NVIDIA’s $588 billion in off-balance-sheet backstops — revenue floors, lease guarantees, and investment-grade credit support — that let neoclouds raise financing without needing a hyperscaler anchor tenant; he frames this as NVIDIA pragmatically diversifying its customer base rather than propping up a bubble, noting NVIDIA’s own internal research use of GPUs is a small fraction of its revenue. On Meta-style cancellation risk, Nanos distinguishes neocloud contracts (five-to-six-year committed terms, hard to cancel) from hyperscaler self-build sites (which could be paused since GPUs, HBM, and optics, not concrete, dominate project cost) and argues current demand still outstrips chip and data-center supply.

05 Security failures and sovereign AI

The guests describe a SemiAnalysis report documenting weak neocloud security, which found the open-source Chinese model GLM 5.3 capable of penetrating dozens of neoclouds; this preceded a tweet from Ilya warning that rogue AI agents could next target neocloud infrastructure to run more copies of themselves. SemiAnalysis says it personally discovered a cluster where it could see other tenants’ data, including what it describes as an intelligence agency from a non-US country with a top-10 GDP. Nanos connects this to the “sovereign AI” framing, arguing nations treat compute as strategic infrastructure akin to telecom buildouts, which explains why government workloads end up sharing clusters in a seller’s market.

06 ClusterMax rankings and lab model withholding

In SemiAnalysis’s ClusterMax 3.0 report, CoreWeave and Nebius rank as top-tier neoclouds, with Oracle and Google Cloud in the second tier; Patel clarifies the rankings measure service quality (reliability, security, storage, network), not stock potential — Oracle rates well operationally despite SemiAnalysis flagging in April that pipeline delays would stall its New Mexico data center for OpenAI, later confirmed by an Oracle force majeure announcement. Nebius, with shorter average contract lengths, has been able to raise prices faster than Oracle or CoreWeave, which locked in fixed-margin long contracts. Separately, Google DeepMind’s share of Google’s total compute has fallen even as Google’s total compute grows, leaving DeepMind with less compute than OpenAI or Anthropic by Patel’s estimate; both guests say Gemini needs to find a product niche since it isn’t winning on coding. Finally, both guests confirm labs are withholding finished frontier models — Anthropic had Mythos ready in February but released only a neutered version called Fable months later, while OpenAI delayed its Astra and Bell models — because keeping models internal protects against distillation and preserves margin on proprietary applications.

Listen · Watch

THE SIGNAL

• The $80B 2030 midpoint sits in a $70-90B range; asked about Jensen Huang’s prediction that Marvell could be the next trillion-dollar company, Murphy says that if you put a multiple on that range four years out, who knows what’s possible (our read: management is not distancing itself from the trillion-dollar framing).
• Murphy corrected the hosts: scale-up optics is the third growth leg, not the fourth, and he hinted at another, unnamed leg, which points to a growth driver Marvell hasn’t disclosed.
• Scale-up optics goes from zero this year to well above the original $300M guide for next year, with the estimate raised in each of the last two quarters, so 2027 consensus may still be catching up (our read).
• Going from $12B+ in custom revenue in 2028 to the $30B 2030 target implies about 2.5x in two years, roughly 58% a year (our read); Murphy says Marvell plays in two separate parts of custom silicon.

01 The long-term guide

Marvell lifted its long-term revenue targets at its investor meeting the day before the interview. Murphy, CEO for ten years and in the industry for 32, describes the company’s forecasts as consistently pragmatic rather than aggressive. He points out that two years ago, at the same venue, Marvell projected $15B in 2028 data-center revenue; sell-side consensus today implies over $30B for that year — roughly double. Cramer cites Morgan Stanley as happy with the new targets but unsure Marvell had to set targets that high; Murphy responds that he avoids framing guidance as conservative or aggressive, preferring to give ranges so investors can judge for themselves, with the $80B company-level midpoint he calls very achievable by 2030. He notes Marvell’s revenue has grown from roughly $2B ten years ago to an expected $20B next year.

