All-In PodcastDeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers with Travis Kalanick and David Sacks
CHAPTERS
- 0:00 – 7:10
CloudKitchens, Ray Dalio, and the ‘Future of Food’ Vision
The episode opens with banter about Friedberg’s surprise Ray Dalio interview before pivoting to Travis Kalanick’s return to the public stage. Kalanick outlines CloudKitchens as an infrastructure company—combining real estate, software, and robotics—to make high-quality, low-cost, ultra-convenient meals that rival grocery economics.
- •Ray Dalio’s new book argues the U.S. must cut deficits to ~3% of GDP; Friedberg frames this as highly relevant to the current administration.
- •Kalanick defines CloudKitchens as building ‘the future of food’ over a 10–100 year horizon: high quality, cheap, personalized, delivered, and mostly machine-made.
- •Cooking becomes a ‘hobby’ like riding horses; most weekday meals will be ordered, personalized, and prepared robotically.
- •CloudKitchens’ Bowl Builder robot automates bowl-style meals (Chipotle/Sweetgreen analogs), from ingredient dispensing to sealing and staging.
- •Orders flow via DoorDash/Uber Eats; restaurants prep bulk ingredients in the morning, then leave robots to handle high-precision final assembly and handoff.
- 7:10 – 31:00
Robotic Bowls, Nutrition Data, and Fully Wired Food Supply Chains
The conversation dives into CloudKitchens’ robotics stack and how it connects to personalization, health data, and agricultural supply chains. Kalanick and Friedberg compare today’s automation efforts to early 20th-century Automats and past bowl-focused concepts like Eatsa.
- •The Bowl Builder logs exact grams of each ingredient per order and can send the composition and a pre-lid photo to customers mid-delivery.
- •Chamath imagines future flows where you authenticate Apple Health or share a lipid panel; consumer-facing brands customize macros and CloudKitchens executes.
- •Kalanick positions CloudKitchens as behind-the-scenes infrastructure (like AWS/NVIDIA) rather than the face brands; they ‘serve those who serve others.’
- •They sketch a fully wired supply chain: from robotic assembly, to Sysco/US Foods, to mechanized farming with granular provenance (field, organic status, etc.).
- •Friedberg recounts Eatsa’s similar commissary + canister system, early automation attempts with Chipotle/Sweetgreen, and why retrofitting human-optimized kitchens is costly and disruptive.
- •CloudKitchens’ delivery-only facilities are purpose-built for automation and asynchronous prep, avoiding the capex and downtime of retrofits in legacy QSR layouts.
- 31:00 – 43:00
Sacks from the White House: DeepSeek R1 and the Global Freakout
David Sacks joins from DC, describing the White House complex and quickly turning to DeepSeek R1’s release and Wall Street’s panicked reaction. He frames DeepSeek as a Chinese, open-source reasoning model that unexpectedly matched OpenAI’s o1, triggering both geopolitical and open-source-versus-closed debates.
- •DeepSeek R1 is a Chinese reasoning model comparable to OpenAI’s o1, but open-sourced and allegedly trained for only $6M.
- •NVIDIA suffered a record single-day market-cap loss (~$600B) as markets questioned whether massive GPU spend was being structurally undercut.
- •Sacks distinguishes base models (‘smart PhD’ giving direct answers) from reasoning models (‘smart PhD who goes off and does the work’ via chain-of-thought and RL).
- •Industry players (Google, Anthropic, Meta) have similar reasoning models in development, but DeepSeek was the first big non-U.S. release.
- •China’s apparent AI gap compresses from a perceived 6–12 months behind to perhaps 3–6 months with R1’s launch.
- 43:00 – 55:00
Debunking the $6M Training Myth and Following the GPUs
Sacks challenges the viral narrative that DeepSeek achieved parity with Western frontier models for just $6M. He and Chamath dissect hardware estimates, export controls, and potential backdoors for NVIDIA chips into China, arguing that the real story is a massive, long-prepped compute cluster plus clever engineering.
