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, SHOP 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 on the GDPx Benchmark, Why Frontier-Model Dollars Will Pool in a Few Hard Domains and Musk’s Compute Squeeze
TBPN, with SemiAnalysis founder Dylan Patel and his colleagues Max Kan and Andrew MegalaaShopify President Harley Finkelstein on the Q3 Sell-Off, Fast-Growing AI Shopping Channels and Why Agents Help Shopify’s Merchants
Squawk on the Street (CNBC), with Sara Eisen and Carl Quintanilla, and Shopify President Harley FinkelsteinSam Altman on the Post-GPT-4 Compute Bet, Launching ChatGPT Over His Team’s Doubts and Why Skipping Equity Was a Mistake
Figuring Out, with Raj Shamani, and Sam Altman, CEO of OpenAIDisruptive’s Alex Davis on His $10 Billion Fund, Running Groq After the NVIDIA Deal and the Coming NeoCloud Shakeout
Sources, with Alex Heath, and Alex Davis, founder of Disruptive and executive chairman of GroqVincent Tessier and Warren Woodward on Google’s Ad Network Growing Again, Meta’s 23% U.S. Price Jump and When Chatbot Ads Arrive
Mobile Signal, with Josh Chandley, Vincent Tessier of 52 Entertainment and Warren Woodward of Uptick
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 Dylan Patel on Ethmos
THE SIGNAL
Kan says xAI sold Anthropic short-term capacity this year at $40B+ per gigawatt versus the $12-15B of typical five-year contracts; Patel says Musk pulled compute from xAI’s own team to sell to Anthropic and Google.
Patel cites spot GPUs at about $15/hour versus roughly $4/hour TCO, and Musk’s compute funding at ~$60B cash plus a $40B raise against a $400-500B, 13GW plan (our read: resale margins are essential).
Patel says safety classifiers leave OpenAI and Anthropic behind Google, Meta and Chinese models on cyber and bio tasks, risking share in the very niches GDPx favors (our read).
Patel’s downside case: if hard-science and adversarial domains fail to monetize, existing tasks saturate, white-collar diffusion stays slow, and labs may miss their compute commitments, a key AI capex risk.
01 GDPx vs GDPval
Max Kan introduces GDPx as SemiAnalysis’s answer to what the team sees as a flawed approach in OpenAI’s GDPval, which selects evaluation domains based on their share of GDP or occupational taxonomies like SOC/ONet. Kan argues this methodology misses where AI revenue will actually concentrate: domains that can adopt new technology quickly and have high enough skill ceilings to extract continued ROI from ever-more-capable models. GDPx instead targets a narrower set — chip design, biology, cybersecurity, financial services, material science, and recursive self-improvement (RSI) — as the areas SemiAnalysis believes will absorb the bulk of frontier-model dollar spend, even if not the bulk of raw token volume.
02 Why diffusion will be slow
Dylan Patel lays out a thesis that broad white-collar AI adoption will lag expectations because companies face structural inertia — compensation structures, job definitions, and organizational habits don’t change quickly, even as model capability does. He contrasts this with domains where intelligence can be absorbed almost without limit: scientific fields far from the human frontier (material science, chip design) and adversarial fields where outcompeting rivals (finance, cybersecurity) creates continuous demand for more capable models. Patel frames this explicitly as SemiAnalysis’s house view, distinguishing it from the more commonly assumed broad-diffusion narrative, while noting the team is bullish that OpenAI, Anthropic and the other leading labs could deliver a trillion dollars of revenue within a couple of years — a timeline he jokes Kan and Megalaa will call bearish.
03 The engineering bottleneck in hard science
Pressed by a host on whether material-science spending could realistically scale given long R&D lag times, Patel argues that the dominant value pool is engineering, not discovery — citing Periodic Labs as a company focused on incrementally improving processes like material deposition rather than inventing new compounds outright. He uses concrete examples to illustrate scale: CPU substrate yields currently run 40-60% and could approach 95%; Intel’s wafer production generates roughly 100 terabytes of data per wafer that mostly goes unused because much of the scientific workforce still relies on Excel and MATLAB; and NVIDIA’s roughly two-and-a-half-year design-to-shipment cycle could compress toward a year with better simulation tooling, potentially accelerating adoption of technologies like co-packaged optics. Patel frames the primary constraint as siloed, slow-moving experimental loops rather than raw ideas or discovery itself.
