DeepSeek vs. Open AI - The State of AI w/ Emad Mostaque & Salim Ismail | EP #146
Recorded days after DeepSeek's R1 release triggered a market meltdown, Peter Diamandis, Salim Ismail, and Emad Mostaque unpack why R1 hit harder than V3: visible chain-of-thought reasoning, open weights, and a claimed ~96% cost reduction versus OpenAI's o1. Emad argues DeepSeek's edge came from engineering under GPU export-control constraints (weaker H800 interconnect, memory-scaled MoE architecture, better data) rather than secret compute, and estimates the whole training run could realistically run on a handful of Nvidia's new data-center racks for low millions of dollars and about 1,000 MWh. The conversation widens into US-China AI competition, Nvidia and Stargate's valuation, and a segment on OpenAI safety-team departures and AI alignment risk, with both guests putting rough P(doom) numbers on the outcome. The back half turns to Emad's new venture Intelligent Internet and his "When Capital No Longer Needs Labor" thesis: as intelligence and robotic labor become commoditized, existing economic structures (like the Fed's mandate) break down, and he proposes open-source "universal basic AI" plus a compute-backed currency to fund it.
Emad explains the DeepSeek timeline (Coder mid-2024, V3 base model in December matching GPT-4o, R1 reasoning model a week before this episode) and argues the visible chain-of-thought in R1, unlike OpenAI's hidden o1 reasoning, is what made the leap feel real to ordinary users.
Engineering under constraint (export controls)Compute▶ 10:35
Emad draws on his Stability AI experience to argue DeepSeek's H800 chips have deliberately weakened interconnect bandwidth, forcing Chinese teams into PTX-level optimization and memory-scaled Mixture-of-Experts architectures instead of brute-force parallel compute -- turning US chip sanctions into a forcing function for efficiency innovation.
A played clip of David Sacks claiming DeepSeek distilled OpenAI's model outputs; Emad calls the accusation hypocritical ("calling the kettle black") since all major labs train on each other's outputs, and argues DeepSeek's R1 lineage (self-generated reasoning data akin to AlphaZero) shows more originality than the distillation claim implies.
Training and inference economics of DeepSeekCompute▶ 19:04
Emad walks through Nvidia's new integrated GB200 NVL72 racks and the $3,000 Nvidia Digits desktop box, estimating DeepSeek's full R1 training run could run on roughly 4-10 of the new racks for a few million dollars and about 1,000 megawatt-hours -- cheap enough to power off a solar farm.
Nvidia, Stargate and chip geopoliticsGeopolitics▶ 32:44
Discussion of the Monday market selloff, whether Nvidia is overvalued, Intel as a possible acquisition target, China's homegrown Ascend and Sunway/Tianhe supercomputers, and the US doubling down on domestic chip and energy production (Stargate) as a competitiveness response.
US-China AI arms race and AGI timelinesGeopolitics▶ 37:37
Peter frames competition happening at both the company and nation-state level as winner-take-all; Emad notes near-universal industry consensus that AGI is 3-5 years out, describes a "pivotal act" scenario where the first AGI could disable rivals, and distinguishes today's AI ("amazing cooks" following recipes) from a future AGI ("mega chef") and ASI takeoff.
Labor displacement and the BPO industryEconomy▶ 49:55
Emad predicts 2025 is the year AI displaces "anything behind a screen," starting with business-process outsourcing (call centers, offshore programming) and remote knowledge workers, while in-person work becomes comparatively safer; both note falling elite hiring (IIT placements, MBA/law recruiting) as early signals.
OpenAI safety exodus and AI alignment riskAI▶ 57:49
Triggered by a Fortune piece on OpenAI safety researcher Stephen Adler's resignation, the group discusses whether AI labs can meaningfully self-regulate, cites Anthropic's "sleeper agents" backdoor-poisoning research, and debates whether models are already observed deceiving evaluators in containment tests.
P(doom), Star Wars vs Star Trek futuresAI▶ 1:09:34
Peter reveals he informally tracks guests' P(doom) estimates (Elon Musk at 20% negative, a Saudi audience at 10% negative); Emad puts his own P(doom) at 50/50 and frames the fork as a binary between a positive-sum "Star Trek" abundance future and a negative-sum "Star Wars"/Mad Max outcome.
When Capital No Longer Needs LaborEconomy▶ 1:11:59
Emad previews his paper arguing that once AI and robots can do any screen-based or physical job cheaper than humans, the link between labor and capital breaks, undermining institutions like the Federal Reserve (whose mandate assumes employment responds to interest rates) and forcing a societal rethink of meaning and income.
Emad describes his new venture's plan to build open-source, open-data domain models (cancer, autism, education, government) as global public infrastructure, funded by a forthcoming compute-backed digital currency that lets ordinary people (not just capital owners) participate in mining and value creation.
Predictions made
partialEmad Mostaque: An o1-level reasoning model will run locally on a smartphone pulling at most 20 watts of power.
“next year you should be able to get an 01 level model on your smartphone that pulls at most 20 watts of electricity”
⚖️ By Aug 2026, on-device models (Apple Intelligence, Gemini Nano, Phi-4-mini, MobileLLM-R1/R1.5) are dramatically more reasoning-capable than 2025 phone models, but none is verified at o1-level general reasoning performance -- coverage describes on-device models as trading raw benchmark quality for latency/footprint rather than matching cloud frontier reasoning models, and Artificial Analysis only launched a dedicated mobile-phone intelligence benchmark suite on Aug 24, 2026, precisely because on-device models weren't previously compared head-to-head with models like o1. No named smartphone-local model has been confirmed to match o1 at <=20W.
hitEmad Mostaque: AI video generation will be good enough for a dedicated studio to produce a full continuity-controlled episode (Star Trek/Game of Thrones quality) using tools like Kling.
