Brad Gerstner: No AI Bubble, Semis Eat the Nasdaq & AI's Take Off Problem

All-In Podcast

AI-assisted summary of Brad Gerstner: No AI Bubble, Semis Eat the Nasdaq & AI's Take Off Problem — original source. Check the original for full context.

Source published: 2026-09-22

Detailed notes

## Background Published on YouTube by the All-In Podcast on September 17, 2026, this episode features Altimeter Capital founder Brad Gerstner. According to the description, the discussion covers initiatives like Trump Accounts and preventative CAC scans, analyzes whether downstream AI revenues can justify massive hyperscaler CapEx, evaluates power and token constraints, and outlines systemic risks ranging from regulation to rising interest rates.

## TL;DR Brad Gerstner presents a data-driven market assessment arguing that the current tech rally is driven by real earnings growth and multiple compression rather than an AI valuation bubble. He explains that sustaining the multi-trillion-dollar infrastructure buildout requires leading frontier labs to rapidly scale enterprise offtake revenues to offset hyperscaler capital expenditures. While physical power constraints and grid delays will likely limit new data center compute below consensus estimates, Gerstner asserts that automating white-collar knowledge work provides a vast total addressable market capable of driving corporate margin expansion. Ultimately, he advises investors to avoid overleveraged speculation and maintain mental agility as macro factors like interest rates and regulatory policy determine the flight path.

## Atmosphere & style The presentation carries the brisk, punchy cadence of an elite insider's market brief, balancing high-conviction tech optimism with sober balance-sheet realism. Speaking warmly as the show's trusted 'fifth bestie,' Gerstner commands the room with infectious energy, transitioning seamlessly from personal health crusades to rigorous financial modeling. The atmosphere feels intensely engaged and pragmatic, contrasting the reckless speculative frenzy of prior bubbles with sharp, disciplined warnings about real-world physical limits and interest rate gravity.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) Brad Gerstner advocates for 'Trump Accounts,' a philanthropic and infrastructure initiative aimed at giving all 70 million American children under 18 a direct ownership stake in the US economy. - [01:01](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=61s) Gerstner promotes $100 coronary artery calcium (CAC) heart scans as a high-ROI preventative health screening that could prevent up to 50,000 fatal heart attacks annually if adopted widely. - [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) The tech market's strong run is driven by fundamental earnings growth rather than multiple expansion, with multiples contracting while tech earnings grew 26%. - [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) Market gains are heavily concentrated in semiconductors and hardware infrastructure, which accounted for 70% of the Nasdaq's gains, directly mirroring hyperscaler CapEx spend. - [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) Frontier AI labs have achieved historic revenue expansion, exemplified by Anthropic scaling its annualized run rate from $2 billion to $11 billion within three months. - [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) To justify hyperscalers' projected $1.5 trillion infrastructure CapEx, downstream AI software offtake revenues must ultimately expand toward $800 billion to $1 trillion ARR. - [08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) Knowledge work represents the largest financial TAM in global history; capturing just 4% (~$1.2 trillion) is sufficient to amortize the AI infrastructure CapEx buildout. - [11:18](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=678s) AI adoption is shifting tech business models toward growing revenues (e.g., Uber targeting 20% and Snowflake targeting 30%) with flat headcount, driving corporate margin expansion from 38 bps toward 100 bps annually. - [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) Regulatory overreach poses a severe risk to US leadership in AI, drawing historical parallels to how premature nuclear plant shutdowns stalled domestic clean energy. - [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) Physical energy constraints, permitting, and labor shortages will likely cap incremental US data center buildout at roughly 25 GW next year, well below third-party estimates of 43 GW. - [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) Rising interest rates act as a major macroeconomic headwind to debt-financed data center projects, directly impacting equity valuations. - [15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) Unlike the 2023–2025 period of broad AI momentum investing, 2026 demands disciplined balance-sheet analysis, mental flexibility, and scenario planning between small, medium, and large portfolio exposures.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) Welcoming Brad Gerstner: "Every Child a Capitalist" and Heart Scans

Brad Gerstner, founder of Altimeter Capital and widely regarded by the co-hosts as their "favorite fifth bestie," receives an energetic introduction. The hosts highlight his distinct pedigree—having founded five companies before managing capital, which gives him a builder's operating mindset fundamentally different from the "classic hedge fund guy." Before diving into the financial markets, the room reflects on Gerstner's relentless political and philanthropic push for "Trump Accounts," a initiative aimed at granting every single one of the 70 million American children under 18 a direct ownership stake in the American economy. Characterizing it not as a government entitlement program but as an infrastructure platform, the hosts commend Gerstner's "absolute dogged determination," noting that "this is the largest unlock of direct philanthropy in the history of the country" and framing the premise around a core philosophical rallying cry: "the antidote to more socialism is more capitalism."

Gerstner takes the stage to enthusiastic applause, echoing the profound response from attendees over the policy proposal. He emphasizes that the initiative directly addresses a foundational battle for the country's economic future by transforming an entire generation into direct asset holders. Beyond economic ownership, Gerstner highlights another personal mission: preventative healthcare. He thanks the dozens of conference attendees who lined up outside for coronary artery calcium (CAC) heart scans provided by the Center for Heart Attack Prevention, an organization he helped launch. Gerstner pitches the $100, 15-minute scan as the single highest-return-on-investment diagnostic in modern medicine—one routinely taken by cardiologists and their families. He stresses that if CAC screening were adopted nationally as the cardiovascular equivalent of a mammogram, it would prevent 50,000 fatal heart attacks annually, an outcome he vividly compares to "ending a Ukraine war in America every year."

[02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) Market Reality Check: Earnings-Driven Growth and Semiconductor Dominance

Pivoting away from his moonshot initiatives, Gerstner frames the core presentation as a classic "tech check and state of the market"—a speed-run analysis packed with charts evaluating whether current equity market valuations remain rational. Despite persistent macro headwinds—ranging from tariff disputes and geopolitical strife to looming artificial intelligence regulation—the broader market has surged 15% year-to-date and 39% since January of the prior year. Contrasting this with the consensus trades among macro investors, Gerstner points out that gold has remained flat and Bitcoin has fallen 10%, while technology fundamentals have surged: Nvidia's top-line revenue doubled, hyperscaler capital expenditure doubled, private valuations for OpenAI and Anthropic doubled, and SpaceX appreciated 2.5x despite a hostile macro backdrop.

Crucially, Gerstner argues that this run is fundamentally misunderstood by skeptics diagnosing an asset bubble: this is not driven by multiple expansion, but by pure earnings growth alongside multiple contraction. Earnings across the tech sector expanded 26%, while valuation multiples on both the S&P 500 and the Nasdaq compressed. Nvidia trades around 14 times forward fully taxed GAAP earnings, while the Nasdaq, S&P, and PHLX Semiconductor Sector (SOX) trade beneath their historical averages. However, Gerstner notes that market breadth is intensely bifurcated. Consumer discretionary, traditional enterprise software, and financials have stagnated, meaning the entire index advance is sustained by the single largest technological supercycle in history. Semiconductor stocks alone have generated 70% of the entire Nasdaq's return, giving rise to public mega-caps delivering venture-capital-tier multiples—such as Dell soaring 5x and peers leaping 9x over an 18-month window. Furthermore, a direct correlation reveals that hyperscaler CapEx corresponds almost dollar-for-dollar to the free cash flow generated by upstream semiconductor and infrastructure providers.

```mermaid flowchart LR accTitle: Hyperscaler CapEx to Hardware Cash Flow Loop accDescr: Diagram illustrating how hyperscaler capital expenditures flow directly into semiconductor and infrastructure free cash flow.

subgraph Hyperscalers["Hyperscale Cloud & Labs"] C["Hyperscaler CapEx\n(Microsoft, Google, Amazon, Meta)"] end

subgraph Infrastructure["Hardware & Compute Infrastructure"] S["Semiconductor & Hardware\n(Nvidia, Dell, Server OEMs)"] F["Free Cash Flow Generation\n(Near 1:1 conversion of CapEx)"] end

C -->|"Capital Deployment"| S S -->|"Converted To"| F ```

[05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) The AI Offtake Revenue Math: From Claude Code to the Trillion-Dollar Question

Gerstner revisits an infamous exchange from an October podcast appearance alongside Sam Altman and Satya Nadella, where he asked Altman how OpenAI could commit to an eventual trillion dollars of infrastructure commitments while generating merely $13 billion in GAAP revenue. Altman's terse response at the time—telling Gerstner to simply "sell your shares"—highlighted the market's prevailing anxiety over return on investment. The turning point arrived with the releases of Opus 4.5 and Claude Code, followed by Anthropic's staggering financial trajectory: scaling its annualized run-rate from $2 billion in January, to $4 billion in February, to $11 billion by March. That exponential ramp acted as an "exclamation point answer" that enterprise monetization had arrived, igniting an explosive market rally across April and May before a mild summer consolidation when Anthropic updated run-rate expectations to $65 billion against inflated street expectations of $75 billion.

```mermaid flowchart TD accTitle: The CapEx Rent and Offtake Revenue Requirement accDescr: Flowchart tracking how hyperscaler infrastructure investments require escalating software offtake revenues from frontier labs to remain solvent.

subgraph CapexSpend["Infrastructure Outlay"] A["Hyperscalers (Mag 5)\nBuild Data Centers & Rent Compute"] end

subgraph RequiredOfftake["Required Offtake Revenue Ramp"] B["Exit 2026: ~$200B ARR"] C["Mid-Term: ~$450B ARR"] D["Terminal Scale: $800B - $1.0T ARR"] end

A -->|"Must Be Funded By"| B B --> C C --> D ```

Gerstner stresses that these revenue trajectories are unprecedented in the history of capitalism, climbing double-exponential curves that defy standard venture capital intuition. While historic enterprise software champions were celebrated for achieving $1 billion in ARR over five years, the top three frontier labs (Anthropic, OpenAI, and SpaceX/xAI) held an aggregate run rate of approximately $100 billion by mid-year. However, for the macroeconomic CapEx cycle to remain intact, Gerstner estimates these three players must collectively expand by another $80 billion to hit $180 billion by year-end. Because hyperscalers like Microsoft, Amazon, and Alphabet are building data centers exclusively to rent them out, downstream offtake revenues must scale from $200 billion to $450 billion, and eventually to $800 billion or $1 trillion. Without those end-user cash flows, the hyperscalers cannot financially sustain $1.5 trillion in annual infrastructure CapEx.

