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 AMZN, NVDA, INTC, CRWV or OpenAI 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:
AWS CEO Matt Garman on a $1 Trillion AWS, Inference Under 1% of Its Eventual Size and Years of GPU Shortages
Sources with Alex Heath, with AWS CEO Matt GarmanCFR’s Chris McGuire on the Jensen Huang Scandal, the Megaspeed Enforcement Gap and $80 Billion of Chips Into Southeast Asia
ChinaTalk, with Jordan Schneider and Chris McGuire, senior fellow for China and emerging technologies at the Council on Foreign RelationsVik Sekar and Austin Lyons on Intel’s Worst Day Since 1971, a 20% PC Slump and the Substrate Squeeze
Semi Doped, with Vik Sekar and Austin LyonsCoreWeave’s Michael Terlizzi on 1.5 Gigawatts Built, an 8-Gigawatt 2030 Goal and Power as the Binding Constraint
Cool Vector, with David Snow, CoreWeave SVP of Data Center Operations Michael Terlizzi and Cushman & Wakefield Vice Chair Bo BondBen Bajarin and Jay Goldberg on the Semi-Apocalypse, Anthropic’s Compute Crunch and AMD’s Supply Edge
The Circuit, with Ben Bajarin and Jay GoldbergOpenAI Chief Economist Ronnie Chatterji on an 8x Frontier-Firm Gap, Agents Logging Triple Human Hours and the First Canaries for AI Job Losses
Macro MATTers on Podzept (Deutsche Bank Research), with Matthew Luzzetti, Matthew Raskin and OpenAI Chief Economist Ronnie Chatterji
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 AMZN on Ethmos
THE SIGNAL
Garman expects AWS to stay capacity-short for the next couple of years and says it could sell out every accelerator landing in 2027 today, so supply, not demand, likely caps near-term AWS growth (our read).
Garman says each 50% cut in inference cost has brought roughly 10x more usage, and he puts inference below 1% of its eventual size, which argues price cuts expand AWS AI revenue rather than shrink it.
Rationing is already happening: AWS sometimes pushes customers from a 2-week to a 4-week delivery, or from the East Coast to the West Coast or India, while landing one GPU every 14 seconds.
Most large-customer commitments run 3 to 6 years, some 8 to 10, and customers often spend above them; Garman also puts on-prem at about 85% of workloads, the market Trainium on-prem sales target.
01 The growth case for $1 trillion
Heath notes Andy Jassy’s framing moved from roughly $600 billion by around 2036 early this year to a trillion floated in Q2 and a firmer Q3 tone; Garman ties it to AWS’s current $182 billion run-rate business growing 39% year-over-year, with year-over-year growth accelerating for six consecutive quarters. He attributes this less to AI hype itself and more to a structural shift: companies that kept core workloads on-premise for two decades are now migrating them specifically so AI agents can access and act on that data, since agents struggle to operate against siloed mainframes or legacy data warehouses. He frames this as opening an entirely new pool of revenue and business creation for customers, rather than cloud share simply reallocating spend from one place to another.
02 Capacity, risk, and who gets the GPUs
On CapEx discipline, Garman describes a staged investment approach modeled on Amazon’s supply-chain roots: buying long-lead-time assets (land, power contracts, permits) years in advance while deferring purchases of servers and networking gear until closer to need, which limits downside if demand assumptions prove wrong since land can be resold. He says large customers often sign five-to-ten-year capacity commitments that give AWS visibility, and that demand still exceeds supply despite heavy spending. On allocation, Garman says AWS deliberately avoids selling 100% of future capacity to large labs, reserving blocks for enterprises needing small GPU counts and for startups that don’t exist yet, plus capacity set aside for its Bedrock service. Asked about systemic compute-spending risk tied to neo-clouds and data-center financing, Garman argues AWS’s exposure is structurally different: its customer base is broad and not predominantly high-rate debt-financed, and even if one model provider — Anthropic or OpenAI, both close AWS partners — stumbled, inference demand would simply shift to other models rather than disappear.
03 Inference, training, and open vs. closed models
Garman calls misleading a stat Heath cites, that only about 2% of Americans have paid for AI, as a sign the market is small, comparing it to how few people directly pay for database software despite relying on it daily through other products. He argues inference — not training — is the dominant and still-nascent AWS AI workload, embedded invisibly in applications like ad serving, trading, and product search. Training, by contrast, he describes as valuable but concentrated among a small number of customers and a minority of AWS’s business. On open versus closed weights, Garman speculates frontier labs may slow open-weights progress by locking down models against distillation (hiding reasoning tokens, encryption), though he expects other players, including in China, to keep building open models from first principles; he also notes OpenAI and Anthropic have begun releasing cheaper distilled versions of their own frontier models rather than ceding that market to open weights.