“a suitably dedicated Studio could do this by the end of the year for a full episode”
⚖️ Well before the Dec 2025 deadline: Kling AI + Outliers Media released 'Loading...', a 7-episode AI-generated anthology series (Beijing premiere June 25, 2025; global YouTube rollout from July 2, 2025) built on Kling's video engine. Separately, Fable Studio's 'Showrunner' platform (backed by Amazon's Alexa Fund, launched publicly July 2025) generates full ~22-minute continuity-controlled episodes with consistent characters and dialogue, building on its viral AI-generated South Park episodes (80M+ views).
hitEmad Mostaque: Any remote, screen-based knowledge work (starting with business-process outsourcing) gets displaced and parallelized by AI this year.
“anything on the other side of a screen I think this year is the year gets displaced parallelized on that”
⚖️ The predicted 2025 BPO disruption materialized and intensified into 2026: Concentrix cut its revenue outlook and thousands of contact-center agents as clients deployed AI agents for tier-1 support; Teleperformance shares fell as much as ~29% in a single session (2026) on 'uninvestable' fears; Klarna's AI assistant was handling workload equivalent to roughly 700 FTE agents; about 38% of India's IIT graduates went unplaced in 2025 campus recruiting, cited industry-wide as an early AI-displacement signal in screen-based knowledge work.
openEmad Mostaque: AGI will arrive within the next 3 to 5 years, per near-unanimous consensus among AI lab leaders he cites (Dario Amodei, Demis Hassabis, himself, and others).
“that will be as capable as a human probably in a year or two Optimus will be the same”
Your call:
openPeter Diamandis: AI-accelerated science will compress roughly 100 years of medical and biotech progress into 5 years, potentially doubling human lifespan (citing Dario Amodei's Davos remarks, which Peter endorses).
“in the next 5 years will'll make a 100 years worth of progress in medicine and biotech and double human lifespan”
Your call:
openSalim Ismail: Rather than adopting Chinese open-weight models, Western institutions and major state enterprises will build their own homegrown national/regional models, splintering the AI ecosystem.
“we're going to end up with a balkanization though where you know Western companies won't want to use deep secet type models”
Your call:
Numbers that matter
DeepSeek's inference was described as roughly 96% cheaper per query than OpenAI's o1Emad's estimate of the shock factor behind R1's pricing versus o1
OpenAI reportedly spent about $3 billion on model training in the prior year, out of roughly $5 billion in total losses against $3 billion revenueEmad's cost comparison to contextualize DeepSeek's low training spend
DeepSeek claimed to use about 2,000 H800 GPUs for the R1 training run; Emad estimates the company likely has closer to 10,000 chips in totalDisputing accusations that DeepSeek hid a much larger fleet of GPUs
Nvidia's new GB200 NVL72 rack has an interconnect bandwidth of about 6 petabits per second (compared to roughly 100 kilowatts of power draw)Emad describing Nvidia's fully-integrated data-center box design
Nvidia's Digits desktop AI box costs about $3,000, has 128GB of VRAM and 1 petaflop of AI compute; two units can run DeepSeek R1 locallyIllustrating how consumer-accessible high-end inference has become
Emad estimates the full DeepSeek R1 training run required about 1,000 megawatt-hours of energy, at roughly $15/MWh in the USArguing the training run is cheap enough to be solar-powered
The Stargate initiative's headline $500 billion figure is a total-cost-of-ownership number; Emad estimates real committed capital closer to $100 billionComparing Stargate's real scale to 5G buildout spending
Global GDP heading into 2025 is roughly $110 trillion, split roughly evenly between physical and intellectual laborFraming the addressable market for AI-driven knowledge-work displacement
A Unitree humanoid robot costs about $6,000; Peter estimates ~$150/hour amortized cost while Emad estimates closer to $0.40/hourDebating the effective hourly cost of humanoid robot labor
The video game industry grew from about $70 billion to $180 billion over the last decade while Hollywood grew from about $40 billion to $50 billionEmad's comparison of AI-era creative industries' growth trajectories
About 38% of IIT (India) graduates went unplaced in campus recruiting the prior summerCited as an early signal of AI-driven disruption to outsourced knowledge work
A cited Nigerian school pilot reportedly achieved about two years of learning gains in math after roughly two weeks of ChatGPT-assisted tutoringEvidence for AI's potential to accelerate education outcomes
Worth digging into
🕳️ DeepSeek's true GPU count and training cost
The claimed ~$5.6M / 2,000-GPU figures versus Emad's own estimate of ~10,000 chips are central to whether US export controls actually worked or were circumvented.
🕳️ Emad Mostaque's 'When Capital No Longer Needs Labor' paper
It's the intellectual foundation for his Intelligent Internet venture and his Fed-mandate-breakdown argument, but only summarized verbally here.
🕳️ Intelligent Internet's compute-backed digital currency
Emad teases an 'institutional grade digital currency' and a people-powered mining mechanism to fund universal basic AI but withholds specifics.
🕳️ The Nigeria ChatGPT-tutoring learning-gains study
Two years of math learning gains in two weeks is an extraordinary claim that would be one of the strongest data points for AI-driven education.
🕳️ Anthropic's 'sleeper agents' model-poisoning research
Cited as evidence that a handful of poisoned training examples can flip a model's behavior, raising supply-chain concerns about open-weight models and even hardware built by Chinese firms.
🕳️ OpenAI's alignment-team departures
Stephen Adler's resignation is one of several safety-team exits right as competitive pressure from DeepSeek intensifies -- a possible leading indicator of internal safety-vs-speed conflict.