02 Custom silicon versus connectivity

Murphy argues Marvell’s differentiation in this AI cycle is connectivity, not custom XPU silicon alone. He describes three waves: compute companies, then a memory wave led by three trillion-dollar memory firms, and now a gap in the connectivity layer needed to link installed compute and memory so nothing sits idle — a gap he says grows more acute with agentic AI and reasoning models requiring more diverse compute types. This was his thesis at Computex in June, which he says Jensen Huang validated on stage. On custom silicon specifically, the TAM estimate he cites grew from $40B in April 2024 to $235B by 2030; Marvell’s $30B custom revenue target implies about 13% market share, which Murphy says is not a stretch given the company already expects $12B+ in custom revenue by 2028.

03 Breaking the copper wall

Murphy describes a structural shift inside data centers: links that used to be passive copper are moving to optical due to rising bandwidth and data-rate requirements, which has already benefited Marvell. He distinguishes scale-out networking (connecting racks) from scale-up networking (coherently connecting GPUs/XPUs within a rack), noting that today intra-rack connections are still all copper but will begin moving to optical starting next year. Marvell’s scale-up optics sub-business is $0 in revenue this year; Murphy says he guided it at $300M two quarters ago, raised that estimate again last quarter, and it has moved up further since — which Murphy called the third leg of Marvell’s growth stool.

04 Data-center politics and risk

Asked about political moratoriums and risk around data-center buildouts given what Cramer calls a $400 billion addressable market, Murphy says the company monitors the external environment constantly but currently sees no effect on its market potential or near-term revenue growth. He attributes the next-year guidance increase from $18B to $20B primarily to connectivity demand, framing hyperscaler customers as executing long-standing buildout plans with visibility Marvell can track.

05 Insider buying

Hosts note Murphy faced press criticism twice and responded by buying Marvell stock in the open market on two occasions in the $70s, roughly a million dollars of stock by Cramer’s account, which he and the hosts frame as a direct answer to doubters.

Listen

THE SIGNAL

• Field names design-system adherence as the top roadmap priority after GA, aiming for better-than-human use of components from local canvas context and past work; this would tie agent quality to customers’ existing Figma libraries, strengthening lock-in (our read).
• Field says companies hit the AI-shock phase at different times: some finished it late last year while others are only starting, which could make enterprise demand for AI design tooling and seat decisions uneven across customers (our read).
• Field admits Figma believed its quality-tracking for non-deterministic models was already dialed in but it wasn’t, so the beta-to-GA period went largely to rebuilding the basic per-change checks of what improved or regressed (our read).

01 Agent launch and beta lessons

Figma’s design agent moves from beta (launched in May) to general availability, with Field citing improvements in speed, context understanding, Figma-specific workflow comprehension, and design-system adherence. He says the team learned that designers and product managers dislike working in a terminal or IDE, reinforcing the need for agent-native, canvas-based workflows that feel more like home. Building with non-deterministic AI models required returning to basics: rigorously tracking whether each change made quality better or worse, task by task, rather than assuming prior tuning was sufficient.

02 The floor-versus-ceiling thesis

Field’s central framing is that AI democratizes creation (lowers the floor) but has not raised the ceiling of achievable quality — a correction from an earlier, more optimistic view that he says proved wrong. He notes AI performs best on formally verifiable domains (math, cybersecurity) where answers are binary, but design has no single correct solution, so models tend toward center-of-distribution, average output, which he calls slop — though he notes the definition of slop keeps shifting as models improve. He argues craft reflects a creator’s thinking and care, and that people perceive this even if they can’t articulate it, making well-crafted work stand out regardless of AI’s rising baseline.

03 Augmentation, not replacement

Field repeatedly frames the agent as empowering designers rather than replacing them: it should automate tedious, hours-long tasks while preserving distinctly human skills like lateral thinking, cultural context, and emotional judgment that models lack. He describes organizations moving through an AI adoption cycle — initial shock at model capability, followed by a return to valuing employees’ deeper reasoning, judgment, and forward planning. The risk, in his view, is letting AI do the thinking entirely rather than using it to learn faster and experiment more boldly.

