Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter
In a Nutshell
Socialists swept three NYC congressional primaries with DSA-backed candidates, signaling a radical left takeover of safe Democratic seats via organized turnout, migrant votes, and rank-choice voting. Chinese open-source model GLM 5.2 has reached near-frontier performance on coding and long-context tasks at 85% lower cost through distillation and Huawei hardware, while Micron's HBM revenue exploded 4x to $42B with supply sold out through 2026. AI infrastructure faces severe memory and power bottlenecks, driving modular/orbital compute solutions and a potential $4T wave of AI IPOs.
These notes were generated by AI and may contain inaccuracies.
Freeberg took a mental health day after the socialist sweep in New York City. Guests include Travis Kalanick from Adams and Gavin Baker of Atreides Management.
Gavin Baker describes the SpaceX IPO as a culmination of decades of hard work and references Bill Belichick's Cincinnati comment about treating every game as 0-0.
On Tuesday, New York City Mayor Mamdani went three for three in the candidates he endorsed, and they all won their primaries. In the 10th district, Brad Lander defeated two-term incumbent Dan Goldman. The 10th district is one of New York's richest districts and includes the West Village, Wall Street, Dumbo, Cobble Hill, Carroll Gardens, and Park Slope. In New York's 13th district, Shiovalier beat a five-term incumbent backed by House Speaker Hakeem Jeff. Shiovalier is a 32-year-old Democratic Socialist with a history of spicy remarks representing Harlem and the West Bronx. In New York's seventh district, Claire Valdez won the open seat over the handpicked successor. The seventh district is a DSA stronghold known as the Kami Corridor including Bushwick, Williamsburg, Long Island City, and Greenpoint.
According to Polymarket, Mamdani's sweep chances were just 26% before election day. These DSA candidates performed better with younger, college-educated, and high-income voters. These are all safe Democratic seats, so the DSA will likely win. The DSA caucus is described as rich people pretending to be poor.
AI is described as the greatest economic leveler that creates equality and evens the starting line for every person on earth. AI takes the world's knowledge from the internet and Google search and transforms it into expertise and intelligence, allowing every person to have an equivalent to a super-founder as a co-pilot. There is no gatekeeping preventing access to this capability. The only limitation is directing that energy into something productive. Claims about jobs being lost and water consumption are described as fabrications and misinformation created to help specific actors in the AI race.
Some key congressional races were described as a referendum on AI. Anthropic-funded anti-AI groups competed against OpenAI-funded pro-AI groups, with pro-AI groups winning narrowly in Utah and New York.
The DSA platform includes abolishing the Senate, abolishing the carceral state including police forces and prisons, abolishing ICE, granting amnesty for all, ending all deportations, replacing the president and Supreme Court with an executive and judiciary subordinate to Congress, abolishing the electoral college, replacing the two-party system with multi-party democracy, expanding the House of Representatives, implementing proportional representation and ranked choice voting, supporting free Palestine, pursuing public ownership of major corporations, and defunding the Department of War.
Shiovalier is an unemployed 32-year-old PhD candidate who has been in college for 10 years. She has declared she wants to end Western civilization, used the American flag as a napkin, attended a rally one day after October 7th celebrating the slaughter of Israeli civilians, calls white women ugly colonizers, calls for defunding the police and abolishing all prisons and borders, calls US service members war criminals, says the US is a disgrace, and has written favorably about communism and seizing the means of production.
DSA co-chair Josh Block stated the organization is using the Democratic Party as a ballot access vehicle, building their own organization, getting elected under the Democratic label, caucusing with Democrats when useful, and pushing their own agenda from the inside. They view the Democratic establishment as an obstacle, not a home.
Mamdani would not have won the mayoral election if only native-born New Yorkers voted. The mass migrant vote swung the election. The combination of recent college graduates as organizers and the migrant movement is taking over the Democratic Party. Democratic policies on open borders and cracking the melting pot through multiculturalism and wokeism enabled this shift. Immigrants are no longer being assimilated into the American system.
