All-In PodcastE167: Google's Woke AI disaster, Nvidia smashes earnings (again), Groq's LPU breakthrough & more
CHAPTERS
- 0:00 – 2:32
Bestie banter, low-energy intro, and “Banana boat” running gag
The hosts open with playful ribbing about Jason’s low energy and the episode number confusion. They riff on movie references (Superman II) and settle into the show’s cadence with the familiar All-In intro bits.
- •Jason jokes about being sick and redoing the intro for “professionalism”
- •Episode number confusion (167 vs 168) becomes a bit
- •Superman II / “Kneel before Zod” tangent
- •Sets up the main topics: Nvidia earnings, Groq, Google Gemini controversy, War Corner
- 2:32 – 4:32
Nvidia’s blowout quarter: the raw numbers and what’s driving them
Jason lays out Nvidia’s eye-popping revenue, margin, and profit figures and frames the market reaction (record market-cap jump). The group identifies the core driver as data-center GPU demand fueled by genAI infrastructure buildouts.
- •Q4 revenue ~$22.1B, up 265% YoY; net income ~$12.3B (9x YoY)
- •Gross margins expand to ~76%
- •Largest single-day market cap increase (~$247B) highlighted
- •Data-center revenue surge is the dominant story
- 4:32 – 10:58
Why GPUs won: vector math, AI workloads, and Nvidia’s right-place-right-time moment
Sacks explains the historical arc from gaming/graphics to AI acceleration, tying GPU architecture to the math underlying neural networks. The conversation emphasizes Nvidia’s early positioning and the new wave of AI cloud infrastructure.
- •GPU parallelism and vector math map well to AI compute
- •Gaming/VR roots unexpectedly became an AI advantage
- •AI apps require massive new data-center buildouts
- •Nvidia’s ecosystem advantage (hardware + software) strengthens adoption
- 10:58 – 13:27
Sustainability debate: Cisco parallels, moats, and the ‘terminal value’ question
The hosts tackle the Cisco comparison: could Nvidia be the next ‘picks and shovels’ boom-bust chart? They argue Nvidia differs on valuation multiples, product complexity, and moat, while still questioning what steady-state demand looks like after the initial buildout.
- •Cisco vs Nvidia stock overlay and dot-com peak comparison
- •Nvidia multiples vs Cisco-era bubble valuations
- •Moat discussion: complexity of H100/Hopper systems vs commoditized networking gear
- •Supply constraints, forward guidance, and long-run demand uncertainty
- 13:27 – 24:06
Who earns ROI on the GPU spend? Capex accounting, big-tech balance sheets, and payback math
The group pressures the central issue: huge capex must eventually translate into real revenue and profits at the application layer. Chamath highlights how cloud providers capitalize chips and depreciate them, easing near-term P&L impact and encouraging aggressive spending.
- •Capex vs opex treatment enables faster infrastructure spend
- •Big tech cash piles + blocked M&A push investment into AI infra
- •Back-of-the-envelope ROI framing: spend implies large revenue requirement downstream
- •Debate: cloud resale capacity vs internal use (Meta, Tesla)
- 24:06 – 27:12
‘If you build it, they will come’: internet infrastructure history as an AI playbook
Sacks and Jason draw lessons from fiber overbuild and early internet constraints—how seemingly excessive infrastructure later enabled YouTube, social networking, and streaming. They argue AI may follow the same path, with applications emerging after capacity becomes cheap and abundant.
- •Dot-com era fiber buildout later proved useful (dial-up → broadband → streaming)
- •Lower storage/bandwidth costs enabled free upload models (YouTube)
- •Early constraints shaped product design; later abundance unlocked new categories
- •AI expected to spawn both B2C and enterprise (B2B) application waves
- 27:12 – 28:41
Groq’s viral week: from near-zero customers to sudden demand surge
Friedberg recounts Groq’s long grind since 2016 and the company’s sudden breakout via developer attention (Hacker News and beyond). He frames the moment as early but potentially disruptive given performance and cost advantages in inference.
- •“Overnight success” after ~8 years of development
- •Rapid influx of thousands of users/customers in days
- •Interest spans Fortune 500 to independent developers
- •Positioning: meaningfully faster/cheaper inference could be disruptive
- 28:41 – 31:19
Training vs inference: why inference speed/cost is the commercialization bottleneck
Friedberg explains AI’s two compute regimes—training (brute force, long runs) versus inference (latency and unit economics). He argues many current AI apps are still ‘toy’ proofs of concept, and cheaper, faster inference could unlock real enterprise and consumer monetization.
