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 DELL or CRWV on Ethmos to get a brief whenever their executives go on a podcast. TMTB readers get 14 days free and 25% off their first month.
Today’s Essential Episodes:
Dell CEO Michael Dell on Why There’s Too Little Supply for an AI Bubble, a Tighter Memory Market Next Year and Data-Center Pushback
CNBC’s Squawk on the Street, with David Faber and Michael Dell, founder and CEO of Dell TechnologiesCoreWeave EVP Chen Goldberg on Forge, Opening CoreWeave’s GPUs to Self-Serve Engineers and the Fragmentation Risk in AI Agents
ThursdAI, with Alex Volkov and Chen Goldberg, EVP of product and engineering at CoreWeavePaul Tudor Jones on a Market That Still Feels Like 1999, a Year-End Rally and Why He Expects to Be Short in 2027
CNBC’s Squawk Pod, with Andrew Ross Sorkin and Paul Tudor Jones, founder and chief investment officer of Tudor Investment CorporationIneffable’s David Silver and Lasse Espeholt on Learning Without Human Data, a Vera Rubin Cluster and Superintelligence in Three Years
Training Data, with Sequoia’s Sonya Huang and David Silver and Lasse Espeholt, co-founders of Ineffable Intelligence
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 · Watch · Follow DELL on Ethmos
THE SIGNAL
Dell expects the supply-demand gap to worsen next year, especially in memory, and possibly to last four years given fab lead times; that supports memory pricing but pressures server OEM input costs (our read).
Dell says H100 rental prices hit a one-year high about a week ago, on a chip Dell started selling four years ago, which undercuts the bear case that GPUs lose economic value quickly and supports longer depreciation lives (our read).
Dell says frontier labs drove the first wave and enterprise demand is only now starting, which could broaden Dell’s AI customer mix beyond the frontier labs (our read).
Dell rations supply by vetting buyers on capital, ready data center space, and their downstream customers and applications, which suggests a quality-screened backlog but some deferred revenue (our read).
01 Data centers under pressure
Faber opens on whether community opposition and moratoriums threaten the AI buildout pace. Dell characterizes the increased community engagement — over siting, power sourcing, and water-efficient liquid cooling — as a positive development, and says Dell has a large project backlog with no material delays. He notes data centers and workload demand are now global, not just US-concentrated, and sees mitigation options like more efficient systems and modular data centers. Pressed twice on whether this threatens 2027 guidance, Dell says he sees no impediment and no change to expectations, declining to speculate on whether post-midterm politics will ease pressure.
02 Safety debate and government’s role
Faber frames the AI safety spectrum between Huang’s lighter-touch, lab-responsibility view and Gates’s characterization of AI as among the most dangerous technologies ever. Dell compares AI to fire — capable of great benefit or harm — and says labs are already thinking hard about safeguards, with a role for government, bounded permissions for AI agents, monitoring, and human oversight. He stresses open-source and smaller models give customers choice beyond frontier models. On US-China competition, Dell argues that opting out of building AI and data centers would hurt the US economy while benefiting rival economies, favoring staying ahead while finding a path to safe, powerful systems. He says he isn’t a policymaker but engages with them, and places himself generally in Huang’s camp while noting Dell builds infrastructure, not models — infrastructure itself has no awareness of its own use. Asked directly if he worries about existential AI risk, he predicts things will be fine a few years out and says humans are good at creating countermeasures to anticipated risks.
03 Pricing, supply and the bubble question
On GPU economics, Dell distinguishes smaller systems running on less powerful infrastructure from frontier training systems requiring the latest, costlier semiconductors and memory. He says demand for compute remains greater than supply, describes the industry as still early in its growth curve, and acknowledges any tech cycle mixes excess, opportunity, and skepticism. Asked directly if this is a bubble, Dell says there isn’t enough supply for it to be one, and that Dell prioritizes orders based on evidence customers will actually use the capacity rather than taking every speculative order. His key bubble indicator is watching rental prices for older chips like the H100 — a one-year-high print signals genuine compute scarcity, not pre-positioning. He flags the roughly four-year fab lead time as a structural constraint, implying the supply gap could persist for years, with memory in particular tightening further ahead.
04 Compute-to-intelligence shift and Dell’s own transformation
Dell elaborates on his public claim that only 10-15% of companies grasp the organizational rewiring AI requires. He says firms typically try to bolt new technology onto legacy processes rather than reimagining outcomes, data architecture, and simplified/standardized processes before applying AI. Dell says his own company has executed this transition over three years, reflected in improved operating leverage, with further changes expected ahead. Faber connects this to Dell’s headcount falling by 11,000 to roughly 97,000 even as revenue rose 19%, though Dell reframes the benefit as speed and capacity expansion rather than just headcount efficiency.
05 Financing and philanthropy
Faber cites figures showing Dell earned $4.1 billion in income last quarter but only $1 billion in free cash flow, with receivables rising as Dell finances customer purchases. Dell says most of this financing is short-term (three to four months) and gets handed off to banks and lenders. On philanthropy, Faber notes Dell and his wife have donated $6.25 billion directed to certain ZIP codes; Dell says the foundation he created in 1999 has focused on children in urban poverty for twenty-seven years and confirms more giving is planned both domestically and internationally (97% of children live outside the US), leaving open whether future giving extends to Invest America accounts.
Listen · Watch · Follow CRWV on Ethmos
THE SIGNAL
Goldberg says the latest ClusterMax report raised its criteria and dropped several vendors a tier, while the host notes CoreWeave kept its Platinum rating; the gap among GPU clouds may be widening (our read).
The host, who joined CoreWeave through its Weights & Biases acquisition, says nine of ten big labs run on CoreWeave, citing the Anthropic and OpenAI deals; self-serve Forge and GPU sandboxes now give individual engineers access, opening a long-tail funnel beyond lab concentration (our read).
