All-In · 2025-01-31 · 109 min

DeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers with Travis Kalanick and David Sacks

Travis KalanickDavid Sacks

YouTube ↗

TL;DR

DeepSeek’s R1 release cut the market’s estimate of China’s AI lag from six-to-12 months to roughly three-to-six months, challenging the scarcity premium around frontier models. David Sacks called R1 comparable to OpenAI’s o1 and its API roughly one-twelfth the cost, but rejected the “$6 million versus $1 billion” framing: $6 million reportedly covered only the final training run, while one analyst estimates DeepSeek and its affiliated hedge fund control 50,000 Hopper GPUs worth more than $1 billion.
NVIDIA’s 17.7% plunge and roughly $600 billion market-cap loss exposed the core semiconductor debate: does cheaper model training destroy compute demand or multiply it? Chamath Palihapitiya argued DeepSeek’s use of GRPO and bare-metal PTX showed that constrained teams can route around memory-heavy orthodoxy and CUDA lock-in. Travis Kalanick offered the counterweight: “When AI gets cheap…there’s going to be a lot more AI,” while Sacks invoked what he called Jin’s Paradox.
The DeepSeek story combines genuine engineering innovation with unresolved evidence of distillation from OpenAI. Sacks said V3 identified itself as ChatGPT-4 in five of eight tests and noted that DeepSeek disclosed roughly 800,000 reasoning samples without clearly explaining their source; he nevertheless called the team technically strong. The innocent explanation is training on publicly posted ChatGPT output, while the disputed one is mass use of OpenAI’s API—an uncertainty the panel refused to smooth over.
AI equity value may migrate from interchangeable models toward routing layers, applications, proprietary data and physical execution. Chamath would first build a “shim” that can hot-swap OpenAI, Claude, Llama, DeepSeek or a future R2; Jason Calacanis argued models are becoming infrastructure like storage or GPS; and Kalanick separated the opportunity into “wrapper,” tools and vertical-specialist businesses. Sacks pushed back that OpenAI’s o3 frontier remains ahead of R1 and that declaring closed-model returns dead is “a little premature.”
OpenAI’s reported attempt to raise $40 billion at a $340 billion pre-money valuation is the cleanest test yet of whether more capital creates a moat or institutional softness. Kalanick called access to capital a strategic weapon but warned that overcapitalization can make companies “too bureaucratic, too loose, too weak, too soft” while “a thousand flowers” bloom in open source. His experience with Masa: refuse the money and it may subsidize competitors; accept it and assume the investor’s intelligence will inform other bets.
Export controls may slow China while simultaneously forcing the engineering workarounds that make it more self-sufficient. Chamath highlighted unverified claims that up to a quarter of NVIDIA revenue flows through Singapore, whose roughly 100 data centers consume about 876 MW, and questioned where the chips ultimately land. The panel’s harder policy problem: restrictions could push China toward domestic fabs, simpler process nodes and AI-designed chips—“constraint as a feature, not a bug.”
DOGE’s claimed $1 billion of daily savings matters most through the Treasury market, not the headline arithmetic. The panel framed the fiscal objective as cutting roughly $1 trillion to $1.1 trillion from a $2 trillion annual deficit to get below 3% of GDP; faster cuts could lower inflation and long-term yields, reducing future interest expense. But courts must determine how much congressionally mandated spending the executive can stop, while Social Security, Medicare and Medicaid remain far harder than leases, headcount or discretionary procurement.
Cheap AI could accelerate autonomous transport and robotic food production, but electricity and real estate may become the binding assets. Kalanick’s Bowl Builder was rolling out with five customers in April, while his Waymo thesis is that “cheap good AI makes cheap good autonomy”; yet converting all California ride-share miles to EVs could, by his rough calculation, require doubling the state’s energy capacity. If autonomous fleets need roughly one-tenth as many parked cars, 20%-30% of urban land could be repriced—making charging depots, grid access and repurposable parking more consequential than another model benchmark.

Digest · the substance, structured for research

1. Food automation needs purpose-built infrastructure, not a robot wedged into an old restaurant

Kalanick’s hundred-year endpoint is food that is high-quality, inexpensive, convenient and precisely matched to each person’s dietary preferences, with machines making and delivering it at a cost approaching—or surpassing—the grocery store. People will still cook, but as a soulful hobby: “I love horses, but I don’t ride a horse to work.”
CloudKitchens is his real-estate, software and robotics stack for that transition—“either the AWS or the Nvidia…for food.” Restaurants remain customer-facing; CloudKitchens serves “those who serve others,” handling the infrastructure that could eventually connect an order to health preferences, ingredient quantities, provenance and a photograph of the finished bowl.
The Bowl Builder lets a restaurant prep ingredients in the morning and leave. DoorDash or Uber Eats orders then trigger hot and cold dispensing, sauces, bagging, utensils, sealing, conveyor movement and an automated locker; a courier waves a phone at a camera and only the correct compartment opens. Five customers were scheduled to begin using the machine in April.
Friedberg had pursued a similar canister-based system years earlier, with working demos, but it never reached production. Kalanick’s distinction is deployment: retrofitting a human-optimized QSR can require major capex and two-to-three months closed, while delivery-only kitchens were designed around the machine. One 800-square-foot restaurant did about $3 million annually and reportedly handled roughly 800 orders an hour at lunch.

2. DeepSeek changed the perceived AI race before anyone resolved its cost claims

The immediate market verdict was brutal: NVIDIA fell 17.7%, erasing roughly $600 billion, while TSMC, Arm and Broadcom also sold off. The catalyst was DeepSeek’s claim that R1 reached performance comparable to leading Western reasoning models after a final training run costing about $6 million on only 2,000 GPUs.
Sacks argued the story became unusually explosive because two conflicts landed together: United States versus China and closed models versus open source. Remove either dimension and it probably does not become a global story capable of erasing close to a trillion dollars in market value.
His substantive update was narrower but important. Before R1, industry estimates placed China six-to-12 months behind; because R1 followed o1 by roughly four months, he would now call the gap three-to-six months. A Chinese company becoming the second major public release of a reasoning model was “legitimately surprising.”
The $6 million comparison, however, mixed a final run with fully loaded Western R&D, hardware and operating costs. Sacks cited an estimate of 10,000 H100s, 10,000 H800s and 30,000 H20s across DeepSeek and its founder’s hedge fund—a 50,000-plus-Hopper cluster costing more than $1 billion—while stressing that outsiders cannot empirically validate either total spending or undisclosed hardware.

3. Constraint produced engineering paths the compute-rich West had little reason to explore

Chamath’s first specimen was reinforcement learning. Rather than follow the prevailing PPO approach, DeepSeek used GRPO, which he said requires less memory while remaining highly performant. His inference was not that Western engineers lacked the ability, but that surplus compute gave them little reason to abandon orthodoxy.
His second specimen was DeepSeek’s route around CUDA. The team worked through PTX, which Chamath compared to writing assembly and controlling the NVIDIA hardware closer to bare metal. That weakens the assumption that CUDA’s high-level ecosystem is an unassailable moat and points toward a more heterogeneous hardware and software environment.
The disagreement remained useful: semiconductor bulls have incentives to discredit the $6 million claim, challengers have incentives to celebrate it, and neither side can fully audit DeepSeek. Jason called the team “badass,” while Sacks agreed that some of the work was technically brilliant and made constraint look like “a feature, not a bug.”

4. The distillation evidence is substantial, but it does not erase DeepSeek’s own work

Friedberg described distillation as using a large model’s answers to train a smaller, cheaper one. A demonstration in which DeepSeek began discussing China before deleting the answer illustrated post-generation censorship, but was not by itself proof of which model supplied the underlying reasoning.
Sacks’s stronger evidence came from V3: when asked its identity, it reportedly answered ChatGPT-4 in five of eight cases. That implies substantial exposure to ChatGPT output, but leaves two explanations—DeepSeek could have crawled publicly posted conversations, or it could have made large-scale API calls that violated OpenAI’s terms.
DeepSeek disclosed roughly 800,000 reasoning samples used in moving from V3 to R1 but remained hazy about where they originated. Sacks called the small sample count itself remarkable; Jason emphasized the white paper’s science and thoroughness, while Sacks agreed that some of the work was technically brilliant.
Jason’s pushback—worth keeping—was that OpenAI itself abandoned its original open-source posture while training on publishers, artists and other internet content, then objected when a Chinese rival allegedly reused its outputs. Sacks was sympathetic to OpenAI’s operational shock, while Chamath warned against answering it with universal KYC or cloud restrictions that would slow legitimate innovation.

5. Open source is simultaneously a community benefit and a Chinese catch-up strategy

Chamath put the immediate burden on Meta: the next Llama must “embrace and extend” DeepSeek’s improvements, attract developers and exceed Gemini and R1. In his framing, dependence on one chip, one high-level framework and one model family has become an unacceptable single point of failure.
Sacks resisted the leap from R1 to inevitable model commoditization. R1 was comparable to o1, released four months earlier and trained internally perhaps nine or ten months earlier; OpenAI had already moved to o3, while Google, Anthropic and Meta had their own reasoning work. “It’s a little premature to conclude that there’s no reward for being at the frontier.”
Nor did Sacks accept the image of a purely altruistic “plucky upstart.” A Chinese company that is behind has a strategic reason to open-source capable models and undercut leading American firms. Jason’s reconciliation was cleaner: DeepSeek can be advancing open access and pursuing geopolitical self-interest at the same time.

6. Model churn moves the investable moat upward, downward and sideways

Friedberg asked where equity value belongs if capable models become fast, cheap and broadly available. His electricity analogy: the utilities did not capture all the gains from electrification; much of the value accrued to the factories, products and wider economy that cheap power made possible.
Chamath’s first startup requirement would be a model-neutral “shim.” Engineers should not become locked to Claude, OpenAI or Llama; a company must be able to substitute R1, a future R2, an Alibaba model or another leader without rebuilding its application whenever benchmark leadership changes.
Jason pushed hardest toward applications, comparing models with storage beneath YouTube or GPS beneath Uber. He revived Gavin Baker’s phrase that “the fastest depreciating asset in the world was a large language model,” arguing that application distribution and real-world hardware will retain value longer.
Kalanick divided the opportunity into wrapper, tools and specialized vertical AI, then challenged the simple NVIDIA bear case: “When AI gets cheap…there’s going to be a lot more AI.” Sacks connected that to what he called Jin’s Paradox—lower unit cost can unlock so many economically feasible uses that aggregate spending rises rather than falls.

7. China’s copy cycle has already flipped into operating-model innovation

Kalanick remembered Uber China as an “all-out war” in which a painstaking product launch could be copied within two weeks, then one. Uber staffed an entire San Francisco floor with roughly 400 Chinese nationals and advertised in Chinese along Highway 101 to recruit people “to serve the homeland.”
His conclusion changed over time: extreme copying compresses the gap until there is nothing left to copy, after which the practiced organization turns toward creativity. For the future of online food delivery, he would now study Shanghai rather than New York; features reaching Uber Eats or DoorDash may have existed in China three or four years earlier.
The best specimen was the office-building handoff. Hundreds of perimeter lockers receive food and packages from couriers, while a second class of runners completes delivery inside the building—an “epically efficient” system adapted to local labor economics.

