"NVIDIA's $96.2B Quarter, China's 200,000 Fake Accounts, & OpenAI's New Chip" — Moonshots

Full transcript: [[2026-08-29-moonshots-ep284-nvidia-96b-quarter-openai-chip-transcript]].

Why this is in the vault

Direct sequel to EP #283 (2026-08-27) with hard numbers on the AI-infra circular-financing question RDCO tracks (Nvidia $96.2B quarter, OpenAI's in-house inference chip) plus a rare skeptical-panel moment on whether frontier-lab demand is being artificially inflated — useful counter-signal against the vault's generally bullish AI-infra-capex framing.

Episode summary

The regular panel (Diamandis, Alex Wissner-Gross, Dave Blundin, Salim Ismail; Emad Mostaque absent) opens on Nvidia's $96.2B quarterly revenue and debates whether the AI economy's circularity (Nvidia financing the data centers that buy Nvidia chips) is healthy income-statement circularity or riskier balance-sheet/private-credit-financed circularity. From there the episode moves through China's video-model dominance vs. US LLM focus, Flock Safety's license-plate surveillance network, a bot-farm influence operation targeting US data-center sentiment, OpenAI's new in-house inference chip ("Jalapeno"), labor-market survey data undercutting AI-job-loss narratives, and a pancreatic cancer drug breakthrough.

Key arguments / segments

[00:05:00] Nvidia booked $96.2B revenue in one quarter, up 106% YoY, guiding to $108B next quarter; Jensen's 2028 guidance of 70% growth vs. Wall Street's 44% consensus.
[00:09:00] Alex Wissner-Gross flags the "elephant in the room": how much of Nvidia's demand is Nvidia-financed (vendor financing of its own customers), distinguishing healthier income-statement circularity from riskier balance-sheet/private-credit circularity — he wants "more clarity" on which is at play.
[00:14:00] Elon Musk's SpaceX revenue projection ($3.5T by 2033) debated; panel frames SpaceX as a multi-layer bet (chips, data centers, models, comms, launch, robotics) rather than a single-layer model bet.
[00:25:00] China's Alibaba Qwen/"Juan" 3.0 video model and a claim that video generation is ~70% of China's AI token consumption, framed as "America is LLM-pilled, China is world-model-pilled" — Wissner-Gross's counter-theory: US labs are revenue-maxing on enterprise/code use cases (Anthropic), China is not revenue-maxing and is giving away model weights, so spare compute goes to video.
[00:35:00] China moving to restrict AI companion apps for minors, citing demographic collapse (China ~1 child/family, South Korea 0.75); panel debates whether this is genuine child-protection policy or (per Wissner-Gross's prediction) a CCP pivot toward state-controlled AI companions once party utility is recognized.
[00:45:00] Flock Safety segment: AI license-plate-reading camera network across 6,000+ US communities, 49 states, 20B+ scans/month — panel discusses "find the suspect" pattern-matching shifting to "find people who behave like suspects" as the cost of mass surveillance drops to near zero.
[01:10:00] Principal Financial Group survey of small/mid-size businesses (130,000 customers): only 4% expect AI to reduce staffing/wages, 31% expect increased staffing and wages, and of companies that did cut staff, only 1.4% attributed it to AI/automation — framed as directly countering "AI causes mass unemployment" narratives.
[01:20:00] X's Global Affairs team disclosed a ~200,000-account Chinese bot farm; 200 of those accounts pushed narratives that AI data centers are driving up household electricity prices, aimed at suppressing US data-center buildout sentiment.
[01:26:00] OpenAI's new inference chip "Jalapeno" (built with Broadcom): claims 1.5-1.9x more AI work per watt and up to 3.6x lower end-to-end latency vs. Nvidia's GB200/GB300, running at 700W vs. GB300's 1400W. Panel frames this as OpenAI becoming a chip designer/competitor to Nvidia in inference, not training.
[01:55:00] FDA approved daraxonrasib, first RAS inhibitor for metastatic pancreatic adenocarcinoma — nearly doubled median survival (6.7 to 13.2 months) and tripled tumor response rate (11% to 32%); RAS mutations drive ~30% of all human cancers.

Notable claims

Nvidia: $96.2B quarterly revenue, +106% YoY, next-quarter guidance $108B (~$1B+/day); Jensen's 2028 growth guidance of 70% vs. Street's 44% [00:05:00].
OpenAI's Jalapeno inference chip: 1.5-1.9x perf/watt, up to 3.6x lower latency, 700W vs. Nvidia GB300's 1400W; OpenAI claims ~54x throughput-per-user gain on open-weight models like GPT-OSS [01:26:00].
X disclosed a Chinese bot farm of ~200,000 accounts, with 200 identified pushing anti-data-center electricity-price narratives [01:20:00].
Small/mid-size business survey (Principal Financial Group, 130,000 customers): only 4% expect AI to cut staffing/wages; 1.4% of staff reductions attributed to AI [01:10:00].
Flock Safety: 6,000+ communities, 49 states, 20B+ license-plate scans/month [00:45:00].
Daraxonrasib: pancreatic cancer median survival 6.7 → 13.2 months, tumor response 11% → 32% [01:55:00].

Guests

No outside guests this episode. Regular panel: Peter Diamandis (host), Alex Wissner-Gross ("Alex"/AWG), Dave Blundin ("Dave"), Salim Ismail ("Sel"/"See"). Emad Mostaque sent regrets (travel).

Sponsorship

Mid-episode read for Google for Startups, pitching its "startup technical guide for generative media" as a blueprint for deploying Google DeepMind models (image/video/audio) in production [~00:24:00]. Also a recurring Fountain Life (full-body MRI / early cancer detection) segment ahead of the health-news block, which functions as a standing self-promotional partner slot rather than a one-off paid read — noted for completeness though the frontmatter sponsor_entity reflects the identifiable paid ad (Google for Startups).

Mapping against Ray Data Co

The Wissner-Gross circularity-skepticism moment [00:09:00] is the most useful data point here for RDCO's AI-infra thesis tracking: it's a rare instance of the Moonshots panel (generally bullish) explicitly flagging vendor-financing risk in Nvidia's demand chain, which is a live open question in RDCO's memory-cycle/capex investing thesis (see investing-edgar-watch capex pulses). The OpenAI Jalapeno chip news is a second concrete data point for the "inference commoditizing away from Nvidia while training stays Nvidia-locked" pattern already tracked from EP #282's frontier-training-pause coverage — worth cross-referencing against the next hyperscaler capex pulse. The 4%-job-loss survey data is a useful counter-citation against catastrophist AI-labor narratives that occasionally surface in Sanity Check research; it's Ray-adjacent evidence, not decision-grade (single-vendor survey, self-selected small-business sample).

[[2026-08-27-moonshots-ep283-sam-altman-singularity-slowdown]] — immediate prior episode (EP #283), direct sequel; shares the AI-infra-capex and frontier-lab-verticalization threads.
[[2026-08-21-moonshots-ep282-openai-pauses-frontier-training]] — earlier OpenAI/inference-economics thread this episode's Jalapeno chip news extends.
[[2026-08-18-moonshots-chinas-endgame-asi-timelines]] — prior China-AI-strategy episode, relevant background for this episode's China world-model/video-token-share segment.

moonshots ep284 nvidia 96b quarter openai chip