2025-10-08
OpenAI vs. Grok: The Race to Build the Everything App w/ Emad Mostaque, Dave Blundin & AWG | EP #199
A rapid-fire 'WTF just happened in tech' episode triggered by OpenAI Dev Day, where the panel dissects Sam Altman's token-growth numbers, the new Apps SDK/Agent Builder push toward an 'everything app,' voice-controlled Codex, and Sora 2's sketch-to-video capability. The conversation pivots to compute economics: the AMD-OpenAI 6GW chip deal, BlackRock's data center buying spree, Corning's optics bet, and the looming energy bottleneck pushing data centers off-grid. It closes on robotics, with Tesla's FSD 14.1 and Optimus, and Google's Gemini Robotics ER 1.5, framed as converging toward unified vision-language-action models that the panel believes are close to solving physical labor. Throughout, the group argues we are at a 'tipping point' within six months across AI, compute buildout, and robotics simultaneously.
Topics
OpenAI Dev Day token growth AI ▶ 3:02
The 'everything app' battle for attention AI ▶ 9:01
Agent Builder and the shift away from code AI ▶ 16:12
Voice-controlled Codex and AV automation AI ▶ 20:24
Sora 2 sketch-to-video ('mind to materialization') AI ▶ 28:02
OpenAI's monetization pivot Economy ▶ 33:51
Anthropic nears superhuman computer use AI ▶ 37:38
Grok Imagine and Elon's gaming push AI ▶ 41:32
AMD-OpenAI 6GW chip deal Compute ▶ 47:14
Data center capital buildout Energy ▶ 52:02
Data center energy constraints Energy ▶ 56:52
Tesla FSD 14.1 and robotics convergence Robotics ▶ 58:52
Gemini Robotics ER 1.5 and robot safety benchmarks Robotics ▶ 1:00:58
Predictions made
open Dave Blundin: The AI/compute/robotics tipping point discussed across OpenAI, China's open-source models, and robotics will arrive within the next 6 months.
“We're just at this tipping point and the tipping point is in the next 6 months.”
Your call:
open Emad Mostaque: OpenAI's token volume will roughly 5x to about 30 quadrillion tokens/year next year, then by the following year OpenAI alone will process more tokens than all humans speak in a year.
“next year OpenAI is probably going to be at 30 and then by themselves the year after they'll overtake in terms of tokens all the human words spoken every single year”
Your call:
open Dave Blundin: A startup building AI-controlled, foolproof AV/presentation systems (voice-controlled cameras, lighting, screen-sharing) could ship a product within six months and become a major success.
“Very doable. You could get that product out the door in like six months or less and just crush it.”
Your call:
open Alex Wissner-Gross: Anthropic (and likely other frontier labs) will reach superhuman performance on the OSWorld computer-use benchmark by the end of the year or within the next few months.
“probably again by the law of straight lines perhaps by the end of this year in the next few months we're going to see... superhuman performance at the ability to control computers”
Your call:
open Dave Blundin: The robotic/physical-world version of an always-available AI assistant (an 'Iron Man suit' AI) is one to three years away, though the virtual-world version already exists today.
“The robotic version of it where it makes your Iron Man suit might be a year or two or three out”
Your call:
open Alex Wissner-Gross: Tesla's FSD/VLA tech stack will converge with the Optimus robot tech stack within roughly the next 2-3 years, and other frontier labs will follow with their own cross-embodiment VLA models.
“I would expect we're going to see over the next few years maybe 2 to 3 years... the Optimus tech stack converge with FSD maybe in some future version”
Your call:
open Alex Wissner-Gross: The US will need approximately 250 gigawatts of data center capacity by the early 2030s, and may lack the human labor to build it without robotic construction labor.
“if we're going to achieve 250 gawatts by the early 2030s, we may not have the human labor to accomplish that”
Your call:
open Emad Mostaque: OpenAI and other labs will replicate a 'Macro Hard'-style strategy of fully autonomous AI workers as a product, priced roughly $10,000-$100,000, to justify the massive data center capex being committed.
“I think in the next few years you're going to see basically OpenAI and others replicate the whole macro hard strategy of fully autonomous workers... they will cost like 10, 20,000, 30,000, $100,000”
Your call:
Numbers that matter
- OpenAI API throughput grew from 300 million to over 6 billion tokens per minute since 2023
- OpenAI has 4 million developers building on its platform and over 800 million weekly ChatGPT users
- 6 billion tokens/minute = about 3 quadrillion tokens per year for OpenAI alone
- Sora 2 API base pricing is 10 cents per second, i.e. $360 per hour
- OpenAI generated $4.3 billion in revenue in the first half of the year, versus a $200 billion global video game market last year
- AMD stock rose about 30% on news of the OpenAI chip partnership
- OpenAI can earn up to 10% ownership of AMD via warrants, at effectively no cost, if deployment milestones are hit
- Combined Nvidia (10GW) and AMD (6GW) commitments imply roughly $800 billion, potentially over $1 trillion, in data center buildout at ~$50 billion per gigawatt
- BlackRock is buying up to 78 data centers totaling 5 gigawatts in a $40 billion deal
- Roughly half of all US GDP growth this year is attributed to AI
- The US may need approximately 250 gigawatts of data center capacity by the early 2030s
Worth digging into
🕳️ Video as a first-class chain-of-thought modality
🕳️ DeepMind's 'Azimov' robot safety benchmark
🕳️ Off-grid, co-located power for data centers ('Colossus' model)
🕳️ The AMD-OpenAI warrant deal as a repeatable playbook
🕳️ Codex's origin in a CMU/SEI feature request
🕳️ Elon's 'Macro Hard' autonomous AI software company