- **Hyperscaler CapEx vs. Semi Cash Flow**: Cloud provider capital expenditures convert on an almost 1:1 basis into hardware supplier free cash flow. - **Anthropic Run-Rate Progression**: Jumped rapidly from $2 billion (January), to $4 billion (February), to $11 billion (March), later leveling out around $65 billion. - **Collective Frontier Run-Rate**: Top three foundation model developers sit near $100 billion collectively, requiring an additional $80 billion by year-end to validate market infrastructure assumptions.

[08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) Compute Constraints and the Massive Knowledge Work TAM

To maintain this aggressive expansion, labs are accelerating their infrastructure commitments in preparation for recursive self-improvement loops. Citing Dylan Patel's research at SemiAnalysis, Gerstner walks through the staggering scale of energy and data center deployment: in 2026, the industry added roughly 19 gigawatts of capacity, with approximately 7 gigawatts reserved strictly for the two leading labs. For the following year, forecasts project an addition of 43 gigawatts—14 gigawatts dedicated to the top two players alone. Gerstner emphasizes the sheer magnitude of this buildout: the incremental power scheduled to come online in a single year matches the entire cumulative base of data center compute operational across the United States today. Furthermore, as David Sacks highlighted, current development trends indicate that by 2028 more than half of the nation's total computing capacity will be concentrated under the direct control of just two private labs.

Gerstner dismisses concerns over total addressable market (TAM) limitations, asserting that knowledge work represents the largest financial TAM in global history. Spanning digital advertising, software engineering, customer operations, and white-collar enterprise workflows across millions of organizations, capturing merely 4% of this total economic surface—roughly $1.2 trillion—is sufficient to fully amortize the projected CapEx spend. Drawing a parallel to Jensen Huang’s widely questioned forecast on the podcast two years prior that inference workloads would expand by 1,000,000,000x, Gerstner notes that autonomous agentic architectures have validated that projection, driving token production to 47 quadrillion. User growth for agentic developer tools like Codex and Claude Code has increased 40-fold over eight months, while the median enterprise budget allocation for knowledge-work AI models has expanded 17x over the prior 18 months, demonstrating massive organic demand pull across small and mid-sized businesses alike.

[11:18](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=678s) The Productivity Dividend and Margin Expansion

Brad Gerstner explains that for enterprises of every scale—including his own firm, Altimeter Capital—buying AI has shifted from an experiment to an operational necessity. Looking back at historical performance from 2015 to 2025, the Nasdaq generated roughly 10% annual earnings per share (EPS) growth, driven by 6% top-line revenue growth and approximately 38 basis points of margin expansion per year. The central economic question for technology investing now is whether AI can accelerate that margin expansion from 38 basis points to 100 basis points annually.

```mermaid flowchart LR accTitle: AI Productivity Engine accDescr: Diagram illustrating how AI tooling enables revenue growth with frozen headcount, expanding corporate margins. subgraph Input["Cost Structure"] H["Human Engineering Headcount (Largest Single Cost Input)"] end subgraph OperationalShift["Operating Leverage"] G["Revenue Growth (e.g., Uber +20%, Snowflake +30%)"] F["Flat Headcount (Slowing Net New Hires)"] end subgraph Outcome["Economic Dividend"] M["Nasdaq Margin Expansion (Target: 38 bps -> 100 bps/yr)"] end H --> F G --> M F --> M ```

Gerstner argues that the answer is "obviously yes." The primary operational cost across tech companies is engineering and human talent. Rather than executing mass layoffs, tech giants and leading cloud platforms are freezing headcount while expanding revenue aggressively. Uber has guided toward 20% revenue growth without adding headcount, while Snowflake is targeting 30% top-line growth on a flat employee base. At the consumer tier, specialized personal agents in everyone's pocket—recalling Gerstner's friendly wager with Bill Gurley about autonomously booking hotels—are emerging rapidly through apps like MultiOn and others, creating an entirely new trillion-dollar product category that will consume massive amounts of tokens.

- **Historical Baseline (2015–2025):** Nasdaq delivered ~10% annual EPS growth (~6% revenue growth + 38 bps margin expansion). - **Target AI Trajectory:** Pushing annual margin expansion from 38 bps up to 100 bps. - **Enterprise Precedents:** Uber growing 20% and Snowflake growing 30%, both holding net headcount flat.

---

[12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) The Three Major Headwinds: Regulation, Atoms, and Interest Rates

While the long-term potential of AI is immense, Gerstner outlines three critical real-world bottlenecks: regulatory overreach, physical power constraints, and rising interest rates. On the regulatory front, he rejects binary extremism, emphasizing the necessity of common-sense, pragmatic standards that reassure everyday voters and families that the technology is safe. He highlights Elon Musk's proposal for industry peer-review mechanisms as a sensible middle path. Gerstner issues a severe warning against activist-driven panic, pointing directly to the cautionary tale of nuclear energy in the United States, where the premature shutdown of 67 fission reactors crippled domestic clean energy and amounted to unilateral technological disarmament against China.

```mermaid flowchart TD accTitle: AI Macro Risks and Bottlenecks accDescr: Overview of the three primary headwinds facing the AI infrastructure buildout. subgraph Risk1["Regulatory Risk"] R1["Activist Pushback & Safety Fears"] R2["Historical Nuclear Precedent (67 Fission Reactors Shut Down)"] end subgraph Risk2["Physical Constraints"] P1["Power, Interconnection & Permitting"] P2["Sold-out Equipment & Labor Shortages"] P3["Buildout Caps Out at ~25 GW (vs. 43 GW Est.)"] end subgraph Risk3["Macro & Capital Costs"] C1["Rising Hurdle Rates on Debt"] C2["Warren Buffett: Rates are to stocks what gravity is to matter"] end R1 --> R2 P1 --> P3 P2 --> P3 C1 --> C2 ```

Moving to physical limits, Gerstner asserts that "atoms and energy are hard." The entire United States currently operates on less than 40 gigawatts of total data center compute. Dylan Patel's forecast projecting the addition of 43 gigawatts of new capacity in a single year appears overly optimistic given local permitting battles, multi-year grid interconnection delays, acute shortages of skilled electrical labor, and completely backlogged power equipment. Gerstner projects that the country will realistically stand up closer to 25 gigawatts next year, with Anthropic and OpenAI consuming roughly half of that new footprint. However, because Anthropic generates an estimated $100 billion to $110 billion run-rate on merely 1.5 gigawatts of compute, scaling up by another 4 to 5 gigawatts is more than sufficient to fulfill forward revenue expectations. Finally, macro capital costs represent a looming drag, as over 90% market odds point toward rate hikes. With massive data center projects financed through debt, escalating hurdle rates directly dampen equity valuations, underscoring Warren Buffett's maxim that "interest rates are to stocks what gravity is to matter."

- **Existing US Compute Capacity:** Under 40 GW across the entire country. - **Capacity Forecast Comparison:** Dylan Patel / SemiAnalysis projection of 43 GW vs. Gerstner’s estimate of ~25 GW. - **Compute Allocation:** Roughly half of the anticipated 25 GW buildout will be absorbed by OpenAI and Anthropic alone. - **Efficiency Metric:** Anthropic reported generating $100B–$110B in revenue on roughly 1.5 GW of compute power.

---

[15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) The Flight Path for 2026: Staying Mentally Flexible

Reviewing Altimeter’s portfolio posture, Gerstner notes that while the Nasdaq is up roughly 15%, investors face a wide fan of outcomes that dictates whether to remain positioned small, medium, or large. The primary catalyst for a full "takeoff" scenario—and potential high-profile IPOs—will be whether monthly AI lab revenues climb toward the $8 billion run-rate mark. Conversely, downside risks loom if 10-year Treasury yields advance toward 5.5%, if oil shocks reignite broader inflation, or if regulatory scrutiny delays key public market listings like Anthropic.

Gerstner contrasts the present market environment with the preceding era. Between 2023 and 2025, successful investing required getting only one core macro thesis right: recognizing AI as the greatest supercycle in tech history and pushing all your chips into the trade. By contrast, the market in 2026 has already priced in widespread awareness of AI. Success now demands disciplined balance-sheet analysis, scenario planning, and intellectual flexibility. He cautions against reckless speculation, humorously warning the audience not to "YOLO" with 4x leverage like a peer up north who blew up and surrendered his capital to Citadel. Altimeter remains at a disciplined "medium" position, ready to press bets if monthly revenues accelerate and energy prices abate, or dial back exposure if macro conditions deteriorate.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) this is the largest unlock of direct philanthropy in the history of the country

> [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) the antidote to more socialism is more capitalism. — Brad Gerstner

> [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) ending a Ukraine war in America every year. — Brad Gerstner

> [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) sell your shares — Sam Altman

> [11:18](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=678s) obviously yes. — Brad Gerstner

> [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) atoms and energy are hard. — Brad Gerstner

> [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) interest rates are to stocks what gravity is to matter. — Warren Buffett

> [15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) YOLO — Brad Gerstner

## Mentioned Brad Gerstner, Altimeter Capital, Trump Accounts, Center for Heart Attack Prevention, Nvidia, OpenAI, Anthropic, SpaceX, S&P 500, Nasdaq, PHLX Semiconductor Sector, Dell, Sam Altman, Satya Nadella, Opus 4.5, Claude Code, xAI, Microsoft, Amazon, Alphabet, Dylan Patel, SemiAnalysis, David Sacks, Jensen Huang, Codex, Uber, Snowflake, Bill Gurley, MultiOn, Elon Musk, Warren Buffett, Citadel

## Video description (00:00) Welcome Brad Gerstner! (01:01) Trump Accounts, Every Child a Capitalist & The CAC Scan (5:07) Can AI revenue pay for the CapEx? (8:53) The Build Out Issue: Gigawatts, TAM, Token Growth, and Margin Expansion (12:30) The risks: AI regulation, the nuclear precedent, power limits, and rising rates

Thanks to our partners for making this possible! IREN is a vertically integrated AI Cloud platform, delivering data centers, compute and software for AI training and inference. https://iren.com/

