Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback
In a Nutshell
Anthropic is targeting a $2T IPO in October on a $100-120B ARR trajectory growing 10x YoY, while Nvidia is structuring $500B in asset-backed GPU financing to unlock compute demand. Zuckerberg's open-source manifesto directly challenges Anthropic's centralized AI safety vision, arguing for decentralized models and individual empowerment over concentrated power. The AI market is shifting toward open-source alternatives at 90% lower cost, with physical constraints around energy and compute becoming the primary bottlenecks rather than demand.
These notes were generated by AI and may contain inaccuracies.
David Sacks and Gavin Baker join the All-In podcast for a short crew but long docket. Gavin is in SpaceX afterglow, describing it as the biggest win of his career, though noting there's only one SpaceX and quoting Bill Belichick: "it's on to Cincinnati." The SpaceX story continues with more ahead.
The hosts discuss a recent All-In interview show with Rahm Emanuel that got "super spicy." Rahm came in "full guns blazing" and the interview is described as "really great." They note Rahm represents a throwback to the Clinton-Obama DNC era. The verdict is that Rahm has "no chance whatsoever in the Mamdani DSA Democrat party." This moment mirrors the civil war within the Republican party, with hope that moderates have a chance against "lunatic democratic socialists."
Gavin notes Rahm's "incredible amount of experience" having worked for Obama and Clinton, served as mayor of Chicago, and ambassador to Japan. Rahm was "awesome on foreign relations" and went "totally hard" against democratic socialists, calling socialist policies "a road to nowhere."
At time of recording on Thursday, Anthropic is targeting a $2 trillion IPO, which would break SpaceX's record-setting $1.75 trillion IPO. The Financial Times reported Wednesday night that investors expect Anthropic to go public in October (six to eight weeks away). Polymarket gives an 80% chance of Anthropic IPOing this year, with Brad Gerstner also signaling it would happen sooner rather than later.
Anthropic's annualized run rate is expected to end the year between $100-120 billion, representing "extraordinary ramp up of revenue that we have never seen in Silicon Valley." At $2 trillion valuation, this represents 16-20 times sales. Revenue growth over the last year has been 10x, with Anthropic growing 10x year-over-year for the last 3 years.
Anthropic still has AI's top model, but Grok is "catching up quickly" and OpenAI is "not too shabby as well." Polymarket gives a 67% chance that at year-end, Anthropic will have the top model. On Thursday, Anthropic was reported in talks to buy AI startup DART for $6 billion, whose software "lowers the cost of AI training and inference by making chips more efficient."
Gavin is not currently an Anthropic shareholder, noting the Anthropic-SpaceX datacenter deal came too late. Elon decided to rent excess compute from Colossus to Dario and Anthropic after getting comfortable with Dario's morals, stating they "don't set off my evil detector."
The rumored numbers are "exceptional" and unprecedented in capitalism history. April and May moments were "some of the most extraordinary moments in the history of capitalism." On the margin, Anthropic is losing share to OpenAI, open source, and Grok, yet still growing exceptionally fast. The AI pie is "getting ginormous" - even if Anthropic loses percentage share, the real number continues growing. AI broadly accelerated in July, contrasting with public stock market action.
Gavin is "a little skeptical" of the $2 trillion valuation. With SpaceX, bankers jockeying for lead left leaked to press claiming $2 trillion when the actual target was $3-4 trillion. These leaks come from bankers who lost lead left and want to make lead look bad. If Anthropic prices at $2 trillion, it means testing the waters and roadshow went "incredibly well."
Gavin notes that responsible IPO pricing allows absorption of lockup without creating volatility that's "very disconcerting for employees." The goal is not distracting employees at this critical moment. If priced at $2 trillion and trading to $3 trillion, that would be expected.
Sacks questions whether the revenue ramp can continue, noting if growth rate continues, they'd hit $1 trillion ARR by end of next year. Physical constraints include compute and energy availability. The TAM appears big enough with demand from new AI applications, particularly agents. The question becomes whether they can physically meet demand and maintain market share.