8. Export controls may be creating the very workarounds they were meant to delay

Chamath highlighted claims—explicitly still speculation—that up to a quarter of NVIDIA revenue was associated with Singapore and that chips could be moving onward to China. Singapore is only about 250-260 square miles; its roughly 100 data centers consume approximately 876 MW, and the entire local data-center industry was described as only a $1.5 billion-to-$2 billion revenue business.
The inference was not proof that every chip leaves Singapore. It was a policy question: if export controls have an easily used shell-company route, the administration must determine where the hardware actually ends up and whether closing one channel changes outcomes or merely redirects trade.
Friedberg’s second-order objection was that sanctions encourage China to reproduce fabs, supply chains and eventually ASML-dependent capabilities. If any modern industrial system can coordinate the required effort, he argued, China is a plausible candidate.
Chamath made the workaround more concrete: AI can help design chips for older, simpler manufacturing processes. Groq’s choice of 14 nanometers—“VHS and Beta” technology in his analogy—showed that useful architectures need not depend on cutting-edge two-nanometer yields. The alternative Chinese strategy is centralized frontier-model capex, followed by subsidized distillation for domestic firms.

9. OpenAI’s proposed mega-round tests capital as both weapon and anesthetic

The reported transaction was $40 billion at a $340 billion pre-money valuation, potentially led by Masa, alongside the Stargate narrative of Sam Altman, Larry Ellison and large-scale infrastructure. Jason framed the consumer contest as OpenAI versus Meta: ChatGPT has more than 1 billion monthly active users and hundreds of millions of daily active users, while Meta has billions of eyeballs and daily active users.
Kalanick corrected the record that Uber did not take Masa’s money during his tenure, despite years of pursuit. He regarded Masa as a “promiscuous investor”: information gained from backing one company could inform investments in every competitor. Uber instead secured a $3.5 billion Saudi investment before the Vision Fund existed.
The catch was that refusing the capital did not neutralize it; Masa’s money flowed into competitors including DoorDash and subsidized their markets. Kalanick’s rule: when access to capital is a competitive weapon, “you must play ball,” while understanding that the resulting intelligence and influence may be deployed elsewhere.

10. More data centers do not guarantee the strongest long-run moat

Kalanick’s warning for OpenAI was that excess capital can produce an organization that is “too big, too bureaucratic, too loose, too weak, too soft.” Against a decentralized open-source ecosystem where “a thousand flowers” bloom, sheer infrastructure may be an overwhelming advantage in some full-stack sectors and irrelevant in others.
Friedberg questioned the premise of one giant do-everything model. His alternative resembles a mixture of experts: repeatedly shrink copies into many smaller models, some specializing in mathematics, reading or writing, until a coordinated network uses less energy and time than a monolith.
Chamath nominated proprietary content as the durable leverage—Reddit, Quora, newspapers or Disney—while Friedberg argued text is only a fraction of the prize. He claimed YouTube’s video library could be 100-200 times larger than the rest of the internet’s and cited Tesla’s years of camera data as the kind of domain-specific moat that improves a product and generates still more proprietary data.
Friedberg’s caution was conditional: eventually data or algorithm quality becomes “the long pole in the tent,” and additional compute no longer repairs the constraint. Jason’s own licensing example was a blanket $2,500 offer to put his book into Microsoft’s training pool, which he favored to help establish a properly licensed market.

11. DOGE begins with people and buildings, but the large savings sit inside systems

Jason said DOGE was claiming approximately $1 billion of savings per day—about $3 per American daily or $1,000 annually—with an ambition to triple the pace. The early program combined return-to-office mandates, an eight-month resignation offer expected to attract perhaps 5%-10% of federal workers, hiring restraint and lease cancellations.
Friedberg saw execution of a previously advertised playbook, not a surprise; Bill Clinton had also used buyouts during his deficit-reduction program. The unsettled issue is authority: courts will decide how far the executive can suspend spending that Congress required by statute and what must return to the legislature.
Jason focused on controllable friction. Even where spending is authorized, the executive can make hiring, procurement and replacement harder, impose competency reviews and slow administrative processes. Return-to-office itself may produce substantial attrition if buildings are as empty as the panel had heard.
Chamath’s “three-layer onion” was people, physical infrastructure, then IT and services. The first two save money, but engineers embedded at the nucleus of DOGE teams can enter systems of record and forensically trace payments; he suspected that identified waste could ultimately exceed $2 trillion.

12. The Treasury market makes speed more valuable than the nominal size of each cut

The panel’s benchmark was a roughly $2 trillion annual deficit and Ray Dalio’s recommendation to bring it below 3% of GDP, implying approximately $1 trillion to $1.1 trillion of cuts. One comparison suggested that applying 2019 spending to 2024 revenue would yield a $500 billion surplus instead of a $1.5 trillion deficit—a $2 trillion swing.
Friedberg’s feedback loop was central: credible spending cuts reduce inflation pressure and increase confidence in 30-year repayment, lowering Treasury yields and future interest expense. “The faster you make the cuts, the less you have to cut.”
The market was not yet convinced. The 30-year yield had touched 5% on January 13 and eased to about 4.77%, while nearly 30% of federal debt was expected to refinance that year around current rates; Chamath noted one auction had barely two-times coverage, and a senior capital-markets contact expected 5.5% before relief.
Political constraints remained the bearish case. Congress members still wanted benefits for their own districts, while Social Security, Medicare and Medicaid cannot be slow-rolled like discretionary procurement. Chamath’s hope was that Elon Musk could popularize “use a pencil” solutions and name-and-shame waste; Jason argued that dividing each saving by 330 million Americans makes the benefit legible.

13. Waymo has crossed the psychological threshold from experiment to transportation

Kalanick contrasted his first Uber autonomous ride in Pittsburgh—leaving him unable to stand straight from adrenaline, with only a large red stop button—with Waymo now: “You’re not even thinking twice.” Normalization is itself part of deployment because riders increasingly experience the system as ordinary rather than theatrical.
He still viewed today’s retrofitted vehicles as transitional. Tesla’s Cybercab represents the destination: no steering wheel and an interior designed around passengers rather than a backup human driver. He said Tesla’s newer FSD models had improved perhaps tenfold in miles per intervention over one three-month period.
His causal chain was concise: cheap AI makes cheap autonomy, which could in turn commoditize driving intelligence. Manufacturing then becomes a harder bottleneck and a Tesla advantage, while Waymo’s manufacturing partners and Uber’s collection of autonomous-vehicle partners determine who can deploy physical fleets at scale.
On safety, Kalanick argued autonomous vehicles are becoming “provably safer” despite individual failures and also remove interpersonal risk between riders and drivers. Regulation should ease as cities accumulate ordinary experience, though Chinese systems such as BYD may face a separate barrier: whether the United States permits the technology to enter at all.

14. Electricity, depots and obsolete parking may matter more than the autonomous-driving stack

Kalanick’s back-of-the-envelope calculation was that electrifying all California ride-share miles could require doubling the state’s energy capacity. Even 10%-20% additional capacity would be a five-to-ten-year challenge in many places, leading to his contrarian possibility: combustion-engine autonomous vehicles may scale faster than EV autonomy.
Electric fleets therefore need urban land for charging, cleaning, maintenance and robotized servicing; the panel compared these depots with data centers that require their own substations. Permanent magnets and China’s rare-earth position add another physical dependency.
Autonomy also attacks parking economics. Kalanick estimated cars on the road could be utilized about 15 times more than before, allowing perhaps ten times fewer vehicles after preserving rush-hour capacity. Because parking occupies roughly 20%-30% of urban land, even a partial transition could release an extraordinary amount of property.
Jason flagged the downside before the redevelopment opportunity: falling land values could affect household retirement accounts and pension-fund balance sheets. Kalanick imagined housing, charging infrastructure or even hydroponic “farm-to-table” production, but repeated that grid upgrades are already the long pole in many CloudKitchens developments.

15. The DCA tragedy exposed a safety system still dependent on voice radio and human reaction

Jason relayed a commercial pilot’s warning that DCA was “the sketchiest airport we fly into,” with controllers playing “fast and loose” and helicopter traffic nearly impossible to see while moving at 150 mph. The pilot said both crew members remain on “red alert” during every approach.
Wisk CEO Brian Yutko’s proposed inversion was that collision-avoidance software should be permitted to take control even in piloted commercial aircraft. Some fighters already use automatic ground-collision avoidance when pilots lose consciousness; comparable authority could prevent a commercial crew from continuing into a collision the aircraft already predicts.
The second failure was infrastructure: critical instructions still travel through VHF voice communications, while air-traffic control runs on technology the panel described as dating to the 1960s. Better data links, tower automation and advanced pilot training were presented as investable upgrades after a tragedy that “should never have happened.”

Jason Calacanis

All right, everybody, welcome back to the All-In podcast. Check out Friedberg’s surprise drop with his hero Ray Dalio, live on all platforms today. Dalio discusses how countries go broke and recommends getting the U.S. to roughly 3% of GDP as its net deficit, net of all expenses including interest expense.

We are super delighted to have in the red throne Travis Kalanick. He is the co-founder and CEO of CloudKitchens. He also worked in the cab business for a little bit as the co-founder and former CEO of Uber. We had a great interview at the All-In Summit last year, and he’s back from his media hiatus. He’s been in the lab working on CloudKitchens. How are you doing, brother?

Travis Kalanick

I’m doing really well. I’ve got to say, just like at the Summit, Jason, it’s an honor to be in the presence of such a prominent Uber investor.

Jason Calacanis

Absolutely. Finally, somebody has recognized my contribution to greatness.

Travis Kalanick

Absolutely.

Jason Calacanis

I’ll mention it 3 or 4 times and we’ll be all good.

Travis Kalanick

I’ll give you the props. You don’t have to do it for yourself anymore.

Jason Calacanis

Thank you. I appreciate it. Give everybody a little overview of CloudKitchens, the business, and how it’s going, because people are obviously addicted to ordering food at home. It’s quite a trend.

Travis Kalanick

At a high level, the way to think about it is that it’s about the future of food. What does the future of food look like? You go, “In 100 years, we’ll start way out there.” In 100 years, you’re going to have very high-quality food, very low-cost food, and incredibly convenient food. There are going to be machines that make it, machines that get it to you, and it’s going to be exactly to your dietary preferences, your food preferences, and so on.

It just comes to you, and it’s so inexpensive that it approaches, or has surpassed, the cost of going to the grocery store. That’s more of a today analog, so you go 100 years. Of course, that’s the thing. Nobody’s going to be making food. What about 20 years? What about 10?

The company is real estate, software, and robotics. It’s all about the future of food. If you can get the quality there and get the cost down to start approaching the cost of going to the grocery store, you do to the kitchen what Uber did to the car. That’s the thing. It’s like a lot of bits and atoms. In the Uber world, this is like 5 times more atoms per bit. It’s heavy-duty industrial stuff, probably more along the lines of where Elon goes with some of his companies. They’re super-interesting technologies, but you’ve got to grind out those atoms.

Jason Calacanis

Do you see people actually cooking in the future, or does it become a centralized service optimized to people’s health? What do you think the implications for the food supply are if your vision holds? How do you think about all those things?

Travis Kalanick

People will cook in the future as a hobby. I sort of make a joke at the office. I say, “I like horses. I love horses, but I don’t ride a horse to work.” It’s going to be a little bit like that. You can cook. It’s a soulful thing to do. It’s very human.

But it’s late, Mom gets home late from the office, and needs to get the kids a nutritious meal. She doesn’t have to cook it now, and she won’t have to cook it. She won’t have to go to McDonald’s either. It will be high-quality, convenient, and low-cost, all at the same time. Yes, dietary preferences and everything else, because it’ll be hyper-personalized—like the internet is for content, plus plus plus in terms of your specific preferences for what you want.

Jason Calacanis

You’ve got these computers rocking—these robots rocking, I think, in Philadelphia somewhere, in the lab, where they’re making bowls.

Travis Kalanick

We’re out of the lab at this point. We have our machine. We have a machine called a Bowl Builder that basically makes different cuisine types with bowls. Think of Sweetgreen. We’re not working with these brands specifically, but I’ll use them as a good analogy. Think of Chipotle, Cava, or Sweetgreen. You get the idea.