Trump wanted to delete agencies like the IRS, USAID, Department of Education, and EPA but did not want to delete the US Constitution or Western culture. Travis Kalanick references truth and justice as the immune system for society. When the immune system is suppressed, social ills flare up. Indicators include whether truth is winning or losing and whether justice is winning or losing. Communism exists in all humans through laziness and wanting something for nothing. The difference is making it a way of life when ecosystems allow it without consequence.
The DSA represents a fusion of factors. The Democratic Party was historically the party of working-class people trying to create opportunities. The DSA voting base consists of relatively wealthy white liberals who are downwardly mobile. They are losing votes with working-class people, poor people, Black Americans, and Hispanic Americans. The DSA is described as dangerous for electing anti-American candidates who have called for violence to eradicate America. A class of people went to elite schools, grew up in nice circumstances, and went into the NGO nonprofit machine instead of industry. Elon Musk has done more to decarbonize the planet than every activist combined, times 10.
Government outsourcing of funding to NGOs is a major problem. The Curley effect describes pursuing policies that drive out rivals so jobs can be given to friends and allies. In California and New York, per capita spending on homelessness has more than doubled or quadrupled and gone to NGOs while outcomes have gotten worse. Tens to hundreds of billions flow to these NGOs.
Zohran Mamdani is described as one of the most talented politicians seen, capable of great speeches, strong in interviews, and a chameleon who can shift. He is the reason the DSA is ascendant, not their ideas. Two lost generations feel they will do worse than their parents due to housing, college debt, healthcare, and a belief the system is rigged. They have no conception of what socialism or communism produced historically in Germany or during the Red Scare. The DSA wraps communism in the branding of democratic socialism, described as Coke Zero of socialism.
Canada, the UK, and Australia veered far toward socialism before the United States. All three countries show unmitigated disaster on reasonable measures of progress despite virtue signaling on social issues, immigration, and climate change. Each country now bans social media for those 16 and under. This is compared to making sure a kid does not get addicted to a drug too early. Florida has already implemented this in the United States.
Social media bans starting at age 16 are already in place in Canada and may come to the US. While social media is acknowledged as harmful to kids and adults, the real purpose of age-based restrictions is to force adults to identify themselves and deanonymize, enabling a full-scale censorship regime similar to what is being contemplated in the UK. Censorship is not about harmful content but about content that people in power disagree with.
Age gating is really about controlling adults rather than protecting kids. A social media ban under 16 would be beneficial but comes at a high cost of restricting anonymous accounts on X that say things EU powers do not like. Free speech and X have prevented a much worse world.
Communism has failed utterly and ended in death and misery wherever tried across different cultures and mechanisms, yet each generation may still need to experiment with it. The saving grace is preserving free speech. DSA policies and super progressive Democratic policies lead to measurably bad outcomes.
If you care about black lives, electing a Republican DA drops all-cause mortality for young black men in that city by 7%. Progressive environmental regulations prevent building solar, which is the only thing that matters since the world will run on sunlight. Progressive education approaches such as getting rid of elite schools and eliminating math in California are profoundly disadvantageous to low-income people. On crime, 0.1% of people commit 70% of crimes, and dealing with that 0.1% would effectively eliminate crime.
The United States has been a very stable political and economic entity geographically for a long time. If socialists get total control of cities and drive out productive people, recovery becomes uncertain. This is not just a New York issue. In LA, a DSA-aligned candidate beat Spencer Pratt thanks to ballot harvesting, and DSA now holds roughly half the city council seats while growing. DSA thrives in low-turnout elections because they are highly organized and know how to exploit ballot harvesting and rank-choice voting systems.
Mandani is now the spiritual leader of this movement, more radical than AOC or Bernie. The movement is building a big tent including Bernie, Mandani, and others. Established progressive leaders and the Democratic establishment are bending the knee to this energy because they fear primary challenges. Three major congressional races saw DSA candidates win, unseating strong incumbents. Every congressional race in blue districts will now tilt toward DSA positions to avoid similar upsets.