Anthropic made a "huge bet on coding" that proved prescient. They watched Cursor's utilization and decided to "take that business," moving into verticals where users were building successfully on their platform. OpenAI has now pivoted to coding with GPT 5.6, and their growth rate has accelerated from 3-4x annually to over 20% month-over-month.
Gavin reports internally Anthropic is "very confident" with Dario saying Anthropic might be "the only private company in the world at some point" in an Anthropic-maximalist vision with only governments as competitors. Gavin would "take the under" on this prediction. When Anthropic was at $60 billion ARR, they thought they could reach $600 billion in one year - now they're at over $80 billion without slowing down.
If Anthropic reaches $200 billion, that's "an amazing outcome" making them peers to Microsoft's Office franchise or Google's search franchise. If they reach $4-500 billion, they'd be the biggest software company.
TAM estimates range from $25-65 trillion in knowledge work, with the $25 trillion figure considered potentially low by an AI institute head at one of the three largest investment banks. The question is whether AI represents labor substitution or accelerating growth. Evidence suggests accelerating growth, with more job openings for software coders today than a year ago. The epicenter shows increased demand for coders.
Supply constraints represent "where the rubber meets the road." Data center construction is compared to the TV series Landman - it's "atoms, not bits" requiring orchestration of thousands in remote locations under 110° heat. America excels at solving these problems once geared up.
Turbine blade constraints are manufactured in only two facilities (America and Europe), but facilities now run 24-hour shifts. Elon bought a turbine company personally to go fast. Private jet residuals are strong as engines are repurposed as turbines for data centers.
Texas Governor Abbott announced an energy audit this week - not a moratorium, but requiring solutions for grid impact. The narrative against AI includes false claims about water usage (off by factor of 100,000), electricity costs (actually lowered), and other issues. Communities embracing data centers see 10x tax revenue increases, better schools, and improved services.
The Chinese Communist Party runs active anti-AI, anti-data center campaigns. Media hysteria about data centers does disservice to low-income, small-town, and minority Americans. Water usage is manageable - less than a golf course, with water recirculating.
Every aspect of the data center story in media is described as wrong. Data centers don't use up community water, lower electricity costs when providing own power generation, and sell excess back to grid while paying for upgrades. Rising electricity prices in blue states result from decarbonization and green regulations.
Noise pollution is legitimate but manageable with proper siting. Temperature increases of a couple degrees in valleys are monitorable. Almonds and golf courses use far more water than data centers.
Anthropic's growth rate will slow considerably next year. Tripling to $300-400 million is possible, but reaching $1 trillion next year is not. GLM-2 versus Claude Opus pricing shows 90% cheaper alternatives. Corporate America will embrace open source solutions despite setup complexity, following the pattern from 20 years ago.
Twenty years ago, corporate America faced concerns about open source adoption, fearing they would get rug pulled or held hostage by large tech corporations. Instead, they recognized that open source was the safer bet because it prevented technology lock-in. This same dynamic is playing out today in corporate America and startups, making open source increasingly valuable.
Doresh, a podcaster close to the AI scene, posted that Anthropic's AI and constitution are wired to do what Anthropic thinks is best for humanity. While acknowledging that Anthropic's founders are genuinely good people doing what they believe benefits humanity, the concern centers on unintended consequences in an unpredictable world. Open source ensures a rich variety of AIs and diversity of AI systems, which creates a better outcome for humans and America than having one, two, or three dominant models.
Elon Musk will be judged kindly by history for dedicating Grok to the pursuit of objective truth. Musk explicitly believes in open source and that technology should be shared for the common good, which was his original motivation for backing OpenAI.
Mark Zuckerberg released a 6,500-word essay titled "The Future is for Everyone: The Path to a Positive AI Future," which paradoxically took up OpenAI's original manifesto. The essay presents an extremely optimistic AI worldview where Meta would release open source models, create agents for everyone, make them free, and provide everybody in the world with their own tutor and unlimited abundance. Zuckerberg has picked up the abundance crown that Sam Altman dropped when OpenAI became a private company. He argues that super intelligence is invention, not automation, and that AI safety should be based on the balance of power rather than singular centralized intelligence.