We created test brands that were like those things and built the machine at the same time as we were building an actual restaurant. We built that restaurant to prove that the machine works. Now we have customers touring the facility, and we’re rolling out with 5 customers in April that are using the machine.

The way it’s going to work is that they will come into one of our facilities. Of course, we have the real estate. We have tens of thousands of kitchens around the world. They will come into one of our kitchens in a facility. It’s a delivery-only restaurant. They’ll prep the food in the morning and then leave.

The machine will receive an order from DoorDash, Uber Eats, or whoever. They’ll order online the way they do now: build your own bowl exactly as you want it. The bowl gets all the ingredients dispensed, hot or cold, with sauce and everything else. The bowl goes into a bag, the utensils go into the bag, the bag is sealed, and then it comes out on a conveyor belt.

The machine gets the bag and takes it to the front of the facility. It gets put into a locker. That locker is sitting there, and the DoorDash driver waves their phone with an app in front of a camera. It pops open the locker that has the food they’re supposed to get.

Jason Calacanis

That’s so cool. If you’re a restaurateur, the grind of the on-demand meal—which is the restaurant world—goes away. You basically prep, and that’s asynchronous from when people order food. The machine does the final assembly, or what’s known as plating, essentially.

Do you think there’s a service in the future where I can share my physiology with you through CloudKitchens, and you just keep optimizing my food based on what you know is good or bad for me?

Travis Kalanick

First, what we do is serve the restaurants. You’ll be sharing your dietary preferences with Uber Eats, DoorDash, Sweetgreen, or somebody else. Our customer promise at the company is, “We serve those who serve others.” Put another way, we’re infrastructure for better food.

We’re the AWS or the NVIDIA, or whatever you want to call it, but for food. We’re behind the scenes. We’re the infrastructure. You’ll give your preferences. It should be a brand like Sweetgreen or Chipotle that says, “Share with me an encrypted hash of your dietary restrictions, needs, whatever your lipid panel is, and I’ll customize this thing.”

It’s pretty close, JCal. You can authenticate your Apple Health. When these bowls come off the line—and it’s like an assembly line, bowls come off the line—the label on the bowl says how many of every ingredient are in it, plus a picture of what it was before we put the lid on. That can be sent to the person while the bowl is on its way via a courier.

Chamath Palihapitiya

What do you think, Travis, about this whole MAHA movement and the food supply itself? How does that change things? Do restaurants embrace more farm-to-table stuff?

Travis Kalanick

I think what we’ve seen with supply chains in a bunch of different industries is that everything is going to get super-wired up. Right now, we’re at the point of manufacturing, but what happens next? You go, “Okay, we’re doing assembly.” Then you go, “Okay, what about prep?” You go further upstream and ask, “What about the supply chain? What about Sysco and US Foods?”

Then you go further up and ask, “How does mechanization occur on farms and in agriculture? How does that all get wired up to serve the customer and what they’re looking for?” You really can know exactly what kind of wheat was put into that food, whether it was organic for real or not, and what the actual field was that it came from. You could imagine really getting tight about supply chains as they relate to dietary issues and MAHA.

As for MAHA, hell to the yes. I ordered a couple of different things. I went to the RFK Jr. website, and he has MAHA merch. I have the green MAHA merch hat. I should have worn it today. I’m all about it.

David Friedberg

Travis, your Bowl Builder—you tried to do this, right?

We had a Bowl Builder 10 years ago, in 2015 or 2016. Diego saw it; he actually visited it when we built it. We designed the system around a canister mechanism. All the food prep was done in a similar sort of modular model. It was loaded in bulk and then put into little canisters.

There were 30 slots in the canister dispenser. The canisters would move down the device, open up, and you could assemble bowls with rice, beans, and all sorts of things. The whole thing was automated. We were in the process of building out our first automated store when I took a medical leave of absence from it, and ultimately the company did not get it into production.

We had great working demos, though. It was definitely a no-brainer.

Jason Calacanis

This is the original Automat in New York.

David Friedberg

Yes, in the early 20th century. They had a commissary behind that wall and made plates of food. You put a quarter in, turned the knob, and got your meal out. It’s the classic artificial intelligence.

Travis Kalanick

It’s like the Mechanical Turk.

Here’s the nuance that’s super-interesting about automation in QSR restaurants. They have an existing brick-and-mortar facility that’s built a certain way. That layout is meant for humans and for those humans to work through certain processes in a very specific way. Every square inch of that kitchen and space is dialed in.

When you put a machine like this in, it changes the whole thing. Just to get going, you’ve got to change everything. If you’re going to replace the frontline at Chipotle, you’ve got to take out that frontline, demo it, and put in a new machine. That’s the challenge they’ve all had.

Now it’s a huge amount of capital expenditure. Your store is down for 2 or 3 months, and the economics start not to work. By the way, you still have to have humans in that brick-and-mortar facility.

We have a different take. We’re in the delivery-only model. This is true infrastructure for making food behind the scenes for delivery. You don’t have those issues. Our infrastructure—these kitchens—is designed for these kinds of machines to be in them, and vice versa. We’ve designed the machine to be in them.

When we did this early at ITA, food delivery was very early. We built these e-restaurants that were smaller footprints. We had an 800-square-foot restaurant doing $3 million a year in revenue, and it had a handful of people working in it. We were putting about 800 orders an hour through that restaurant during the lunch rush, ordering custom bowls.

Jason Calacanis

That was by One Market, right?

Travis Kalanick

Right, One Market.

Chamath Palihapitiya

By the way, did you guys notice that JCal was plugging his product there in the background, even though it had absolutely nothing to do with what Travis was saying?

Jason Calacanis

Welcome back to the show. Nothing’s changed. Sacks is here. No one else even noticed that. I just heard this voice from above. It was the czar of AI and crypto. I was like, “Wow, send it.”

David Sacks

It’s good to be back.

Jason Calacanis

Sit back and listen. The czar is back. Sacks, any anecdotes you want to share about life in D.C.? How exciting has it been in the administration during the first week?

David Sacks

It’s been amazing. It’s hard to believe it’s only been a week.

Jason Calacanis

Are you in the White House or in that building next to it? Do you have an office in the building next to it?

David Sacks

You mean the Treasury Building?

Jason Calacanis

I know somebody was talking about a building next to it. I don’t know.

David Sacks

I have an office in the Old Executive Office Building, otherwise known as the Eisenhower Building. I have a pass where I can just walk over to the West Wing if I want to.

There’s a whole White House complex behind the gates. The West Wing is part of it, along with the Eisenhower Building and a couple of other buildings in that complex. It’s really cool. It’s really neat to show up for work at the White House.

Jason Calacanis

That’s awesome. It’s like being in a movie or a TV show.

David Sacks

It is really cool.

Jason Calacanis

Any interesting meetings you can talk about? I know we’re here today to talk about DeepSeek, but are there any interesting meetings or anecdotes from the vibes and walking around? What’s the coffee like? Is there a commissary where you run into anybody interesting?

David Sacks

There is a commissary in the White House called the Navy Mess. I think they’re just opening up for business now. One of the cooler things you could do is take people to lunch at the Navy Mess.

Jason Calacanis

I look forward to it.

David Sacks

I look forward to taking Chamath and Friedberg there.

Jason Calacanis

I’ll wear my MAGA hat.

Well, let’s get started. You’re here because the world is ending, Jason. The Western world is ending, and David Sacks is going to save it.

We had a little bit of a freakout last week regarding DeepSeek. If you don’t know, it’s a Chinese AI startup that released a new language model called R1. It’s basically on par with some of the best models in production in the West, such as OpenAI’s o1 model.

They claim—and you can trust claims coming out of China for what they’re worth—that they did this all for $6 million on only 2,000 GPUs. For comparison, OpenAI reportedly spent $100 million to train GPT-4, which you’re all using now, and Sam claims they’re going to spend $1 billion training GPT-5. That’s about 7% of the cost of GPT-4.

There are obviously export restrictions on NVIDIA H100s to China, so there’s a big debate about whether they actually have H100s or not. Monday was a bloodbath in the stock market. NVIDIA had the worst day in the history of the stock market in terms of total dollar amount of market capitalization lost. It was down 17.7%, or $600 billion. TSMC, Arm, and Broadcom were also down.

Everybody is asking the question: How did they do this? Did they do it? There’s also a debate about whether they stole from OpenAI, which is rich coming from OpenAI, since it got caught red-handed stealing everybody else’s content. Now they’re crying foul that the Chinese stole or trained through what’s called distillation of their model in order to build theirs.

Sacks, obviously you are the czar of AI. I’m curious what your take on all of this is, and thanks for coming.

David Sacks

One of the really cool things about this job is that when something like this happens, I get to talk to everyone. I don’t know if I’ve talked to everyone in AI, but I’ve talked to all the top people in AI. It feels like I’ve talked to most of them.

There are definitely takes all over the map on DeepSeek, but I feel like I’ve started to put together a synthesis based on hearing from the top people in the field.

It was a bit of a freakout. It’s rare that a model release is going to be a global news story or cause $1 trillion of market-cap decline in 1 day. It’s interesting to think about why this was such a potent news story.

I think there are 2 things about that company that are different. One is that it’s obviously a Chinese company rather than an American company, so you have the whole China-versus-U.S. competition. The other is that it’s an open-source company, or at least it open-sourced the R1 model. You have the whole open-source-versus-closed-source debate.

If you take either one of those things out, it probably wouldn’t have been such a big story. But the combination got a lot of people’s attention. A huge part of TikTok’s audience, for example, is international. Some of those people like the idea that the U.S. may not win the AI race, that the U.S. is getting a comeuppance here. I think that fueled some of the early attention on TikTok.

Similarly, there are a lot of people rooting for open source, or who have animosity toward OpenAI. They were rooting for the idea that there was an open-source model that was going to give away what OpenAI had done at one-twelfth the cost. All of those things provided fuel for the story.

Now, what should we make of this? I think there are things that are true about the story and things that are not true, or should be debunked.

The true thing here is that if you had said to people a few weeks ago that the second company to release a reasoning model along the lines of o1 would be a Chinese company, people would have been surprised. There was a legitimate surprise.

To back up for people, there are 2 major kinds of AI models now. There’s the base large-language-model type, like ChatGPT 4o. DeepSeek’s equivalent is V3, which they launched a month ago. That’s basically like a smart Ph.D. You ask it a question and it gives you an answer.

Then there are the new reasoning models, which are based on reinforcement learning—a separate process from pretraining. o1 was the first model released along those lines. You can think of a reasoning model as a smart Ph.D. who doesn’t give you a snap answer but actually goes off and does the work.

You can give it a much more complicated question, and it’ll break that complicated problem into a subset of smaller problems. Then it’ll go step by step to solve the problem. That’s called chain of thought.

The new generation of agents is based on this idea that an AI model can sequentially perform tasks and figure out much more complicated problems. OpenAI was the first to release this type of reasoning model.

Google has a similar model it’s working on called Gemini 2.0 Flash Thinking. It has released an early prototype of this called Deep Research 1.5. Anthropic has something, but I don’t think it has released it yet. Other companies have similar models to o1 either in the works or in some sort of private beta.

DeepSeek was really the next one after OpenAI to release a full public version. Moreover, it open-sourced the model. That created a pretty big splash.

It was legitimately surprising that the next big company to put out a reasoning model like this would be a Chinese company, and that it would open-source it, give it away for free, and offer API access at something like one-twelfth the cost.

All of these things drove the news cycle, and I think for good reason. If you had asked most people in the industry a few weeks ago how far behind China was on AI models, they would have said 6 to 12 months. Now they might say something more like 3 to 6 months, because o1 was released about 4 months ago and R1 is comparable to it.

