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 AMD, a16z, SNDK or Anthropic 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:
AMD CEO Lisa Su on Demand Outrunning Supply, Over-80% Growth Next Year and the $1 Trillion Market After the Election
Squawk on the Street (CNBC), with Jim Cramer, David Faber and Carl Quintanilla, and AMD chair and CEO Lisa Sua16z’s Martin Casado on Frontier-Model Economics, Why Capital Is AI’s New Scaling Lever and the Cursor Deal
Invest Like the Best, with Patrick O’Shaughnessy, and Martin Casado, General Partner at Andreessen HorowitzDan Ives on Year Three of the AI Build-Out, Its 5-6x Capex Multiplier and Strong Appetite for Anthropic’s IPO
Pitch The PM, with Doug Garber, and Dan Ives, tech analyst at Yorkville IvesAsianometry’s Jon and Tim Culpan on High-Bandwidth Flash: More Capacity Than HBM, an Endurance Problem and a Fit for KV Cache
Supply Chained, with Jon of Asianometry and Tim Culpan of CulpiumBenedict Evans and Toni Cowan-Brown on Anthropic’s S-1: Exploding Revenue, Thin Profits and Commitments That Act Like Leverage
Another Podcast, with Benedict Evans and Toni Cowan-Brown
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 AMD on Ethmos
THE SIGNAL
Su said demand exceeds supply and declined to size unconstrained upside, so the over-80% 2027 figure is supply-capped (our read).
Server CPU is a second AI growth leg: Su dates real agentic workloads to the past three to six months, with customers asking monthly for more CPU over 12-18 months (our read: Venice upside).
Data-center revenue rising 50% sequentially costs only about one point of gross margin (56% to 55%), implying Helios/MI455 is less dilutive than feared (our read); Su sees margins holding in 2027 plus opex leverage.
01 Results and demand acceleration
Su says AMD grew over 40% year over year in a third quarter that beat its own expectations, with data center extremely strong, and says the company is seeing even stronger demand looking forward. She guides data-center revenue to grow 50% sequentially into Q4, and frames this moment as an early stage of an adoption cycle for AI and what she terms super intelligence, requiring CPUs, GPUs and accelerators together — AMD’s specialty, in her telling.
02 Raising the long-term target
Asked by Jim Cramer to revise the 60% data-center compound annual growth rate AMD gave at its November 2025 Analyst Day, Su describes very strong data-center growth this year and next, and says AMD’s business will grow over 80% next year. She attributes the upgrade to accelerating AI demand, customers validating AMD’s newest products, the Helios platform launch in Q3, and a surge in agentic workloads over recent weeks, citing Meta’s Muse product adoption as a concrete signal. She says the majority of 2027 volume is already locked in via contracts with both customers and supply-chain partners.
03 Supply chain as the real constraint
Pressed by Faber on where the tightest bottleneck lies, Su declines to name a single constraint, instead describing an end-to-end chain — wafers, substrates, advanced packaging, memory, data centers, land, power and shell — that must all come together, and says AMD is working each link rather than facing one chokepoint. She says AMD plans its roadmap and data-center build-out three to five years in advance with partners including OpenAI, Anthropic, Microsoft, Meta and Oracle, and notes AMD has already landed first racks of its newest systems at all of these customers.
04 Competitive position vs Intel and NVIDIA
Su says she favors AMD’s competitive position against both rivals. On CPUs, she points to the new Venice server chip, just shipping in Q3, describing it as the best CPU in the industry based on customer demand for performance and cost efficiency. On AI accelerators, she highlights Helios and the MI455 platform, claiming leadership in memory bandwidth, and argues AI itself is accelerating customer adoption of new hardware. Cramer calls Intel a loser in both PCs and data-center racks; Su answers only that she likes AMD’s competitive position.
05 Election risk and margins
On the midterm election and local opposition to data-center siting, Su argues the industry needs clearer dialogue with communities about AI’s value proposition, but maintains AMD has strong diligence and visibility into where its forecasted capacity is being built, and reaffirms the $1 trillion data-center TAM by 2030 is unchanged. She does not directly address Cramer’s follow-up on whether debt-funded builders like Hut 8, CoreWeave and IREN can still build, instead reiterating that AMD partners with large, well-resourced frontier labs. On margins, Su notes gross margin exited Q3 at 56% and is guided to 55% in Q4 as the AI mix grows faster than CPUs, and expects margins to hold steady into 2027 aided by operating-expense leverage.
Listen · Watch · Follow a16z on Ethmos
THE SIGNAL
Casado says model charts in pitches show a frontier model’s commercial life is just 3-6 months, so one slipped release could cost a lab the 80%+ dollar-weighted share leaders have held for three years.
Casado says coding may be near saturation, where users have automated ~95% and need no more tokens, but that doesn’t mean cheaper models: competing startups will always pay 10x more for the smartest one.
Casado says Snowflake-type data infrastructure growing 30-40% faces no real AI disruption, because operating it takes a long tail of features and performance tuning, and more vibe-coded software lifts infra demand.
Casado says AI cluster networking has new entrants for the first time since Arista, with changes in build materials and a shift toward optical (our read: watch optical suppliers).
01 Investor mindsets and moats
Casado outlines four investing archetypes he says are routinely conflated: financial investors who proxy business health through SaaS metrics and market size; productive investors (his own camp) who diligence technical approach, adjacent markets, and value-capture speed; frontier investors, a 1960s-70s-style bet purely on founders and talent (he cites Chris Dixon’s bets on Nootropics, Soylent, Oculus, and Coinbase); and speculative investors who trade assets without caring about underlying business metrics, common in crypto and secondaries. He warns that without clarity on which archetype you are, you can end up justifying any deal, or none, based purely on personal bias.
On network effects, Casado says the term is used inconsistently — some mean Metcalfe’s-law super-linear growth (true network effects, historically rare, exemplified by telcos, LinkedIn, and GitHub), others mean mere product improvement with scale, and others mean self-reinforcing “flywheels” without knowing if they’re actually super-linear. He argues the real question is whether a moat’s growth is super-linear, linear, or sub-linear relative to competitor spending — sub-linear moats let challengers catch up regardless of brand or mechanism. He adds a stronger claim: in growing markets, defensibility debates are largely moot because companies find new white space rather than fight head-on; this, he says, is why years of investor skepticism toward models and Neo Clouds as undefensible missed how much value they’d still accrue.
02 Infrastructure under maximum disruption
Casado frames infrastructure investing around five classic layers — compute, network, storage, databases, and (newly) models, which he says behave more like ASICs than software given their capital intensity and competitive pressure. He says agentic software is forcing a rebuild of sleepy categories: OLTP databases are being reinvented because agents don’t need human-friendly query languages; networking (his PhD field) is seeing its first new entrants since Arista due to massive, rigid-performance AI clusters; and compute abstractions (containers vs. VMs vs. durable workflows like Temporal) are being rethought for agent workloads.