Google I/O 2026, Karpathy Joins Anthropic, and Cerebras’ $95B IPO | EP #256
The panel dissects Google I/O 2026: Sundar Pichai's staggering scale numbers (3.2 quadrillion tokens/month, ~$185B CapEx, 900M+ Gemini users), the new Gemini Omni multimodal model, Gemini 3.5 Flash, Anti-Gravity 2.0, Gemini Spark, SynthID content-authentication, Universal Cart shopping, and Android XR audio glasses, largely concluding Google is a fast-follower leveraging distribution rather than pushing the raw frontier. They cover Andrej Karpathy's move from independent researcher to Anthropic's pre-training team, and Elon Musk's lawsuit loss to OpenAI. The back half is a long interview with Cerebras CEO Andrew Feldman following its record IPO (up 68%, $95B market cap), covering the wafer-scale engine's SRAM-based architecture, its 15-20x inference speed advantage over GPUs, the company's near-death first years, a $20B+ OpenAI deal, and skepticism toward Elon Musk's Terafab timeline. A closing AMA touches on compute economics, why money alone can't buy AI/chip leadership, and reports of China reselling discounted US AI tokens.
Google I/O 2026 keynote: scale numbers and CapExCompute▶ 5:39
Sundar Pichai's keynote numbers: monthly token processing jumped from 9.7 trillion (2 years ago) to 3.2 quadrillion today; Google now has 13 products with 1B+ users; CapEx is rising 6x from $31B (2022) to ~$185B, powering a 1M+ TPU global training cluster.
Demis Hassabis unveils Gemini Omni, a model family (built on VEO, Nano Banana, Genie) that generates realistic video/image content from any combination of text, photo, video, and audio input, including real-time conversational video editing.
Gemini 3.5 Flash launch and Google's frontier positioningAI▶ 18:39
Google launches Gemini 3.5 Flash as the new Gemini app default, emphasizing throughput (4x faster output tokens/sec than rivals) and agentic/tool-use benchmarks rather than raw frontier intelligence; panel calls it 'solidly mid' versus GPT-5.5 and positions Google as 'tier one and a half.'
SynthID has watermarked 100B+ images/videos and 60,000 years of audio; OpenAI, Kakao, and Eleven Labs join Nvidia in adopting it. Panel frames this as the start of industry self-regulation (cryptographic 'proof of reality') substituting for slow-moving government regulation.
Google's Windsurf-derived Anti-Gravity 2.0 goes fully agent-first, hiding the code editor by default. Panel calls it a slow-follow copy of Cursor's agent-first pivot; nobody they know uses it for primary development versus Claude Code or Codex.
Gemini Spark: always-on agent and ecosystem lock-inAI▶ 40:26
Gemini Spark is Google's always-on personal agent (Google's answer to 'Open Claw'/Hermes-style agents), running 24/7 on a GCP VM and deeply integrated with Gmail, Sheets, and other Google products. Panel calls it a 'lazy copycat' but agrees Google's default distribution (Android, Search, Chrome) makes it unstoppable regardless of frontier ambition.
Universal Cart: agentic shopping vs. AmazonEconomy▶ 55:46
Google launches a cross-merchant 'Universal Cart' that tracks products from Search, YouTube, Gemini, and Gmail, monitoring price drops and stock. Panel views nearly every Google shopping move through the lens of competing with Amazon, and discusses a shift from human shopping to agents transacting on a person's behalf.
Android XR audio glasses vs. Meta's smart glasses leadRobotics▶ 1:07:01
Google announces audio-only (no display) smart glasses with Samsung, Warby Parker, and Gentle Monster, arriving fall 2026. Panel notes Google abandoned Google Glass and ceded the smart-glasses market to Meta, and debates the societal cost of ambient AI whispering in your ear all day.
Gemini for Science and Isomorphic LabsBiotech▶ 1:14:47
Demis Hassabis announces Gemini for Science tools (literature synthesis, hypothesis generation, hurricane-path simulation via 'Weather Next') and updates that Isomorphic Labs is in preclinical stage on treatments for immune disorders and cancer.
Andrej Karpathy joins Anthropic's pre-training teamAI▶ 1:24:31
Karpathy (OpenAI co-founder, ex-Tesla Autopilot lead, ex-Eureka Labs) joins Anthropic to accelerate Claude's own pre-training research using Claude, citing a need to stay inside a frontier lab so his judgment doesn't 'drift' from what's actually being built.
Elon Musk loses lawsuit against OpenAIGeopolitics▶ 1:29:44
A federal jury unanimously rejected Musk's lawsuit against OpenAI in two hours, ruling he waited too long (statute of limitations). Musk's team plans to appeal, though panel and guest Andrew Feldman expect it to fail and call the whole fight a costly distraction from building.
Cerebras' record IPO and wafer-scale engine origin storyCompute▶ 1:33:00
CEO Andrew Feldman recounts founding Cerebras in 2015 on two contrarian bets (dedicated AI silicon, clean-sheet non-GPU design), building a chip 58x larger than any before it, packed with SRAM for memory bandwidth, achieving 15-20x faster inference than GPUs after years with near-zero sales before the 2024-25 inference boom hit.
US chip manufacturing crisis and Musk's TerafabCompute▶ 1:45:50
Feldman details why US fabs (TSMC, Samsung) and the packaging supply chain lag by years/billions of dollars, spanning multiple presidential administrations and local ordinances (e.g., Samsung's Texas fab fire-code redesign). He calls Musk's Terafab a 15-20 year project, not 5-10, while praising Musk's track record elsewhere.