Oracle connects the data, applications, and infrastructure that turn AI into business outcomes—with the flexibility, choice, and control to optimize as AI evolves. http://oracle.com/ai

EY helps tech innovators scale from startup to exit to megacap. You build the future. We’ll handle the rest. http://www.ey.com

Meta believes the future is for everyone. We're focused on giving every person the tools to reach their full potential and making sure the benefits of technology are distributed to all. http://www.meta.com

Keel Infrastructure owns the power, land, and connectivity that HPC and AI run on - backed by secured energy assets and established grid interconnections across North America. https://keelinfra.com/

Airwallex - Agentic Global Business Accounts. Open local accounts in 70+ countries to accept payments, earn yield, pay globally, and manage spend. http://airwallex.com

PayPal has been revolutionizing commerce globally for more than 25 years. Creating innovative experiences that make moving money, selling, and shopping simple, personalized, and secure, PayPal empowers consumers and businesses in approximately 200 markets to join and thrive in the global economy. For more information, visit https://www.paypal.com

Google for Startups connects founders with the right people, products, and best practices to help startups build faster and go further. https://startup.google.com/

Explore ideas, industries, and technologies worth understanding with Chamath every week on Learn with Me: https://research.socialcapital.com/allin

Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg

Follow on X: https://x.com/theallinpod

Follow on Instagram: https://www.instagram.com/theallinpod

Follow on TikTok: https://www.tiktok.com/@allin

Follow on LinkedIn: https://www.linkedin.com/company/allinpod

Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg

#allin #tech #news

[Source](https://www.youtube.com/watch?v=PJrntzMA4iQ)

中文简报

Brad Gerstner：AI并无泡沫、半导体吞食纳指与AI起飞难题 Brad Gerstner提出了一份基于数据的市场评估，认为当前的科技股上涨行情是由真实的盈利增长和估值倍数压缩驱动的，而非由AI估值泡沫所致。他解释称，要维持数万亿美元的基础设施建设，领先的前沿实验室必须迅速扩大企业端营收规模，以抵消超大规模云服务商的资本支出。尽管物理电力限制和电网延迟可能会使新增数据中心算力低于市场一致预期，但Gerstner坚称，白领知识工作的自动化提供了一个庞大的潜在市场空间，足以推动企业利润率扩张。最终，他建议投资者避免过度加杠杆的投机行为，并保持思维敏捷，因为利率和监管政策等宏观因素将决定未来的行进路线。 Brad Gerstner倡导设立“特朗普账户”（Trump Accounts），这是一项慈善兼基础设施倡议，旨在让全美7000万18岁以下的儿童直接拥有美国经济的股权份额。 (0:00) Gerstner推崇100美元的冠状动脉钙化（CAC）心脏扫描，认为这是一种高回报率的预防性健康筛查，如果广泛普及，每年或可预防多达5万例致命的心脏病发作。 (1:01) 科技市场的强劲表现是由基本面盈利增长而非估值倍数扩张驱动的；在科技企业盈利增长26%的同时，估值倍数实际上出现了收缩。 (2:16) 市场涨幅高度集中在半导体和硬件基础设施领域，这两者贡献了纳斯达克70%的涨幅，直接映射出超大规模云服务商的资本支出状况。 (2:16) 前沿AI实验室实现了历史性的营收扩张，典型的例子是Anthropic在三个月内将其年化营收运行率（run rate）从20亿美元拉升至110亿美元。 (5:08)

Reading summary

Altimeter Capital founder Brad Gerstner argues that the artificial intelligence market cycle is supported by real corporate earnings and multiple compression rather than a speculative bubble, though sustaining the buildout requires immense downstream software monetization alongside managing severe energy and capital constraints. Earnings-Driven Growth and Infrastructure Cash Flows (2:16) Gerstner rejects characterizing the tech market advance as an asset bubble, pointing out that broad equity gains have been accompanied by 26% tech earnings growth and valuation multiple compression. Nvidia trades at approximately 14 times forward fully taxed GAAP earnings, while indices like the Nasdaq, S&P 500, and PHLX Semiconductor Sector trade below their historical valuation multiples. However, market performance remains sharply bifurcated. Traditional enterprise software, financials, and consumer discretionary sectors have stagnated, leaving semiconductor equities to generate roughly 70% of the entire Nasdaq's return. Gerstner observes that hyperscaler capital expenditures convert almost dollar-for-dollar into upstream free cash flow for semiconductor and hardware infrastructure providers. The Offtake Revenue Imperative and Market Capacity (5:08) To justify roughly $1.5 trillion in projected annual hyperscaler infrastructure spending, downstream software revenues must expand significantly. While the leading frontier developers—Anthropic, OpenAI, and SpaceX/xAI—reached an aggregate run rate of approximately $100 billion by mid-year, Gerstner estimates they must scale to $180 billion by year-end, with downstream software offtake ultimately needing to reach $800 billion to $1 trillion to support cloud infrastructure commitments. Gerstner asserts that this required revenue expansion is achievable because white-collar knowledge work represents the largest total addressable market in global history. He estimates that capturing approximately 4% of this surface—equivalent to $1.2 trillion across digital workflows, coding, and enterprise operations—would be sufficient to fully amortize the infrastructure capital expenditure. Margin Expansion Through Headcount Decoupling (11:18) A critical component of Gerstner’s operational thesis is that enterprise AI adoption enables companies to decouple top-line revenue growth from labor costs. Between 2015 and 2025, the Nasdaq averaged 10% annual earnings per share growth, supported by 6% revenue growth and roughly 38 basis points of annual operating margin expansion. Gerstner contends that widespread AI integration can accelerate this margin expansion to 100 basis points per year. Rather than relying on sweeping layoffs, companies are freezing headcount while expanding their core businesses. Gerstner highlights corporate precedents such as Uber guiding to 20% revenue growth and Snowflake targeting 30% growth while keeping their respective net employee bases flat. Energy Bottlenecks, Regulation, and Capital Costs (12:30) The physical and financial execution of the AI buildout faces three major bottlenecks: physical power constraints, regulatory pushback, and elevated interest rates. While SemiAnalysis projects the addition of 43 gigawatts of data center capacity in 2027, Gerstner considers this overoptimistic given local permitting delays, grid interconnection backlogs, and specialized labor shortages, forecasting a realistic ceiling closer to 25 gigawatts against an existing national base of under 40 gigawatts. Regulatory intervention presents another risk; Gerstner cites the historical shutdown of 67 US nuclear fission reactors as a warning against alarmist policies that could impair technological competitiveness. Finally, with large-scale data center projects relying on debt financing, rising capital costs and high market odds of further rate hikes impose higher hurdle rates that directly pressure asset valuations. Gerstner concludes that while 2023 to 2025 rewarded an aggressive, broad-based exposure to the AI theme, 2026 demands portfolio discipline and scenario planning. Investors face a wide fan of outcomes governed by whether frontier lab monthly revenues maintain their acceleration or macro pressures—such as 10-year Treasury yields approaching 5.5% and persistent energy deficits—stall the infrastructure flight path.

精炼阅读摘要

Altimeter Capital 创始人 Brad Gerstner 认为，当前科技股行情并非由估值倍数扩张推动的资产泡沫，而是由半导体与人工智能基础设施的大幅盈利所支撑。然而，超大规模云服务商数万亿美元的资本支出最终必须由下游软件终端收入承接，同时行业还面临物理电力供应、监管过度与利率高企等严峻制约。业绩驱动而非泡沫：半导体支撑反弹 (2:16) Gerstner 指出，尽管面临关税争端、地缘冲突以及监管预期等宏观阻力，美股市场年初至今依然上涨 15%，自前一年 1 月以来上涨 39%。然而，与黄金走平、比特币下跌 10% 形成鲜明对比的是，科技板块基本面显著爆发：Nvidia 营收翻倍，超大规模云厂商的资本开支翻倍，私营企业 OpenAI 与 Anthropic 的估值翻倍。Gerstner 强调，市场并未出现泡沫典型的估值倍数扩张，反而是科技板块整体盈利增长了 26%，而 S&P 500 和 Nasdaq 的估值倍数出现收缩；Nvidia 的远期全税 GAAP 市盈率仅约 14 倍，主要指数估值均处于历史均值下方。不过，市场广度呈现出极度分化。传统非必需消费品、传统企业软件与金融板块表现低迷，指数涨幅主要由半导体板块独力支撑，半导体类股票贡献了 Nasdaq 全部回报的 70%。与此同时，超大规模云服务商的资本开支（CapEx）与上游半导体及硬件基础设施提供商的自由现金流之间呈现出近乎 1:1 的直接转化关系，并推动 Dell 等成熟上市硬件厂商在 18 个月内实现 5 倍乃至 9 倍的巨额市值跃升。算力开支闭环与前沿实验室收入兑现 (5:08) 随着超大规模云厂商持续斥巨资采购基础设施并对外出租算力，市场最关心的核心问题在于下游软件收入能否形成闭环。Gerstner 回顾了此前向 Sam Altman 提出的质疑：在 OpenAI 仅有 130 亿美元 GAAP 营收时，如何承担高达万亿美元的基础设施承诺。随后 Anthropic 展现了快速的年化经常性收入（ARR）扩张：其年化运行率从 1 月的 20 亿美元跳升至 2 月的 40 亿美元，3 月达到 110 亿美元，随后稳定在 65 亿美元左右。 Gerstner 估算，目前排名前三的前沿实验室（Anthropic、OpenAI 与 SpaceX/xAI）合计年化运行率已达约 1000 亿美元，但若要支撑起云厂商每年高达 1.5 万亿美元的资本支出，这三家企业年底前必须再增加 800 亿美元收入，使其总规模达到 1800 亿美元。长期来看，下游终端软件收入必须从 2000 亿美元攀升至 4500 亿美元，最终达到 8000 亿至 1 万亿美元。Gerstner 认为这一增长空间并非不可企及，因为全球知识工作构成了历史上最大的潜在市场；只要人工智能渗透数字化工作流并获取其中 4% 的经济价值（约 1.2 万亿美元），便足以全额摊销当前的算力资本投入。知识工作自动化驱动企业利润率扩张 (11:18) Gerstner 强调，部署人工智能工具已从小规模试验转变为企业运营的硬性需求。回顾 2015 至 2025 年的历史轨迹，Nasdaq 实现了年均约 10% 的每股收益（EPS）复合增长，主要由 6% 的营收增长与每年约 38 个基点的利润率扩张共同驱动。当下科技投资的核心命题在于，人工智能能否推动企业年度利润率扩张幅度从 38 个基点提升至 100 个基点。在 Gerstner 看来，这一拐点正在通过“冻结员工人数、实现业绩增长”的经营杠杆成为现实。以科技巨头与云平台为例，工程与人力成本是其最大的支出项，而目前 Uber 已给出在不增加净新增员工的情况下实现 20% 营收增长的指引，Snowflake 也制定了员工规模持平而营收增长 30% 的目标。此外，随着 Codex 和 Claude Code 等智能体工具的用户量在 8 个月内增长 40 倍，企业在知识工作模型上的中位数预算在 18 个月内扩大了 17 倍，自主个人智能体的普及正在开辟万亿美元级的新产品赛道。物理电网、监管审查与资本成本三重瓶颈 (12:30) 尽管软件需求迅猛，但人工智能扩张正迎头撞上严酷的现实约束。首先是物理电网与数据中心建设的局限，Gerstner 直言“原子与能源极其艰难”。针对 SemiAnalysis 分析师 Dylan Patel 预计全美单年将新增 43 吉瓦（GW）算力容量的预测，Gerstner 持怀疑态度，指出当前全美数据中心总算力还不足 40 吉瓦。受制于地方审批纠纷、长达数年的电网互联延迟、熟练电气工人短缺以及变电设备严重缺货，Gerstner 预计来年美国实际新增容量更可能接近 25 吉瓦，其中约一半将被 OpenAI 和 Anthropic 消化。另外两大阻力则来自监管与资本成本。在监管方面，Gerstner 警告过度的激进主义恐慌可能重蹈覆辙，正如美国历史上过早关停 67 座核裂变反应堆从而削弱了自身清洁能源根基一样，可能导致对外部竞争对手的单方面技术退让；他倾向于 Elon Musk 提出的行业同行评审机制。在宏观层面，市场普遍押注加息预期，这大幅推高了依靠债务融资构建的数据中心项目的资本门槛；借用 Warren Buffett 的名言“利率之于资产如同重力之于物体”，不断上升的资金成本将直接对权益资产估值形成下行压制。 Gerstner 总结指出，相较于 2023 至 2025 年只需全力押注人工智能超级周期即可获利的宏观单边行情，2026 年的市场已经充分消化了人工智能预期，投资转而取决于对资产负债表与微观场景的精细研判。面对多空交织的复杂局面，Altimeter 当前选择维持“中等”敞口定位：若前沿实验室月度收入年化规模迈向 80 亿美元且能源通胀受控，将伺机加大配置；反之，若 10 年期美债收益率逼近 5.5% 或监管持续延缓重点企业上市，则需果断收缩风险敞口。