04 Nova reset, safety, and automated reasoning
Garman says Amazon remains just as committed to frontier in-house models despite recent Nova layoffs and pausing, characterizing it as a strategic reset to fix talent, data, and infrastructure approaches, not a retreat. On agent safety, he highlights automated reasoning — neuro-symbolic AI combining mathematical proofs with AI systems — as a major internal investment, citing AWS’s IAM Access Analyzer as a working example, with ambitions to let agents (e.g., processing insurance claims) operate under provably correct guardrails rather than needing human-in-the-loop review; he expects some such tools to ship next year.
05 Stack strategy, external chip sales, and data-center politics
Garman says AWS intends to compete at every layer of the cloud/AI stack — chips and networking, managed services, and end-user applications (Connect, Quick, Kiro) — while treating inference providers like Fireworks, Baseten, and Modal as largely complementary partners building atop AWS, akin to Snowflake or MongoDB. He confirms AWS is now selling Trainium chips for on-premise and sovereign-AI deployments, noting a large share of workloads remain on-prem globally and that NVIDIA already sells broadly beyond the major hyperscalers. Asked about the political backlash to data centers after the midterms and the Texas governor’s race, Garman argues AWS is unfairly lumped in with less responsible builders, pointing to AWS’s carbon-free energy investments, a 2030 water-positive commitment, and his statement that AWS’s largest data centers use less water than 40 US households, while acknowledging the industry needs better community engagement and codified standards with state governments. He dismisses space- and ocean-based data centers as impractical for a decade or two given launch costs, heat dissipation, and radiation, but sees desert locations as plausible given cooling and land advantages.
Listen · Follow NVDA on Ethmos
THE SIGNAL
Singapore, Malaysia and Thailand rank second, fourth and fifth among AI chip importers this year, together taking about $80B; McGuire suspects much is China-bound, making this demand the most exposed to third-country licensing (our read).
McGuire says Megaspeed got about $4.8B of chips and is still buying today, and Schneider notes it is still listed as an official NVIDIA cloud partner, a live exposure if Commerce or DOJ acts (our read).
McGuire expects the Match Act, which would cut off DUV tools and push broader equipment controls, to become law eventually, a risk to toolmakers’ China sales (our read); Overwatch would make no-Blackwell-to-China law.
McGuire says NVIDIA overstated Huawei’s chip output, so the scandal undercuts its main Washington argument for China sales while US buyers like Musk and Jassy remain chip-short.
01 Why the affair matters strategically
Host Jordan Schneider and guest Chris McGuire (CFR senior fellow, former NSC deputy senior director) frame the alleged relationship between Huang and a figure tied to Chinese intelligence as significant not because of infidelity but because of what Huang knows and controls. McGuire says Huang has personal, direct communications with President Trump on AI and export-control policy, and that NVIDIA has more visibility into global compute allocation than the U.S. government itself. Because demand for advanced chips vastly exceeds supply, Huang effectively arbitrates who receives them and when — making any compromise of him a potential channel for adversary access to both policy intelligence and allocation decisions.
02 The enforcement decision at the core
McGuire identifies the real inflection point as Commerce’s May 2025 choice not to enforce the Biden administration’s third-country license requirement, which he says was made at Huang’s request, with the administration promising a replacement rule that never materialized. The result, he says, is an unclear, selectively enforced patchwork of controls that Chinese buyers are exploiting through Southeast Asian intermediaries — a bigger practical loophole, in his view, than the long-running debate over H200 sales to China directly. Legal gaps compound this: an Inspur-linked U.S. subsidiary reportedly supplied Megaspeed despite Inspur’s entity-list status, illustrating how listed-company restrictions can be circumvented domestically.
03 Commerce’s non-response and the Kessler report
Both speakers question why Commerce allegedly held intelligence about Megaspeed’s leadership links to Chinese intelligence yet took no action to block the shipments or warn Huang. McGuire calls the absence of a defensive briefing a deliberate choice, speculating it could reflect a separate ongoing investigation or a desire to preserve intelligence sources, though he stresses this is speculation. McGuire cites a New York Times report that under secretary Kessler discussed Commerce’s Megaspeed investigation directly with NVIDIA — behavior McGuire calls a clear breach, comparable to FBI leadership briefing a suspect’s business partner — and he says DOJ then took the probe over from Commerce, disputing NVIDIA’s public claims that investigations had cleared Megaspeed of diversion.