DeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers with Travis Kalanick and David Sacks
Speakers
Chamath Palihapitiya
— CEO, Social Capital
Travis Kalanick
— Founder & CEO, Atoms
David Friedberg
— CEO, The Production Board
Jason Calacanis
— Angel Investor / Founder, LAUNCH
David Sacks
— General Partner, Craft Ventures
Summary
Travis Kalanick joins the All-In hosts, and White House AI and crypto czar David Sacks drops in, for an episode dominated by DeepSeek's R1 release and the record Nvidia selloff it triggered. Sacks argues the $6 million training claim is misleading and that the frontier still has value, while Chamath focuses on DeepSeek bypassing CUDA with PTX and on Nvidia chips flowing to China through Singapore. The second half covers OpenAI's reported $40 billion raise, DOGE's first ten days and its effect on federal spending and long-end treasury yields, and the economics of autonomy, electric power and urban real estate.
- DeepSeek's R1 release triggered the largest single-day market cap loss in Nvidia's history, with TSMC, Arm and Broadcom also selling off.
- David Sacks argues the $6 million training figure is not an apples-to-apples comparison and that DeepSeek sits on roughly 50,000 Hopper GPUs, and that some distillation from OpenAI almost certainly occurred.
- Chamath highlights DeepSeek's GRPO algorithm and its bypass of CUDA via PTX as evidence that Nvidia's software moat can be engineered around, pointing to a more heterogeneous compute future.
- Debate over whether model companies get commoditized, with Jevons paradox cited as the reason cheaper AI could raise aggregate AI spending rather than reduce it.
- Chamath's Singapore analysis suggests a large share of Nvidia revenue booked there cannot be consumed locally, making export-control enforcement a live risk, while Friedberg warns controls push China to build its own fabs and copy ASML.
- Data and content ownership are framed as the durable moat: YouTube's video library, Tesla's camera fleet, and buying publishers such as Reddit, the New York Times and Disney outright.
- DOGE's first ten days cover buyouts, return-to-office, canceled federal leases and building sales, with Friedberg arguing spending cuts would pull long-end treasury yields down and Chamath warning about the refinancing wall and weak auctions.
- Travis Kalanick on autonomy: cheap AI makes cheap autonomy, Tesla holds the manufacturing edge, electric grid capacity is the binding constraint, and freed parking land could reprice urban real estate.
Ideas
CUDA moat broken; compute goes heterogeneous
DeepSeek bypassed CUDA entirely by writing close to the bare metal with PTX, and it also threw out the orthodox PPO reinforcement-learning approach for its own memory-light algorithm (GRPO). Chamath has repeatedly said CUDA is Nvidia's biggest moat but also its biggest lock-in threat factor, and this proves the moat can be engineered around. With so much money and risk riding on a single chip and a single high-level framework, he expects the industry to move toward a more heterogeneous compute environment rather than accept one point of failure.
Meta must embrace and extend DeepSeek
The pressure now sits on Meta: it has to ship a next iteration of Llama that beats Gemini and exceeds R1 so the West has a credible open-source counterweight to whatever China releases next. More importantly, Meta should embrace and extend what DeepSeek published rather than simply buying tens of thousands of Nvidia GPUs, adopting the lower-level approaches such as PTX instead of high-level frameworks that DeepSeek proved out, and making itself the place developers and applications gather.
Cheaper AI drives much more demand
Countering the bearish read on Nvidia: when AI gets cheap there will simply be a lot more AI. Travis argues price elasticity here is positive, so as the cost per unit of intelligence falls, usage, revenue and total spend all rise. He compares it to cheap oil in the United States enabling factories and the industrial revolution, and expects AI to also specialize by vertical task, which creates more distinct workloads rather than fewer.
Export bans push China to self-supply
As with sanctions on Russia and earlier sanctions regimes, closing the floodgates just pushes the buyer to create a market somewhere else. Friedberg argues that cutting China off from Nvidia chips and US exports invites the second-order effect of China using stolen and copied IP to build out its own fabs and to find ways to replicate ASML's lithography technology. If any group in history could pull that off it is modern China, so export controls risk manufacturing the competitor instead of containing it.
AI-designed chips need simpler lithography
It is worse than China copying the toolchain: today's models are already capable of designing chips that do not rely on ASML's most complicated machines. Chamath points to Groq deliberately designing its chip at 14 nanometer, an old and simple process, and questions whether yields at leading-edge 2 nanometer nodes are good enough to justify the spend. If China is forced to engineer around the restrictions, it will use AI to design chips that can be manufactured with simple equipment, which undercuts the case for the most advanced lithography.
Own content IP, not more GPUs
Past a certain point more H100s stop conferring any advantage, and the durable moat becomes owning proprietary IP and content. Jason's play is to buy the owners of large proprietary corpora, naming Reddit, Quora, the New York Times, the Washington Post and Disney, then withhold that content from other model trainers and litigate, patent-troll style, against anyone who has already absorbed it. The most proprietary library wins.
