Tae Kim Says Market Misjudges Imminent AI Compute Boom From Self-Improving Models
TBPN
Upcoming recursive self-improvement and agentic AI models will trigger exponential demand for compute and memory infrastructure rather than lower capital expenditure requirements.
## Background
TBPN published this video interview on YouTube on July 29, 2026. According to the page description, the episode features a wide-ranging conversation with Tae Kim about misconceptions surrounding the AI trade, why models like Kimi and open-weight architectures increase compute demand, Meta's capital expenditure plans, NVIDIA's strategy and CUDA moat, and the rapidly approaching milestone of Recursive Self-Improvement (RSI).
## TL;DR
Tech analyst Tae Kim addresses recent market volatility and Wall Street narrative shifts, arguing that fears of an AI spending slowdown or compute glut are heavily overblown. He explains how breakthroughs in open-weight and reasoning models actually accelerate GPU demand, why Meta's ad-engine integration provides immediate ROI, and how upcoming Recursive Self-Improvement (RSI) will trigger the next massive compute wave. Kim also debunks sensationalized headlines regarding data center leasing costs and NVIDIA's financial backing, pointing to robust supply chain signals and AMD's expanded market forecasts as evidence of sustained long-term growth.
## Atmosphere & style
The conversation carries a relaxed, peer-to-peer banter dynamic mixed with sharp, high-conviction market analysis. The hosts and Tae Kim maintain an upbeat, reassuring, and slightly sarcastic tone toward media sensationalism and panicking retail investors, grounding their points in hard enterprise data and semiconductor realities. It feels like an insider debrief among tech enthusiasts who are calmly looking past short-term market noise to focus on fundamental technological shifts.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=0s) Market sentiment dropped significantly due to retail and hedge fund investors buying near a parabolic top, combined with chip stock volatility heavily driven by geopolitical events in the Middle East.
- [01:45](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=105s) Media interpretations claiming Meta would cut AI capital expenditures were proven false, with Meta instead planning dramatic capex increases this year and next.
- [05:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=350s) Meta holds a unique structural advantage because its frontier AI model improvements are integrated directly into its core ad engine, generating immediate ROI without relying on separate software revenue.
- [07:27](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=447s) NVIDIA successfully built an open-weight coalition representing over $18 trillion in market capitalization to defend against domestic regulatory overreach that could disadvantage American tech firms.
- [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) Highly efficient open-source models like China's Kimi actually spark massive compute demand rather than creating a glut, requiring large GPU clusters that quickly overloaded servers at launch.
- [12:04](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=724s) Recursive Self-Improvement (RSI)—where AI models autonomously design and refine subsequent AI generations—is much closer than the public realizes and will trigger an unprecedented surge in compute demand.
- [13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) SK Hynix faces hardware demand 5x to 6x higher than its capacity, while AMD drastically raised its AI data center TAM forecast from $120 billion to $220 billion in only three months.
- [13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) Morgan Stanley estimates hyperscaler inferencing workloads yield 60% to 80% profit margins, directly refuting comparisons to unprofitable dot-com era infrastructure.
- [16:36](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=996s) The focus on near-term AI ROI ignores the massive operating cash flows generated by cloud hyperscalers like Azure and Google Cloud, which are experiencing dramatic top-line revenue growth.
- [16:36](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=996s) Enterprise AI adoption is still in its infancy outside of the San Francisco startup bubble, leaving massive headroom for expansion within the $6 trillion corporate IT and knowledge management TAM.
- [16:36](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=996s) Companies must consider 'Return on Revenue' or return on fear, as failing to adopt agentic AI risks losing market share to competitors who actively deploy it.
- [20:56](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1256s) Sensational media reports over Nvidia backstopping data center leases or OpenAI-related ventures omitted critical context, misrepresenting standard operational or tenant leasing commitments as systemic panics.
- [25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) Nvidia's true moat extends beyond CUDA's decade of software optimization into balance sheet power, utilizing aggressive prepayments to corner the supply chain for TSMC wafers and HBM memory.
- [25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) Jensen Huang stated that the chip supply chain has enough capacity to double Nvidia's revenue year-over-year, which significantly outpaces current Wall Street consensus estimates.
- [28:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1705s) Nvidia's expanding balance sheet investments in ecosystem partners like CoreWeave, Lumentum, and Coherent serve to aggressively catalyze the supply chain and ensure infrastructure meets incoming demand.
Tae Kim returns to TBPN after previously "bottom-ticking" the market during his last appearance. The hosts joke about keeping an "emergency strategic reserve of Tae Kim appearances" on standby to rescue falling stock prices, noting how his previous bullish calls on CPUs, memory, and NVIDIA were followed by massive upside moves.