It has definitely moved up people’s time frames for how close China is on AI.

Now let’s take the claim that they did this for $6 million. I’m with Palmer Luckey, Brad Gerstner, and others. I think that number should be debunked. First of all, it’s very hard to validate a claim about how much money went into training a model. It’s not something we can empirically discover.

But even if you accept it at face value, that $6 million was for the final training run. When the media is hyping up these stories and saying that this Chinese company did it for $6 million while these dumb American companies did it for $1 billion, that’s not an apples-to-apples comparison.

To make an apples-to-apples comparison, you would need to compare the final training-run cost by DeepSeek to that of OpenAI or Anthropic. The founder of Anthropic said—and Brad has said something similar, being an investor in OpenAI and having talked to them—that the final training run cost was more in the tens of millions of dollars about 9 or 10 months ago.

It’s not $6 million versus $1 billion.

Jason Calacanis

The $1 billion number might include all the hardware they’ve bought and the years they’ve put into it—a holistic number, as opposed to the training number.

David Sacks

Exactly. It includes running it. It’s not fair to compare a fully loaded number by the American AI companies to the final training run by the Chinese company.

Jason Calacanis

Real quick, Sacks: You’ve got an open-source model, and the white paper they put out is very specific about what they did to make it and the results they got out of it. They don’t give you the training data, but you could start to stress-test what they’ve already put out there and see if you can do it cheaply.

David Sacks

I think it is hard to validate the number. If we assume that we give them credit for the $6 million number, my point is less that they couldn’t have done it and more that we need to compare like with like.

If you were going to look at the fully loaded cost of what it took DeepSeek to get to this point, then you would need to look at the R&D cost to date of all the models, experiments, and training runs they’ve done, as well as the compute cluster they surely have.

Dylan Patel, the leading semiconductor analyst, has estimated that DeepSeek has about 50,000 Hopper GPUs. Specifically, he said they have about 10,000 H100s, 10,000 H800s, and 30,000 H20s.

Chamath Palihapitiya

Is that DeepSeek, or DeepSeek plus the hedge fund?

David Sacks

DeepSeek plus the hedge fund, but it’s the same founder. By the way, that doesn’t mean they did anything illegal. The H100s were banned under export controls in 2022, and then they did the H800 in 2023. This founder was very farsighted and ahead of the curve. Through his hedge fund, he was using AI to do algorithmic trading, so he bought these chips a while ago.

In any event, you add up the cost of a compute cluster with 50,000-plus Hoppers, and it’s going to be over $1 billion. This idea that you have this scrappy company that did it for only $6 million is simply not true.

They have a substantial compute cluster that they use to train their models. Frankly, that doesn’t count any chips they might have beyond the 50,000 that they may have obtained in violation of export restrictions, which obviously they’re not going to admit to. We don’t know the full extent of what they have.

I think it’s worth pointing out that this part of the story got overhyped. It’s hard to know what’s fact and what’s fiction, because everybody on the outside who is guessing has their own incentive. If you’re a semiconductor analyst who is massively bullish on NVIDIA, you want it to be true that it wasn’t possible to train this on $6 million. If you’re the person who makes an alternative that disrupts NVIDIA, you want it to be true that it was trained on $6 million.

All of that is speculation.

Chamath Palihapitiya

The thing that struck me was how different their approach was, and TK just mentioned this. If you dig into not just the original DeepSeek white paper but also the subsequent papers that have refined some of the details, I do think this is a case where necessity was the mother of invention.

I’ll give you 2 examples where I read these things and thought, “These guys are really clever.” Let’s put a pin in whether they distilled o1, which we can talk about in a second. At the end of the day, these guys were asking, “How am I going to do this reinforcement-learning thing?”

They invented a totally different algorithm. There was an orthodoxy—this thing called PPO—that everybody used. They said, “No, we’re going to use something else called GRPO.” It uses a lot less computer memory and is highly performant.

Maybe they were constrained by some amount of compute, and that caused them to find this. You may not have found it if you had a total surplus of compute available.

The second thing that was crazy is that everybody is used to building models and compiling through CUDA, which is NVIDIA’s proprietary language. I’ve said a couple of times that CUDA is NVIDIA’s biggest moat, but it’s also the biggest threat factor for lock-in.

These guys worked totally around CUDA. They did something called PTX, which goes right to the bare metal. It’s controllable and effectively like writing assembly.

The reason I’m bringing this up is that we in the West, with all the money we’ve had, didn’t come up with these ideas. It’s not that we’re not smart enough to do it; we weren’t forced to, because the constraints didn’t exist.

I wonder how we make sure we learn this principle. When an AI company wakes up and rolls out of bed and some VC gives it $200 million, maybe that’s not the right answer for a Series A or a seed round. Maybe the right answer is $2 million so that they do these DeepSeek-like innovations.

Constraint makes for great art.

David Friedberg

I think it also enables a new class of investment opportunity, given the low cost and speed. It highlights that maybe the opportunity to create value doesn’t sit at that level in the value chain but further upstream.

AppLovin made a comment on Twitter today—I think it was referring to the wrapper. It was something like, “It turns out the wrapper may be the moat,” which is true. At the end of the day, if model performance continues to improve and models get cheaper, and the market becomes commoditized much faster than anyone thought, then the value is going to be created somewhere else in the value chain.

Maybe it’s not the wrapper. Maybe it’s the user. When electricity production took off in the United States, it wasn’t the companies making all the electricity that captured most of the value. It was the rest of the economy that accrued a lot of the value.

Jason Calacanis

You’re about to see a big test of this. OpenAI is raising $40 billion at a $340 billion valuation. That just hit the wire.

The underwriting is exactly what you just said, Friedberg. It’s the wrapper. ChatGPT is the next killer app. It’s getting to more than 1 billion monthly active users and hundreds of millions of daily active users. It’s competing for consumer usage.

That puts it on a collision course with Meta. Meta is the only company that could really impact that, because it’s the only company right now that has billions of eyeballs and daily active users.

Zuck said this in his earnings release. He said there’s only going to be one company that brings AI to 1 billion-plus people, and it will be us. That was essentially the quote in his earnings release yesterday.

Microsoft showed weakness in the cloud, and Microsoft is down 6% today. I think this is a window for OpenAI to say, “We’re going to go up against Meta. This is it. We’re going to be the players.”

Everyone is kind of ignoring Google. What do you guys think is happening right now between OpenAI and Microsoft? If it’s true that this distillation thing actually happened, there’s only one place where you could have distilled the o1 model: Azure. What the hell is going on over there?

David Friedberg

R1 is supported on Azure.

Jason Calacanis

Explain distillation real quick.

David Friedberg

When you have a large, high-parameter model, the way you get to a smaller, more usable model is through a process called distillation. The big model feeds the little model. The little model asks questions of the big model, you take the answers, and then you refine the smaller model.

Chamath Palihapitiya

You can see this. Nick, I sent you a clip. Show the clip of the DeepSeek run where it gives the China answer and then deletes it.

Jason Calacanis

What was Winston’s job in 1984?

Chamath Palihapitiya

It starts to go through a whole summary, and then the person asks, “Are there any actual states that currently do that?”

Jason Calacanis

Hold on, here it goes.

Chamath Palihapitiya

It says North Korea, then China. Watch this—boom. It deletes it.

The reason this is happening is that you’re seeing the chain of thought, the several layers, and then it’s catching the answer after the fact. We know this was distilled from some other model.

My only point, tongue in cheek, is that when you use OpenAI right now, you’re using it while it sits in an Azure instance somewhere. It’s Microsoft’s cloud infrastructure that runs it.

It begs the question: It’s not necessarily OpenAI’s fault that this distillation happened, and I’m not trying to assign blame. But typically, if this were to happen, you’d look to your cloud provider and say, “How are you letting this happen?” I don’t think anybody has had a good answer for that.

Jason Calacanis

The cloud provider is hosting R1 now, so they’re literally undercutting ChatGPT and OpenAI at the same time.

Chamath Palihapitiya

To clean that up, they’re hosting their own copy of it, right?

Jason Calacanis

R1 is open source.

Chamath Palihapitiya

When you say “they,” who are you referring to?

Jason Calacanis

Microsoft.

Chamath Palihapitiya

Microsoft is hosting its version of R1, which means it is actively subverting its partner OpenAI and pushing people to a cheaper model.

Jason Calacanis

Amazon is going to host its own version of R1. Groq has a version of R1. Just try it out. It’s open source now. Who has R1 on his laptop?

Travis Kalanick

Exactly.

Jason Calacanis

If it was stolen, as Sam is claiming, and the IP was stolen, you’d think Sam could call Sacks and say, “Can you not put the stolen IP on your server and promote it to everybody at a lower cost?”

It just shows that Microsoft has no loyalty to OpenAI.

Travis Kalanick

But you think they would have loyalty?

Jason Calacanis

They have loyalty.

Travis Kalanick

What it would take to distill o1 through brute force wouldn’t be, “I can’t believe it was distilled.” It would be such a massive number of calls against an API or something that you would notice it.

Jason Calacanis

They did come out and say they blocked some suspicious activity recently.

David Sacks

They’re always doing that. That’s constant. You’re always doing that.

Jason Calacanis

Go ahead, Sacks.

David Sacks

Let me address the distillation point. I mentioned a few days ago on Fox News that I thought it was likely or possible that distillation had occurred. There was some evidence for it, and it became a news story. I didn’t even realize that saying it would be news, because it’s kind of an open secret in Silicon Valley.

Everyone I’ve talked to is doing some level of distillation, because you need to test your model against other models anyway. Every single person I’ve talked to has basically agreed that there was some distillation here from OpenAI.

That doesn’t mean it was the only thing going on. To be sure, the DeepSeek team is very smart, and there were some innovations. But there was also some distillation.

This wasn’t even a fresh news story from the point of view of Silicon Valley. A month ago, we had a press cycle when DeepSeek’s V3 model came out. DeepSeek V3 was self-identifying as ChatGPT. When you asked it who it was or what model it was, 5 out of 8 times V3 would tell you it was ChatGPT-4. There are lots of videos and examples of this online.

The point is that we knew a month ago that V3 had been trained on a substantial amount of ChatGPT output, because V3 was self-identifying as ChatGPT.

There are 2 ways that could have happened. The innocent explanation is that DeepSeek crawled the web, found a lot of published ChatGPT output, and trained on that. That wouldn’t violate OpenAI’s terms of service or its IP.

The other explanation is that they used the OpenAI API and basically went to town.

Jason Calacanis

They went to town.

David Sacks

There’s no way, based on what we know, to prove one way or another. But I know what most people think happened.

At the end of the day, OpenAI can probably figure it out. It has indicated that it thinks there was some improper distillation here.

The Financial Times says OpenAI has found evidence that the Chinese artificial-intelligence startup DeepSeek used the U.S. company’s proprietary models to train its own open-source model. That’s what I’m referring to.

You have to be sympathetic to OpenAI in this, because if you’re building a startup and trying to raise money, there’s momentum. We’ve all gone through this cycle, guys. You celebrate momentum internally. That’s what gives you the energy to push your team even further and harder.

Then all of a sudden, it turns out that some portion of that momentum was bad. As Travis said, there’s probably a chart inside OpenAI’s offices showing how many times the APIs are being hit and how many times the endpoints are being hit. It all looks positive, and then you realize that some portion of it was actually bad and trying to undercut your value.

That’s a hard pill to swallow. You have to course-correct very quickly and lock things down.

Chamath Palihapitiya

Security is exactly what we haven’t talked about. You have to lock these models down now. You have to lock the endpoints down.

In the Biden administration, if this had happened, the first conversation would have been, “We need to KYC the people who use these models.” What are you talking about? We don’t KYC the cloud.