The Israel issue is now hugely salient in Democratic primaries. Dan Goldman, a two-time congressman who led the Trump impeachment effort and checked every progressive box, lost to Brand Lander primarily over Israel. Lander went to a mosque to denounce what he called the genocide in Gaza. 80% of Democrats now disapprove of Israel. Young Democrats are highly motivated by this issue.
Within the Republican Party, views on Israel split by age. Older establishment Republicans and Fox News viewers maintain high approval ratings for Israel, but under-50 Republicans show 57% disapproval. Tucker Carlson left the party over this issue, and Megyn Kelly has expressed strong opposition.
Parallels exist between the post-9/11 response and current events. When a country suffers enormous injury, it grants leaders authority to set things right, as seen with the Patriot Act after 9/11. People are conflating Jews, Israelis, and Netanyahu, which is unfortunate. Reasonable criticism of Netanyahu's leadership does not justify folding an entire country and religion together. Israel has a giant PR problem and needs young, English-fluent, social-media-savvy spokespeople under 35 for America and Europe. Netanyahu's wife reportedly had a promising young spokesperson fired after October 7th, and no adequate replacement has been found.
A major reason many Democrats voted for Trump was hearing what their children were being taught during COVID. The American educational system should not include overtly anti-American content. America has made mistakes but is about as good as it gets. Slavery was endemic worldwide except in some East Asian countries and is not a uniquely American problem. The melting pot is a beautiful thing that allows people to keep their culture while joining American culture. America is the only vaguely successful multicultural society on Earth.
DSA co-chair Gustavo Gordillo stated the group is part of the Democratic Party caucus but does not agree with how the party runs its apparatus. They are building independence through volunteer-led movements, believe everyone should be trained to participate in politics, and say the Democratic Party is funded by billionaires while claiming to represent the working class. They view this as forcing a choice between the billionaire class and the working class. Mandani is applying the Trump playbook of building a big tent and taking over the party from within through superior communication. Socialism is communism and leads to suffering.
Z.AI released GLM 5.2, a frontier-class open-source model with 744 billion parameters, a 1 million token context window, and MIT license. It is fully self-hostable with no regional restrictions. It scored 51 on the Artificial Analysis Intelligence Index, the highest score of any open-weight model. It beat GPT 5.5 on the Frontier SWE coding benchmark and trails Claude Opus 4.8 by less than 1 percentage point. API usage costs 85% less than GPT 5.5 for comparable performance. Z.AI's founder told Elon Musk that open-weight fable capabilities will arrive sooner than Q1 2027.
Distillation is occurring at massive scale. Chinese operations use tens of thousands of phones, iPads, and computers with masked accounts to query frontier model APIs, harvest reasoning traces from every token, and feed those traces back into their models during reinforcement learning and pre-training. This allows models to get very close to the frontier at a fraction of the cost. GLM 5.2 is now good enough to perform its own RL, potentially putting the advantage permanently out of reach.
The future belongs to composable models. Enterprises will maintain their own open-weights model trained on their data alongside frontier models. A router will direct queries: 85% or more may go to the open-source model, with only the hardest queries escalated to frontier models. This creates a council of LLMs where each model handles what it does best. Open-source models shift economic value from frontier lab margins to infrastructure providers but are not bad for AI overall. Frontier tokens may continue capturing 90% of economic value while open-source tokens represent over 80% of tokens processed.
Nvidia is effectively the American open-source champion because it can release models like GLM 5.2 or better whenever desired.
Jensen Huang faces a classic channel conflict by competing with customers. Nvidia competes with Elon Musk on self-driving while Musk develops his own chips. Frontier model companies should carefully consider A6 and the incentive it creates for Nvidia. OpenAI launched their Jalapeno chip this week, built by Broadcom, signaling a direct challenge to Nvidia. Nvidia could respond by entering the model space and starting an OpenAI competitor.
