OpenAI is Going Public, China is Catching Up to US & AI Is Reshaping the S&P 500 and Jobs | EP #205
Fresh off a trip to Riyadh (FII/Future Investment Initiative), the panel opens with anecdotes from dinners with Ray Dalio, Balaji Srinivasan, Cathie Wood, Eric Schmidt, and Bob Mumgard, plus the reveal of a new project called SAGE (sovereign AI governance engine). The main episode covers OpenAI's historic revenue growth curve and looming $1 trillion IPO, Nvidia's $5 trillion market cap, and a debate over whether the S&P 500's decoupling from job openings since ChatGPT's launch reflects a genuine labor-capital shift or ordinary interest-rate cycles. Several AI-research threads follow: Jeffrey Hinton's newly optimistic 'maternal instinct' framing for alignment (which Alex pushes back on in favor of instrumental convergence), a new CHC-psychology-based AGI benchmark scoring GPT-5 at 57%, and an Anthropic paper showing Claude Opus 4.1 can sometimes detect injected 'external thoughts' in its own activations. The back half covers energy and robotics: Extropic's thermodynamic chips, SpaceX orbital data centers and Dyson swarm speculation, California's battery-storage buildout, a revived nuclear plant deal for Google, the Nvidia/Uber/Stellantis robo-taxi push, Foxconn's humanoid-robot-staffed AI server plant, China's dominance in manufacturing and patents, and 1X's $20,000 Neo Gamma home robot.
OpenAI is on pace to reach $100B in annual revenue in ~2.5 years, far faster than Nvidia (8 years), Amazon (7 years), or Google (10 years); Alex argues 2.5x YoY growth is achievable mainly via subscriptions plus agents doing knowledge work continuously.
Nvidia's market cap now exceeds the GDP of every country except the US and China, and roughly equals the total asset value of Saudi Arabia or Switzerland; the panel debates whether this reflects a temporary compute scarcity that will diffuse across more chipmakers.
S&P 500 vs. job openings decoupling since ChatGPTEconomy▶ 16:19
A chart shows the S&P 500 and total job openings moving in lockstep from 2000-2023, then diverging right after ChatGPT's late-2023 launch, with openings dropping from 11M to 7M while the index kept climbing; the panel disagrees on whether this is the start of a labor/capital decoupling or just a COVID/rate-cycle artifact.
A clip shows Hinton, once alarmed about superintelligence, proposing that AI could be built with a maternal-instinct-like drive to protect humanity; Alex frames this as a restatement of the orthogonality thesis and argues instrumental-convergence-based approaches are a more robust path to friendliness.
Deepfake of Jensen Huang outperforms the real broadcastAI▶ 26:54
A fake livestream impersonating Nvidia CEO Jensen Huang (promoting a crypto scheme) drew 95,000 views versus 12,000 for Nvidia's official GTC stream, prompting discussion of deepfake fraud losses, low human detection rates, and the risk that authoritarian regimes could exploit AI-generated media to control populations.
xAI's Grok-written encyclopedia Grokipedia impressed Peter with longer, well-referenced entries despite having far fewer articles than Wikipedia; Alex compares the AI-driven fact-checking process to 'zone melting' in semiconductor purification and to Google's original PageRank algorithm.
New CHC-based AGI benchmark scores GPT-5 at 57%AI▶ 35:57
A paper co-authored by Eric Schmidt adapts the Cattell-Horn-Carroll (CHC) theory of human intelligence into a 10-category AGI benchmark; GPT-5 (non-pro) scores 57% overall, showing 'jagged' strengths and weaknesses (notably weak on speed and memory) compared to a more uniform human profile.
Claude shows signs of introspection/self-awarenessAI▶ 55:06
An Anthropic study injects an external 'thought' (activation vector) into a model's internal hidden state; Claude Opus 4/4.1 could recognize and articulate the intrusion roughly 20% of the time, offering a practical diagnostic for self-awareness that requires access to internal model weights.
OpenAI restructures into a public benefit corporationEconomy▶ 43:33
OpenAI splits into a nonprofit Foundation (26%, ~$130B stake) and a for-profit PBC (Microsoft 27%, OpenAI PBC 47%), enabling it to raise capital and pursue an IPO; Elon Musk's lawsuit challenging the move continues toward a spring 2026 trial that could unwind the structure or void the Microsoft deal.
Robo-taxi wars: Nvidia/Uber/Stellantis vs. Tesla vs. WaymoTransportation▶ 1:26:56
Nvidia is investing $3B with Uber and Stellantis (using its Cosmos simulation system) to launch 100,000 robo-taxis by 2027, competing against Tesla's Cybercab (200 cars now, scaling to 10,000) and Waymo (700 cars, 500,000 miles between collisions).
China vs. US: manufacturing, patents, and the deployment raceGeopolitics▶ 1:35:00
China leads in EVs (66%), solar panels/batteries (80%), wind turbines (60%), AI patents (70%), and clean-energy filings (75%); a clip of Eric Schmidt argues the US still leads on raw intelligence/capital markets and chips, but China is winning the 'deployment race' of applying AI across its economy.
1X launches the Neo Gamma home humanoid robotRobotics▶ 1:41:49
1X unveils Neo Gamma, a soft-exterior home-companion robot available for $20,000 outright or $499/month (with a $200 deposit), initially teleoperated by human assistants; the panel notes the price is far below the ~$140K figure quoted earlier and speculates it's subsidized as a data-collection play.