04 Design as the new code, and what’s next

Field’s one-year prediction is that design becomes the new code: if coding models keep improving, the differentiating decisions become architecture, brand, marketing, and point of view rather than code itself, which becomes raw material. He expects designers who step up to gain larger roles shaping products, while acknowledging not everyone will embrace this shift. Figma’s codelayers initiative extends this idea literally — letting designers move from a creative canvas to contributing to a pull request, engineering permissions allowing.

05 Ten-year speculation and FigJam as cognition

Asked about a decade out, Field speculates about 10,000 logical-qubit computers, the start of a Mars colony, working brain-computer interfaces, and a mapped human connectome, while suggesting everyday experience may still feel similar unless people opt into radically different interfaces. The conversation closes on FigJam as a nonlinear thinking tool, raising the open question of how humans might communicate spatially with AI agents, and agents with each other.

Listen

THE SIGNAL

• Goldberg says the up-to-$42B Broadcom loan commitment comes from a preliminary Anthropic IPO S-1 that Reuters obtained, which suggests Anthropic is moving toward listing and that Broadcom’s credit exposure is tied to one lab’s TPU ramp (our read).
• Goldberg says agentic AI is lifting CPU demand, which tightens supply of cheaper LPDDR too; he reads reports of flat consumer memory prices as a deliberate supplier choice but heard nothing on Micron’s call to confirm it.
• On HBM, Goldberg sees a reverse node shrink: about four consumer-grade wafers’ worth of capacity yields one HBM part priced roughly 100x higher, a strong incentive to starve the consumer and edge segments.
• Bajarin says margins normalizing in the sixties would still be fine, against the past-cycle swings from 80% to 30% gross margins that Goldberg describes; a phone forecaster Goldberg spoke with expects a decade of no growth as global upgrade cycles stretch past about four years.

01 Memory’s structural floor vs cyclical crash

Following Micron’s earnings — a strong quarter and guide, with the stock roughly flat for the week — Bajarin and Goldberg debate whether the memory market’s new $600-800B revenue floor represents permanent structural change or merely a bigger base before a traditional bust. Bajarin notes the market appears to be pricing in demand destruction and margin compression arriving around 2029, while bit growth and some pricing leverage could plausibly continue past that point. Goldberg lays out the bear case: DRAM and NAND bit growth is running only in the low-to-mid twenties percent through 2027-28 even as prices have roughly doubled this year, suggesting supply-demand equilibrium by 2028, with the market potentially turning pessimistic ahead of that. He notes optimists believe long-term agreements and take-and-pay structures will permanently dampen the industry’s historic boom-and-bust swings.

China’s CXMT is a wildcard both agree matters: it’s pulling forward capacity targets and can build fabs in roughly 18 months versus 2.5-3 years in the US, Korea, or Japan, implying a glut risk concentrated around 2028. Bajarin separately argues the industry may now resemble TSMC’s model of monetizing legacy nodes (DDR4) for extended periods rather than clearing clean rooms for newer product, since enough bit-growth demand exists across both premium (HBM) and commodity segments to justify it — a genuine structural shift if true. Goldberg counters he sees no clear signal memory makers have internalized that discipline, citing despeccing of memory in Tesla’s robots due to supply constraints and the strong economic temptation to redirect capacity toward HBM, where roughly four trailing-edge wafers’ worth of capacity could instead produce one far more profitable leading-edge HBM chip. Separately, Goldberg relays a contact’s forecast that smartphones face a decade of essentially flat growth as upgrade cycles lengthen amid rising component costs and weaker phone improvements.

02 HPE’s networking repositioning

Bajarin recounts attending HPE’s networking investor event, where executives described the company (post-Juniper acquisition) as effectively a networking company, with networking serving as the lead hook pulling through compute, software, and services sales, analogous to how Dell leads with storage. He highlights that enterprise customers are reportedly accelerating network upgrades primarily due to agentic-AI-driven security vulnerability fears on aging infrastructure. HPE showcased its Helios rack with six liquid-cooled trays using Broadcom Tomahawk switches over Ethernet for scale-up domain switching, built with AMD — a genuine engineering challenge, in Bajarin’s view, though not using HPE’s own Juniper IP for that specific function. A notable moment: when investors pressed on HPE’s unique IP given the reliance on Broadcom switches and third-party components, the answer was reportedly uncomfortable. Goldberg, a longtime Juniper analyst, says he disagrees with many points of HPE’s investor thesis, while Bajarin notes HPE did raise its guidance.

