"Google I/O 2026, Karpathy Joins Anthropic, and Cerebras' $95B IPO" - Moonshots EP #256

Why this is in the vault

Cross-references the Google IO 2026 announcements (capex 6x to $180-190B, Gemini 3.5 Flash, anti-gravity 2.0, Gemini Spark, universal cart) with two structural AI-infra stories the same week: Karpathy joining Anthropic to accelerate Claude's own pre-training research, and Cerebras' $95B IPO. Anchor data for the AI-infra capex thesis, the frontier-lab two-horse-race narrative (OpenAI/Anthropic vs Google's distribution play), and the Cerebras-vs-NVIDIA inference competitive dynamic.

Episode summary

The mates recap Google IO 2026 ("Google was cooked? Google is disrupting the disruptors"). Highlights: capex 6x to $180-190B, Gemini 3.5 Flash (throughput-maxed not frontier), anti-gravity 2.0 (Windsurf-rebuild copying cursor), Gemini Spark (OpenClaw response), Synth ID adoption by OpenAI/Cacao/Lean Labs, universal cart vs Amazon. They then cover Karpathy's move to Anthropic (focused on using Claude to accelerate Claude's pre-training - recursive self-improvement), the Elon-vs-OpenAI verdict, and close with a long interview with Cerebras CEO Andrew Feldman after a 68% IPO pop landing at $95B market cap (third-biggest tech IPO in history).

Key arguments / segments

• [00:05:00] Sundar's IO opening: token volume jumped 7x YoY to 3.2 quadrillion/month, 13 products at 1B+ users, Gemini app at 900M MAU (up from 400M), 8.5M monthly developers. Capex growing from $31B in 2022 to $180-190B in 2026 (~6x). Training distributed across 1M+ TPUs globally, new chips at 2x perf/watt.
• [00:13:00] Gemini Omni unveil (Demis): single model spanning video, image, audio, interactive sim. Alex's framing: GDM is the only American frontier lab still betting on multimodality (OpenAI deprioritized Sora; Anthropic was always codegen-focused). The thesis: modality-scaling as a different path to superintelligence, possibly extending to bio/protein sequence data.
• [00:18:00] Gemini 3.5 Flash: "solidly mid" per Alex - not competing with GPT-5.5 or pro-tier, but optimized for throughput and tool use. Sundar showed throughput-vs-perf instead of tokens-vs-perf (flattering metric choice). Dave's read: SF "two-horse race" between OpenAI and Anthropic for raw frontier, Google's bet is distribution + integration depth.
• [00:25:00] Synth ID watermarking: 100B+ images, 60K years audio. OpenAI, Cacao, Lean Labs adopting. Alex frames as cryptographic chain-of-custody for reality emerging from the synthesis side, not the capture side. Dave: this is the first major move of AI-industry self-regulation because Congress can't keep up with the cadence.
• [00:33:00] Anti-gravity 2.0: Windsurf team's rebrand, agent-first IDE. Alex/Dave: copycat of cursor's agent-first pivot, "nobody at Google DeepMind uses it - they want Claude Code." Useful data point on internal Google tooling preference.
• [00:40:00] Gemini Spark: Google's "OpenClaw at home" - always-on agent on GCP VMs powered by 3.5 Flash. Alex: minimum-viable strategic response, not the "art of the possible" Google should have shown. Dave: but distribution wins - half the world uses Google, one-click onboarding will dwarf power-user installs of OpenClaw/Hermes.
• [00:48:00] AI search mode + intelligent search box: persistent search agents that monitor for new listings/price drops, autocomplete that nudges toward profitable queries. Dave's TripAdvisor parallel: default ordering captures 80% of intent. Alex's strategic note: Google's main obstacle to self-disrupting search was technical (latency/cost of gen models in a search-economics envelope), not business-side - which explains the 3.5 Flash throughput focus.
• [00:56:00] Universal cart: cross-merchant cart aggregating from search, Gemini, YouTube, Gmail. Salim: "intent to agent to transaction" replaces "human to website to cart to checkout"; every CMO needs to figure out how to market to agents. Alex: elephant in the room is Amazon - they will not comply with Google's cart standard.
• [01:01:00] Notebook LM still siloed, Spark/Flash/Omni/anti-gravity all separately branded - Alex/Dave call out Google's peanut-butter problem (Brad Garlinghouse's Yahoo memo); ship-and-abandon culture rewards launches over maintenance.
• [Karpathy segment - approximately mid-episode after IO recap]: Karpathy joins Anthropic to "accelerate Claude's own pre-training research" - explicit recursive-self-improvement framing. Major signal for the AI-infra thesis: the labs are now publicly stating they're using current-gen models to design next-gen.
• [Elon-OpenAI verdict segment]: jury verdict in Musk's lawsuit against OpenAI (covered briefly, transcript context implies Musk lost).
• [Cerebras IPO + Feldman interview]: 68% pop on IPO day, $95B market cap, third-biggest tech IPO in history. Feldman in person from Cerebras HQ. Feldman covers wafer-scale-engine economics, the inference market positioning vs NVIDIA, customer wins.

