Nvidia's $60B quarter, two labs taking half of new compute, and a 6% death-spiral warning
The week after a sales-process issue, the tape snaps back to power and money: SemiAnalysis's Dylan Patel says Anthropic and OpenAI already outbid the world for watts, and four days later All-In reads Nvidia's most-profitable-ever print as the AI-capex-bubble story dying — while Chamath puts a 6% long bond as the start of years of fiscal pain.
Aug 24 – Aug 30, 20262 episodes2 shows50 claims on the record20 new predictions
01
The print and the grab: Nvidia's quarter, two labs' watts
Patel’s mechanism is a margin flip: Anthropic now clears as much as $50M of revenue per megawatt against a $10-15M cost, and recycles the spread into training. Signed contracts, he says, put Anthropic and OpenAI at 40-50% of all new compute next year — half the incremental watts by the end of 2027, most of the world’s usable flops by the end of 2028. Four days later Sacks treats Nvidia’s guide as the same tape from the other side: the AI-capex-is-a-bubble story is being shredded, and the boom has legs. Chamath reads Jensen’s reported Hugging Face and Poolside deals as an open-source stack grab after Altman poked him with AMD. One week, two layers: the labs taking the watts, the chip vendor printing the cash.
“In the case of Anthropic, the revenue has gone as high as 50 million dollars per megawatt.”
The SaaS apocalypse gets shredded — keep the system of record
Salesforce ripping ~20% after a beat is the All-In panel’s evidence that ‘SaaS is dead’ was a simple extrapolation: agents can code, therefore software goes to zero. Sacks had been saying core CRM will not be vibe-coded away; Chamath, who called the stock oversold in May, now frames three AI phases — models, harnesses, then context-trained agents — and says the horizontal systems of record that play with those models sit in a special place. Friedberg tried to rebuild CRM over a weekend and bought Salesforce. The product implication Sacks draws from Benioff’s Claude deal: stop defending the UI, build the agent interface, and let the model own the front end while you remain the source of truth.
“Well, this narrative of the SAS apocalypse was totally overdone.”
Patel models about $11 trillion of CapEx through 2029, roughly $5T of it debt, and Dwarkesh takes the crowding-out to sovereigns — a second Volcker shock. All-In’s fiscal hour is the domestic version of the same constraint. Friedberg walks a 30-year around 5.2% and a $10 trillion refinance in twelve months and says Bessent cannot buy the long end into submission; Chamath names the buybacks yield-curve suppression and puts 6% as the start of a years-long death spiral only Congress can stop. Sacks calls spending a tragedy of the commons, then offers the escape hatch the Nvidia print is supposed to fund: only AI creates the exponential growth that grows the country out of the hole.
“In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029”
Patel's non-consensus call: labs will allocate less compute to inference over time and more to research — he says they are already doing it. DWARKESH →
Friedberg: Moderna’s neo-antigen ‘vaccine’ is immunotherapy priced at $500,000 after public-funded research clinics already run for ~$50,000. ALL-IN →
China AI-optimism poll cited on All-In: over 80% vs about 30% in the US — Sacks’s gap that can make America shoot itself in the foot. ALL-IN →
Sacks on Ukraine: both sides denied the Ratcliffe-NATO-warning leak; Zelenskyy ‘doesn’t have the cards’ and should take a deal. ALL-IN →
Jason treats an AI-written Druckenmiller WSJ op-ed as Adele lip-syncing; Chamath says you’d have to be dumb not to use the tool. ALL-IN →
SpaceX sold a gigawatt of compute to Google for about $40 billion — Patel still expects most watts under $20B/GW by the end of next year. DWARKESH →
The record
Every episode this week
The issue above is written from these structured records — nothing else. Each card is the full breakdown: claims, predictions, quotes, all timestamped.
20 falsifiable calls entered the ledger this week — the most recent below, all of them filterable in the Prediction Ledger.
David Friedberg · ALL-IN · over the next couple of years · source 1:33:13
Personalized cancer immunotherapy proliferates overseas at much lower cost over the next couple of years via right-to-try and cheap clinic processes.
David Sacks · ALL-IN · in five years · source 33:12
In five years the large tech companies will each have their own cloud, models, silicon, and data centers — soup to nuts.
David Sacks · ALL-IN · well into the future / next year+ · source 30:21
AI capex continues well into the future rather than ending soon; Nvidia's guide is the tell.
David Friedberg · ALL-IN · 2030-2032 · source 51:26
Between 2030 and 2032 Social Security cannot pay and multiple states go bankrupt, forcing Social Security and state bailouts/restructurings.
Chamath Palihapitiya · ALL-IN · if/when the long bond hits 6% · source 44:37
If the long-term Treasury yield prints 6%, that is the beginning of a years-long fiscal death spiral unless Congress gets the budget under control.
Dwarkesh Patel · DWARKESH · by end of the decade · source 1:07:36
By the end of this decade, a single frontier lab holds more effective AI labor than there are people on earth.
First mentions
New to the graph this week
Subjects no indexed episode had discussed before this week.
AI compute×12US national debt×5AI capex×4AI value capture×4SaaS apocalypse×4Interest rates×3Stanley Druckenmiller×3US Treasury×3Congress×3Moderna×3neo-antigen immunotherapy×3ASML×2Jane Street×2Scott Bessent×2Volodymyr Zelenskyy×2mRNA×2
How it's built
Written at ingest, never at query time
The week's episodes are diarized into structured records; one editorial pass writes this issue from those records when the week's ingest lands. A grounding gate greps every quote above against its episode record — a quote that isn't verbatim fails the build.
the data behind this issue
# this issue, as structured data — stories, citations, receipts
curl 'https://aipod-two.vercel.app/digest/2026-w35.json'# the records it was written from
curl 'https://aipod-two.vercel.app/api/episode/<id>'# any id in episode_ids