YouTube video data is Google's moat
Text is only a fraction of what matters; video is where the scarce training data sits, and Google's YouTube content library is probably 100 to 200 times larger than the rest of the internet combined. Friedberg says Google holds the rights to a good chunk of that user-generated content, is already using it and is doing so legally, which is a data moat that compounds into better products and therefore more data.
Tesla's camera fleet data compounds advantage
Tesla is the other example of a proprietary data body: it was pressed years ago to put cameras on everything, and that fleet data is what lets it build models that do self-driving. Friedberg argues data advantages like this, which arise in specific industry segments rather than uniformly, are where the moat lies, and that the moat produces better products that in turn gather more data, a more persistent advantage than owning the biggest data center network.
Deficit cuts pull long yields down
The US runs roughly a $2 trillion annual deficit and needs to cut about a trillion to get the federal deficit below 3% of GDP. Friedberg argues that as spending is cut, the risk premium on US debt falls and long rates come down: the 30-year is near 5% even while the Federal Reserve is cutting, because investors keep selling treasuries over doubts about 30-year debt service. The 30-year already eased from its 5% January 13 peak to about 4.77% as the administration took action, and he thinks it can fall significantly further if DOGE cuts survive the courts. The faster the cuts are made, the fewer cuts are ultimately required.
DOGE lease cancellations flood office market
Federal buildings are next-level empty, and unlike startups, which landlords force into five or ten year terms, the government is treated as such a reliable client that it sits on rolling one-year leases, so DOGE can simply cut them. Combined with return-to-office attrition and the buyout take-up, that terminated space plus the property the government intends to sell will flood the market, which is why he is glad not to own buildings leased to the federal government.
Commoditized autonomy leaves Tesla's manufacturing edge
Cheap AI makes cheap autonomy: as good AI proliferates, self-driving gets easier and eventually commoditizes the way models are commoditizing, which Travis already sees in Tesla FSD improving roughly 10x in miles per human intervention inside a three-month window. Once the software advantage compresses, the binding constraint becomes building the hardware at scale, and that is where Tesla has a huge advantage over rivals that still have to line up manufacturing partners to get volume on the road.
Power grid is autonomy's binding constraint
The dark horse constraint on autonomous ride-hailing is electricity. Travis's back-of-the-envelope math is that moving all California miles to EV ride-sharing would require doubling the state's energy capacity, and even adding 10-20% is a five to ten year exercise; his own affluent Los Angeles neighbourhood loses power constantly because the grid is only patched as it breaks. His hot take is that combustion-engine AVs may therefore scale first, and he says energy storage, grid upgrades and modular capacity additions are now the long pole in the tent for construction and development in most of the cities where his facilities operate.
Uber's AV bet needs charging real estate
Uber's bet is that cheap democratized AI produces cheap democratized autonomy, which is why it takes every AV provider into the network rather than owning the stack. Travis says the follow-on requirements are lining up the physical hardware partners among car manufacturers and then solving electricity, so there is a large real estate and fleet-management play in electrifying parking lots and configuring them so robots can clean, charge and service the cars efficiently.
Autonomy frees parking land, crushing prices
If autonomous ride-sharing scales, car ownership drops like a knife and the cars that remain on the road are utilized roughly 15x more, so the fleet shrinks by something like 10x. Parking, which takes 20-30% of all land in a city, is then largely unnecessary, releasing about a fifth of urban land into supply. Travis's conclusion is that land prices have to come crashing down, with city planning rebuilt around constraints that no longer exist, and the surplus physical inventory repricing value that sits in household 401ks and pension fund balance sheets. The caveat is timing: the electric grid could make it a slow burn, and you have to evade the falling knives first.
Refinancing wall may push yields higher
The long end of the yield curve is still telling us there is a chance of inflation, and the deeper problem is the refinancing wall left behind: so much short-term paper was issued that nearly 30% of the debt has to be refinanced this year at about 5%, while the last auction barely had 2x coverage, which could take a lot of the energy out of the market. Chamath cites a senior capital-markets contact who expects the 30-year to reach 5.5% before it comes down, and argues that in total dollars of interest, 5.5-6% today is equivalent to 10-11% rates twenty years ago, so things start to break well before those levels. Only fast, credible deficit cuts pull the market back.
Up Next
This All-In Podcast video, published January 31, 2025, features Chamath Palihapitiya, Travis Kalanick, David Friedberg, Jason Calacanis discussing NVDA, META, ASML, RDDT, NYT, DIS, GOOGL, TSLA, TLT, Office real estate, Electric grid infrastructure, UBER, XLRE. 15 trade ideas extracted by AI with direction and confidence scoring.
Speakers: Chamath Palihapitiya, Travis Kalanick, David Friedberg, Jason Calacanis · Tickers: NVDA, META, ASML, RDDT, NYT, DIS, GOOGL, TSLA, TLT, Office real estate, Electric grid infrastructure, UBER, XLRE