Zuckerberg takes a shot at effective altruists and Anthropic, questioning why they are rushing to create a future they believe is dystopian where everyone is out of work and AI goes rogue. The fundamental contradiction he identifies is that if they truly believe this is such a bad future, why are they excited about creating it. The answer is that they believe the future is dangerous and only under their enlightened control can humanity be protected.
Zuckerberg argues that the notion that AI is so dangerous that the only safe path is extreme concentration of power is inherently problematic. Historically, hoping that absolute power will benevolently provide for humanity has not led to safe or positive outcomes. The major frame in the AI debate is not just about open versus closed, but also centralized versus decentralized.
Former Biden administration officials who moved to Anthropic's government affairs team, along with think tank people, held the view that AI was inherently dangerous and required centralized control. Their vision included limiting AI to the United States by not exporting chips or models, forming an atomic energy commission-style cartel to control the industry, and deeply embedding AI with government power. This represented an extremely centralized vision merging corporate and state power.
The preferred approach is maximum empowerment of the individual through decentralized outcomes, open models, and data sovereignty. While proprietary frontier models from Anthropic and OpenAI can exist, the goal is creating decentralized, empowering outcomes for individuals. Anthropic and effective altruists believe the technology is too dangerous to distribute, while Zuckerberg, Elon, and Jensen believe it is too dangerous to centralize.
The right to bear arms serves as an analogy for the right to have personal AI aligned with individual values rather than someone else's judgment of what is best. AI Jesus (referring to Dario Amodei) is cursed with resting smiling face while explaining the Terminator scenario and claiming only he can save humanity, even while acknowledging a 20% chance he might not succeed.
Effective altruists believe competition will create a race to the bottom and lack of safety, while the counterargument is that competition brings out the best. They want to centralize control and have less competition, while the opposing view favors decentralization and more competition.
Effective altruists view AI literally as super intelligence and the nuclear bomb, believing they will create the one money-printing machine and save everyone. This represents massive delusions of grandeur. The idea that you can keep the genie in the bottle is absurd, and the technology will diffuse anyway. Unlike nuclear weapons, every consumer and business needs AI, making it fundamentally different as a consumer technology before it becomes a military technology.
Banning open source and allowing centralized control would result in losing AI, geopolitics, and everything to China. American patriots should support the distribution, democratization, decentralization, and individualization of this technology.
Zuckerberg addresses job displacement by noting that while automation might outpace individual capability growth, there is no rule that AI must increase automation faster than it increases individual capabilities or demand for new skills. Recent statistics suggest it may be more likely that individual capability growth could match or outpace automation. The framing is that people will not lose jobs to AI, but to someone using AI.
Zuckerberg argues that all AI models are derived from human knowledge and that it is important to protect the principle that you can learn from anything you can observe. He preemptively states his desire to scan the world's data. However, there is hypocrisy in this position since Meta has sued every startup that has tried to index or use the social graph at Instagram or Facebook, including hundreds of lawsuits against people trying to distill information about top influencers or social connections.
If Dario Amodei had gotten his way with highly regulated AI apparatus including an FAA for AI, Anthropic would not be positioned for a successful IPO. Their pricing power depends on being six months ahead of open models, but regulatory approval would eliminate this advantage. The FAA takes minimum five years to certify new airframes, and the FDA takes years to approve new medicines. Applying this framework to AI would commoditize models and harm US competitiveness against Chinese competitors.
Frontier model companies like Anthropic and OpenAI are on a hamster wheel where they must stay in the lead to justify premium pricing. However, this pricing power may be sustainable because a significant subset of the market (similar to Apple versus Android) is willing to pay premium for true frontier intelligence. If the top 20% is willing to pay 10x for the best models, the companies remain profitable.
CEOs are increasingly going direct on X: Nikesh from Palo Alto Networks started tweeting and doing long form, Jensen opened an X account for long form content, and Zuckerberg activated his account to write his own missive. Zuckerberg's 6,500-word essay was 100% human-written according to AI detection. This represents a maturation where CEOs are deciding to write their own messages rather than relying on corporate PR departments.
CEOs without their own voice on X are at risk, as corporate PR departments cannot protect them indefinitely. Having your own voice provides protection, power, and armor. It is essential for public figures and CEOs to speak their own truth in authentic words.