If you’re trying to use an EC2 endpoint or an S3 bucket, you don’t have to prove who you are. You just use a credit card and go. That’s the whole point of why proliferation can happen so quickly.

If we take the wrong takeaway from this period, there will be a bunch of people who clamor to lock these folks down and make innovation go much slower. I think that would be bad.

Jason Calacanis

Here’s the other side. You go through the white paper and see what they did and what they innovated on. You see the science behind it and the thoroughness, and you’re like, “These guys are badass.”

It does not feel or sound like somebody who just took something. When you get through it, it could be that OpenAI wrote the white paper for them, but it’s real innovation and strong technology. You’re like, “This is legitimate.”

David Sacks

I agree with that, but in the paper they’re hazy about where the data is coming from. They’re fairly transparent about everything else they did, but they’re not clear about the data.

Specifically, they say that to get from V3, which is the base model, to R1, the reasoning model, they had about 800,000 samples of reasoning. They’re quite unclear about where those reasoning samples came from.

It is remarkable that you can get from a base model to an R1 with just 800,000 samples. But this is a problem. We—the Western AI community—have been trudging along on a path where we’ve been very orthodox. We’ve said the only way you can do reinforcement learning is through PPO.

Is that true? It turns out that if you’re a really smart team that has no other choice, you move away and invent your way out of it.

We have to learn from that example, too. I think some of the things they’ve done are technically brilliant, but they also use constraint very much as a feature, not a bug. The Western AI economy has been the opposite so far.

Jason Calacanis

The best part of this is that Sam Altman was supposed to be doing open source. He made it a closed-source company, stole everybody’s data, got caught red-handed, and is being sued by The New York Times for all of that.

Now the Chinese have come and open-sourced all the stuff he stole, and he has a real competitor on the original mission of what OpenAI was supposed to do. I have zero sympathy for him or the team over there. I’m glad this is going open source. It should have been open source, and it’s better for humanity.

The fact that the Chinese did it to Sam Altman, after he stole everybody else’s content, is my position.

David Sacks

There you have it.

Jason Calacanis

I don’t have strong opinions on it.

David Sacks

It’s hilarious.

Jason Calacanis

Does nobody see the irony? He was supposed to be doing open source.

Chamath Palihapitiya

I will say the models are closed source.

Jason Calacanis

You’re right. There was also the lawsuit with Scarlett Johansson for stealing her voice after she said no. There’s a real question about The New York Times lawsuit, and now there’s a question about YouTube data being used to train the video models.

There’s a lot of pressure on OpenAI. It’s on its heels a little bit, so I definitely see your point.

David Sacks

I think the real pressure right now is on Meta. Meta has to show up with the next iteration of Llama that beats and exceeds Gemini and R1. That’s going to be crucial for us to have a counterweight to whatever China puts out after this.

Chamath Palihapitiya

But it’s open source. Embrace and extend. Meta has to embrace and extend everything these guys have shown.

Meta is buying tens of thousands of NVIDIA GPUs. Great. But what did this show? It shows that CUDA—and high-level languages in general—suck. We’ve all known that. We’ve been going through it thinking it’s the right thing to do.

DeepSeek throws it out the window. They use something called PTX. What Meta does next is critical to understand. It needs to embrace this stuff.

That will create a more heterogeneous environment. There’s too much money and risk on the line to go through a single point of failure: a chip and a high-level framework to get to that chip. That’s nuts.

That kind of “the emperor has no clothes” moment is upon us.

David Sacks

Let me ask you another question. Let’s assume we start the world of AI today. There’s no legacy from the last 3 years, and you wake up today with an open-source model that has 670 billion parameters. You can run it on your desktop computer, it’s completely available, and everything is transparent.

If I ask you the question—forget about all the big companies involved and everyone’s historical strategy—what’s the model today to build value here? Where do you build equity value as a business if you’re starting a company? Where do you invest as an investor?

Chamath Palihapitiya

The first thing is that you have to build a shim. A shim is critical because there’s so much entropy at the model level. You can’t pick any one model.

The problem is that the people who manipulate these models—the machine-learning engineers and so on—become too oriented toward understanding how to get high-quality output using one thing. It shouldn’t be the case that we have engineers who can only use Claude, or only OpenAI, or only Llama.

You need flexibility to hot-swap models as they change. If you’re starting a company today, the first technical problem I’d want to solve is that. Tomorrow, if it’s R2, Alibaba’s model, or Llama, I want to be able to rip one out, put another one in, and have everything work. Right now, we can’t do that.

Jason Calacanis

The answer to your question, Friedberg, is the application layer. This is all going to become storage. It’s like YouTube being built on top of storage, or Uber being built on top of GPS. All these innovations are being commoditized, and this one is happening faster than all the rest.

Do you want to be in the storage business, or do you want to be in the YouTube business? Do you want to be in the Uber business, or do you want to be in the GPS-chip business?

They’re both decent businesses, but Gavin Baker came on this podcast and said the fastest-depreciating asset in the world was a large language model. He’s been proven right.

They’re not worth anything. They’re all going to be open-sourced and commoditized, and that’s for the best of humanity. Now we’re going to be on the application level, and on the hardware level with robots. I think that’s where the opportunity is.

Travis, what company do you start today, given where the world is and given the open-source models?

Travis Kalanick

I’m getting excited.

The first degree out is what you’ve got. Is there a wrapper company? Of course, maybe those companies already exist. Then is there a tools company?

In a funny way, even though Meta could be the wrapper, it has a tools business that DeepSeek is challenging by going fully open source and putting something out there that’s really good.

Meta has to decide, “We are going to embrace and extend this. We’re going to make sure all the developers come to us and all the cool applications get built here.”

I think there’s a tools business, a wrapper business, and then, when AI gets cheap, here’s what’s going to happen: There’s going to be a lot more AI.

I don’t think the price elasticity is negative here. I think the price elasticity is actually positive. As the price goes down, usage and revenue are going to go up.

This has been the history of technology forever. Bill Gates once said, “I don’t know what to do with more than 640 kilobytes of memory.” The question is, did we have cheap oil?

Cheap oil in the United States drove the Industrial Revolution. When we started discovering oil, we were suddenly able to build factories and make things we had never imagined possible.

AI is going to get cheap. It’s going to be oil, but it’s also going to be specialized for different tasks. You’re going to get into nuances. What does investor AI look like? What does autonomous-car AI look like? What does Google Search AI look like?

Jason Calacanis

So you could go vertical and siloed.

Travis Kalanick

Siloed—understand what I’m saying.

David Sacks

There’s a thing called Jin’s Paradox, which speaks to this concept. Chamath actually tweeted about it.

It’s an economic concept where, as the cost of a particular use goes down, aggregate demand for all consumption of that thing goes up. The basic idea is that as the price of AI gets cheaper and cheaper, we’re going to want to use more and more of it.

You might actually get more spending on AI in the aggregate, because more applications will become economically feasible.

That’s a powerful argument for why companies are going to continue to innovate on frontier models.

You guys are taking a very strong point of view that open source is definitely going to win, that the leading model companies will all get commoditized, and that there will be no return on capital for continuing to innovate on the frontier. I’m not sure that’s true.

The R1 model is basically comparable to o1, which OpenAI released 4 months ago and was training internally 9 or 10 months ago. OpenAI is on o3 now. Its frontier is ahead of where R1 is. Anthropic and Google have things in the works, and even Meta may have models ahead of R1.

It’s not clear that R1 is the frontier. Those frontier-model companies, having seen what might have happened with distillation, have a strong incentive to make sure it doesn’t happen again. They’ll be taking countermeasures.

There’s a question of how much they can do to stop it, but I think it’s premature to conclude that there’s no reward for being at the frontier.

Jason Calacanis

Does anybody have any other questions for Sacks before we drop him off to go back to serving the American people?

David Sacks

One final point on the open-source-versus-closed-source issue: I’m not going to take sides, but I think it’s a mistake to view what happened here as this plucky upstart doing the community a huge service out of the goodness of its heart.

It’s open-sourcing everything because it’s a Chinese company trying to catch up. If you’re behind and trying to catch up, open source is a strategy that makes sense.

They’re trying to undercut the leading American companies. I don’t think they did it with $6 million. They have massive resources behind them.

Some of the pro-DeepSeek vibes are a little naive. In Silicon Valley, people are supporting DeepSeek because they think it’s doing a huge service for the community. I think it’s a little more self-interested than that.

Travis Kalanick

It could be both. There’s a theory that they’re trying to undercut and neuter the leaders. At the same time, there are a bunch of people who believe in open source and don’t think anybody should control this—and certainly not Sam Altman.

Jason Calacanis

Two things can be true at the same time.

David, thank you so much for coming on. We appreciate it.

David Sacks

Thank you for having me.

Jason Calacanis

Thanks to David Sacks for coming in. Let’s open up the aperture and talk a little bit about relations with China.

We’re obviously in a bit of a cold war with China. We have tariffs, Taiwan, and the trade war going on over exports of H100s. Where do we want to start?

Travis, you have deep experience. You’re one of probably 5 American entrepreneurs who ran an at-scale business involving Uber and DiDi in China, so you have a unique position. Tim Cook and Elon are probably the only other 2 people who have really had at-scale businesses there. Maybe Disney, since it has Disneyland there.

What’s your take on the relationship and what’s going on in China? How is China going to operate differently from the United States? From your experience and point of view, tell us a little bit about the culture and business ethics in China, particularly as they relate to AI.

Travis Kalanick

I had this experience almost 10 years ago, during the Uber days, when we were running Uber China. I cannot express the frenetic intensity of the copying they would do on everything we rolled out in China.

It was so epically intense that I developed a massive amount of respect for their ability to copy what we did. We would do hard work, make something, dial it in, and roll it out. It would be epic and awesome. Then 2 weeks later, boom, they had it. A week later, boom, they had it.

Of course, I used that to drive our team. There are so many great stories. We had about 400 Chinese nationals in Silicon Valley at our offices in San Francisco. We had a whole floor for the China growth team, and it was primarily Chinese nationals.

We had billboards on Highway 101 in Silicon Valley in Chinese—Uber billboards recruiting people to join our team in Chinese to serve the homeland. It was an all-out war. It was really epic.

When you went to that floor in our office, you were in China. They had Red China-style density. The desks were literally smaller. The density of the space was China.

What happens when you get really good at copying and the time gets tighter and tighter and tighter is that you eventually run out of things to copy. Then it flips to creativity and innovation.

At the beginning, it’s all over the place. The kind of innovation you see early on is like, “Really?” But as they exercise that muscle, it gets better and better.

If you want to know about the future of food and online food delivery, you don’t go to New York City. You go to Shanghai.

Jason Calacanis

What’s an example? Doesn’t Meituan do drone delivery?

Travis Kalanick

Here’s an example. If you went to offices in Shanghai, Beijing, or any of the major cities, the office buildings would have hundreds of lockers around their perimeter.

Everything you get, whether it’s food or anything else, gets dropped off by couriers in these lockers in the office buildings. Then there’s another group of people—interoffice runners—who bring it to your office.

It’s epically efficient. They’re taking advantage of their labor economics and things like that. It wouldn’t work exactly that way here, but a lot of the innovation you see coming out on Uber Eats or DoorDash today existed 3 or 4 years ago in China, maybe longer.

Eventually, you cross that threshold of copying, and you start innovating and leading. We see that in a whole bunch of different places.

Jason Calacanis

Here’s a look at these smart lockers. They’re available for sale online. These things are crazy.

Travis, you’ve experimented with those as well, right? Didn’t you have a commissary concept in downtown Los Angeles?

Travis Kalanick

We have a couple of things. In every one of our facilities—and we’ve got hundreds of them—we’ll have lockers. The courier waves their phone in front of a camera, the right locker pops open, and they get the food.

