All In Podcast: China's Open-Source AI Models Challenge U.S. Frontier Labs as Memory Bottleneck Drives Industry Inflation
Episode 278 featuring Gavin Baker and Travis Kalanick, recorded June 2026
Chinese AI labs are rapidly closing the gap with U.S. frontier models through open-source releases that now match or exceed the performance of OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8, raising fundamental questions about whether American self-imposed safety restrictions still serve any strategic purpose. The release of GLM 5.2 by China's Z.AI marks a watershed moment, as the model achieves the highest score of any open-weight model on key benchmarks while costing 85% less than comparable closed models from U.S. companies.
China's Open-Source Breakthrough Renders U.S. Safety Theater Obsolete
GLM 5.2 scored 51 points on the artificial analysis intelligence index, the highest ever for an open-weight model, and beat GPT-5.5 on the Frontier SWE coding benchmark while trailing Claude Opus 4.8 by less than one percentage point. The model features 744 billion parameters and a 1 million token context window, released under the MIT license with no regional restrictions. Z.AI's founder told Elon Musk that open-weight models with Fable-level capabilities will arrive sooner than Q1 2027.
David Sacks, speaking from his role in the administration, warned that the current approval process for U.S. frontier models is handing China a competitive advantage. "We do not have months to give away in this race," Sacks said. "The Chinese have been 9 months behind our models, plus or minus 3 months depending on capability. But when they know there's been a breakthrough around something like cyber, they can deploy more resources against that particular problem and catch up faster."
Sacks blamed Anthropic CEO Dario Amodei for creating the regulatory environment that led to Fable being rolled back. "I've been saying from the beginning that we are in a very competitive situation with China. We cannot afford to do things unnecessarily that slow our companies down." He described the situation as a potential self-fulfilling prophecy, noting that "Dario got hoisted on his own petard here" by advocating for government approval processes that now constrain his own company.
Gavin Baker acknowledged that GLM 5.2's quality "has challenged some of my beliefs," but emphasized that distillation from frontier models played a major role. "There's been an immense amount of distillation. No question," Baker explained. "Picture tens of thousands of phones, iPads, and computers that are asking the cloud API through masked accounts very specific questions, and these reasoning traces are being harvested. Those reasoning traces are then fed back into the model during the reinforcement learning process and probably during the pre-training process."
However, Baker noted that now that GLM 5.2 is "so good, it is good enough to do its own RL and the cat may be out of the bag." The critical question is whether upcoming models from OpenAI, Mythose, and SpaceX will reopen the gap.
DRAM Bottleneck Creates Industry-Wide Inflation and Strategic Vulnerabilities
Micron's blowout earnings revealed the extent to which high-bandwidth memory has become the defining constraint in AI infrastructure, with revenue up 4x year-over-year to $42 billion and the company's entire 2026 supply already sold out. Baker called DRAM "the most important bottleneck" in the AI stack, noting that "memory capacity and bandwidth are foundational to the performance of every AI model."
The supply shortage is driving dramatic price increases across consumer electronics. Apple raised prices on its MacBook Neo by 14% to $799 and Mac Studio by 25%, passing on costs it had previously absorbed. "Inflation has come to the desktop," as the hosts described it. Microsoft raised Xbox prices, and Nintendo Switch and PlayStation are expected to follow.
Baker explained that only three companies globally can produce the specialized HBM DRAM needed for AI servers: Micron, SK Hynix, and Samsung. "This is as close to magic as science can get," he said, emphasizing the difficulty of stacking 8, 12, or 16 DRAM dies together with the necessary packaging. A fourth player, China's CXMT, is going public and "will flood the market with cheap consumer-grade DRAM," but lacks the capability to produce the customized chips required for AI infrastructure.