Goldberg says demand for an agent-improvement loop only appeared in the last 12 months, and self-serve waited until security controls were ready; the host says CoreWeave became a GPU, CPU, sandbox and services provider within a year.
Forge connects Mission Control and observability to CoreWeave’s inference and training services and carries context across them; that is a software layer that could raise switching costs on top of compute (our read).
01 Forge: origin and architecture
Forge is CoreWeave’s new unified platform for building and running agentic AI systems, announced at the Fully Connected conference. It stitches together previously separate assets — Weights & Biases models, Weave, inference/serverless, training, and OpenPipe — into one workflow following a loop of run, observe, curate, improve, evaluate. Goldberg says the idea emerged organically from two internal projects: CoreWeave’s own agentic research system (internally referenced as ARIA) and a speculative-decoding optimization effort for its inference service. In both cases, engineers — she references one named Aaron — found that manually stitching tools together across the loop left the human in the loop worn down, and that machine-learning engineers and developers were no longer cleanly separated roles. CoreWeave frames its advantage as full-stack, purpose-built-for-AI optimization, arguing that just as training jobs spanning thousands of GPUs required unified tooling, agentic inference workloads (always-on, running on large GPU fleets) demand the same approach. Forge is available to all users immediately, without the traditional sales and invoicing cycle — a notable shift for a company historically reliant on enterprise sales relationships, alongside a related self-serve sandboxes product that now includes GPU access.
02 Fragmentation risk in the agent ecosystem
Goldberg’s central cautionary argument, drawn from a fireside conversation with Dylan Patel of SemiAnalysis, is that the AI industry risks repeating — or worsening — the standardization failures Kubernetes avoided. She argues Kubernetes succeeded in part because the community deliberately built shared APIs and best practices that prevented fragmentation; in today’s agentic AI world, by contrast, rapid model and chip churn means teams are forced to rewrite systems every few months with no equivalent investment in shared standards. She says she’d like to see frontier and open-weight models coexist within composable, flexible frameworks rather than each vendor building closed, bespoke stacks — an aspiration she presents as a hope, not a prediction.
03 Engineering identity in the agent era
Goldberg discusses an identity crisis she sees among software engineers now that much code is agent-generated: engineers feel less ownership and recognition for work when agents produce it, and worry about ideas being copied once agentized. She argues the professionals thriving most are domain experts — critical thinkers who can communicate well through prompts because of deep subject knowledge — but notes a looming problem: expertise is traditionally built through roughly 10,000 hours of hands-on repetition, a pathway increasingly bypassed as agents take over execution. She credits a CoreWeave inference-team engineer, Rita, with emphasizing that teams need to stay curious rather than lazy to keep developing expertise. Host Alex Volkov (who said on tape he joined CoreWeave via the Weights & Biases acquisition) shares his own view that he once expected to become redundant as an engineer but now believes expertise is more valuable than ever, since most people are effectively managing fleets of agents like a hybrid product-manager-and-engineering-manager role.
04 Infrastructure proof points and partnerships
Goldberg points to Caterpillar’s Physical AI VP, Brandon Hootman, as an example of enterprises choosing to build AI capability in-house on CoreWeave after an earlier attempt to partner externally failed, leveraging Caterpillar’s machine and site data. Volkov says CoreWeave was the first company to deploy Vera Rubin NVL72, citing a claimed 4.8x improvement in token throughput over Blackwell, and calls out Cognition’s Devin coding agent; Goldberg adds that the ecosystem grows when teams use tools others build rather than everyone building everything themselves.
05 Sentiment on AI and personal agent use
Both speakers discuss a broader shift in public and enterprise sentiment, with Goldberg noting that organizational demand for AI tools is now driven bottom-up by teams eager to experiment rather than top-down mandates, across customers like Caterpillar, MasterClass, and Canva. She acknowledges public concerns about security, data sharing, and job impact as valid, but frames platforms like CoreWeave’s as helping make AI adoption secure and accessible. On personal use, Goldberg jokes about running multiple AI assistants and expresses hope for future consolidation and efficiency rather than everyone building redundant personal agent stacks, noting that people increasingly become the limiting factor for their own agents rather than the other way around.
THE SIGNAL
Jones says he no longer holds the hyperscaler and AI tech stocks he was buying last May; his plan is near-term weakness, a year-end rally possibly helped by Anthropic’s listing, then likely being short for all of 2027.
Crowding-out framework: heavy hyperscaler debt issuance already pushed down corporate bond prices, and he sees the same supply now moving into equities through SpaceX, OpenAI, Anthropic and later unlocks, which he compares to 2000-02.
He puts 60/40 odds on a sovereign-debt breaking point; AI build-out debt near 3.6% of GDP would be 50% above railroads and triple highways, which is why he thinks this time differs from past can-kicking.
Oil long rationale: a gentle Iran fix would have happened before the midterms, so he sees a tougher post-midterm resolution hanging over oil, which supports upside (our read: a hedge against his equity-bearish calls).
01 The 1999 analogy and market positioning
Jones, speaking live to CNBC’s Andrew Ross Sorkin at Good Shepherd Services, reiterates a view he first offered a year earlier: the market still feels like 1999, bifurcated between companies embracing AI and those lagging. At a recent conference of roughly 200 people, Jones asked how many had been trading in 1999; only about 20% had been, underscoring how few current market participants have lived through a comparable cycle. He says he can’t see professionally much beyond December 31, so his near-term playbook is tactical: short-term weakness, strength into year-end tied to midterm-cycle seasonality (markets typically bottom late October/early November, then rally into Q4), followed by a much tougher 2027.