The Manhattan Project required approximately 2,000 physicists and engineers despite having only 10 Nobel laureates at the top. Having more affordable intelligence makes frontier intelligence more valuable rather than less. A plausible outcome sees frontier tokens representing 65-85% of economic value while open source tokens represent 80% of volume.
With Anthropic at $100 billion revenue in 2026, the group debated 2027 revenue projections. Estimates ranged from $400-500 billion as realistic and achievable targets. Wall Street consensus estimates were expected to be conservative, with underwriters potentially sandbagging revenue projections. The group noted that Anthropic had grown from $10 billion to $400 billion in just 10 months, representing 40x growth.
Physical limitations around energy and compute infrastructure may constrain growth. There are concerns that frontier AI companies may need to slow down open source development to avoid commoditization of their business. However, the United States lacks jurisdiction over China, which would continue accelerating development. Global compute governance agreements with China were considered unrealistic pipe dreams.
Human brains are remarkably energy efficient compared to AI systems. To approximate human brain function, a data center would need to consume power equivalent to a million American homes for 6-12 months. Each AI query requires power equivalent to a thousand American homes. AI systems are improving inference efficiency by 5-10x every 12-18 months.
Mark Zuckerberg is positioned as the open source champion from America, potentially developing the best open source model within the next year. His competitive approach mirrors past successes at copying, mimicking, and competing effectively. The panel discussed Nvidia's position, noting that while Jensen Huang prefers not to be considered in the open source business, one hyperscaler recently switched from a $100 billion frontier model to open source alternatives.
Nvidia is partnering with Goldman Sachs, BlackRock, and other firms to raise $500 billion in AI compute financing. The initiative treats GPUs as a financeable income-producing asset class similar to mortgage-backed securities. Instead of companies paying billions upfront for chips, they can borrow money to purchase Nvidia systems and generate revenue through compute rental to repay loans.
Major asset managers including Blackstone, KKR, Goldman Sachs, and Apollo are validating GPU compute as a financeable asset. Nvidia serves as a matchmaker connecting customers needing compute with lenders. The company provides residual value guarantees after 3-4 years, guaranteeing GPUs can be rented at certain rates. Nvidia bears approximately 25% of the risk while the world's leading underwriters handle the remainder.
CoreWeave reports renting Ampere GPUs profitably through 2029, indicating 9-year lifespans for 2020 chips. Open source models can be optimized to run on older hardware, extending useful life. The panel discussed 6-8 year operational periods as realistic expectations.
The primary risk is not demand-side but rather a glut of compute leading to oversupply. Historical parallels to dark fiber after the dotcom crash suggest potential disaster if GPU demand doesn't materialize. However, political headwinds around data center construction may prevent overbuilding relative to exponentially growing demand.
In exchange for residual value guarantees, Nvidia receives revenue share above agreed floors, functioning effectively as royalties. Morgan Stanley noted Nvidia could become a capital-light cloud business. The financing initiative alleviates capital constraints that could otherwise limit TAM growth. Elon Musk's planned 6-8 gigawatt expansion next year would require $300-400 billion in capex.
The financing model resembles aircraft financing where planes have inherent value regardless of airline creditworthiness. Nvidia must standardize reference designs to enable Wall Street to package and securitize GPU investments. This creates asset-backed securities similar to mortgage-backed instruments.
Anthropic and OpenAI serve as dead man switches in the financing ecosystem. OpenAI scaled back ambitions from $1.4 trillion to $600 billion in buildout plans. If demand signals weaken or compute efficiency improves dramatically, these companies can reduce purchasing, preventing market collapse. Anthropic's IPO will provide quarterly earnings as the industry's most important demand signal.
Anthropic generates approximately $100 billion per gigawatt of compute, enabling payments of $50 billion per gigawatt to SpaceX and $30 billion to Nvidia for chips. The entire AI supply chain depends on Anthropic's quarterly results as the leading indicator. The S-1 filing will demonstrate profitability and cash generation, contradicting assumptions that tokens are subsidized.