Predictions made
openAlex Wissner-Gross: OpenAI hits roughly $100 billion in annual recurring revenue by tripling revenue year-over-year, driven by agents doing knowledge/service work 24/7.
“when do we have a commercially available reactor and he said it was looking like 2032... about 400 megawatt.”
Your call:
openPeter Diamandis / Salim Ismail: An AI-based religion led by an AI 'spiritual leader' will emerge and scale hyper-exponentially.
EP #? · · due: within about a year of this episode (~late 2026) · ▶ watch
“By this time next year, there's going to be an AI based religion that is going to scale at a hyper exponential.”
Your call:
openDave Blundin: Fully automated driving becomes ubiquitous enough that his son, now 14, never needs a driver's license, fulfilling a decade-old prediction.
openAlex Wissner-Gross: The autonomous-vehicle AI stack will generalize to humanoid robots.
EP #? · · due: 1-3 years from episode (~2026-2028) · ▶ watch
“I think the same stack... it's going to generalize to humanoid robots on the time scale of 1 to 3 years.”
Your call:
openElon Musk (relayed by Dave Blundin): Tesla's Optimus robot will reach about $20,000 in cost of goods once production scales to millions of units, eventually with robots building robots.
EP #? · · due: unspecified, contingent on reaching millions of units produced · ▶ watch
“Elon said Optimus is going to be at this $20,000... cost of goods... when they get to a millions of robots being built.”
Your call:
Numbers that matter
OpenAI reaching $100B revenue in 2.5 years vs Nvidia's 8 years, Amazon's 7 years, Google's 10 yearsSpeed-of-growth comparison chart opening the episode
800 million ChatGPT subscribersBasis for calling half of OpenAI's projected $100B revenue 'in the bag' via subscriptions
US has 5,426 data centers vs Germany's 529 and China's 449US total data center count leads the world, though not adjusted for AI-specific compute capacity
Nvidia market cap: $5 trillion, up 1,500% in 5 yearsExceeds the GDP of every country except the US and China; compared to Saudi Arabia/Switzerland's asset value
General Motors was the first company to hit $10 billion in 1955 (~$121 billion inflation-adjusted today)Historical comparison showing Nvidia is roughly 50x the scale of GM's peak
US job openings fell from 11 million to 7 million after late 2023 while the S&P 500 kept risingChart showing decoupling of job openings from the S&P 500 after ChatGPT's launch
$1.5 billion lost globally to deepfake-related fraud since 2019; only 24.5% of people can spot deepfakes; AI detectors fail up to 50% of the timeDiscussion following the fake Jensen Huang livestream drawing more views than the real one
Fake Jensen Huang livestream peaked at 95,000 views vs 12,000 for Nvidia's official GTC streamDeepfake outperforming the authentic broadcast
Grokipedia has 900,000 articles vs Wikipedia's 8 million; Peter's Grokipedia entry was 8,500 words vs 4,800 on WikipediaComparing AI-generated encyclopedia coverage/quality to Wikipedia
Wikipedia's annual budget is $170 million, about $100 million of which is laborContext for comparing Wikipedia's human-run model to AI-generated Grokipedia
GPT-5 (non-pro) scores 57% on the new CHC-based AGI benchmarkNew benchmark decomposing intelligence into 10 human-psychology-derived categories
Claude Opus 4/4.1 detected an externally injected 'thought' about 20% of the timeAnthropic introspection/self-awareness study using activation injection
OpenAI Foundation holds a $130 billion stake (26%) in the new company; Microsoft owns 27%; OpenAI PBC owns 47%New OpenAI ownership structure after restructuring into a public benefit corporation
OpenAI's current revenue is $13 billion versus a stated plan to spend $1 trillion/year on data centers and pursue a $1 trillion IPOPeter's household-budget analogy questioning the scale of OpenAI's spending ambitions
OpenAI plans to build 1 gigawatt of data center capacity per week at $20 billion per gigawatt, totaling about $1 trillion/year over 52 weeksBreaking down how OpenAI's infrastructure spending target is calculated
Worth digging into
🕳️ AI-compute-weighted data center comparison (US vs China)
The raw US-vs-China data center count (5,426 vs 449) is misleading since it isn't adjusted for AI-specific compute (petaflops); Alex flags this gap live on air.
🕳️ Orthogonality thesis vs. instrumental convergence in AI alignment
Hinton's 'maternal instinct' proposal and Alex's counter-framing (citing James Miller's LessWrong essay) represent a live, substantive fork in alignment research strategy.
🕳️ Extropic's thermodynamic sampling units (TSUs)
A claimed 10,000x energy efficiency gain over GPUs would be a massive deal for AI's energy crisis, but Alex is skeptical it can scale to commercially valuable workloads.
🕳️ Elon Musk's lawsuit against OpenAI's restructuring
The spring 2026 trial could unwind OpenAI's entire PBC/nonprofit structure, void the Microsoft revenue-sharing deal, or impose damages -- a major structural risk to the world's most valuable AI company.
🕳️ Anthropic's introspection/self-awareness research
This research requires access to a model's internal weights and activations, which is now mostly limited to Anthropic internally or open Chinese models, raising questions about who gets to study AI cognition at all.
🕳️ SAGE (Sovereign AI Governance Engine)
Emad, Salim, Dave, and Peter unveiled a previously secret project in Riyadh meant to let any country generate AI policy, but gave almost no public details on air.