Claude Code Ends SaaS, the Gemini + Siri Partnership, and Math Finally Solves AI | #224
Recorded during CES 2026 and the World Economic Forum in Davos, the panel opens on the 'Cambrian explosion' of humanoid robots and robot-hand companies at CES, then digs into Nvidia's Cosmos/Alpamayo/Vera Rubin keynote and the DRAM/memory shortage reshaping the PC and data-center markets. A large chunk of the episode covers the 'job singularity' (McKinsey's human-plus-agent workforce, Robinhood CEO Vlad Tenev's job-creation thesis) and the claim that Claude Code plus Opus 4.5 ('Clopus') is a watershed moment that could dissolve SaaS and vibe-coding business models. The group debates whether Google powering Siri and the Universal Commerce Protocol will kill the open web, whether OpenAI's compute-vs-revenue chart proves or merely correlates with a real business, and whether frontier labs will consolidate or get broken up by antitrust action. Later segments cover AI bulk-solving open Erdős math problems, Cerebras/SRAM chips as an inference alternative to Nvidia GPUs, xAI's Colossus 3/Macrohard, the US-China energy production gap, and a rapid-fire AMA on human agency, post-work capitalism, founder differentiation, and AI liability/personhood.
CES 2026: the Cambrian explosion of humanoid robots and robot handsRobotics▶ 4:16
Peter recaps CES 2026 (148,000 attendees, ~38 humanoid robot companies, ~12 robotic-hand manufacturers) and the panel debates whether the robotics field will consolidate the way the US auto industry did (253 automakers in 1908 down to 44 by 1929) or the tire industry did (278 tire makers), with Salim joking the eventual winner will be an 'octopus arm' multi-limb design.
Nvidia's Cosmos, Alpamayo, and Vera Rubin keynoteAI▶ 10:33
A clip of Jensen Huang's CES keynote introduces Cosmos (a physical-world foundation model that generates synthetic training video), Alpamayo (an end-to-end reasoning autonomous-vehicle model), and Vera Rubin (a six-chip GPU-accelerated supercomputing architecture). The panel discusses whether synthetic simulation erases Tesla's real-world data moat, and frames Nvidia's strategy as 'commoditizing the complement' the way Intel once did with software SDKs.
DRAM/memory shortage and compute becoming the new client deviceCompute▶ 17:23
Alex argues Vera Rubin signals Nvidia becoming a fully vertically integrated hardware provider (CPU+GPU+memory+interconnect), and that AI-driven DRAM demand is making local PCs uneconomical. Dave adds that Elon believes TSMC isn't building fabs fast enough and is building his own fab capacity, with GPU/DRAM demand described as heading 'to infinity' rather than being cyclical.
The job singularity: McKinsey's human+agent workforce and the rise of the soloreneurEconomy▶ 28:13
A clip of McKinsey CEO Bob Sternfels describing the firm's 60,000-strong workforce (40,000 humans, 20,000 agents, up from 3,000 agents 18 months ago) sparks debate on whether consulting survives AI disruption. Salim counters that ratios will reach ~100 agents per human, and the group discusses Robinhood CEO Vlad Tenev's 'job singularity' concept of micro-corporations, solo institutions, and billion-dollar single-person startups.
Claude Opus 4.5 / 'Clopus' as a watershed for coding and the death of SaaSAI▶ 35:27
Alex describes the combination of Claude Code and Opus 4.5 (nicknamed 'Clopus') as pushing autonomy time horizons hyper-exponentially, comparable to GPT-5.2-to-Codex, enabling projects like entire browsers written in Rust from scratch. Dave describes running 5-10 concurrent Opus 4.5 coding agents, a Claude bill in the tens of thousands of dollars a day, and a debate over whether frontier-model code generation will let anyone rebuild Salesforce, SAP, or Stripe and collapse the SaaS and vibe-coding business models.
Google powers Siri; the Universal Commerce Protocol and 'web extinction' debateAI▶ 46:56
Google will reportedly power a rebuilt Siri via its TPUs and custom models. The panel discusses Scott Stanford's framing of the Universal Commerce Protocol (UCP) as shifting from 'a search box that gives information to a magic box that gives action,' and debates whether AI agentic checkout will make websites and typing obsolete. Alex argues UCP is just a JavaScript-based e-commerce standardization layer within chat agents, not a web-killer.