Courier pickup is asynchronous from food production. You don’t have lines anymore. There are no more lines, which speeds up delivery, shortens the amount of time, and reduces how much money you spend on couriers.

We have another thing called Picnic. If you’re in an office building, you go to a website and order food from 100 different restaurants. Those restaurants happen to be in our facilities.

One courier goes to one of our facilities and picks up 50 orders at a time. The courier brings them to an office and puts them on a shelf on every floor. You get notified when your food arrives, and it arrives at the same time every day.

You go to the shelf on your floor, get your food, and go right back into your meeting. It saves people time at the office and gives them a lot of selection, especially in food deserts.

Even if there’s a Sweetgreen right downstairs from my office, I could save 20 minutes by using our own service instead of going downstairs. You get it at the same price, because the courier is delivering 50 orders at a time. Courier costs essentially go to zero.

Jason Calacanis

What do we think of the export controls? Should we be banning more H100s or other chipsets from going to China, or is that futile?

Chamath Palihapitiya

I don’t know the answer to that. I think Sacks and President Trump will make a good decision.

Here’s the curious case of the export controls. The first thing people are claiming is that DeepSeek is getting access to a bunch of NVIDIA GPUs using Singapore as a back door. Essentially, you create a Singaporean shell company, place an order with NVIDIA, NVIDIA fulfills it into Singapore, and then the chips go somewhere else.

There are examples suggesting that up to a quarter of all NVIDIA revenue goes into Singapore. The speculation is that 100% of those chips then go into China. That’s an enormous claim, because it’s a huge amount of NVIDIA’s revenue.

The interesting thing is that if you try to understand whether the chips might simply be sitting inside Singapore, the theory starts to unravel.

Singapore is about 250 or 260 square miles. It’s a small place. I tried to find out how many data centers are in Singapore, and it’s about 100.

You would think, “What does that mean? A hundred could mean anything.” But then you look at the energy data, which they publish. All of those 100 data centers consume about 876 megawatts.

These are small data centers. The entire industry is a $1.5 billion to $2 billion revenue business.

I think Sacks and the administration are going to have to dig into this and figure out what their opinion should be. There are clearly a ton of these chips going into Singapore. I don’t think anybody knows where they end up.

The question is, what does America think about that? Why did we implement these export controls in the first place? If there’s a simple back door, how do you want to react?

David Friedberg

If the U.S. finds a path, I mean, let’s talk about what happened with sanctions in Russia and other prior sanctioning efforts around the world. If we cut off access to NVIDIA chips and U.S. exports, aren’t we recognizing that the second-order effect will be that China takes IP and copies it, develops and builds out its own fabs, and finds ways to copy the ASML technology?

At the end of the day, there’s a lot to put together. I know it’s deeply technically complex, but if ever there were a group of people in the history of human civilization to pull it off, it’s probably the modern Chinese. They’ll say, “Let’s go build our own infrastructure.”

Chamath Palihapitiya

It’s worse than that. The models today are capable of designing chips for you that don’t rely on the most complicated technologies ASML creates.

One of the luckiest things that happened to Groq was that we designed our chip at 14 nanometers, which is effectively in the spectrum of technologies like VHS and beta. We chose a simple technology stack.

The latest cutting-edge chips at 2 nanometers use these complicated ASML machines. It’s not clear that the yield is actually that good. Why would you spend all that money?

If China is forced to engineer around those constraints, it can use these models to design chips that can be manufactured in simple ways and make simple stuff. This doesn’t solve the problem.

David Friedberg

It doesn’t. This is why I think the real problem is how we incentivize people in America to out-engineer and out-innovate the competition.

Chamath Palihapitiya

Or AI ensures that we’re entering an era of extraordinary abundance, and that abundance ultimately reduces the drive for conflict.

David Friedberg

There’s another possibility. China could simply bear the cost, as a central authority, of building an incredibly capable model. It could spend all the money and then tell Chinese companies, “You can distill from this model for free,” because the government has a golden vote and a seat on the board of every major company anyway.

There’s that possibility, where one central authority bears the capital expenditure of creating something that everybody else can draft off.

Jason Calacanis

Let’s talk a little bit about OpenAI. They’re in Washington asking for money now. Is that the concept? Should the government back it?

The rumor today is that OpenAI is raising $40 billion at a $340 billion pre-money valuation, with Masayoshi Son potentially leading it.

I would love Travis’s read on this, because he’s taken large amounts of money from Masayoshi Son in the past and has been through this. How does he think about the decision?

We all know about the meeting I had with him last summer. Someone said I should meet with Masa, so I sat down with him and started talking. He looked at me and said, “This is not generative AI. I only do generative AI. I think your company will be very successful, and you will be very successful. Goodbye.”

He walked me out, and that was the end of everything.

Travis Kalanick

I need to bust a myth. I did not take money from Masayoshi Son. He begged me to take money for years, and we did not take it.

He’s a promiscuous investor. Once he invests in you, you should probably count on him using your information and investing in all of your competitors. At least historically, that’s what he’s done.

I didn’t go there, but then he kept investing in all my competitors. They kept subsidizing these markets, and I started thinking, “Maybe I should have just soaked up the money that was there.”

When you look at whether OpenAI should take a lot of money from a Masayoshi Son-type situation, it’s a double-edged sword. If you don’t take the money, it goes somewhere else. But if you do take the money, understand that whatever intelligence they get while giving you the money, hanging around your board, or whatever else, may be used to do other things.

That is the nature of the Masayoshi machine. You’re damned if you do and damned if you don’t, but you have to pick.

If the money is flowing and access to capital is a strategic competitive weapon or advantage, you must play ball.

We were able to do things with the Saudis before the Vision Fund existed. They wrote a $3.5 billion check when that was the biggest thing that had ever happened. We were okay not having Masayoshi’s money.

But that money then went to our competitors, including DoorDash.

Jason Calacanis

In this OpenAI context, Travis, knowing what you know about AI, is raising $40 billion going to be a competitive advantage for Sam? Where does it go when he’s up against China, Microsoft, Alphabet, and Meta?

Travis Kalanick

If constraint is the mother of invention—or whatever the aphorism is—then you get into a strange situation when you’re overcapitalized.

In the Uber model, the war was about subsidizing rides for market share. It was necessary. You’re screwed if you don’t do it.

The question is whether you get to a place where you’re overcapitalized—too big, too bureaucratic, too loose, too weak, too soft. When you have an open-source model that’s very smart and a thousand flowers are blooming, with lots of innovation happening everywhere, that could be an overwhelming force.

Different sectors will treat it differently. Going full-stack in certain industry sectors will matter. In other places, having a chaotic system where everybody does a small piece may be fine.

We could spend dozens or hundreds of hours talking about the nuances.

Jason Calacanis

It seems like there’s some relationship between the Stargate announcement, with Masa and Sam standing up there with Larry, and Sacks showing up in the conversation, as well as this fundraise.

The idea is that more hardware and infrastructure, faster, creates a moat. That’s the real thing you have to believe, which becomes harder to believe after what happened over the last week.

David Friedberg

I’ve thought for a while that it doesn’t make sense to have one large, do-everything model. This mixture-of-experts architecture could be the future.

You could think about taking a large model, making 2 copies of it, and shrinking each model down to whatever is necessary. You run 2 models less frequently, so that combination uses less power and takes less time. Then you do the same thing again and shrink it down to 4, and then 12.

Eventually, you have a lot of smaller models. Some are experts at one thing, such as mathematics, reading, or writing.

The reality is that we don’t know whether humans have thought about the world in the right way. AI may resolve into smaller expert models that we don’t understand, where we don’t know why one model is the expert at something, but you have a network of small systems that work together.

That ultimately leads to commoditization—not just in model cost, development, and runtime, but also in what’s needed. Do you really create much of an advantage by having all these data centers?

Chamath Palihapitiya

That’s the key point. You’re not going to get an advantage by having more H100s at a certain point. The real advantage is going to be in IP and owning content.

The smart thing to do would be to buy Reddit, Quora, The New York Times, The Washington Post, and Disney, and then not allow other people to use the content. Sue the hell out of them every time they try.

Jason Calacanis

Take The Washington Post off that list.

David Friedberg

The New York Times comes off the list, too.

Jason Calacanis

All of those would be great because you could then, like a patent troll, tell anybody who has absorbed New York Times stories historically—or Disney content—that you own it and sue the hell out of them. You’d have the best, most proprietary model.

David Friedberg

You’re describing text content, which is a fraction of where this is important. Video is probably more important.

Google’s YouTube content library is probably 100 to 200 times larger than the rest of the internet combined.

Jason Calacanis

You’re such an old-school copyright guy. You’re such an old-school media guy.

I believe in artists and their right to their content. We’ve had a series of conversations, and I feel very confident telling you that they have the rights to a good chunk of that content.

They don’t have the rights to a lot of the copyrighted content that the big media companies have given them, but they have the rights to a lot of user-generated content. They are legally using it.

Then there’s the separate body of content that comes, for example, from Tesla. Tesla has an extraordinary advantage because it was pressed to put cameras on everything years ago. That gives it the ability to build models that do self-driving.

There’s a lot more data advantage in certain industry segments than others. That’s where the moat will lie. That moat will allow you to build better products, which gives you a more persistent advantage in gathering more data.

That’s ultimately where I think this resolves. It may not be about who has the biggest data-center network.

David Friedberg

At some point, the amount of data becomes the long pole in the tent. At another point, the quality of the algorithms becomes the long pole in the tent. More compute isn’t going to change that.

I don’t think we’re there yet. That’s the one thing that counters the idea that cheap AI means more AI. Is there enough data or enough algorithmic innovation to make more AI work?

I do agree with siloing it and getting more expert and better in these ways, but there’s an interesting tradeoff between all these variables.

Jason Calacanis

I just got offered $2,500 to put Angel into an AI-training dataset, because HarperCollins did a deal with Microsoft. I think it’s $500 per year for 3 years.

They did this blanket license for every book. They didn’t look at sales or how desirable the book was. It was just a blanket deal: Everybody gets $2,500 per book for 3 years.

I think I’m going to do it just to support proper licensing, so people can start going down this path.

Let’s get into DOGE. We’re about 10 days into this administration, and Trump formally established DOGE, the Department of Government Efficiency, through an executive order.

Apparently, Elon has been spending a lot of time at the offices. There are a bunch of wins DOGE is claiming on the internet. They say they’re saving American taxpayers around $1 billion a day—$3 for every American, every day, or about $1,000 a year in savings for each U.S. citizen. They claim they can triple that.

We have $36 trillion in debt. Have fun with some numbers there if you like. For a family of 5, that would be about $155,000 a year, maybe $60,000 during Trump’s second term.

The key announcement was similar to the Twitter execution: The ability for people to resign, done in a very kind way. Eight months of severance is being offered to federal workers. They expect 5% to 10% of federal workers to take the buyout.

This could be something like $100 billion in savings. Eight months of severance is not actually a legal concept that you can simply implement, so these are some sort of buyouts. There’s obviously a lot of hand-wringing about it.

I think they’re off to a good start. They’ve also been canceling leases. As we discussed before the election, there’s a lot of space that isn’t being used. The federal government is terminating a lot of leases and selling or consolidating properties.

At the same time, everyone has to return to the office. Who wants to go first with the first 10 days of DOGE?

Chamath Palihapitiya

I see some eggplant emojis in the group chat.

Jason Calacanis

What’s that about?

Chamath Palihapitiya

I’m adding you right now. How are you not in the group chat?

Jason Calacanis

I’m adding you right now. Every time one of these things hits the group chat, it’s hilarious. There are eggplants everywhere.

David Friedberg

The eggplant always comes from Friedberg first.