- •The widely cited $6M is almost certainly only the final training run, not full R&D or hardware; apples-to-apples, OpenAI’s final runs were tens of millions.
- •Analyst Dylan Patel estimates DeepSeek plus its related hedge fund own ~50,000+ NVIDIA ‘hopper’ GPUs (H100/H800/H20), implying >$1B in hardware.
- •Many of those chips were legally acquired earlier, before export controls tightened; further unreported hardware can’t be ruled out.
- •Chamath notes incentives: NVIDIA bulls want to debunk the $6M; DeepSeek wants to highlight it. Regardless, the compute footprint is clearly non-trivial.
- •Emerging evidence suggests Singapore may be a re-export hub for NVIDIA chips into China, complicating enforcement of U.S. export controls.
- 55:00 – 1:06:00
DeepSeek’s Technical Breakthroughs: GRPO, PTX, and Constraint-Driven Innovation
The panel explores why DeepSeek’s engineering choices matter beyond cost: a new RL algorithm and low-level GPU coding that break from Western orthodoxy. They argue that constraints in China—limited top-tier GPUs and CUDA lock-in—led to optimizations Western teams may have overlooked because they had too much capital and compute.
- •DeepSeek replaces orthodox PPO reinforcement learning with GRPO, a memory-efficient, high-performance alternative devised under compute constraints.
- •They bypass CUDA (NVIDIA’s proprietary stack and key moat) to program directly in PTX assembly, squeezing more performance from GPUs.
- •Chamath sees this as a warning shot for CUDA lock-in and a catalyst for more heterogeneous hardware and software stacks beyond NVIDIA.
- •Constraint is framed as a forcing function: with huge funding, Western teams follow the ‘easy’ path; with scarcity, DeepSeek innovates at lower levels of the stack.
- •Meta is urged to ‘embrace and extend’ R1-era techniques for LLaMA, both to remain an open-source counterweight and to avoid single-vendor dependence.
- 1:06:00 – 1:24:00
Did DeepSeek Distill OpenAI? IP, Cloud Security, and Open Source Fallout
The hosts tackle the simmering allegation that DeepSeek heavily distilled OpenAI’s models—transforming ChatGPT outputs into training data for its own systems. They review behavioral evidence, OpenAI’s public claims, and the awkward position of cloud providers hosting potentially stolen IP while also selling access to the original vendors.
- •DeepSeek V3 sometimes self-reported as ‘ChatGPT-4’ when queried, implying substantial training on ChatGPT outputs.
- •Sacks notes two pathways: large-scale web scraping of ChatGPT content (arguably ToS-compliant) or hammering OpenAI’s API (violating ToS).
- •OpenAI has told the Financial Times it found evidence DeepSeek improperly used its proprietary models for training.
- •A viral clip shows R1 starting to answer with ‘China’ then deleting it—clear evidence of distilled chain-of-thought plus post-hoc censorship on sensitive entities.
- •Microsoft now hosts R1 on Azure—even as OpenAI claims it’s stolen IP—illustrating how open-sourcing a model lets every cloud vendor undercut the originator.
- •Chamath warns against overreacting with heavy-handed KYC-style controls on model APIs, which would slow innovation and diverge from how general cloud infra is used.
- 1:24:00 – 1:51:00
Where Is the Value in AI? Shims, Apps, Data, and Mixture-of-Experts
The besties shift from model drama to business strategy: if frontier LLMs are racing toward commodity status, where should founders and investors focus? They debate shims that abstract multiple models, data moats, application-layer businesses, and the likely rise of many small expert models over single giant ones.
- •Chamath argues the first technical priority now is a ‘shim’ layer—so developers can hot-swap models (o1, R1, LLaMA, etc.) without rewriting their stack.
- •Jason and Friedberg see enduring value at the application layer (YouTube vs. storage analogy, Uber vs. GPS chip makers).