The DRAM shortage has created a peculiar dynamic where it now constitutes 30% to 40% of all hyperscaler capex. Baker suggested this might actually benefit society by slowing the AI arms race. "This may give us as a society time to adapt," he said, referencing what Brad Gerstner calls "the social contract." The high cost of building gigawatt data centers—now $35 billion for semiconductors plus $25 billion for power and cooling—means "even for the hyperscalers, economics matter."
Chamath Palihapitiya noted that energy constraints are compounding the memory bottleneck. "Since 2021, about 40% of all data centers get contested," he said. "I think that number is going to go up. Whatever forecasted energy consumption we are looking at in AI is grossly imbalanced. There is very meager supply and there's effectively infinite demand."
Distributed Compute and Orbital Infrastructure Emerge as Solutions
The terrestrial bottlenecks are accelerating interest in both distributed computing and orbital data centers. Baker outlined the economics that make space-based compute increasingly viable. To build a gigawatt data center on Earth costs $60 billion total—$35 billion in Nvidia semiconductors and $25 billion in power and cooling equipment. Once Starship becomes rapidly reusable, launch costs to put equivalent compute in orbit would fall to just $5 billion, bringing total costs to $40 billion.
"The $25 billion that is power and cooling is clearly inflationary," Baker explained. "It may be that in three or four years it's $70 billion terrestrially versus $40 billion in space. And that $5 billion as Starship becomes rapidly reusable is likely deflationary." The orbital approach eliminates the need for massive cooling infrastructure, with racks in space linked by lasers rather than traditional data center architecture.
The hosts discussed Tesla's newly trademarked "Megapod" system, describing it as modular data center hardware that could be deployed at Supercharger stations where power infrastructure already exists. Palihapitiya described the appeal of prefabricated compute modules that can be manufactured in warehouses and deployed in 90 days rather than the typical two to three year timeline for traditional facilities.
Travis Kalanick highlighted that his company Adams is exploring distributed inference networks, where surplus compute from various sources can be aggregated. "If you're a corporation, you're going to have a router and every query that somebody comes in, every task that needs to be done at your company, that router is going to send it to your version of DeepSeek or whatever open-source model you're using," Baker explained. "Then at some point in the workflow, a frontier model may or may not come in to check it."
The disaggregation of inference into prefill and decode operations enables this distributed approach. Baker explained that prefill—understanding the question and context—is memory capacity bound, while decode—generating the next token—is memory bandwidth bound. Companies like Groq (which Nvidia acquired) and Cerebras enable older H100 or A100 GPUs to be paired with specialized decode chips, extending the useful life of existing hardware to seven or even 12 years.
DSA's Socialist Insurgency Captures Democratic Party Infrastructure
New York City Mayor Zohran Mamdani's Democratic Socialists of America swept all three congressional primaries where he endorsed candidates, unseating established Democrats in a development that Sacks described as the left's version of Trump's populist takeover. "The choices of the future are going to be communism or socialism in the Democrat party or nationalism in the Republican party," Sacks said. "Those are the two populist directions."
The victorious DSA candidates support abolishing the Senate, eliminating police forces and prisons, dissolving ICE with amnesty for all, replacing the president and Supreme Court with bodies subordinate to Congress, and implementing proportional representation with ranked choice voting. "This is a total makeover of our constitutional system," Sacks noted. One winner, 32-year-old Shioalier, has declared she wants to "eradicate Western civilization" and attended a rally celebrating Israeli civilian deaths the day after October 7.
The DSA co-chair Josh Block was explicit about strategy: "We're using the Democratic Party as a ballot access vehicle. Not because we share its goals. We build our own organization, get elected under the Democratic label, caucus with Democrats when it's useful, and push our own agenda from the inside. We see the Democratic establishment as an obstacle, not a home."
Baker argued that Mamdani himself is "one of the most talented politicians I've ever seen in my lifetime," comparing him favorably to AOC. "He can give a great speech. He's good in an interview. He can tap into all of this. He's kind of a chameleon who can shift." However, Baker emphasized that the DSA's actual voting base is "relatively wealthy white liberals who are downwardly mobile" rather than working-class, poor, Black, or Hispanic Americans whom Democrats traditionally represented.