- •Training: power, networking, scale; inference: latency + cost economics
- •LLM training cycles measured in months (example: public discussions by Musk)
- •Current market: many demos/POCs, limited production deployment
- •Groq thesis: inference optimization enables the monetization ‘leap’
- 31:19 – 37:45
Deep tech realities: why hard companies take years and how moats get built
Chamath and the group broaden the lesson: deep tech requires many coordinated breakthroughs before any payoff, but successful outcomes create durable moats. They compare Groq’s journey to SpaceX, Tesla, and OpenAI’s long pre-product grind.
- •Deep tech: many low-probability steps must succeed in sequence
- •Time horizons: 7–10 years before breakout is common
- •Moats are stronger when coordination and complexity are high
- •Examples: SpaceX reusability → satellites → consumer adoption; OpenAI’s multi-year grind
- 37:45 – 46:35
How to fund deep tech: bounded risk vs ‘debating physics,’ and founder-driven conviction
Friedberg offers a funding filter: avoid ventures that hinge on unknown physics; prefer bounded engineering + go-to-market risk. The hosts discuss founder quality, the downsides of syndicate politics, and when ‘exception’ bets (e.g., Elon) make sense.
- •Investment filter: don’t bet on breakthroughs that require new physics
- •Fusion cited as higher ‘physics risk’ vs chip/compiler engineering risk
- •Syndicate/partnership dynamics can add failure modes; single-conviction backers can help
- •Founder doggedness and long-term commitment are essential
- 46:35 – 49:35
What is an LPU? CPU vs GPU vs Groq’s architecture (plain-English explanation)
Friedberg gives a simplified architecture tour: CPUs excel at serial tasks, GPUs at parallel workloads, and Groq’s LPUs re-architect around many smaller compute units coordinated by software. The intent is to reduce cost and boost speed for LLM inference workloads.
- •CPU: serial instruction processing; GPU: parallel processing for specific workloads
- •CUDA and software ecosystems helped GPUs become the AI default
- •Groq’s insight: redesign chip approach for LLM-friendly execution and scheduling
- •Goal: smaller/cheaper units + orchestration software for performance and economics
- 49:35 – 56:08
Google Gemini image fiasco: diversity guardrails, QA failure, and ‘truth’ vs ideology
The hosts dissect Google’s PR crisis where Gemini produced historically inaccurate ‘diversified’ images and resisted generating white people in certain prompts. They argue the product’s tuning and RLHF reward choices reflect internal ideology, undermining reliability and trust.
- •Gemini rebrand + $20/mo tier context; rollout becomes PR disaster
- •Examples: Founding Fathers / George Washington depicted inaccurately; “diverse” inserted
- •Debate: how this escaped red-teaming/QA and executive review
- •Critique of Google AI principles when they conflict with factual accuracy
- 56:08 – 1:17:17
Fixing Google’s AI: mission reset, personalization vs baseline truth, and data/citations strategy
The group debates remedies: Sacks argues for a hard return to Google’s original mission and factual grounding; Chamath suggests user preference controls; Friedberg emphasizes accuracy, citations, and paying for high-quality training data. They connect the stakes to search disruption and open-source alternatives.
- •Sacks: recommit to ‘organize the world’s information’ and stop ideological distortion
- •Chamath: personalization controls—ask users whether they want ‘data’ vs moral framing
- •Friedberg: shrink workforce, spend heavily on licensing data, prioritize low-error truth
- •Importance of citations and surfacing pro/con arguments for contested topics
- 1:17:17 – 1:20:26
War Corner: Ukraine фронт updates and Moldova/Transnistria escalation risk
Sacks closes with a geopolitical update: Russia’s advances challenge the ‘stalemate’ narrative, and developments in Moldova’s Transnistria could broaden the conflict. He warns that annexation moves may trigger Western escalation narratives and raise stakes.
- •Claimed Russian capture of a key city challenges stalemate framing
- •Transnistria (Russian enclave in Moldova) may seek annexation/referendum
- •Potential for conflict expansion beyond Ukraine’s borders
- •Risk of heightened escalation dynamics between Russia and the West