Notable claims

• Google capex 2022 = $31B; 2026 = $180-190B (~6x in 4 years). [00:05:00]
• Gemini 3.2 quadrillion tokens/month across products; 19B tokens/min through APIs. [00:05:00]
• Gemini app 900M MAU (more than doubled from 400M YoY). [00:05:30]
• AI Mode (Google search) at 1B MAU after one year. [00:05:30]
• 50B images generated via nano banana models. [00:06:00]
• Cerebras IPO: third-biggest tech IPO in history, +68% day-one, $95B market cap. [00:04:30]
• Karpathy at Anthropic explicitly tasked with "using Claude to accelerate Claude's own pre-training research" - recursive self-improvement made operational. [intro segment]
• Synth ID: 100B+ images watermarked, 60K years of audio. [00:25:30]
• OpenAI raised ~$120B in cash; Anthropic on similar trajectory - Google's distribution-via-capex moat is narrowing as competitors fund out. [00:23:00]
• Salim's framing: "scarcity = abundance - trust" (Jerry Macowsky). Authenticity becomes infrastructure as intelligence becomes abundant. [00:29:00]

Guests

• Andrew Feldman - Co-founder and CEO, Cerebras Systems. Wafer-scale-engine pioneer, just IPO'd at $95B market cap. Long-form interview in back half of episode covering IPO experience and Cerebras competitive positioning vs NVIDIA in inference.

Mapping against Ray Data Co

Strong for the AI-infra capex thesis ([[hyperscaler-capex]] anchor): clean confirmed datapoint of Google capex 6x to $180-190B. Cross-references the [[2026-04-15-thariq-claude-code-session-management-1m-context]] context-rot discipline (more compute does not equal better outputs above a threshold). The "Karpathy recursive self-improvement at Anthropic" line is the structural story to track - if labs publicly state they are using current models to bootstrap the next generation, the timeline-acceleration thesis updates.

Medium for the Cerebras-as-inference-challenger thesis - Feldman interview deserves a dedicated cross-check against the NVIDIA-moat assumptions in any prior memory-stack thesis work. Note: Cerebras IPO'd at $95B which is asymmetric upside compressed vs entry; useful as comparable to the AI-infra picks-and-shovels lens, not as an entry.

Weak for the consumer-AI-product layer - this is product news, useful as competitive context for any sc.raydata.co AI-tooling posts, but no direct RDCO bet exposure.

Related

• [[hyperscaler-capex]]
• [[2026-04-15-thariq-claude-code-session-management-1m-context]]
• [[ai-infra]]
• Prior moonshots episode: [[2026-05-16-moonshots-anthropic-spacex-leopold-singularity-economy-ep255]]