Jason Calacanis

I’m outing him as an eggplant guy.

David Friedberg

I’m a big DOGE eggplant guy.

Jason Calacanis

Friedberg, tell us how much you love eggplant.

David Friedberg

There’s nothing particularly surprising in the first week. A lot of this was discussed leading up to the inauguration. Vivek and Elon published their piece in The Wall Street Journal a couple of weeks ago, and they talked about the mechanisms of action they could use to drive reductions in cost.

One was returning to the office. Another was giving people a buyout offer. The buyout offer isn’t new. Bill Clinton did the same thing during his presidency.

If you remember when Clinton tried to balance the budget and get to a surplus—which he successfully did—his intention was to reduce U.S. debt to zero by 2013. He had a specific economic and fiscal plan for doing that, which he put into place.

We’re seeing them take the actions they said they would take. They said they would demand that federal employees come back to the office, and they assumed some degree of attrition from that. Now there’s the buyout offer.

The big question is how much authority the executive branch has to stop spending that has been mandated by statute. There will be a lot of lawsuits over the next couple of months. The courts will adjudicate how far DOGE’s intentions can go.

There’s also a separate set of efforts around legislative action. There’s about a $2 trillion annual deficit in the United States federal government right now.

If you look at Ray Dalio’s book on why countries go broke, there’s a simple arithmetic principle: The U.S. needs to get its federal deficit below 3% of GDP. That means we need to cut roughly $1 trillion to $1.1 trillion of spending.

If we can do that, we’re in a more economically sound place.

An important point from the Dalio interview is that as you cut spending, interest rates will come down. There’s currently a significant selloff in Treasuries and a lot of risk associated with the U.S.’s ability to deliver on its debt obligations over the next 30 years.

That’s why 30-year Treasuries are at 5%, even though the Federal Reserve is cutting rates. The rate on Treasuries is going up because people are still selling them.

That’s also inflationary.

Jason Calacanis

It’s inflationary.

David Friedberg

As we cut spending, we’ll see less inflation, and the U.S.’s ability to pay back its debt obligations over the next 30 years will improve. Rates will come down.

There’s a nice cyclical effect as these cuts come into play. The speed at which you make the cuts affects the amount of cuts you have to make. The faster you make the cuts, the less you have to cut.

That’s a key principle going into this. We should expect a big whirlwind of cutting over the next couple of months, or at least an attempt to do that. The courts will adjudicate what needs to be legislated, and then the administration will go to Congress and try to get some of these cuts enacted.

I will tell you, once again, that after our visit to D.C. last week, there wasn’t a single member of Congress I spoke with who viewed cutting spending as a mandate. They all have a very different agenda from DOGE.

Jason Calacanis

This is one of those interesting things where the difference between the legislative and executive branches is that DOGE is bringing it to life. What powers and controls does the executive branch have to spend—or not spend—especially when spending has been legislated?

There’s no law that says you can give a bunch of people 8 months of severance, let them go, and not replace them. There’s no law that says the executive branch can’t make it harder to hire people or procure certain things.

Do they fight bureaucracy with bureaucracy, making it harder to spend, harder to hire people, and harder to procure certain things that you’re supposed to spend money on? You can reduce spending through a lot of very interesting, nuanced friction rules that they’re in control of.

Chamath Palihapitiya

Some friction could slow things down.

Jason Calacanis

They’re talking about competency tests, reviews, and perhaps making people meet certain standards.

If you force people to come back to the office, you’re going to lose 5% or 10% of them. If 10% take the buyout, all of a sudden you’re saving money.

It’ll be interesting to see if it’s 5% or 10% with return-to-office. It could be a lot more. What I’m hearing about these buildings is that they are extremely empty—next-level empty.

Let’s just say I’m really glad I don’t own a bunch of office leases to the federal government right now.

David Friedberg

The interesting thing about those leases is that the government is such a reliable client that they’re all on 1-year leases. With startups, landlords force them to do 5- or 10-year leases because they know the company could go out of business.

The federal government is on rolling year-over-year leases, so it can cut them. That’s going to flood the market.

Jason Calacanis

Chamath, what are your thoughts on the stopping of payments? They’re obviously going for it. They stopped all payments, which is part of the playbook.

I saw this up close and personal at Twitter: Turn off subscriptions and see if anybody is using them. A judge got involved in that, but there’s also aid going to other countries.

We’re starting to look at what we’re actually sending to other countries and for what purpose. There’s a naming-and-shaming strategy, appealing to the public through social media and asking, “Do you want this money going here when we have tragedies in our own country that need to be solved?”

We have health care, houses burned down, and infrastructure problems. Maybe you could talk about winning hearts and minds and your general take so far.

Chamath Palihapitiya

We have to remember that we’re only 9 or 10 days into DOGE. The fact that we’re already at $1 billion a day is incredible. There has been no discernible impact yet. There’s been a lot of misinformation and fake news, but the real impacts have been negligible to nonexistent since they started making these cuts.

I think DOGE is a 3-layer onion.

Layer 1 is the people. We’ve now given federal workers a pretty generous offer. Elon said it was essentially the maximum allowed under these contracts, but they tried to do a good thing there.

Layer 2 is the infrastructure: all the buildings and physical plants the government owns and operates that may be empty or idle. Getting them back into private hands so they can be repurposed is going to save a lot of money.

But both of those layers will pale in comparison with the third layer, which is the IT, services, and spending.

At the center and nucleus of every DOGE team is an engineer. I think the reason is that engineers can get into the systems of record and start tracing where the money is going.

When you uncover through forensic analysis where these dollars are going and how they’re being spent, that’s how you close the gap from $1 trillion. I suspect it could be more than $2 trillion when it’s all said and done.

That is an enormous amount of waste, and it’s unproductive. I’m very excited about what happens over the next little while. The transparency is going to be incredible.

Just for kicks, check this out. If we took 2019 spending—the year before COVID—and put it up against 2024 revenues, we would have a $500 billion surplus.

David Friedberg

That’s crazy.

Chamath Palihapitiya

That’s a $2 trillion swing against the $1.5 trillion deficit, on roughly a $4 trillion budget.

David Friedberg

A lot of that is the fact that we have $1 trillion a year in interest payments now.

Jason Calacanis

This is the thing. We need 2 deflationary forces. One is DOGE, and the other is what AI could do if it really delivers. That will keep us in an okay spot economically.

But this spending has to go, or we’re in Greek territory.

The popular support for this is pretty incredible. I’ll go through some numbers on what people agree with Trump doing early on and what they disagree with.

We talked about Trump pardoning the January 6 protesters and ending requirements for government employees to report gifts. Those are tremendously unpopular.

Then you look at downsizing the federal government, imposing a hiring freeze, and requiring federal employees to return to an office. Those are incredibly popular.

Right now, Trump is at the apex of his political popularity, and these issues are incredibly popular in a very polarized time. He also has done an incredible job with the border, which is another consensus-based issue.

Trump now has downsizing the government, controlling immigration, and removing violent immigrants as incredibly popular parts of his mandate.

That’s a big win for him. If you look at his popularity, Trump is much more popular than he was the first time around. He’s at 49% compared with 44% last time. He’s still historically the least popular president ever, but my point is that the extra 10% of people supporting him right now includes me and other people who are looking at the people he put around him.

He has to stay with the Trump 2.0 agenda, as hard as it is, and stay away from the Steve Bannon agenda and the grifting. Those are the things that will take this apart.

That’s my appeal to them. I give them a B so far. They could do better, but it’s pretty good. Less of the meme coin and less of the drama. We have to make sure we’re not dragging dishwashers, teachers, and people who have been here for 20 years out of the country.

It’s going to be a very important approach if this is going to be sustained. I think it’s a coin toss whether Trump will be able to maintain his popularity.

What he did today with Pete Buttigieg—attacking him over this tragedy—is the kind of thing people don’t want. Less of that, please. More of the DOGE.

Travis, can I ask you about Waymo? Have you taken a production Waymo?

Travis Kalanick

Yes.

Jason Calacanis

What do you think about it? Do you think that’s the future of transportation, and how does Uber play into the self-driving-car business?

Travis Kalanick

It’s funny, because back in 2015, 2016, and 2017, we had our own autonomous vehicles out there.

I remember the first one of ours that I took. I got in the back, and all I had was a big red stop button that I could push if things got weird.

This was in Pittsburgh, where we had our robotics and autonomy division at Uber. I got out of that car and literally felt like I had gotten off a roller coaster. My legs were wobbly. I couldn’t stand straight because I was so freaked out and my adrenaline was pumping.

You get in a Waymo today and you don’t think twice. You’re just like, “It’s all good.” You get in and get out. Part of it is normalization. It’s just working, and that normalization matters psychologically.

Is it an optimized experience for ridesharing? No. The Cybercab is the ultimate destination for what it means to be transported across a city in a vehicle that isn’t meant for a human to drive. There’s no steering wheel. People could potentially face each other. There are a lot of different formats.

The technology works. We know there are different ways to get to the technology. One of the most interesting things to think about is that cheap AI makes cheap autonomy.

As cheap AI gets out there, proliferates, and becomes broadly distributed, autonomy should get easier and easier. You see some of what’s happening with Tesla and FSD. Its new models have improved perhaps 10 times in a 3-month period in terms of performance, measured by the number of miles per human intervention.

That’s what Elon is seeing right now. Cheap, good AI makes cheap, good autonomy. We need to connect those dots.

Then you go one level past that and ask about the hardware. Manufacturing is hard, and that could be a long pole in the tent. That could be a place where Tesla has a huge advantage.

Then you look at who the partners are. Are they set up to do the right kind of manufacturing and get cars out at scale?

There’s another dark horse that nobody is talking about: electricity and power.

I did some quick back-of-the-envelope calculations. If all the miles in California were EV ridesharing miles, you would need to double California’s energy capacity.

Let’s not even talk about what it would take to double the energy capacity of California’s grid. Getting 10% or 20% more is going to be a gargantuan 5- to 10-year exercise.

I live in Los Angeles. It’s a nice area, and we have power outages all the time because the grid is messed up. They upgrade it as things break. That’s where we are in one of the most affluent neighborhoods in Los Angeles.

The dark-horse hot take is combustion-engine autonomous vehicles. I don’t know how you can roll out autonomous vehicles really, really massively and quickly with the electric grid as it is.

Jason Calacanis

What do you think about regulation? Obviously, there was the Cruise incident where a person got hit by a regular car and was dragged. The whole thing imploded.

At Uber, we had the tragedy in Arizona, where somebody was playing Candy Crush while they were supposed to be a safety driver.

What’s your outlook as this rolls out and somebody gets hurt? You brought Uber to tens of thousands of cities. How receptive are those cities going to be, and what do you think the regulatory framework will look like?

Travis Kalanick

Similar to how the technology becomes normalized, it’s like getting used to getting in a car. It becomes normalized psychologically and in the public mindset.

We’re getting to a place where these vehicles are demonstrably safer than human-driven vehicles. There are mistakes, but they are provably safer. People are getting used to it.

We’re moving out of the hysteria and into a place where people say, “It’s just great.” Talk to people who are using these vehicles, and they feel safer.

Of course, I feel like we’re going to get into fewer accidents. But I also feel safer because there’s less chance of an interpersonal problem, especially late at night when people are out partying.

There’s a level of safety on many different dimensions for the rider.

Jason Calacanis

What do you think about BYD? You mentioned everybody getting to autonomy at the same time. Waymo has the biggest lead, Tesla is behind it, and BYD and about 10 other providers are out there doing this.

Do 10 players get there at the same time, and then is it just about who can incorporate these vehicles into their network?

What do you think of Uber’s strategy of having 8 partners, taking everybody into the network, and managing the vomiting in the back of cars, cleaning them, and charging them?