03 Balance sheet as a service

The hosts describe a new competitive dynamic: NVIDIA’s recent non-deal roadshow deck to major sell-side banks reportedly included, for the first time formally, a slide touting use of its balance sheet to fund ecosystem growth — essentially presenting its 10-K footnote obligations, which critics cite negatively, as a selling point. Goldberg frames this as NVIDIA turning a point of investor concern into a boast. Broadcom has now followed with a reported commitment to loan Anthropic up to $42 billion, tied to Google TPU production routed through Broadcom, extending the same vendor-financing pattern.

The hosts debate why fabless/XPU designers rather than foundries like TSMC take on this financing risk. Goldberg argues TSMC sits too far from end demand to forecast it well, while NVIDIA and Broadcom — with higher gross margins and more cash generation — are better positioned to monetize and absorb that risk; he likens TSMC to a skilled industry investor being asked to pick individual stocks outside his expertise. He also invokes Broadcom’s Hock Tan’s reputation for running the company like a private equity fund as a reason to trust its risk-reward judgment on these deals. On competitive implications, Goldberg states NVIDIA generates roughly 20x the free cash flow per quarter of AMD, leaving AMD unable to compete on balance-sheet terms (citing roughly $10B cash for AMD vs $60B for NVIDIA), while Broadcom can partially offset its smaller relative balance sheet by tapping private-equity capital pools. Both agree this dynamic — unprecedented in an industry historically averse to debt given its high-fixed-cost, cyclical nature — is likely to expand further.

04 Quantifying tokenomics

Goldberg introduces a push to make AI economics rigorous rather than vague, arguing that available token pricing data can now reveal real cost structures — for instance, the reasoning-model pricing premium that existed roughly a year ago has disappeared as the feature got bundled into frontier models, while input and output tokens carry very different pricing given different underlying compute costs. This analytical approach, he says, helps explain decisions like NVIDIA buying Groq for $20 billion. Bajarin connects this to OpenAI’s DevDay announcement of a faster inference tier, initially speculated to run on Cerebras though later clarified to be a low-batch-optimized configuration of NVIDIA GPUs rather than specialized accelerators, suggesting a greenfield opportunity for premium-priced, higher-speed token tiers. Goldberg is more cautious about a “premium tokens” framing, suggesting pricing will instead be segmented by customer and specific product — such as a dedicated cybersecurity-focused model — rather than a simple tiered token system, while agreeing speed-based pricing is plausible. Both anticipate this kind of economic transparency becoming increasingly important if frontier AI labs eventually go public.

POWERED BY Ethmos for TMTB

Get briefs like these for every company you follow.

TMTB Podcast Intel picks a handful of episodes a day. With Ethmos, research agents listen to every podcast that touches the names in your book and file briefs like these every morning.

+ NVIDIA · + Marvell · + Figma · + Micron · + Your own list

Start your 14-day free trial

TMTB readers get 25% off their first month. Cancel anytime.

TMTB Podcast Intel is produced by Ethmos for TMT Breakout. Each brief is an AI-assisted analytical summary of a publicly available podcast episode, prepared for information only; it is a condensed analytical summary, not a substitute for the episode: it is not intended to reproduce the programming and may miss important details, context or nuance, and it may contain errors, so verify against the original episode before relying on it. Each episode and its recording remain the property of their creators. Nothing here is investment advice or an offer or solicitation of any kind. Neither Ethmos nor TMT Breakout is affiliated with, sponsored by, or endorsed by the shows, hosts, guests, or companies discussed; show and company names are the property of their respective owners. Publishers who prefer not to be covered, and anyone with a correction, can write to support@ethmos.ai; publisher opt-outs are honored.