OpenAI's compute-vs-revenue chart and the frontier-lab survival debateEconomy▶ 52:00
OpenAI CFO Sarah Friar's paper shows compute scaling from 0.2GW (2023) to 1.9GW (2025) alongside revenue scaling from $2B to $20B over the same period, which Peter suspects is framed to justify a future IPO. Alex counters this is 'correlation not causation' and that trillions in capex will require far more expensive inference consumption from consumers and enterprises to be justified, kicking off a broader debate over which of Microsoft, Apple, Google, Amazon, Meta, Tesla, OpenAI, Anthropic, and xAI qualify as true 'frontier labs' and can survive/consolidate in the next 3 years.
AI bulk-solving open Erdős math problemsAI▶ 1:08:57
Alex reports that GPT-5.2 Pro, paired with formalization/verification tools like Harmonic's Aristotle, is now solving well-known open Erdős math problems multiple times a week. He argues math was the easiest domain to 'bulk solve' because problems are easy to verify and enumerate, and predicts this pattern will spread into physics, chemistry, materials science, biology, and medicine.
Inference/training decoupling and SRAM-based chips (Cerebras, Groq) vs NvidiaCompute▶ 1:13:17
OpenAI's partnership with Cerebras is read as a sign that inference and training compute are decoupling, with 80-90% of compute now used for inference. Alex and Dave explain SRAM-adjacent-to-compute architectures (Cerebras, Groq, acquired by Nvidia for $20B) offer much higher throughput/lower latency for long tool-call chains, and Dave flags that a breakthrough training algorithm chunked for SRAM could become a real vulnerability to Nvidia's GPU/DRAM-centric architecture that underlies most Americans' 401k exposure.
A clip on xAI's Colossus 3 buildout (a 2-gigawatt, $20 billion data center) is discussed alongside Elon's 'Macrohard' project, a tongue-in-cheek Microsoft competitor aiming to replace enterprise software (and the employees who run it) with generative AI, which Alex frames as XAI's answer to 'the dissolving of SaaS' and the capex-justification problem shared across frontier labs.
A chart shows China now generating roughly 40% more electricity than the US and EU combined (~10,000 TWh vs the US's flat ~4,000 TWh), with China's solar generation up 46% in 2024 and 48% in 2025. The panel argues energy, not chips or people, is the binding US constraint on AI, debates US regulatory fear of nuclear/solar/fossil energy, and notes 20 African countries importing 2GW of Chinese solar panels in a single month as part of a broader China-led energy/AI-inference export strategy.
AMA: human agency, post-work capitalism, founder differentiation, and AI liability/personhoodAI▶ 1:32:28
Closing subscriber AMA covers preserving human agency via demonetized AI tools (Salim), whether capitalism survives a post-work world and the possibility of a new 'economics 2.0' (Alex), what differentiates a great founder when execution is automated (Dave, Salim), how fast robotaxi fleets can scale (Peter), and a lightning round on AI liability and personhood, including Salim's rant that trolley-problem ethics debates shouldn't be allowed to slow down life-saving automation.
Predictions made
openSalim Ismail: Google will beat Nvidia's market capitalization.
“Google is going to make an attempt to buy Anthropic. I think that's what leapfrogs them over everybody else.”
Your call:
openAlex Wissner-Gross: Microsoft and Apple will not qualify as, and OpenAI, Anthropic, xAI, and Google/DeepMind will survive as, frontier labs, with OpenAI additionally expected to offer humanoid robots.
“I expect open AI to offer humanoid robots. Anthropic... I expect them to survive and thrive in IPO and XAI.”
Your call:
openAlex Wissner-Gross: No broad M&A recombination/knockout consolidation among the named frontier labs (Microsoft, Apple, Google, Amazon, Meta, Tesla, OpenAI, Anthropic, xAI) will occur, aside from a possible Tesla/xAI/SpaceX combination.
“OpenAI confirmed it's going to unveil its first hardware device in the second half of this year.”