Detailed notes — zh

## Background 该视频由知名播客频道《All-In Podcast》于 2026 年 9 月 17 日发布在 YouTube 平台。节目简介显示，本期特邀嘉宾为 Altimeter Capital 创始人兼首席执行官布拉德·格斯特纳（Brad Gerstner），核心议题涵盖儿童投资账户与 CAC 心脏扫描两大公益计划、AI 营收能否支撑万亿级 CapEx 资本开支、算力电网建设与 TAM 增长，以及监管、电力瓶颈与加息环境带来的潜在系统性风险。 ## TL;DR 格斯特纳在演讲中坚决反驳了“AI 泡沫论”，通过硬核财务数据指出当前行情纯粹由企业爆发式盈利驱动，半导体巨头已近乎 1:1 地将云厂商 CapEx 转化为自由现金流。他指出，以 Anthropic 极速飙升的年化营收为例，全球仅需渗透 4% 的知识工作市场就能完全覆盖基础设施开支，但头部三大实验室年底前仍需背负冲刺 1800 亿美元总营收的重任。与此同时，物理层面的供电落差（实际交付或从预期的 43 GW 缩水至 25 GW）、严苛监管重蹈“核能关闭”覆辙的历史风险以及高利率压制构成了核心外部硬约束。他呼吁投资者在 2026 年告别无脑加杠杆的投机冲动，保持客观敏捷的中等仓位以应对宽幅波动的未来走势。 ## Atmosphere & style 整场演讲信息密度极高且节奏凌厉，充溢着顶级科技投资人兼连续创业者特有的敏锐与宏观掌控力。现场观众全神贯注于一页页硬核数据图表的冲击，格斯特纳巧妙地将家国情怀般的公益登月计划、直面行业巨头的内幕对话与微观财务模型串联在一起。演讲既有对 AI 开启人类史上最大超级周期的澎湃激情，又带有直面物理电网硬约束与宏观利率引力的冷峻克制，带来一种既振奋人心又极度清醒的视听体验。 ## Bulletin - [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) 本轮 AI 科技牛市本质由企业利润爆发驱动而非估值倍数扩张，纳斯达克 70% 的涨幅由半导体板块贡献，云厂商 CapEx 曲线与半导体巨头自由现金流呈现一比一的镜像转化。 - [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) Anthropic 年化营收在第一季度由 20 亿美元狂飙至 110 亿美元，有力打破需求质疑；但头部三大实验室年底总年化收入需冲刺至 1800 亿美元，以证明万亿级基础设施开支的合理性。 - [13:57](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=837s) 物理电网与工程交付构成最大硬约束：分析师预测的 2027 年全美新增 43 GW 算力电力极可能缩水至约 25 GW 实际交付，其中约半数将被 OpenAI 与 Anthropic 瓜分。 - [08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) 全球知识工作市场空间深厚，AI 仅需渗透 4%（约 1.2 万亿美元）即足以支撑全行业 CapEx；企业端正借助 AI 冻结人员增量，推动纳斯达克上市公司利润率每年扩张 100 个基点。 ## Key takeaways - [01:01](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=61s) 通过政策推动全美 7000 万未成年人设立儿童投资账户，以及推广仅需 15 分钟与 100 美元的 CAC 心脏钙化扫描每年拯救 5 万人生命，是兼具资本主义精髓与高 ROI 的公益登月倡议。 - [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) 当前市场并非 2000 年网络泡沫重演，标普与纳指市盈率估值实际上低于历史均值，企业盈利激增 26%，英伟达远期 GAAP 市盈率仅维持在约 14 倍。 - [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) 科技行情呈现极端的结构化集中特征：传统软件、非必需消费品及金融板块原地踏步，而算力基建供应链极为紧俏，半导体贡献纳指 70% 涨幅，戴尔等 IT 硬件商走出 5 至 9 倍涨幅。 - [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) Claude 3.5 Sonnet 与 Claude Code 等工具发布后，Anthropic 年化营收在三个月内由 20 亿跃升至 110 亿美元，以破资本主义历史记录的速度粉碎了市场对 AI 实际需求的质疑。 - [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) 头部大模型实验室承担着算力承销重任，当前三大巨头约 1000 亿美元的年化收入需在年底冲至 1800 亿美元，并在未来迈向数千亿乃至万亿美元，否则无法闭环云巨头的折旧与资本开支。 - [08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) 算力供给呈现双寡头垄断趋势：2026 年新增 19 GW 算力中头部两家实验室占 7 GW，预计到 2028 年全美超过一半的先进制程数据中心资产将被区区两家实验室牢牢掌控。 - [08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) 企业知识工作替代是全球规模最大的单一市场，AI 仅需渗透该市场 4%（约 1.2 万亿美元）即可完全覆盖万亿基建开支；年产 47 亿亿 Token 验证了黄仁勋关于推理需求十亿倍扩张的预言。 - [11:18](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=678s) AI 正在深刻重构科技巨头的成本结构，例如 Uber 和 Snowflake 实现业务增长 20% 至 30% 却冻结人员招聘，推动纳斯达克利润率扩张速度从历史均值 38 基点提升至 100 基点。 - [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) 激进监管与公众恐慌构成关键外部威胁，必须警惕美国曾因激进环保游说关停 67 座核反应堆导致单方面解除能源武装的惨痛教训，倡导以同行评审等务实手段平衡安全与创新。 - [13:57](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=837s) 受制于变压器短缺、电网并网漫长、熟练电工匮乏与地方审批阻力，明年全美 43 GW 的激进建设预期更可能压缩至约 25 GW，但现有电力已足够支撑头部实验室交付下一个千亿营收。 - [15:05](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=905s) 高利率环境是压制资产估值的地心引力，大规模数据中心建设多依靠高杠杆举债，若无风险利率升至 5.5% 以上，高估值科技板块将面临严峻的资本成本考量。 - [15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) 告别 2023 至 2025 年无脑下注（YOLO）的简单周期，2026 年需密切盯紧实验室月度营收斜率、宏观利率、油价及头部公司 IPO 进程，以中等仓位保持高度客观与心理弹性。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) 开场引见与两大公益登月计划：儿童投资账户与 CAC 心脏扫描大会在一片热烈的欢呼声中开场，主持团队隆重请出“All-In 播客的第五位挚友”——Altimeter Capital 创始人兼首席执行官布拉德·格斯特纳（Brad Gerstner）。主持人特别强调，格斯特纳并非传统意义上只懂看盘的对冲基金经理，他曾亲手创立过五家公司，兼具创业者的进取野心与顶级投资人的宏观敏锐度。更重要的是，他在推动公共议题上展现了惊人的执着：格斯特纳坚信全美 7000 万 18 岁以下的孩子都应当拥有专属的投资账户，正是凭借他坚韧不拔的意志，这项被前总统特朗普采纳并全力推动的政策倡议才得以成为现实。这不仅是一项救助计划，更是一个撬动美国历史最大规模直接慈善资金的平台，“消除社会主义阴霾的唯一解药，就是推行更彻底的资本主义”。登台后的格斯特纳感谢了现场听众的热情回应，并动情地阐述了这背后关乎“美国灵魂之战”的意义——让 7000 万年轻一代直接成为国家财富的所有者，是战胜资本匮乏与意识形态极端化的最根本路径。随后，格斯特纳还将话题引向了他发起的另一项登月计划：心脏猝死预防中心在会场外设立的冠状动脉钙化扫描（CAC Scan）体检点。他指出，这项仅需 15 分钟、花费 100 美元的心脏扫描，是目前医疗保健领域投资回报率最高的事情。全美顶尖的心脏病专家几乎都会为自己和至亲安排这项筛查。如果能将其普及为如同乳腺钼靶检查一样的全民预防标准，每年就能拯救至少 5 万名美国人的生命，这相当于每年在本土终结一场惨烈的现代战争。 - **儿童投资账户愿景**：旨在让全美 7000 万未成年人直接拥有股权资产，成为国家财富的直接股东。 - **CAC 心脏扫描普及**：仅需 100 美元与 15 分钟筛查，有望每年挽救全美 5 万人免于突发心脏病死亡。 --- ### [02:16](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=136s) 市场现状与估值解构：由真实利润驱动的超级周期，绝非 2000 年泡沫格斯特纳迅速将焦点转回他在播客中最擅长的核心板块——宏观市场与前沿科技趋势评估。他提醒现场观众备好相机与笔记本，因为接下来将是一连串硬核的数据与图表冲击。尽管地缘政治冲突、关税隐忧以及对人工智能监管的担忧持续笼罩市场，标普与纳指今年仍分别录得 15% 的涨幅，较去年 1 月更是大涨 39%。反观曾经被许多人寄予厚望的避险资产，黄金走势平平，比特币甚至录得约 10% 的跌幅。驱动本轮行情的真正火车头清晰无比：英伟达营收实现翻倍（2x）、超大规模云厂商（Hyperscalers）资本支出翻倍（2x）、OpenAI 与 Anthropic 估值翻倍（2x），连身处恶劣外部环境下的 SpaceX 估值也暴涨 2.5 倍。格斯特纳特别强调，这轮行情的本质绝非估值倍数扩张（Multiple Expansion），而是纯粹由盈利爆发驱动的牛市。事实上，全市场的市盈率估值反而出现了收缩；虽然在 AI 基础设施强劲拉动下企业整体盈利激增了 26%，但纳斯达克与标普 500 的实际交易估值却低于历史平均水平。以英伟达为例，其目前的股价仅对应明年度完全纳税后 GAAP 净利润的 14 倍左右，与 2000 年千禧年科技泡沫时期的非理性繁荣有着天壤之别。然而，这并非普涨行情：非必需消费品、传统软件和金融等大板块几乎原地踏步，整个科技行业的繁荣高度集中——单单半导体板块就贡献了纳斯达克指数高达 70% 的涨幅。格斯特纳分享了一张极具视觉冲击力的图表：超大规模云厂商的资本支出曲线，与半导体巨头的自由现金流（FCF）曲线几乎实现了“一比一”的惊人重叠。云巨头花出去的每一块钱算力资本开支，几乎分毫不差地直接转化为了英伟达等基础设施公司的纯现金流利润。在算力基础设施供应链极度紧缺的红利下，甚至连戴尔（Dell）这样庞大的传统 IT 巨头，都能在短短 18 个月内走出 5 倍乃至 9 倍这种原本只属于早期风险投资的指数级回报。 ```mermaid flowchart LR accTitle: AI 资本支出与利润流向闭环 accDescr: 展示超大规模云厂商资本支出如何转化为半导体自由现金流并依赖终端模型商业化变现的闭环流程 subgraph CapitalExpenditure ["资本投入端"] A["超大规模云厂商 (Microsoft / Google / AWS)"] end subgraph Infrastructure ["基础设施提供商"] B["半导体与硬件巨头 (Nvidia / Dell / TSMC)"] end subgraph Labs ["前沿大模型实验室"] C["AI 头部实验室 (OpenAI / Anthropic / xAI)"] end subgraph Monetization ["终极变现端"] D["企业级知识工作 TAM (代码生成 / 白领工作流 / 广告)"] end A -- "每年数千亿美元 CapEx" --> B B -- "转化为等额自由现金流 (FCF)" --> B A -- "租赁高性能 GPU 算力集群" --> C C -- "Token 生成与 API 订阅收费" --> D D -. "1.2 万亿美元终极买单支撑".-> A ``` - **盈利与估值错位**：全市场盈利增长 26%，但指数估值倍数反而收缩，英伟达远期 GAAP 市盈率仅约 14 倍。 - **半导体板块一骑绝尘**：贡献了纳斯达克 70% 的涨幅，戴尔等基建股在 18 个月内实现 5 到 9 倍涨幅。 - **资本流向镜像效应**：云巨头 CapEx 增长曲线与半导体厂商自由现金流产生完全重合的镜像对应关系。 --- ### [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) 收入能否撑起万亿资本开支？Anthropic 爆发与头部实验室的业绩追赶格斯特纳随后抛出了悬在所有人头顶的核心疑问：AI 的实际收入到底能否撑起数万亿美元的资本支出？他回忆起去年 10 月与萨姆·奥尔特曼（Sam Altman）及萨提亚·纳德拉（Satya Nadella）录制播客时的尖锐交锋。当时他直截了当地问奥尔特曼：“当你的年 GAAP 营收只有区区 130 亿美元时，你怎么敢承诺投入上万亿美元的资本开支？”本以为这能为市场解开迷思，结果奥尔特曼却半开玩笑地怼了他一句：“那你大可卖掉你的股票。” 然而，随后的现实却给出了惊心动魄的答案。12 月初，Claude 3.5 Sonnet 与 Claude Code 等划时代模型与编码工具横空出世；进入新的一年后，Anthropic 的年化运行营收（Run Rate）呈现出令人瞠目结舌的几何级狂飙：从 1 月的 20 亿美元、2 月的 40 亿美元，一跃攀升至 3 月的 110 亿美元！当华尔街和科技圈看清 Anthropic 令人窒息的营收增速时，所有关于“AI 商业化是否有真实需求”的疑虑被瞬间砸得粉碎。正是这一剂强心针点燃了 4 月和 5 月美股波澜壮阔的暴涨行情。虽然此后由于 Anthropic 披露的年化预期从传闻的 750 亿美元微调为 650 亿美元、加之开源模型步步紧逼的压力，市场在六七月间陷入了横盘休整，但这种超指数级的增长速度已然打破了资本主义历史的一切既有记录。格斯特纳指出，头部三大实验室（OpenAI、Anthropic、xAI/SpaceX）目前的总年化营收规模大约在 1000 亿美元左右，但要维持住整个 AI 宏观交易逻辑不坍塌，这三大巨头必须在年底前集体冲刺到至少 1800 亿美元，净增整整 800 亿美元。在传统企业软件软件领域，一家 SaaS 公司用四到五年做到 10 亿美元收入就足以跻身全球前 5% 的殿堂，而眼下市场给 AI 实验室定价的预期却在以百亿乃至千亿美元为步长跳跃。如果微软、谷歌、亚马逊等云巨头每年要在 CapEx 上砸下数千亿乃至 1.5 万亿美元，那么作为算力租户的前沿实验室就必须在未来几年实现从 2000 亿到 4500 亿、甚至 8000 亿至 1 万亿美元的真实承销收入（Offtake Revenue），否则这台空前庞大的算力印钞机就将面临停摆风险。 - **Anthropic 营收三级跳**：从 1 月的 20 亿美元、2 月的 40 亿美元，极速蹿升至 3 月的 110 亿美元年化营收。 - **年底 1800 亿美元军令状**：三大前沿实验室当前年化总收入约 1000 亿美元，需在年内再增 800 亿美元以匹配估值支撑。 - **租户承受力假说**：超大规模云厂商仅是重资产租赁中介，若终端模型开发者无法兑现万亿美元级承销收入，高昂的 CapEx 终将失去经济学根基。 --- ### [08:53](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=533s) 算力扩张的物理鸿沟与万亿级知识工作 TAM 为了满足模型自迭代演进（Recursive Cycles）的极度饥渴与爆发式终端需求，前沿实验室正不惜代价地圈占算力。格斯特纳引用了知名研究机构 SemiAnalysis 分析师迪伦·帕特尔（Dylan Patel）的数据：2026 年全美共计新增约 19 吉瓦（GW）的数据中心计算容量，其中约 7 吉瓦被两家顶尖实验室垄断。而根据预测，明年全美计划新增的计算电力容量将直接飙升至 43 吉瓦，其中头部两家实验室就将瓜分 14 吉瓦。这意味着仅明年一年新增的算力规模，就已经相当于全美国截至今年底累计部署的算力总和。到 2028 年，正如戴维·萨克斯（David Sacks）所警告的那样，全美超过一半的顶级算力资产将完全被区区两家实验室牢牢掌控。面对如此庞大的开支吞吐，终端市场的潜在空间（TAM）究竟是否足够大？格斯特纳给出了无比坚定的回答：知识工作（Knowledge Work）是人类历史上规模最庞大的单一市场，它不仅涵盖代码编写、企业白领工作流替代，还囊括了海量消费级交互与广告生态。全球知识工作市场的总体规模深不见底，AI 生态系统实际上只需要渗透这一总盘子的约 4%（约 1.2 万亿美元），就完全足以消化并支付全部基础设施的折旧与 CapEx。格斯特纳特别提到了黄仁勋两年前在播客中的预言——推理计算（Inference）的需求终将迎来十亿倍（1 Billion-x）的膨胀。当时无数分析师讥讽老黄信口开河，但如今进入智能体（AI Agents）时代后，这一预言正在全方位兑现。今年全球预计将产生多达 4.7 亿亿（47 Quadrillion）个 Token，Codex 等代码生成工具的用户体量在过去短短 8 个月内激增 40 倍，中大型企业在企业级知识工作 AI 上的中位数支出在过去 18 个月内飙升了 17 倍。需求端从未衰减，真正的瓶颈正在不可逆地转移至物理世界——即电网与供电能力究竟能否支撑明年新增 43 吉瓦的算力奇迹。 - **电力与算力供给激增**：2026 年新增 19 吉瓦算力，2027 年规划新增 43 吉瓦（相当于全美现有算力存量总和）。 - **双寡头集中度加剧**：两家顶尖大模型实验室正在包揽全美近半数算力增量，预计 2028 年控制全国过半先进制程算力。 - **十亿倍推理与 4% 渗透定律**：年产 47 亿亿（Quadrillion）Token 验证了黄仁勋的推理膨胀论；全球仅需 4% 知识工作预算（1.2 万亿美元）即可完全覆盖万亿 AI 资本支出。 ### [11:18](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=678s) 生产力红利与利润率扩张：AI 对企业成本结构的重塑 Brad Gerstner 指出，像 Altimeter Capital 这样的大型企业如今已经到了“不买 AI 服务就无法运转业务”的阶段，因此 AI 支出的飙升并不令人意外。从宏观经济的生产力红利来看，纳斯达克在 2015 年至 2025 年间的每股收益（EPS）复合年均增长率约为 10%，其中 6% 来自营收增长，另外约 38 个基点来自每年的利润率扩张（Margin Expansion）。核心问题在于：借助 AI，企业能否将每年 38 个基点的利润率扩张推高到 100 个基点？答案显然是肯定的。 Gerstner 举例说明，几乎所有他接触过的头部科技公司都在践行这一逻辑：Uber 表示即使业务增长 20%，员工人数也将保持不变；Snowflake 同样表示将在业务增长 30% 的情况下冻结人员扩张。对于这些公司而言，最大的成本构成始终是人力和工程师团队。这并不意味着公司会大规模裁员，而是未来的招聘速度将远低于以往。此外，随身携带的消费者端 AI 代理（Consumer Agents）也将爆发，例如通过 Muse 或 Instinct 实现自动预订酒店等复杂任务。这不仅将催生出价值数万亿美元的全新市场门类，更将持续消耗天量 Token。 - **历史基准数据（2015–2025）**：纳斯达克年均 EPS 增长约 10%（6% 营收增长 + 38 个基点利润率扩张）。 - **AI 驱动的目标**：借助软件自动化推动每年 100 个基点的利润率扩张。 - **头部公司策略**：Uber（20% 增速零增员）、Snowflake（30% 增速零增员），最大化控制人力成本。 --- ### [12:30](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=750s) AI 面临的三大挑战：监管博弈与“核能历史悲剧”的前车之鉴在技术快速推进的同时，Gerstner 列出了未来面临的三大核心风险：监管（Regulation）、电力（Power）以及利率走势（Interest Rates）。在监管层面，当前正处于激烈的拉锯战中。面对外界对其“立场两边倒”的质疑，Gerstner 强调极端的放任或一刀切的封禁都不可行，必须拿出务实且符合常识的折中方案，让普通民众、政客及其选民感到踏实与安全，例如 Elon Musk 前一天提出的“同行评审”（Peer Review）机制就是一种极佳的思路。 ```mermaid flowchart TD accTitle: AI面临的三大核心外部风险 accDescr: 结构化呈现AI行业在扩展期所面临的三大宏观瓶颈与风险因素 subgraph Regulation ["监管阻力"] R1["激进主义与公众恐慌"] --> R2["过度规制/关停风险"] R2 -.->|"借鉴"| R3["美国关停67座核裂变反应堆历史教训"] end subgraph Energy ["物理能源瓶颈"] E1["43 GW 激进预测"] --> E2["许可/电网并网/劳动力短缺"] E2 --> E3["实际落地预计压缩至 ~25 GW"] end subgraph Rates ["宏观流动性约束"] M1["加息预期与资本成本走高"] --> M2["数据中心举债建造成本上升"] M2 --> M3["无风险利率竞争压制美股估值"] end ``` Gerstner 发出严厉警告，要求警惕过度监管的历史重演。过去因为公众恐慌与激进环保主义者的施压，美国曾先后关闭了 67 座核裂变反应堆，在清洁能源领域对竞争对手进行了“单方面解除武装”。这不仅导致美国错失大量清洁能源，反而加剧了污染排放，成为国家层面的战略灾难。绝不能让这种由于过度恐慌而摧毁支柱产业的悲剧在 AI 领域重新上演。 - **三大核心风险**：监管阻力、电力/物理约束、宏观利率上升。 - **核能先例的惨痛代价**：因激进主义游说关闭 67 座核反应堆，造成巨大战略与能源损失。 - **监管平衡之道**：引入类似同行评审的务实制度，在保障公众安全感的同时避免扼杀技术演进。 --- ### [13:57](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=837s) 物理世界的硬约束：从 43 GW 降至 25 GW 的电力现实 AI 从比特走向原子（Atoms），面临着极其严峻的工程与电力瓶颈。针对分析师 Dylan Patel 等人预测明年将建成 43 吉瓦（GW）算力中心的观点，Gerstner 提出了强烈质疑。目前全美现有的总算力规模尚不足 40 GW，试图在短短一年内新增 43 GW 几乎是不可能的任务：地方审批遭遇巨大阻力、电网并网排期漫长、熟练技术工人严重短缺、变压器与电力设备全线脱销，这是美国历史上规模最大的一次基础设施建设。 Gerstner 预判，明年全美实际能建成的算力规模更可能在 25 GW 左右，而其中约一半将归属于 Anthropic 和 OpenAI。即便如此，这并不意味着 AI 行业无法达成营收预期。以 Anthropic 为例，传闻其在仅拥有约 1.5 GW 算力的情况下就做到了上千亿美元的潜在年化营收规模；若再获得 4 到 5 GW 的算力供给，完全足以支持其实现下一个千亿美元的营收增量。因此，算力落地不及预期并不会直接击垮 AI 的商业化变现目标。 - **全美算力现状**：现有总算力存量低于 40 GW。 - **落差预判**：市场预期的 43 GW 明年很难达成，更实际的新增交付量约为 25 GW。 - **关键实体配额**：25 GW 中约 50% 的容量将被 OpenAI 和 Anthropic 瓜分。 - **供应链现实困境**：受制于土地许可、电网接入延误、熟练电工短缺与电力设备缺货。 --- ### [15:05](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=905s) 利率环境与资产定价：巴菲特的“引力定律” 除了物理限制，宏观金融环境也构成了严峻考验。当前市场预期再度升温，隔天加息的可能性已被定价至 90% 以上。这一变化对 AI 基础设施的直接打击在于融资成本（Hurdle Rate）的激增，因为眼下建设这些巨型数据中心所消耗的绝大多数都是杠杆债务资金。 Gerstner 引用沃伦·巴菲特（Warren Buffett）的名言警示众人：“利率之于资产价格，正如引力之于物质”（"interest rates are to stocks what gravity is to matter"）。当无风险资金回报率能够达到 5.5% 甚至 6% 时，投资者承担股票权益风险的意愿就会大幅下降，整个股市尤其是高估值科技板块的估值乘数都将面临不可忽视的地心引力压制。 --- ### [15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) 2026 投资飞行路线：告别无脑 YOLO，保持极度客观与敏捷针对当前的资产组合管理，Gerstner 给出了基于概率分布的“飞行路线图”（Flight Path）。尽管纳斯达克年内已上涨约 15%，但未来的收益分布极为宽阔，取决于三大核心变量：各家 AI 顶尖实验室的月度营收数据、宏观利率与油价走势，以及以 Anthropic IPO 推进情况为代表的监管与资本市场窗口。 | 评估维度 | 乐观情景（Liftoff / 腾飞） | 中性/悲观情景（Downside / 承压） | |:--- |:--- |:--- | | **实验室营收** | 月度收入稳定向 80 亿美元靠近，营收曲线维持陡峭斜率 | 月度营收增长放缓，无法验证资本开支闭环 | | **资本市场事件** | 头部大模型公司（如 Anthropic）年内顺利完成 IPO 并在二级市场大获成功 | 监管审查推迟上市步伐，引发市场避险抛售 | | **宏观环境** | 原油价格回落，十年期美债利率企稳并走低 | 油价推高通胀预期，十年期美债收益率冲向 5.5% | | **仓位对应策略** | 确认营收放量信号后，进一步加大高胜率资产配置 | 削减多头敞口，转入防御姿态，退居轻仓观察 | Gerstner 最后总结，2023 到 2025 年的投资逻辑极其纯粹：只要认准“AI 是科技史上最大超级周期”并把筹码全部推上牌桌就能赚钱。然而到了 2026 年，所有人都已知晓 AI 故事，预期已经完全反映在股价中。眼下必须抛弃单纯的投机冲动（"Don't YOLO"），绝不要像那位在北方“把所有钱都送给 Citadel”的投资者那样去加 4 倍杠杆。Altimeter 目前维持中等仓位（Medium Position），并保持充分的心理弹性：若后续几个月营收数据超预期且油价回落，就继续下注；若出现恶化，则果断把仓位降得更低。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=0s) 消除社会主义阴霾的唯一解药，就是推行更彻底的资本主义 — Brad Gerstner > [05:08](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=308s) 那你大可卖掉你的股票 — Sam Altman > [15:05](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=905s) interest rates are to stocks what gravity is to matter — Warren Buffett > [15:43](https://www.youtube.com/watch?v=PJrntzMA4iQ&t=943s) Don't YOLO — Brad Gerstner ## Mentioned Brad Gerstner, Altimeter Capital, Donald Trump, Sam Altman, Satya Nadella, Dylan Patel, SemiAnalysis, David Sacks, Jensen Huang, Elon Musk, Warren Buffett, Citadel, OpenAI, Anthropic, xAI, SpaceX, Nvidia, Dell, TSMC, Microsoft, Google, AWS, Uber, Snowflake, Claude 3.5 Sonnet, Claude Code, Codex, Muse, Instinct, CAC Scan, S&P 500, Nasdaq ## Video description (00:00) Welcome Brad Gerstner! (01:01) Trump Accounts, Every Child a Capitalist & The CAC Scan (5:07) Can AI revenue pay for the CapEx? (8:53) The Build Out Issue: Gigawatts, TAM, Token Growth, and Margin Expansion (12:30) The risks: AI regulation, the nuclear precedent, power limits, and rising rates Thanks to our partners for making this possible! IREN is a vertically integrated AI Cloud platform, delivering data centers, compute and software for AI training and inference. https://iren.com/ Oracle connects the data, applications, and infrastructure that turn AI into business outcomes—with the flexibility, choice, and control to optimize as AI evolves. http://oracle.com/ai EY helps tech innovators scale from startup to exit to megacap. You build the future. We’ll handle the rest. http://www.ey.com Meta believes the future is for everyone. We're focused on giving every person the tools to reach their full potential and making sure the benefits of technology are distributed to all. http://www.meta.com Keel Infrastructure owns the power, land, and connectivity that HPC and AI run on - backed by secured energy assets and established grid interconnections across North America. https://keelinfra.com/ Airwallex - Agentic Global Business Accounts. Open local accounts in 70+ countries to accept payments, earn yield, pay globally, and manage spend. http://airwallex.com PayPal has been revolutionizing commerce globally for more than 25 years. Creating innovative experiences that make moving money, selling, and shopping simple, personalized, and secure, PayPal empowers consumers and businesses in approximately 200 markets to join and thrive in the global economy. For more information, visit https://www.paypal.com Google for Startups connects founders with the right people, products, and best practices to help startups build faster and go further. https://startup.google.com/ Explore ideas, industries, and technologies worth understanding with Chamath every week on Learn with Me: https://research.socialcapital.com/allin Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@allin Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=PJrntzMA4iQ)

