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, NET 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:
Amazon CEO Andy Jassy on $220 Billion of Capex, the AI Demand Amazon Still Can’t Serve and Why Inference Must Get Far Cheaper
The a16z Show, with Ben Horowitz, Erik Torenberg and Andy Jassy, president and CEO of AmazonRamp Economist Ara Kharazian on Who Actually Pays for AI and Why Spending Growth at the Top Is Slowing
Plain English with Derek Thompson, with Ara Kharazian, lead economist at RampBen Thompson and Andrew Sharp on Netflix’s Growth Problem, the Internet’s Trust Crisis and Anthropic’s Alignment Philosophy
Sharp Tech, with Andrew Sharp and Ben Thompson of StratecheryCloudflare CEO Matthew Prince on the PHP2Shell Flaw, Why AI Helps Defenders for Now and Agent Traffic at 1,000x Humans
TBPN, with John Coogan, Jordi Hays and Matthew Prince, co-founder and CEO of CloudflareSemiAnalysis on ASML’s Hybrid-Bonding Tool, Why HBM Waits Until 2029 and China’s Packaging Push
SemiAnalysis Weekly, with Jordan Nanos and SemiAnalysis analysts Sherry Zhang and HowieOpenAI CFO Sarah Friar on 1.2 Billion Weekly Users, the Model OpenAI Is Holding Back and Going Live on Google Cloud
Pioneers of AI, with Reid Hoffman and Sarah Friar, CFO of OpenAI, live at the Masters of Scale Summit
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
Jassy says AWS capacity is sold out for 2026 and largely sold out for 2027, apart from on-demand reserves, with a good chunk of 2028 already booked, giving multi-year visibility behind the $220B capex.
AWS’s AI business passed a $25B run rate within three years of the generative AI wave, versus about $60M for AWS after its first three years; AWS grew 36.7% last quarter on a $169B run rate.
Core non-AI AWS is also growing fast, forcing a split of scarce memory between HBM for accelerators and DRAM for core servers.
Jassy calls inference still an order of magnitude too expensive for pervasive enterprise use and expects enterprises’ existing production workloads to become the largest absolute segment of AI usage as costs fall, though he won’t predict how fast; cheaper inference favors Trainium (our read).
01 The capex calculus
Jassy explains Amazon’s investment logic splits into two cash cycles: data centers require commitment roughly two years ahead of monetization but carry ~30-year useful lives, while servers and networking need only months of lead time, break even in about three years, and run five-to-six-year useful lives on mostly five-year deals. This visibility into customer demand, he says, lets Amazon scale aggressively rather than guess — though the pattern creates near-term free-cash-flow headwinds while data centers ramp. He argues Amazon could justify investing considerably beyond $220B given unmet demand, though he expects supply constraints in power, chips, and memory to persist for several years, not just through 2027.
02 Pushing back on data-center backlash
Jassy disputes several claims behind political opposition to data centers, arguing Amazon’s facilities use less water annually than golf courses consume in two days, and that backup generators run only a tiny fraction of the year — around 0.1% — for maintenance. He compares the current buildout in scale to the U.S. buildout of highways and bridges in the 1950s, and attributes rising energy rates mainly to underinvestment in grid infrastructure rather than to data centers themselves. He notes Amazon has paid roughly $3 billion in taxes in St. Joseph County, Indiana, and has committed another $1 billion to local communities over five years through its Built Together program.
03 Chips, Anthropic and OpenAI
Jassy traces Amazon’s chip strategy to the 2015 Annapurna Labs acquisition, originally meant to offload virtualization and networking tasks from EC2 servers, which later produced Graviton CPUs — used by 99% of Amazon’s top 1,000 EC2 customers, with about 40% better price-performance than x86 — and Trainium AI accelerators, offering roughly 30-40% better price-performance than comparable processors. He frames this as a response to depending on a single dominant chip partner (following an earlier dynamic with Intel) whose incentives didn’t prioritize price-performance. Amazon’s investments in Anthropic and OpenAI (roughly $18B and $50B, a host’s estimate) are described as extensions of deep Bedrock partnerships: Anthropic has long built on Trainium, while the OpenAI deal let shared customers finally run OpenAI models on AWS via a new stateful runtime environment Jassy expects to underpin how agents are built going forward.