The hosts connected the socialist surge to two generations of Americans who feel economically locked out. "They don't believe that they can participate in the system. They feel the system's rigged," Baker said. "And if somebody comes along who speaks to them and they have no conception of socialism and what happened in Germany or during the Red Scare, they have no idea what socialism or communism is."
Kalanick offered a philosophical frame: "Truth and justice is the immune system for society. When the immune system is suppressed, all the social ills flare up." He added that "communism is in all of us. Communism is in our blood as humans. Have you ever in your life been lazy? Have you ever wanted something for nothing? The difference is, do you make that a way of life?"
Cerebras Breaks Deal Price as Public Market Discipline Returns
Cerebras dropped below its IPO price in its first earnings report as a public company, triggering what Baker described as "price insensitive selling" from portfolio managers who automatically exit positions that break deal price. "There are people who run giant funds who I know personally where if a stock breaks deal price they sell no matter what," Baker said. "Shorts know that if a stock gets close to deal price they short it because they want to break deal price and then make a quick 10% or 20%."
Baker argued the market reaction was primarily about supply chain timing rather than fundamental business issues. He outlined how the transformational contract Cerebras signed with OpenAI in December 2025 wouldn't show revenue impact until Labor Day 2026 at earliest. "It takes 4 months from when we make that order for Taiwan Semi to make the chip, then it takes us 2 months plus or minus to turn that chip into a server. And then if we're lucky and we can find the power, it takes us a month to energize that chip and start making tokens with it."
The key question for Cerebras investors is execution velocity on power deployment. "Outside of the hyperscalers, the only companies that have ever brought on more than a gigawatt are CoreWeave, Crusoe, and SpaceX AI," Baker noted. If Cerebras could add 50 megawatts per month in 2027, they would exit the year at roughly a $9 billion cloud computing run rate against a current market cap below $40 billion.
Baker advised companies going public to "price this in such a way that we're not going to break deal price in our first 9 months as a public company," while Palihapitiya advocated more forcefully for auction-based IPO processes to discover true market clearing prices rather than relying on underwriter judgment.
Anthropic Valued at $3 Trillion Ahead of Anticipated Public Debut
Baker offered a stunning valuation assessment for Anthropic, arguing the company "is worth $3 trillion today" and "that is roughly where it would probably trade as a public company." He based this on revenue projections showing Anthropic ending 2026 "well over $100 billion" in revenue and likely reaching $200 billion to $300 billion in 2028. "It's probably not going to trade at 10 times that number and it will be very profitable at that scale because it'll be inference dominated," Baker said, citing reports of 85% gross margins on inference.
The valuation would make Anthropic larger than OpenAI's expected public valuation and represent one of the largest technology offerings in history. However, Baker downplayed concerns about market absorption. "It's just shifting from private to public," he said. "In the scale of global capital markets, these seem like really big numbers. You're just moving from the private markets to the public markets which are even bigger."
The conversation highlighted how radically valuation scales have shifted in the AI era. Baker and Kalanick reminisced about Uber's Series C in 2014, when a $17 billion valuation was considered groundbreaking and controversial. "We got pilloried in the press for not knowing what we were doing," Baker recalled of pricing Uber at $14 billion at Fidelity. A decade later, Anthropic approaching a $3 trillion valuation barely raises eyebrows among sophisticated investors.
Regarding SpaceX's recent IPO, Baker noted that unlike typical tech IPOs, most SpaceX stakeholders had regular liquidity opportunities over the past decade through tender offers, reducing potential selling pressure. "Almost half the employees at SpaceX bought on the IPO," he said. However, Palihapitiya countered that the 8x appreciation in the final year before going public—from a $350 billion private valuation to public pricing—could still create meaningful liquidity events for long-term holders who experienced years of more modest 20% to 30% annual gains.