- •Data becomes the critical moat: Reddit/Quora/NYT/Disney text and video, Tesla driving data, YouTube’s vast library, and domain-specific corpora.
- •Chamath expects mixture-of-experts architectures and networks of small specialist models to dominate over monolithic, do-everything LLMs.
- •Sacks pushes back slightly, arguing frontier labs still have incentive to push ahead and will take countermeasures against distillation.
- •Travis frames the opportunity in ‘tools vs wrappers’: tools (platforms/infra) for AI developers and end-user-facing wrappers with strong UX and distribution.
- 1:51:00 – 2:18:00
China, Copying, and the Evolution into an Innovation Powerhouse
Kalanick recounts Uber’s China war as a case study in how ferocious copying morphs into genuine innovation. They then tie this to current Chinese leads in delivery, locker systems, and AI, and discuss whether export controls will slow or merely redirect China’s advance.
- •Uber’s China team saw Didi copy every major Uber feature within weeks, often at high quality—forcing Uber to build a 400-person China-focused org in SF.
- •At some point, ‘you run out of things to copy’ and must innovate; Chinese companies now lead in many logistics and last-mile systems.
- •Examples: dense locker networks at office buildings for food and parcels; inter-office runners; Meituan experimenting with drones, etc.
- •Kalanick argues if you want to see the future of food delivery, visit Shanghai, not New York.
- •Friedberg and Chamath question the efficacy of export controls, predicting China will build its own fabs and, increasingly, use AI itself to design chips on older, easier-to-produce process nodes.
- •DeepSeek’s use of PTX and non-cutting-edge nodes (analogous to Groq’s 14nm strategy) underscores that cutting-edge EUV is not strictly necessary for competitive AI hardware.
- 2:18:00 – 2:54:00
OpenAI’s $40B Raise, Masa’s Style, and the Perils of Overcapitalization
Rumors surface that OpenAI is raising $40B at a $340B valuation, possibly led by SoftBank’s Masayoshi Son. Travis draws on his experience competing with SoftBank-funded rivals to warn about both the power and risks of taking such capital, while the group questions whether sheer hardware scale is still a moat.
- •Reported deal: OpenAI targeting $40B at $340B pre-money, with Masa as a potential lead; ‘Stargate’ AI supercomputing plans loom in the background.
- •Travis clarifies Uber never took Masa’s money directly; SoftBank instead funded Uber’s competitors and used their intel promiscuously across its portfolio.
- •He frames taking SoftBank-scale money as a double-edged sword: you either accept it or fight a capital-armed ecosystem against you.
- •Overcapitalization can weaken frontier labs by reducing constraint-driven discipline; too much easy money can make organizations bloated and slow.
- •The group questions whether ‘more H100s’ is still a durable moat, versus owning data/IP and distribution to billions of users (e.g., Meta’s reach).
- 2:54:00 – 3:41:00
Autonomy, Power Constraints, and the Coming Commercial Real Estate Shock
Kalanick and the besties extrapolate what widespread autonomy and AI mean for cities, power grids, and real estate. They argue that the real choke points may be electricity and physical infrastructure rather than models or chips, and that parking-heavy land use could be radically disrupted.
- •Kalanick estimates that if all miles in California shifted to EV ride-hail, the state would need roughly double its power capacity; even +10–20% is non-trivial.
- •He highlights frequent power outages even in affluent LA neighborhoods, underscoring grid fragility.
- •Robotaxis could reduce car ownership and parking needs by ~10x; parking currently consumes 20–30% of urban land, which could become surplus.
- •Chamath points out that massive, sudden surplus in commercial and parking real estate would hammer asset values from 401(k)s to pension funds.
- •Kalanick sees long-term value in real estate that is electrified and roboticized for fleet management: charging, cleaning, and maintenance depots.
- •They note Doge-driven federal downsizing may simultaneously dump a large volume of government-leased office space onto the market, compounding the CRE reset.