China has been strong in open source for some time. GLM 5.2 represents a significant advancement as the best openweight model for coding, software engineering, and long context agent work. It scored just a tick below Opus 4.8 and performed comparably to GPT 5.5. The model sits at the level of previous frontier models from Anthropic and OpenAI.
Fable has been rolled back and GPT 5.6 is navigating new approval processes. Dario Amodei and his communication style primed officials to act quickly on credible jailbreak reports from trusted partners. The current situation leaves Anthropic and OpenAI models in regulatory purgatory while a Chinese openweight model matches currently available Western models.
The U.S. is in a highly competitive situation with China. The AI strategy from the beginning has been about winning the AI race and maintaining global competitiveness. Unnecessary actions that slow U.S. companies should be avoided. Dario Amodei may have gotten exactly what he wanted through a calculated strategy to provoke government action, creating a regulatory moat. He has advocated for a federal regulator and an FAA-style approval process for AI models.
Dario got what he craved: a labyrinthine government approval process that rewards regulatory capture. Fable's rollback shows he may have been hoisted on his own petard. Once the jailbreak issue is resolved, Anthropic should return to market quickly. OpenAI should not face unnecessary delays. Months cannot be given away in this race.
Cyber risks exist but the response should be proactive. White hats must find all vulnerabilities and execute a major upgrade cycle with patches rolled out before exploitation. Clamping down on model use will not slow Chinese development since they operate outside U.S. jurisdiction. The Chinese are already at Opus 4.8 level. Z.AI's founder stated they will reach Fable-level capability before Q1. Chinese models have historically been 9 months behind Western models plus or minus 3 months, but they can accelerate on specific breakthroughs like cyber by deploying more resources.
A shock clock is running. Smart actions are required rather than broad slowdowns that will not affect Chinese development. Tools must reach the cybersecurity industry as a force multiplier. A major upgrade cycle must happen quickly within the remaining months. China is 6 months behind on models and 24 months behind on silicon but only a few months behind overall.
China's government has pushed AI labs to develop and train on Huawei chips. Deepseek V4 was claimed to be trained on Huawei chips. Z.AI states the GLM 5 family was trained entirely on clusters of Huawei Ascend 910b chips. China is executing a strong indigenization push to establish Huawei as the national champion. GLM 5.2 inference is optimized for Huawei chips.
China will package Huawei-optimized models as AI in a box and sell them globally at a fraction of the cost. This follows their pattern of offering better, cheaper, faster alternatives. The U.S. must be pro-export because China will be competitive within one to two years. Lost global market opportunities will result from unnecessary export restrictions to friends and partners.
GLM 5.2 hosted at Z.AI refused to answer questions about Taiwan and Tiananmen Square. The model performed exceptionally on tourism recommendations for Paris but declined to create infographics or 3-day agendas, citing insufficient time and directing users to another model. Chinese models contain censorship and political bias out of the box.
American companies have successfully forked Chinese models and removed censorship. Perplexity did this early on, restoring content on Tiananmen Square and similar topics. Censorship is not a fatal problem because American companies can fix it when forking and customizing open source models.
Micron is one of only three companies worldwide making high bandwidth memory chips that sit on top of Nvidia GPUs. Micron's entire 2026 HBM supply is sold out. SK Hynix and Samsung are the other two HBM manufacturers. Micron's revenue increased 4x year-over-year from $9 billion to $42 billion, beating expectations by 16%. Q4 guidance jumped to $50 billion from $43 billion. The stock is up 10x. Gavin correctly predicted HBM makers like Micron would be the best performing asset, achieving 14x returns.
Apple is passing on DRAM cost increases to customers after previously absorbing them. The new MacBook Neo increased from $699 to $799, a 14% rise. Mac Studio prices increased 25%. Inflation has reached the desktop.