04 Secrecy, benchmark results, and no human baseline
In response to a host’s question about why companies improving manufacturing efficiency with AI don’t publicize it the way software “hustler culture” does, Patel contrasts Silicon Valley’s openness with industrial firms that quietly improve products and let financial results speak for themselves. On actual results, Andrew Megalaa reports that SemiAnalysis’s PhD collaborators were surprised by how far model capability in material science had progressed, with Astra beating a published state-of-the-art benchmark result on a subtask — though the model’s overall GDPx material-science score was just 28%. Asked why GDPx omits human baselines, Megalaa explains that some tasks (like certain chip-design problems) represent work that would take teams of engineers rather than a single person, making a meaningful human comparison impractical; Patel adds that the goal is measuring when models become superhuman, not human-equivalent, since only superhuman performance will command sustained demand.
05 Musk’s compute strategy and the Goldman headline
Asked about Elon Musk’s post that Grok Bot is too expensive to run, Kan argues serving costs for personal AI agents could fall roughly 100x within a year, following the broader trend of declining intelligence costs, making current subsidization economically rational. Patel frames Musk’s broader strategy as turning SpaceX into a neocloud to fund an ambitious 13-gigawatt buildout (an estimated $400-500B), partly financed by reselling GPU capacity at a steep markup over true cost. On a headline a host said he couldn’t confirm, that Goldman is banning internal use of Claude Code and Codex, Megalaa notes that even frontier models score only about 50% on SemiAnalysis’s finance-analyst benchmark task, making open or on-prem alternatives even weaker by comparison; Patel calls such a ban shortsighted, predicting Goldman will still rely heavily on frontier tools for functions like penetration testing given security stakes.
Listen · Watch · Follow SHOP on Ethmos
THE SIGNAL
74% of Q3 AI-driven purchases came from outside Shopify’s top 100 categories, showing agentic traffic leans toward the long tail, where Shopify’s fragmented merchant base is overrepresented compared with head-category catalogs (our read).
AI orders grew 3.3x while AI traffic grew 2.6x, implying better conversion per AI visit; Finkelstein expects this holiday season to mark AI moving from mainly discovery last year to actual conversion, so watch Q4 AI order mix.
Finkelstein says Shopify does not know which agent will win, so it is positioning as the neutral identity and checkout layer: Shop Pay credentials, purchase history, tracking and the Shopify catalog feeding merit-based results inside every agent.
Revenue grew 32% versus 29.5% GMV growth, so monetization per GMV dollar is still rising even as GMV slowed (our read); the miss came against a low-thirties guide that management says it hit.
01 Q3 numbers and the market’s reaction
Finkelstein opens by restating headline figures: GMV of $119 billion (up 29.5%), revenue of $3.7 billion (up 32%), and a free cash flow margin of 22%, which he calls rare at Shopify’s scale. He notes the company hit its own guide of low-thirties revenue growth and that big brands such as MAC, Clinique, Lululemon and Balmain joined Shopify (he later adds Estée Lauder). He also cites that since the start of 2025, Shopify has captured nearly half of all incremental US ecommerce dollars while representing only 14% of the market. Despite this, shares fell roughly 9% on what CNBC calls the company’s first revenue miss since 2022; Finkelstein dismisses the reaction as noise against what he calls a hundred-year-company scoreboard, noting it was his 46th earnings call since IPO.
02 AI channels: real but still small
Pressed by Sara Eisen on how much growth is attributable to AI commerce partnerships with OpenAI, Google, and Meta, Finkelstein says AI channels are becoming real discovery engines but remain a small fraction of total volume, with traffic up 2.6x and orders up 3.3x. He highlights that specificity is rewarded in this agentic environment — niche products like robot vacuums for pet hair or three-child car seats disproportionately benefit Shopify’s long-tail merchant base. He contrasts this with the much larger, already-scaled parts of the business: offline point-of-sale GMV just crossed $200 billion, and international is up 30%, framing Shopify as simultaneously building for an AI future while strengthening its current core.