AMA: compute economics, talent vs. purpose, China's token resellingGeopolitics▶ 2:05:35
Closing audience Q&A covers why compute costs will collapse like past technologies, why money/talent alone can't buy chip leadership (citing Intel's mobile-chip failure), Sam Altman's compute/space strategy versus SpaceX's launch dominance, and reports that Chinese proxies are reselling discounted US AI tokens to harvest reasoning data for training.
Predictions made
openAlex Wissner-Gross: Google will release Gemini 3.5 Pro roughly a month after I/O 2026, per Sundar Pichai's on-stage comment.
“Today, I'm excited to announce that our first audio glasses will arrive this fall.”
Your call:
openDave Blundin: Agentic development tools will move beyond chat-based interfaces into real-time graphical/spatial ('holodeck-style') interaction that lets users manipulate what they're building directly.
“that next iteration is certainly within this calendar year.”
Your call:
openAndrew Feldman: Elon Musk's Terafab (a proposed US chip fab meant to out-produce global chip capacity) will take 15-20 years to realize, not the shorter timeline implied by Musk.
openAndrew Feldman: Wafer-scale-style compute will not be in production use in space for roughly another decade, pending fault-tolerance and orchestration work.
“I've got it sort of out the better part of a decade before we have sort of production in space.”
Your call:
openSalim Ismail: The cost of AI-delivered professional services (e.g., legal help) will keep falling from roughly $1,000/hour toward pennies, following the same cost-collapse curve seen in computing, bandwidth, DNA sequencing, and solar.
openAlex Wissner-Gross: If OpenAI plays a long game on compute, it will pursue building chip fabs in space or on the moon to escape dependence on SpaceX's near-monopoly on heavy launch.
“we may not be addicted to heavy launch 5 to 10 years from now.”
Your call:
Numbers that matter
Google processed 9.7 trillion tokens/month two years ago, growing to 480 trillion last year and 3.2 quadrillion tokens/month todaySundar Pichai's I/O keynote scale metrics
8.5 million+ people building apps monthly with Google's modelsGoogle I/O developer adoption figure
Google model APIs process ~19 billion tokens per minuteGoogle I/O infrastructure scale
Google has 13 products with over 1 billion users each, 5 of which exceed 3 billion usersGoogle I/O keynote product portfolio scale
AI Overviews has over 2.5 billion monthly usersSearch AI feature adoption
AI Mode surpassed 1 billion monthly users within a year of launchSearch AI mode adoption
Gemini app grew from 400 million monthly active users a year ago to over 900 million todayGemini consumer adoption growth
Over 50 billion images generated with Google's Nano Banana modelsImage generation product usage
Google CapEx rising 6x from $31 billion (2022) to approximately $180-190 billion (2026)Google infrastructure investment scale
Google's training cluster scales across more than 1 million TPUs globallyLargest training cluster claim
Gemini 3.5 Flash is 4x faster in output tokens/sec than other frontier modelsGoogle's throughput benchmark for the new model
SynthID has watermarked over 100 billion images/videos and 60,000 years of audio assetsScale of Google's content-provenance system
45% of dementia cases are entirely preventable with lifestyle changes; Fountain Life improved brain age in 46% of members with advanced brain age over 13 monthsFountain Life sponsor segment on cognitive health
Cerebras IPO raised $5.5 billion, stock up 68%, market cap $95 billion — the biggest US IPO since Uber in 2019Cerebras' record public offering
Cerebras' wafer-scale chip is about 58 times larger than any chip built before itWafer Scale Engine size claim
Worth digging into
🕳️ Cerebras' wafer-scale SRAM architecture vs. GPU HBM tradeoffs
Feldman's explanation of why a chip 58x larger than any before it, stuffed with SRAM, solves the memory-bandwidth bottleneck is a genuinely novel hardware story backed by concrete benchmark claims (15-20x inference speed; 1,000 vs 70 tokens/sec on Kimi K2).
🕳️ China's proxy reselling of discounted US AI tokens
Alex references recent public reporting that Chinese proxies are reselling American model access at ~10x discounts to harvest reasoning traces for training competing models — a significant IP/national-security angle only briefly mentioned.
🕳️ Terafab timeline dispute: Feldman's 15-20yr estimate vs. Musk's ambitions
A credible industry insider who has actually solved wafer-scale manufacturing problems directly contradicts Musk's implied faster fab timeline, calling fab-building 'pyramids' that take 5-6 years and $40-50B even for established players like TSMC.
🕳️ SynthID coalition as de facto AI self-regulation
The panel frames SynthID's adoption by OpenAI, Kakao, Eleven Labs, and Nvidia as the first move toward industry self-governance replacing slow legislation — worth checking if it becomes a durable cross-industry standard or stalls.
🕳️ Cerebras' $20B+ OpenAI deal and AWS term sheet
Feldman mentions a multi-year OpenAI deal 'north of $20 billion' (Dec 2025) and an AWS term sheet (March 2026) with no further detail on scope, structure, or delivery timeline — a huge number dropped in passing.
🕳️ Fountain Life's brain-health outcomes data
Dr. Don Musalem cites a striking clinical-sounding claim (46% of members with advanced brain age improved it within 13 months) during a sponsor segment, with no citation to a published study.