04 Reorganizing for agentic AI
Jassy points to internal examples — a host recalled his figure that six engineers rebuilt Bedrock’s inference engine in 76 days, work that would have taken 40 people a year, and he describes four or five people taking a new prototype to 80% shippable over a single weekend — as evidence organizations should shrink and de-specialize. He predicts companies will employ fewer people than they do today for current tasks, with smaller cross-functional teams, and near-total rewrites of existing codebases so agents can patch vulnerabilities within minutes rather than through the current multi-step human patch cycle. He singles out cybersecurity, not exotic AI-takeover scenarios, as the more immediate risk companies should focus on.
05 Reinventing products, advice to young people
On Alexa, Jassy says generative AI made it clear the underlying architecture needed a full rebuild, despite internal resistance from teams attached to prior design choices, and says Amazon is applying similar from-scratch reinvention thinking to retail, Prime Video, and Fire TV, even though the need there is less obvious. Asked for advice to an 18-year-old today, he says AI use in education is now unavoidable and universities shouldn’t fight it, but stresses that human judgment, critical thinking, and the work of synthesizing information into a strategy remain irreplaceable skills. He closes by linking Amazon’s sprawl across AI, robotics, autonomous ride-hailing (Zoox) and space (Amazon LEO, targeting 500 million underserved households) to a decentralized, single-threaded leadership structure that lets differentiated teams pursue multiple frontier bets simultaneously.
THE SIGNAL
Ramp’s per-employee monthly AI spend tiers: median buyer about $12-15, top decile about $800, top 1% over $10,000; the top cohort spends roughly 800x the median per employee and is now actively cutting costs.
Kharazian agrees with the Jevons argument: price wars and spreading spend across providers squeeze model-maker profits but still need more compute, so value goes to NVIDIA/infrastructure and end users rather than OpenAI/Anthropic.
Thompson’s hurdle: model spend has to go from about $200B this year to about $400B, then about $800B; slowing top-1% growth, slower customer additions and falling prices all threaten that doubling pace.
Fable lost to GPT Sol, which was slightly cheaper with little performance loss, because DoD-required prompt logging conflicted with enterprise data-retention duties; compliance friction plus near-substitutes cap frontier premium pricing (our read).
01 Framing the bubble as supply vs demand
Thompson opens by distinguishing the AI bubble’s supply side (chips, data centers, hyperscaler capex now partly debt-funded) from its demand side (actual revenue AI models generate). He argues the supply side looks less alarming than past bubbles because the buyers — Amazon, Microsoft, Google, Meta — are already enormously profitable (unlike thinly capitalized 19th-century railroads), and because nearly every AI chip today is in active, premium-priced use, unlike the vast majority of fiber left dark after the dot-com bust or the quarter of U.S. railroad mileage in bankruptcy by the 1890s. The demand side, by contrast, is largely unverifiable: OpenAI and Anthropic are private, so the public relies on leaks — one report pegged OpenAI’s revenue pace near $70B annually, another (FT) at closer to $50B, a large discrepancy that still moved markets. This opacity is why Thompson frames demand, not supply, as the central question, and why he brought on Kharazian, whose firm Ramp processes over $275 billion in annual business spend across 85,000+ companies, including over $10 billion in AI spend.
02 Revenue history and the Claude Code inflection
Kharazian traces AI enterprise revenue from ChatGPT’s consumer launch, through OpenAI’s early enterprise inroads, to Anthropic’s sharp share gains driven by Claude Code, a coding agent he says effectively took over or heavily augmented software engineers’ jobs and became a major revenue driver that spurred further infrastructure investment. More recently, he says competition has intensified — from SpaceX AI, NVIDIA’s own models, and overseas open-weight competitors — squeezing margins for the American labs.