US corporate and individual AI spending could reach $1 trillion annually, representing 5-10% of total employee salaries ($10-12 trillion). Average employee salary of $80,000 would justify $4,000-8,000 in AI efficiency investments. Even conservative estimates of $500 billion annually align with current company trajectories.
New Jersey's Attorney General sued Amazon over its Delivery Service Partners (DSP) subcontractor model, arguing the company uses subcontractors to shield itself from liability. New York City Mayor Zohran Mamdani supported the lawsuit, demanding Amazon directly employ drivers through municipal law. The DSP system insulates Amazon from union bargaining and labor obligations while creating safety concerns through subcontractor pressure.
A widespread industry practice involves companies like Amazon using networks of independent contractors as drivers through Delivery Service Partner (DSP) arrangements. Amazon has employed this model for years, arguing that restricting independent contractors would raise delivery costs, with one study showing average households could pay hundreds of dollars extra annually. The DSP structure enables rapid scaling during peak periods like holidays, supports dense urban routing, and provides lower costs compared to employing all drivers directly. This decentralized model creates more jobs and entrepreneurship, with DSPs employing thousands of people in New York City alone.
The discussion centers on whether government should insert itself into mutually beneficial arrangements between workers, DSP owners, and Amazon. One position advocates letting freedom of contract and voluntary exchange prevail, noting the market structure evolved because all parties found it beneficial. Another viewpoint suggests Amazon should proactively add a line item to bills called the "Zohran Mamdani fee" to show consumers the direct cost of proposed regulations.
Amazon's service efficiency is highlighted as unmatched compared to government-provided services, with same-day or next-day delivery representing optimized consumer pricing. The DSP drivers currently earn $18-20 per hour, though the model was designed to shift employee costs onto taxpayers rather than purely market evolution. The proposal suggests it would cost Amazon approximately 25 cents per delivery in New York to transition drivers to full-time employment with benefits at $21-22 per hour.
The discussion addresses how companies optimizing for efficiency may create conditions that fuel socialist sentiment among young people. The argument suggests Amazon should get ahead of potential regulatory changes by treating frontline workers as part of the Amazon family, similar to how Howard Schultz elevated barista wages at Starbucks to around $20-22 in major cities. This proactive approach could prevent political attacks during upcoming election cycles.
A middle-ground position suggests maintaining some DSPs for holiday scaling while converting others to direct Amazon employment. Workers in DSPs reportedly want Amazon equity, RSUs, and healthcare benefits. Many drivers work full 10-12 hour shifts rather than flexible arrangements, and some toggle between Amazon and Uber driving. Unionization efforts have faced resistance, with DSP structures potentially serving as tools to complicate organizing.
Converting to direct employment would raise package costs by approximately $5.20 per delivery, costing New York families an estimated $664 more annually while putting around 5,000 jobs at risk. The contrast between voter preferences for socialism and consumer preferences for lower prices creates tension in policy outcomes.
Grok 4.6 represents a significant pricing disruption in the AI model market, achieving frontier-level performance at lower costs. The model demonstrates strong results on Cursor benchmarks and Databricks evaluations, outperforming Claude 3.5 Sonnet in quality while maintaining competitive pricing. xAI's rapid progress stems from acquiring the Cursor team and bringing in SpaceX management talent over six months.
Elon Musk's approach involves building massive compute clusters that serve dual purposes: training frontier models while maintaining the option to sell excess capacity to other labs. The Anthropic deal reportedly includes 90-day cancellation clauses allowing recommitment of compute based on spot pricing dynamics. This creates strategic flexibility whether xAI succeeds as a frontier lab or serves as a major compute provider.
Investors often overlook Grok's capabilities when evaluating SpaceX, despite Grok's competitive performance across benchmarks. The discussion notes Anthropic's valuation trajectory from $1 billion to $50 billion as a potential parallel for Grok's growth trajectory within the broader SpaceX ecosystem.
Silver Lake's potential acquisition of Workday at a 17% premium signals renewed private equity interest in software companies. The software sector experienced significant declines from approximately $300 to $100 per share before recovering modestly. Private equity interest may reflect recognition that open source alternatives and AI efficiency gains create value extraction opportunities in oversold software assets.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