DRAM represents the most important bottleneck because memory capacity and bandwidth are foundational to AI model performance. Elon Musk is focusing the Terrafab on memory for this reason. The bottleneck will persist for some time. Micron announced supply chain agreements with floor and ceiling pricing covering 50% of revenue with just four customers. Floor pricing in these contracts exceeds prior cycle peaks from a gross margin perspective, potentially transforming the industry.
Most semiconductor supply chain segments have rerated with higher multiples. HBM DRAM is increasingly a customized chip. CXMT is going public in China and will flood the market with cheap consumer-grade DRAM. However, only three companies can produce the DRAM required for AI servers. This technology sits at the intersection of magic and science. Terrafab will be an important part of the solution. Memory stocks trade at a discount relative to the rest of AI.
DRAM will likely comprise 30-40% of all hyperscaler capex next year. Hundreds of billions of dollars will flow directly to DRAM. Higher costs for gigawatt data centers may force society to adapt and give time for the social contract around AI to develop. CXMT will address consumer-grade DRAM, but AI server DRAM remains constrained.
HBM, SOCAMM, and LPDDR represent the types of DRAM needed in AI data centers. These are semiconductor-grade products, not consumer-grade DRAM. HBM 3 stacks eight dies high, with stacking increasing to 12 and 16 dies. Stacking and packaging creates advanced technology requirements. Only three companies can execute this process.
Data centers are consuming DRAM at scale, creating price pressure on consumer electronics. One HBM chip uses multiple stacks of DRAM. New capacity takes years to ramp. Micron's New York plant broke ground then faced shutdown over environmental issues. SK Hynix and Samsung are based in South Korea. Micron is the only U.S.-based HBM provider.
Building a new fab is difficult in deep blue states due to regulatory hurdles. New York offered incentives but regulatory barriers proved decisive. Texas may be a more viable location. The incentive dynamic has reversed. States previously courted fabs, but now fabs seek states that can actually permit and build facilities.
A normal fab requires a two to three and a half year process. Elon Musk's track record suggests Terrafab may be stood up faster than historical precedents, though the intersection of magic and science makes the process inherently difficult.
Standing up a 1 gigawatt data center costs $35 billion in Nvidia semiconductors and $25 billion in power and cooling equipment. The $25 billion includes significant human labor for installation, making it clearly inflationary. Orbital compute economics compare $35 billion of silicon per space installation against $25 billion of terrestrial power and cooling. Starship reusability could reduce launch costs to $5 billion for a gigawatt of compute in space, totaling $40 billion versus $60 billion terrestrially.
Since 2021, approximately 40% of data centers have faced contestation. This rate is expected to increase given meager supply and effectively infinite demand. This imbalance pulls forward the economic case for orbital compute. Terrestrial compute from hyperscalers will face economic disadvantages versus SpaceX at scale. Remaining terrestrial capacity will be extremely valuable, particularly in the hundreds of megawatts to gigawatts range.
A trademark was filed June 18, 2026 for Megapod: modular data center hardware for artificial intelligence computing including networks of computer servers, AI processing hardware, network hardware, electric power distribution units, and cooling systems sold as a unit. The filing describes self-contained modular computing hardware systems for AI workloads.
The trademark is interpreted as describing giant battery packs with GPUs at existing supercharger stations. Turning on terrestrial compute requires land, zoning approval, and power access, with power being very difficult. Basis of design documents from Anthropic, OpenAI, CoreWeave, AWS, and GCP range from 50 to 500 pages of technical specifications.
Surge pricing is emerging as a response to compute constraints. Basis of design restrictions are relaxing. Google and Nvidia's liquid cooling requirement for large girth supercomputer racks is being reconsidered. Air-cooled older hardware is being evaluated for utility. Shipping container-based solutions that plug into power on concrete pads are being developed by multiple companies including Dell.
Dell and Vertiv manufacture the racks and modules that enable Tesla's megapods. These prefabricated units can be assembled in a warehouse, trucked to site, craned into place, and powered up within a 90-day build cycle. The discussion centers on whether these megapods will be available for external purchase or remain reserved for Tesla and SpaceX internal use.