Source quotes

[0:00] Brad Gerstner: 消除社会主义阴霾的唯一解药，就是推行更彻底的资本主义 [5:08] Sam Altman: 那你大可卖掉你的股票 [15:05] Warren Buffett: interest rates are to stocks what gravity is to matter [15:43] Brad Gerstner: Don't YOLO

Source quotes

[0:00] this is the largest unlock of direct philanthropy in the history of the country [0:00] Brad Gerstner: the antidote to more socialism is more capitalism. [0:00] Brad Gerstner: ending a Ukraine war in America every year. [5:08] Sam Altman: sell your shares [11:18] Brad Gerstner: obviously yes. [12:30] Brad Gerstner: atoms and energy are hard. [12:30] Warren Buffett: interest rates are to stocks what gravity is to matter. [15:43] Brad Gerstner: YOLO

Original transcript

[0:00] Our favorite fifth bestie in the world, [0:03] the one, the only Brad Bersner from [0:06] Altimter. Brad has had unbelievable [0:08] career starting five companies. So, he's [0:11] got a very different mentality than your [0:13] sort of classic hedge fund guy. [0:15] >> He's an amazingly successful guy and [0:17] he's put up a tremendous amount of [0:19] money. [0:19] >> Every child in America, all 70 million [0:23] kids under the age of 18 deserve to have [0:25] one of these accounts. This would not be [0:27] a law if Brad Gerster did not pursue it [0:31] with absolute dogged determination. [0:34] >> This is not a program. This is a [0:36] platform. It is the largest unlock of [0:39] direct philanthropy in the history of [0:40] the country. [0:41] >> I think the antidote to more socialism [0:43] is more capitalism. [0:46] Please welcome Brad Gersner. [0:54] Woo. [0:56] Wow. [0:58] Let's go. [1:02] Let's go. [1:04] Thank you guys. And thank you for so [1:06] much love yesterday, especially on the [1:08] Trump accounts. So many people came up. [1:12] Everybody gets what this means for [1:14] America. We're in a battle for the soul [1:16] of America. 70 million kids are going to [1:19] be made direct owners in America. That [1:22] is how we beat the scourge of capital or [1:25] socialism. We make every child a [1:27] capitalist. [1:32] >> And also [1:34] thank you to all the people yesterday [1:36] who came up who took the CAC scan, the [1:39] heart scan outside from the center for [1:41] heart attack prevention which we [1:43] started. There's no doubt based on these [1:45] results we uh you know we're going to [1:48] save some lives even yesterday. Okay. [1:51] This is the highest ROI thing you can do [1:54] in healthcare. Every cardiologist I talk [1:57] to does this for themselves, their [1:59] family, and their friends. It's a [2:00] hundred bucks, 15 minutes. Get it done. [2:03] If we turn this into the mammogram for [2:05] the heart, we'll save 50,000 lives a [2:07] year in this country. It's like ending a [2:09] Ukraine war in America every year. So, [2:11] go get your CAC scan done if you have [2:13] it. But today is not about those two [2:14] moonshots. [2:16] Today is kind of a throwback to what I [2:18] used to do on the pod with these guys, [2:20] which is a market check, a tech check, [2:22] state of the market. Where are we? Where [2:24] are we going? What do we have to believe [2:26] to be true in order for the market to [2:28] continue to work? We're going to do a [2:30] bit of a speed round here. So, bear with [2:32] me. Get your cameras out. Get your notes [2:35] out. Some good some good chart candy in [2:37] here for you guys. So markets up 15% [2:41] this year, up 39% since January of last [2:43] year, despite all the concerns about [2:45] tariffs, despite all the concerns about [2:47] geopolitics, [2:49] despite all the concerns about AI [2:51] regulation. [2:52] The scoreboard, we have a lot of people [2:55] in the bestie group who said gold was [2:57] going to be off the charts this year. [2:58] It's flat. Bitcoin's down 10%. But look, [3:01] we have Nvidia revenue up 2x. [3:03] Hyperscaler capex up 2x. Open AI and [3:06] Enthropics valuation up 2x. SpaceX up [3:09] two and a halfx in a pretty nasty [3:11] backdrop. [3:13] This is not about multiple expansion. [3:16] This is an earnings driven market [3:19] expansion. We've seen multiple [3:21] contraction this year. Earnings are up [3:23] 26% of course driven a lot by AI [3:26] infrastructure, but the multiple on the [3:28] NASDAQ and the S&P is actually down. [3:31] Look at Nvidia trading at 14 times next [3:34] year's fully taxed gap earnings. This is [3:37] no bubble like it was in 2000. [3:40] NASDAQ, S&P, Socks, Nvidia, all trading [3:44] well below their average multiples. [3:47] Okay, MAG 7 basically in line with its [3:51] average multiple, but not everybody is [3:53] winning. Okay, at the bottom here, [3:55] consumer discretionary, software, [3:57] financials, huge sections of the market [3:59] have barely moved, right? This is a [4:02] market that is being driven by the [4:04] largest capex buildout, the largest [4:06] super cycle in the history of [4:08] technology. [4:11] Semiconductors are 70% of the NASDAQ's [4:14] return. [4:16] 70% of the return. That's both good and [4:18] bad, and we'll get into that. [4:22] So, who's making the money? The makers [4:23] of the tokens are making the money and [4:26] the buyers of the tokens are basically [4:28] going along for the ride [4:33] because of the tightness in the [4:35] infrastructure market. We have massive [4:38] public companies that look like venture [4:40] capital returns. [4:42] Dell up 5x up 9x in just 18 months. [4:49] I love this chart. In blue you have the [4:52] hyperscaler capex. [4:54] In orange you have the free cash flow of [4:57] the semiconductor companies. Do you [4:59] notice anything? [5:02] Their capex is almost dollar fordoll [5:04] free cash flow to the infrastructure [5:06] companies. [5:08] Okay. Where were we at the start of the [5:10] year? This is a really important [5:14] framework to get your head around. You [5:16] may remember a certain podcast I did [5:19] with Sam Alman and Satcha in October of [5:21] last year. And I asked Sam a very basic [5:24] question that was on everybody's mind. [5:26] How can you commit to a trillion dollars [5:28] in capex when you have $13 billion of [5:31] gap revenue? [5:33] I thought this would be a way to clear [5:35] up some confusion in the market. [5:37] Instead, he told me to sell my shares. [5:42] But then in the beginning of December we [5:44] have opus 4.5 and claude code and then [5:48] in January anthropics revenue was two [5:51] billion and then in February it's 4 [5:53] billion and in March it was 11 billion. [5:56] Do you notice anything that happened [5:57] there in March? People started figuring [5:59] out what anthropics revenue was. We had [6:02] a an exclamation point answer to the [6:05] question will the AI revenue showed up? [6:07] It showed up in a massive way and that's [6:09] why we got this historic run in April [6:11] and May. We ripped off the bottom [6:14] because the fuse was lit by Anthropic's [6:16] monthly revenue. And then in in June and [6:20] July, we had some consolidation. Why? [6:22] Because Anthropic came out and said, [6:24] "Oh, our annual run rate revenue is 65." [6:26] People thought it was 75. So they had to [6:29] revise down their estimates a little [6:30] bit. And we had some concerns about open [6:33] source. Is open source catching up? you [6:35] know, is Anthropic going to continue to [6:36] be able to generate those revenues? And [6:39] so, we've kind of moved sideways since [6:40] then. [6:44] It's hard to get your head around the [6:46] fact that these revenues have never [6:49] happened before in the history of [6:51] capitalism, [6:53] right? We are on parabolic double [6:55] exponential curves around these [6:57] revenues. So the collective run rate, [7:00] let's just call it of the top three [7:01] labs, Anthropic, Open AI, and SpaceX is [7:05] about a hundred billion based upon all [7:08] the rumors that are out there coming out [7:10] of July. [7:12] I think they need to collectively get to [7:15] at least $180 billion by the end of the [7:17] year, add another $80 billion across [7:20] those three labs just to keep the AI [7:23] trade intact. [7:27] So this is the single most important [7:29] data point in the market today, right? [7:32] Is Anthropics monthly revenue is [7:34] OpenAI's monthly revenue going to be 4 [7:36] billion or 8 billion? It's almost hard [7:38] to get your head around. Most of us [7:40] who've been in venture capital for a [7:42] long time, if you added a billion, if [7:44] you got to a billion dollars in software [7:46] revenue over like four or five years, [7:49] you were in the top 5% of software [7:51] companies. But this is what the world is [7:54] now pricing in. So why is this so [7:56] important? It's this slide. If you're [7:59] going to build a a trillion half dollars [8:02] a year in capex, somebody has to pay for [8:05] it, right? Microsoft's not paying for [8:08] it. They're building it to rent it. [8:10] Google's not paying for it. They're [8:12] building it to rent it. Amazon's [8:14] building it to rent it. Well, who is the [8:16] person renting it? We have to have the [8:18] offtake revenues in order to pay that [8:20] rent. So if we exit this year around, [8:23] you know, let's call it 200 billion of [8:25] run rate revenue. That's below the the [8:27] years there. I think you have to go from [8:29] 200 to 450 to 800 or a trillion dollars [8:34] just to keep up. The blue bar is the [8:37] expected capex just from the MAG 5 and [8:40] the orange bar is the offtake revenue, [8:42] the gap offtake revenue that we need to [8:45] see in order to keep this trade intact [8:48] over the course of the next few years. [8:49] Otherwise, we can't build this much [8:51] capex. [8:53] So, the labs are increasingly, [8:56] you know, like the conversation I had [8:58] with Sam last October, they're [9:00] aggressively expanding compute. Why? All [9:02] the things you heard yesterday, we're [9:04] heading into recursive cycles. They're [9:06] seeing, you know, incredible demand for [9:08] the product. And so, they're building [9:10] out compute. This year the total amount [9:12] of uh compute added so 2026 about 19 [9:16] gawatt and about seven of those [9:18] gigawatts went to the two leading labs. [9:21] Next year this is semi analysis so Dylan [9:23] Patel forecast that the work that we're [9:26] going to add 43 gawatt