Travis Kalanick

The big issue with anything Chinese is whether you’ll be allowed to bring it into the United States. You maybe can now, but what happens under President Trump? Will there be blocks on bringing this kind of technology into the U.S.?

That’s a whole issue.

The bet Uber is making, consciously or subconsciously, is whether cheap, democratized AI will happen. If it does, does that make cheap, democratized autonomy?

Then you have to line up your physical hardware partners and car manufacturers. You have to say, “Is the electricity where it needs to be?” Are there other bets to make to ensure you can charge your cars?

There’s a huge real-estate and fleet-management play in figuring out how to electrify these plots of land called parking lots. You also have to set them up so that robots can clean cars in a very efficient way.

There’s a whole system. When you talk about data centers, data centers need their own power substations to meet power demands. If we see a world of robotics and automation, these moving robotic systems will need to have their power demands met in a similar way.

They’ll all go into their recharge buildings and get recharged, whether they’re cars, humanoid robots, food-delivery robots on the sidewalk, or drones.

You need actuators for robots. Do you know what you need for an actuator? A permanent magnet. Do you know what you need for a permanent magnet? Rare earths.

Jason Calacanis

Who’s the rare-earth king?

Travis Kalanick

China.

Jason Calacanis

Greenland.

Travis Kalanick

Greenland.

There are a couple of interesting things here. One is how these companies think about real estate, electrifying that real estate in urban environments, and robotizing it so they can do servicing and maintenance.

It could be manual for a while.

Chamath Palihapitiya

Can I put you on the spot and go one level above it? Merge the last 2 concepts together.

You talked about the federal government and DOGE. Isn’t there the potential for a complete surplus of physical inventory that exists in America?

Travis Kalanick

Yes, big time.

Chamath Palihapitiya

What does that mean for commercial real estate? How do you navigate around that? You’ve got to evade the falling knives first.

Travis Kalanick

Let’s go down the autonomous ridesharing lane.

Car ownership is already dropping, and it drops like a knife. There’s this thing in cities that takes up 20% to 30% of all the land called parking. It’s no longer necessary, because the cars that exist on the roads are getting utilized 15 times more than they were before.

Hypothetically, you need one-fifteenth the number of cars. Maybe you could say one-fifth or one-eleventh, because you want to be able to surge during rush hour. It depends on carpooling and things like that.

Let’s call it 10 times fewer cars and one-tenth the land necessary for parking. Maybe it’s less than that.

Now you’re opening up 20% of the land in a city. That land goes fallow. What should we do with it? Is there demand for it?

Jason Calacanis

Maybe it should be housing. Don’t we have to reevaluate all of today’s city planning? City planning works backward from constraints that are 1.0 constraints.

Here’s the traffic flow. Here are the traffic patterns. Theoretically, those constraints don’t exist anymore, or they exist in a totally different way.

Travis Kalanick

There’s a massive amount of creativity required to figure out what you can do with that land at a high ROI.

Some people say you’re going to have hydroponic farms in urban environments. That’s not a bad idea if you want to have farm-to-table healthy food. It’s literally farm to table—it’s a mile away from you.

There are interesting ideas. Land prices would have to come crashing down, and there are interesting ramifications if that were to happen.

Jason Calacanis

That’s what I wanted you to say—not to try to get you there, but that’s the craziness nobody is thinking about.

This physical built inventory has so much value built up in people’s 401(k)s and on the balance sheets of huge pension funds, but that value could be very different.

David Friedberg

The crazy part is that electricity production and electric capacity on the grid could be the gating factor that makes it a slow burn.

Travis Kalanick

I’m just riffing here.

Jason Calacanis

Right.

Travis Kalanick

It could mean energy storage, electric-grid upgrades, modular energy-capacity upgrades, and energy production. Those things are going to be very important.

If you want to see what happens when you have unlimited land, live in Austin. Look at the distance between San Antonio, Houston, Dallas, and Austin. You get 30 minutes outside the city centers, and there’s unlimited land with less regulation.

Housing prices and rents have come down for 2 or 3 years in a row. This could happen in other major cities. If DOGE means less regulation and you can build more, it could be amazing for Americans to afford homes again.

David Friedberg

Maybe you convert some of this space to energy storage, electric-grid upgrades, modular energy-capacity upgrades, and production.

Travis Kalanick

We do this all the time. We have facilities in every major city in the United States and around the world. Utility upgrades are the long pole in the tent in construction and development in many of our cities—not all cities, but many.

Jason Calacanis

The Fed held rates. They’re getting close to the 2% goal. I think we’re at 2.4% or 2.9% inflation.

Any thoughts on where we are with the Fed deciding not to cut rates? Chamath, any broader thoughts?

Chamath Palihapitiya

The long end of the yield curve is telling us that there’s still a chance of inflation.

The next 30 to 60 days from the administration are critical. If DOGE gets to $3 billion a day faster than people expect, there will be a lot of room for the president to make a valid argument that rates are too high for where they are.

We’ll have much more control over expenses, which means less need to spend.

That doesn’t solve the problem Yellen and Biden created on the way out the door. The biggest problem is that they put America in a very difficult position by issuing so much short-term paper. It’s extremely expensive.

As all of that rolls off, we have to refinance a lot of this debt at 5%. Nearly 30% of the debt will be refinanced this year.

The question is, what will these auctions look like? The last auction barely had 2 times coverage.

Watch the Dalio interview, because this is exactly the topic he covers. As we need to refinance the debt, rates climb and appetite isn’t there. It becomes a spiral.

That’s why we have to cut the deficit quickly to attract the market.

The market has moved a little bit. On January 13, the 30-year Treasury peaked at exactly 5%. Today it’s at 4.77%, so there’s been a little relief since the administration came into office and taken action.

As more of this action is realized, if people appreciate that DOGE is successful and the courts allow reductions in spending, I think we could see that rate drop from 4.77% much more significantly. That will create a great deal of relief.

David Friedberg

It either does that or it really doesn’t. It could get super nasty and bad.

I got a text from somebody senior in capital markets who thinks this is going to go to 5.5% before it goes down. They think there’s going to be a turbulent period ahead.

Jason Calacanis

The problem is that the idea it’s going to get to 5.5% before it comes down can spiral on itself. You have to print money to get to that place, and the printing drives it.

Chamath Palihapitiya

If we go to 5.5%, that’s not just 80 basis points. You have to think about the total tonnage of dollars that need to be paid back.

If you look backward, that’s effectively like 10% rates from 2000. Could you imagine what the economy would have done if you’d brought rates to 10% or 11% 20 years ago? It would have crippled the economy.

We don’t have much room here. If you walk rates up to 5.5% or 6%, a lot of things will start to break.

This is why I think DOGE will be successful. As people internalize all of this, every member of Congress who may have wanted benefits for their community will have to step back. The broader optimization for America has to take priority.

Jason Calacanis

It just doesn’t work like that. I agree with the idea, but I don’t believe any individual member of Congress will take responsibility in that way.

Chamath Palihapitiya

No, they won’t. But the question is whether they can block it.

Jason Calacanis

Or, put another way, the executive branch can slow-roll spending in a lot of ways—except you can’t do that with Medicare and Social Security.

Chamath Palihapitiya

Discretionary spending is about 20%. Mandatory spending—Social Security, Medicare, and Medicaid—is the larger outlay. That’s where we come back to the fact that this will never be addressed until it has to be, because of the political suicide involved.

Jason Calacanis

I agree, but this is where Elon’s fame can be helpful.

There’s a famous Sputnik story that NASA spent millions of dollars engineering a pen that could write upside down, while the Russians simply used a pencil.

That’s what we need to do to the U.S. government. Even if there’s a lot of mandated spending, the real question nobody knows the answer to is whether that spending is useful.

Even if Congress appropriated the money, there has to be a feedback loop that says, “You can just use a pencil. You don’t need the upside-down writing pen.”

If there’s anybody who can broadcast that to the world, it’s Elon. This is where Trump gets enormous leverage by having Elon in the West Wing. Nobody else could give him that leverage. The rest of us would just be chirping into the darkness.

Chamath Palihapitiya

This is the naming and shaming of government waste that will actually work. The DOGE account on Twitter is doing it. They’re saying, “We’re giving foreign aid for this project and that project.”

Will it be perfect every time? No. But you show an empty office, show people not coming to work, and show people wasting money.

David Friedberg

If that’s even real.

Jason Calacanis

There will be back-and-forth, but overall, if you keep naming and shaming each project and do the back-of-the-envelope math for every American, it will work.

Take whatever you saved, divide it by 330 million Americans, and tell every American how much they just paid less in taxes or how much they just saved individually. The naming, shaming, and math are going to work.

Do we want to wrap with the tragedy in D.C.?

We were talking with our friend Sky Dayton, who’s very involved in aviation. He has written several blog posts recently and has invested in a company doing pilot training.

I’ll share 2 things. The first is anonymous. It’s from a friend of mine who is a commercial pilot. He gave me permission to share it, and I posted it.

He wrote: “Honestly, DCA is the sketchiest airport we fly into. I feel like the controllers there play fast and loose, hence the periodic runway incursions. I’ve said to every first officer in my threat briefings that we both need to be on red alert at all times there. DCA calls out helicopter traffic, and vice versa, all the time, but it’s borderline impossible to see them when you’re moving along at 150 miles per hour.”

That’s from a pilot who has no incentive to sugarcoat things.

The second is a message from Brian Yutko, the CEO of Wisk, which is building a lot of these autonomous systems. He said the first point is that autonomous systems already exist, but they don’t take control from the pilot to save the aircraft, even if the software and systems know a collision is about to occur.

That’s the paradigm shift that needs to happen in aviation. Automation should be able to kick in and take over even in piloted aircraft to prevent a crash. That’s the minimum of where we need to go.

Some fighter jets have something called an automatic ground-collision-avoidance system that does exactly this when fighter pilots pass out. It should be possible for commercial aircraft.

His second point was that we need better air-traffic-control software and automation. Right now, we use VHF radio communications for safety and critical instructions. That’s insane. We should be using data links.

The whole air-traffic-control system runs on 1960s technology. They deserve better software and automation in the control towers. It’s totally ripe for change. The problem is that attempts at reform have failed.

There’s a huge opportunity to make this better. This should never have happened.

Sky Dayton has also been pushing hard for the U.S. government to do advanced pilot training. One of the things he constantly says is that a lot of the pushback is union rhetoric about what they perceive to be right for their constituency.

Hopefully this starts a conversation, because people like Sky and Brian are working on the next level of autonomous solutions that can make flying totally safe—safer than it is now.

The crazy statistic is that we haven’t had a commercial-airline disaster in the United States in almost 25 years. Isn’t that incredible?

Chamath Palihapitiya

It’s been 15 years.

Jason Calacanis

It seems like there are questions about why these Apaches are flying around such a crowded airspace. They seem to be shuttling politicians around, and maybe that’s not the best idea in such a dense area.

Chamath, what are your thoughts?

Chamath Palihapitiya

Thoughts and prayers for the families of the people who died. It’s a terrible tragedy.

This is an area where we should invest money and use the private sector and all the incredible innovation available to upgrade these systems and infrastructure.

Jason Calacanis

This has been another amazing episode of the All-In Podcast. Thanks, Travis, for joining us.

Travis Kalanick

Thank you for having me. A lot of fun, guys. This is my first time on a podcast ever.

Jason Calacanis

Right out of the gate.

Travis Kalanick

Yes.

Jason Calacanis

Come back anytime. You were great, man.

Travis Kalanick

I appreciate it.

Jason Calacanis

Very based. The audience is going to like it. Tell us what you think, and we’ll see you all next time. Love you, boys. Bye-bye.