Your call:
openAlex Wissner-Gross: AI's bulk-solving of open problems will spread beyond math into physics, chemistry, materials science, biology, medicine, and the humanities.
“This bulk solving of everything is not going to stay confined to math. It's going to walk out of math into physics and chemistry and material science and biology and medicine and the humanities.”
Your call:
Numbers that matter
148,000 attendees, 4,000 exhibitors, 1,200 startups at CES 2026Peter's recap of CES 2026 scale, with roughly 38 humanoid robot companies and 12 robotic hand manufacturers exhibiting.
253 US automotive companies in 1908, falling to 44 by 1929Salim's historical analogy for the coming shakeout among humanoid robot makers.
278 tire companies in the US in the early 1900sSalim's parallel analogy for the robot-hand/component maker market.
McKinsey: 60,000 total workforce = 40,000 humans + 20,000 agents, up from 3,000 agents ~18 months earlierBob Sternfels (clip) describing McKinsey's internal human+agent staffing ratio.
Elon Musk / xAI deal with Samsung worth a minimum of $16 billion, possibly up to $40 billionDave discussing Samsung's exposure and worry as Elon builds his own chip/memory fab capacity.
OpenAI paid roughly $6.5 billion for its hardware device effortReferenced ahead of OpenAI's planned hardware unveiling in H2 2026.
Alex's personal Claude usage bill running between roughly $10,000 and $100,000+ a dayIllustrating the scale of compute spend from running many concurrent Opus 4.5 coding agents.
OpenAI compute: 0.2 gigawatts (2023) to 0.6 GW (2024) to 1.9 GW (2025); revenue: $2 billion (2023) to $20 billion (2025)Chart from OpenAI CFO Sarah Friar's paper, used to argue compute and revenue are scaling together.
Alphabet valuation hits $4 trillion; stock up 65% in 2025Credited to Sundar Pichai's execution and Google's chip-to-model-to-interface-to-distribution stack.
Nvidia acquihired Groq (with a Q) for $20 billionCited alongside Cerebras as one of the two leading SRAM-accelerated compute vendors.
xAI's Colossus 3: a 2-gigawatt data center, roughly a $20 billion buildBuilt to power Elon's 'Macrohard' enterprise-software-replacement initiative.
China now generates ~10,000 terawatt-hours of electricity vs the US's ~4,000 TWh; China produces roughly 40% more than the US and EU combinedUsed to argue energy, not chips, is the binding US constraint on AI progress.
China's solar generation grew 46% in 2024 and 48% in 2025Cited as evidence of China's rapid non-fossil energy buildout.
20 African countries imported 2 gigawatts of solar panels from China in a single monthCited as evidence of China exporting energy infrastructure and influence to Africa.
Roughly 3,000 people with machine guns lining the roads at DavosDave's description of security at the World Economic Forum during Trump's visit.
Worth digging into
🕳️ Universal Commerce Protocol and 'web extinction'
The panel splits over whether Gemini-on-iPhone-style agentic checkout will actually replace websites and typing, or just standardize a narrow slice of e-commerce inside chat agents.
🕳️ SRAM-based accelerators (Cerebras, Groq) as a threat to Nvidia's architecture
Dave flags that a training algorithm chunked to fit SRAM could undercut Nvidia's GPU/DRAM dominance, which underlies most Americans' 401k exposure via index funds.
🕳️ Elon's independent fab strategy amid the Samsung DRAM deal
Dave suggests Elon is building his own chip/memory fabs while Samsung has a $16-40 billion supply deal with him, creating an odd tension where Samsung may be building its own future replacement.
🕳️ AI bulk-solving open Erdos math problems
Alex frames this as the inflection point where 'problems waiting to be prompted' start getting solved automatically, with implications for physics, chemistry, biology, and medicine.
🕳️ OpenAI's compute-vs-revenue chart ahead of a possible IPO
Alex calls the parallel compute/revenue growth curves 'correlation not causation,' suggesting the chart may be framed to support fundraising and IPO valuation rather than prove a sustainable business model.
🕳️ Possible Google acquisition of Anthropic
Peter's speculative bet that Google will move to acquire Anthropic before its IPO would reshape the frontier-lab landscape and test US antitrust tolerance for AI consolidation.