Supercharger locations present access and security constraints that make many workloads impractical. Liquid cooling requirements and physical security needs limit feasibility at retail-facing sites. One participant referenced 500 properties with existing energy infrastructure, mechanical cooling systems, and natural gas access as potential locations for robotics and physical AI compute deployments.
Running GPUs for personal use in non-traditional spaces like kitchens is feasible, but contributing hardware to a shared cloud requires signing liquidated damages and meeting SLAs that most operators are unwilling to accept. Security architecture in traditional data centers includes man traps with fingerprint access that create layered physical security.
Training workloads suffer dramatic efficiency losses when GPUs are physically separated. Even two kilometers of fiber introduces unacceptable latency for passing jobs between GPUs. Inference workloads are more tolerant of distribution because latency sensitivity is lower, enabling distributed inference clouds.
A startup is developing four-GPU units with battery backup for residential deployment. Homeowners receive power bill discounts in exchange for contributing inference capacity to neighborhood-level services. The units are sealed to prevent tampering.
Modern inference splits into two distinct phases. Prefill involves understanding the question and conversation history; performance improves with greater memory capacity. Decode generates the next token and is memory-bandwidth bound; performance improves with faster token generation speed. This separation enables specialized hardware optimization.
Cerebras and Groq offer decode-optimized chips that can sit in front of older Nvidia GPUs such as H100s and A100s. This approach allows repurposing existing hardware in modular pods or shipping-container data centers. Extended hardware utilization periods of seven to twelve years improve financing economics for the AI buildout.
Distributed inference supports free-speech-oriented and open-source models. While the compensation mechanics for homeowners remain undefined, the concept aligns with broader American priorities around competing with China and maintaining open compute infrastructure.
Bittensor validators verify that contributors deliver promised hardware. Targon enables hourly rental of H200 GPUs at three to four dollars with no permission required. Surplus and unused compute can be contributed to these networks.
A hypothetical offering would bundle GPUs inside every Powerwall, granting owners a discount while Tesla monetizes idle capacity. Combined with Starlink, this would create a distributed network of home battery systems with integrated silicon.
Cerebras completed a successful IPO but traded below deal price in its first two days as a public company. Portfolio managers who treat breaking deal price as a broken promise triggered selling. Short sellers also targeted the name to force a break of the IPO price and capture quick gains.
Anthropic is projected to be worth approximately $3 trillion as a public company. The firm is expected to exceed $100 billion in annual revenue by year-end, with inference workloads driving 85% gross margins. At that scale the business becomes highly profitable.
SpaceX investors and employees have had semiannual liquidity opportunities for the past decade. Roughly half of employees purchased shares during the IPO. Many positions are held through SPVs that are unlikely to distribute soon. Elon Musk is not expected to sell.
A $25 billion OpenAI contract was signed in December. Wafer orders placed with TSMC require four months for production, followed by two months to convert chips into servers and one month to energize and begin token generation. Earliest revenue impact appears around Labor Day, with meaningful contribution building thereafter.
Outside hyperscalers, only CoreWeave, Crusoe, and SpaceX AI have brought more than one gigawatt online. The critical variable for Cerebras is megawatts energized per month rather than competitive positioning. A 50-megawatt-per-month ramp in 2027 would produce a $9 billion annualized cloud run rate by year-end.
Pricing IPOs to avoid breaking deal price in the first nine months is strongly advised. Auctions eliminate the need to guess the right price and reduce influence from multiple stakeholders seeking higher valuations. Facebook and Rivian both traded below IPO price in their first year without indicating fundamental weakness.
Anthropic, OpenAI, and additional AI companies could bring roughly $4 trillion in new public market value. A 5% float of Anthropic alone represents approximately $150 billion. Global public markets are viewed as large enough to absorb these offerings despite the unprecedented scale.
Ten years earlier, pricing Uber at $14–17 billion was considered groundbreaking and drew public criticism. The same round later reached $42 billion. Capital markets have since normalized much larger private valuations before public listing.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