and that about 14 [9:29] gawatt are going to go to the leading [9:32] labs and we'll come back to this [9:33] question of can we really stand up 43 [9:37] gawatts of compute next year. Notice [9:39] that the amount that we're adding next [9:41] year is as much as the total compute we [9:43] have in the United States this year, the [9:46] cumulative compute we have in the United [9:49] States this year. So that's what the [9:51] market is anticipating. [9:53] What happens if that doesn't happen? And [9:56] by the way, if you added those bars up, [9:58] by 2028, to David Sax's point yesterday, [10:01] over half of the total compute in the [10:04] country is controlled by two labs. [10:07] So it does the TAM exist? Does the TAM [10:11] exist? Right? So if we look at the total [10:13] TAM of knowledge work, right? That this [10:16] is a massive massive category. You got [10:18] consumer plus ads plus coding and all [10:21] these white collar workflows, millions [10:24] of enterprises. I would argue it's the [10:25] largest TAM in the history of the world. [10:28] You only have to get to about 4% of that [10:30] TAM or 1.2 trillion in order to pay for [10:33] the capex. So I don't think it's a TAM [10:35] issue. Um, obviously Jensen was on our [10:39] was on the pod and he he he talked two [10:41] years ago that inference was going to 1 [10:43] billionx and remember all the people [10:45] saying he's full of there's no way this [10:47] can 1 billionx. That's exactly what [10:49] we've done in the a age of agents, [10:52] right? We've had this exponential token [10:54] growth this year. 47 quadrillion tokens [10:58] that are going to be produced. So it's [11:00] not a question again of demand. Codeex [11:02] users have grown 40x in the last eight [11:05] months and knowledge work at enterprises [11:08] if you look at the median amount that [11:10] enterprises are spending about 17x over [11:13] the course of the last 18 months and [11:14] I've talked to many people in the [11:16] audience here small businesses medium [11:18] businesses large businesses businesses [11:20] like alimter we can't operate our [11:22] business without buying AI so this is [11:26] not overly surprising to me and when we [11:29] think about the productivity dividend to [11:30] the economy me. You know, if you look at [11:33] 2015 to 2025, EPS growth was about 10%. [11:38] That was 6% revenue plus about 38 bips [11:41] of margin expansion every year out of [11:42] the NASDAQ. So, here's the question. Can [11:45] we turn the 38 bips to 100 bips of [11:47] margin expansion because of AI? The [11:50] answer is obviously yes. Every company I [11:52] talk to, Uber says we're going to grow [11:54] 20%, we're not going to grow headcount. [11:56] Snowflake says we're going to grow 30%, [11:58] we're not going to grow headcount. [11:59] That's what's happening. That is margin [12:01] expansion. The single largest cost input [12:03] to every one of these companies are [12:05] humans and engineers. It's not that [12:07] they're going to fire everybody. They're [12:08] just not going to hire them at the rate [12:09] that they hired him before. And then of [12:11] course, we're going to have consumer [12:12] agents in everybody's pocket. You may [12:14] remember this bet I had with Bill. You [12:16] know, when are we going to be able to [12:17] book a hotel using your consumer agent [12:20] in your pocket? I think we've just [12:21] gotten that with Muse and with Instinct. [12:24] This could be another trillion dollar [12:26] category, but it's definitely going to [12:27] consume massive tokens. Here are the [12:31] three risks and challenges. And sorry [12:32] I'm going so quick, but I want to keep [12:34] it keep it moving. Regulation. We heard [12:37] a lot about this yesterday. Power. And [12:40] then what's going on with interest [12:42] rates? So, you know, this is the [12:43] regulation tugof-war, [12:45] right? Um, and I I've heard from a lot [12:48] of people, they're like, you're on both [12:49] sides of the issue. Here's the fact, [12:52] okay, the answer is not going to be on [12:55] one end or the other. We're going to [12:58] have to have common sense, pragmatic [12:59] solutions that get my mom, my sister, [13:02] and the br brother off the cliff, right? [13:05] We need to give comfort and confidence [13:07] to the people who elect our [13:09] representatives that it's safe. And so, [13:11] you heard Elon yesterday give a great [13:13] suggestion around peer review. So, I'm [13:15] confident that we're going to get there, [13:17] but it's kind of messy the sausage [13:18] making along the way. But we have a [13:21] prior history of excess regulation. When [13:24] people get scared, when activists start [13:27] pushing an agenda, we shut down 67 [13:30] fision reactors in this country. We [13:32] unilaterally disarmed against China. [13:35] It's been a disaster for the country. [13:37] Right? All the clean energy we could [13:39] have gotten. Instead, we've gotten [13:41] non-clean emissions because we had a [13:43] group of activists who were hellbent on [13:45] shutting down nuclear. can't allow this [13:47] to occur to AI. [13:53] So regulation's a threat. It's a risk. [13:57] Atoms and energy are hard. So getting [14:00] back to can we stand up 43 gawatt of [14:02] compute. Our total compute in the [14:04] country is less than 40 gawatt. Doing [14:06] this in one year, we've got to overcome [14:08] permitting and local opposition. You [14:10] guys see all of that. grid [14:12] interconnection delays, skilled labor [14:14] shortages, power equipment is sold out. [14:16] It's the largest buildout in the history [14:18] of the country. I would suggest Dylan's [14:20] forecast to 43 gawatt next year is too [14:23] aggressive. I don't think we're going to [14:25] get there. I think the total amount [14:27] we're actually going to stand up is [14:28] somewhere closer to 25 gawatt. And I [14:31] think of that 25 gawatt, half of it will [14:34] be for anthropic and open AI. [14:38] Um I think that's enough to generate the [14:41] revenue. Remember Anthropics revenue [14:43] reportedly this year 100 billion 110 [14:46] billion. If they do that they're doing [14:48] it with a gigawatt and a half of [14:49] compute. So if they add another four or [14:52] five gawatts of compute that's certainly [14:54] enough to add another 100 billion in [14:56] revenue. So I don't think we need more [14:58] gigawatts to get to the revenue targets [15:01] for next year. But this is my hunch that [15:03] we're not going to get to the 43. [15:05] And then of course we're going to hear [15:07] more tomorrow. I think rate hikes are [15:08] coming. I think it's now over 90% chance [15:10] that we're going to have rate hikes [15:12] tomorrow. Why does this matter? Because [15:13] all of this now borrowed money, right? [15:16] We're borrowing money in order to stand [15:17] up these data centers. So the hurdle [15:19] rate for that money is going up. And um [15:23] so that's not only a challenge for data [15:25] centers, but remember, you know, as [15:27] Warren Buffett says that, you know, [15:29] interest rates are to stocks what [15:32] gravity is to matter, [15:34] right? Right? If you can earn 5 1.5 or [15:36] 6% on your money without taking equity [15:39] risk, then you know it's going to be a [15:41] challenge for stocks. So, here's where I [15:43] think we are and we're going to end on [15:45] this slide and and then then we'll bring [15:47] the guys out and chop it up a little [15:48] bit. But this is the flight path. This [15:50] is how I think about managing the [15:51] portfolio, right? We're up about the [15:55] NASDAQ's up about 15%. But as I sit here [15:58] and think about the risk, do I want to [15:59] be small, medium, or large? This is how [16:02] I think about the fan of potential [16:03] outcomes. If the monthly AI lab revenues [16:06] are closer to that$8 billion number, I [16:09] think it's takeoff. Okay. I think we are [16:11] going to see an IPO this year. Um, and [16:14] so I'm paying very close attention to [16:17] what in fact are those monthly revenue [16:19] numbers. I think the trends are intact, [16:21] but we will see. That's going to be the [16:22] single most important thing as to [16:24] whether or not uh between now and the [16:26] end of the year uh we we have liftoff. [16:28] Second one is rates, the election, oil [16:30] prices. Obviously, rates are following [16:32] oil prices to a certain extent. So, what [16:35] happens there? You know, if rates were [16:37] to go to five and a half on the 10-year, [16:38] that's going to be a big burden uh you [16:41] know, on on the equity market. Then [16:43] finally, the regulation, the anthropic [16:45] IPO, we saw a trade down yesterday [16:47] because people are concerned that maybe [16:49] we're going to have a halt or or [16:50] postponement. That would obviously be a [16:52] major issue. I don't think that's going [16:54] to happen, but again, that would be a [16:56] concern. So that's the fan of potential [16:58] outcomes. From here we're up 15%. [17:01] I think we could go higher uh you know [17:03] through the balance of the year. Uh or [17:05] we could go lower. I'll leave you with [17:07] this. [17:08] The period of 2023 to 2025, you only had [17:12] to get one thing right. That AI was [17:15] going to be the biggest super cycle in [17:16] the history of technology. And you [17:18] needed to shove your chips into the AI [17:19] trade. That's it. Right? If you were in [17:22] the AI trade, you made money. That is [17:24] not where we are in 2026. Everybody [17:26] knows about AI. It's all priced now. [17:29] It's about facts and circumstances. Stay [17:32] mentally flexible. Follow the facts. [17:35] Don't yolo. Okay? Don't go forex like [17:38] our friend up north who uh gave all his [17:40] money to Citadel, right? Um Forex [17:44] levered in this market, you know, very [17:47] dangerous. So, in my estimation, we're [17:50] medium position. We're mentally [17:52] flexible. If we see those revenues come [17:54] in big for these next few months and we [17:57] see oil prices retreat, we're going to [17:59] put more chips on the table. If not, [18:02] we'll, you know, we'll reserve the right [18:03] to go even smaller. With that, thank you [18:05] all. Thanks for having me.