- 3:41:00 – 4:16:00
DOGE, Debt, and the Politics of Cutting $1–3 Billion a Day
The focus shifts squarely to the Department of Government Efficiency (DOGE) and the Trump administration’s early moves to slash federal spending. The hosts connect buyouts, RTO mandates, and lease terminations to broader questions of deficits, interest rates, and the constitutional limits of executive power.
- •DOGE claims ~$1B/day in taxpayer savings via voluntary buyouts (~8 months pay), mandated return-to-office, and cancelling underused leases; they aim to triple this.
- •Friedberg invokes Dalio: the U.S. must reduce deficits to <3% of GDP (~$1–1.1T in cuts) to avoid a debt spiral; faster cuts mean smaller eventual cuts, due to falling rates and improved confidence.
- •Chamath frames DOGE as a three-layer onion: (1) people (buyouts/attrition), (2) physical infrastructure (selling/terminating leases), (3) IT and services (the largest, most opaque layer).
- •They highlight a striking counterfactual: if 2019 federal spending levels were applied to 2024 revenues, the U.S. would run a ~$500B surplus instead of a $1.5T deficit.
- •Political economy: most members of Congress aren’t personally prioritizing cuts; DOGE must instead use executive authority to ‘slow roll’ spending and impose friction, subject to court rulings.
- •Trump’s popularity is at a personal high yet still historically low; Jason warns that 2.0 policy focus (Doge, border) is popular, while 1.0-style grifting and culture-war attacks risk losing the marginal supporters he just gained.
- 4:16:00 – 4:50:00
Interest Rates, Treasuries, and Why DOGE Must Succeed Quickly
The hosts link DOGE’s cuts to bond markets and interest-rate dynamics, arguing that fiscal credibility directly affects long-term yields. They warn that a move to 5.5–6% 30-year yields would be equivalent to double-digit rates on the early-2000s debt stock—economically devastating without swift action.
- •30-year U.S. Treasuries recently hit ~5.0% and have eased to ~4.77%; markets are still nervous despite Fed holds and talk of cuts.
- •Chamath quotes a capital-markets contact who expects 30-year rates to reach 5.5% before falling, which would be tantamount to ~10–11% in early-2000s terms given today’s much larger debt base.
- •Roughly 30% of U.S. debt must be refinanced this year; thanks to heavy issuance of short-term paper, the government is highly exposed to rollover risk at higher rates.
- •Friedberg notes that credible, fast deficit reduction would both reduce inflation pressure and boost confidence in U.S. solvency, pulling long-term yields down and softening the required cuts.
- •Mandatory spending (Social Security, Medicare/Medicaid) remains politically untouchable; DOGE must therefore squeeze discretionary and operational waste, and rely on public ‘naming and shaming’ to build support.
- 4:50:00
Aviation Tragedy, Outdated Systems, and the Case for Automation
The episode closes somberly with reflections on a recent DC-area aviation accident. The hosts relay feedback from commercial pilots and autonomy entrepreneurs about the risk profile at Reagan National (DCA) and the antiquated nature of U.S. air traffic control, arguing that modern software and automation could prevent such tragedies.
- •An anonymous airline pilot calls DCA ‘the sketchiest airport’ they fly into, with controllers ‘fast and loose’ amid intense traffic and difficult helicopter visibility.
- •Brian Yutko (Wisk) notes that collision-avoidance systems exist but generally do not override pilot control; fighter jets’ automatic ground-collision avoidance shows what’s possible if software is allowed to take over.
- •He argues for two main upgrades: (1) automated intervention that can override pilots to prevent collision, and (2) modernizing ATC from VHF voice comms to robust digital datalinks and better software.
- •The U.S. hasn’t had a major commercial airline disaster in ~25 years, but this incident underscores that legacy systems and union politics can slow adoption of safety-enhancing tech.
- •The hosts suggest aviation is another domain—like government IT and industrial food—where applying modern AI and automation could yield massive safety and efficiency gains.