Addressing the current high level of fear, uncertainty, and doubt (FUD) in the market, Kim notes that market sentiment is extremely negative following a parabolic run. Retail investors and hedge funds piled in near the top, leading to a recent unwind. Kim points out that macro factors—specifically escalations in the Middle East—heavily influenced chip stock volatility:
* Stock prices rallied during initial ceasefire talks, but plummeted when military actions resumed.
* Media outlets and viral hot-take commentators have amplified narrative-driven FUD during the downturn.
[01:45](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=105s) Clearing the FUD: Meta's Capex and "Agentic AI" Misinterpretations
Kim directly addresses a July 2nd Reuters report quoting Meta CEO Mark Zuckerberg regarding AI agents. The article claimed that agentic AI tech was progressing slower than expected and highlighted internal organizational changes, prompting media speculation that Meta would cut its capital expenditure (capex).
The hosts and Kim critique the narrative spin, pointing out how rapidly executives like Meta's CTO Andrew Bosworth ("Boz") and Scale AI CEO Alexandr Wang stepped in to set the record straight:
* The original quote stemmed from an internal town hall rather than a public strategic retreat.
* Alexandr Wang continuously reaffirmed Meta's aggressive push to rebuild open-source AI models, emphasizing their commitment at recent events like Y Combinator.
* Reports following the initial scare indicated that Meta actually intends to raise capex dramatically this year and next, completely refuting the pull-back narrative.
* Semianalysis published a detailed bull case for Meta, highlighting its structural advantage in deploying compute directly into core ad products without the internal friction seen at competitors.
```
Key Meta Drivers:
• Direct integration into core ad engine yields immediate ROI.
• Massive capex increases planned for this year and next.
• Focus on frontier open-source development despite media noise.
```
[05:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=350s) Meta's Monotonic Ad Engine vs. The Search for Direct AI Revenue
The discussion turns to Meta’s overall AI monetization strategy. While hyperscalers and developers face questions about where their first $1 billion or $10 billion in direct API revenue will originate, Meta holds a unique position. The hosts emphasize that Meta's core ad engine absorbs AI improvements immediately, driving higher return on investment (ROI) without requiring stand-alone software sales.
While the market frets over the massive capex necessary to compete in the frontier model race, Meta's continuous model releases show rapid progress. A model released just weeks prior performed significantly better than market expectations, proving that its re-aligned AI team is gaining momentum.
[07:27](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=447s) NVIDIA's Strategic Open-Weight Coalition and Regulatory Pressure
The conversation transitions to NVIDIA’s recent open-letter strategy regarding open-source and open-weight AI models. Kim highlights how NVIDIA effectively unified almost the entire tech sector against Anthropic's restricted view on open weights within days, amassing companies representing over $18 trillion in market capitalization.
```
Industry Alignment on Open Weights:
• Major Signatories: OpenAI, Google, Microsoft, Amazon.
• Notable Holdout: Apple (a notable anomaly given their potential benefit from local open-weight models).
```
Addressing whether NVIDIA is acting out of fear, Kim clarifies that the push is primarily a defense against regulatory overreach. He references statements from NVIDIA CEO Jensen Huang and recent White House signals:
* Policymakers expressed concerns over model "distillation" rather than open weights themselves.
* Anthropic CEO Dario Amodei has pushed restrictions under the banner of AI safety.
* Hyperscalers like Microsoft, Google, and NVIDIA contend that domestic tech firms already face strong brand and financial incentives to prevent safety incidents.
* The primary concern is that over-regulating domestic open-weight models disadvantages American companies, while foreign entities face few barriers to releasing models without safety guardrails.
[10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) Kimi, DeepSeek, and Why Open Models Increase Compute Demand
Addressing the recurring narrative that efficient open-source models (such as China's Kimi) will lower compute requirements and shrink chip demand, Kim draws a direct parallel to the "DeepSeek moment" from the previous year.
Contrary to fears that efficient models create a compute glut, Kim explains why capability breakthroughs actually spark exponential demand surges:
* Kimi is not a lightweight model; it is a massive 2.8-trillion-parameter model requiring enormous compute infrastructure.
* Kimi's official documentation notes it runs best on clusters of at least 64 GPUs, leading to overloaded servers on launch day.
* Just as reasoning models like DeepSeek ultimately accelerated compute demand, advanced agentic models will do the same.
> "When you have more capable models that come out, people find uses for them... the next six to nine months are going to be bigger than anyone believes."
[12:04](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=724s) Recursive Self-Improvement (RSI) and the Next Compute Surge
Kim highlights upcoming catalysts in Agentic AI and Recursive Self-Improvement (RSI), citing recent statements by OpenAI's Sam Altman at a Y Combinator event. Altman noted that the next six months will bring more dramatic AI capability advances than the previous two years combined.
Kim posits that RSI—where models autonomously design and refine the next generation of AI—is far closer than public markets realize:
* Leadership and researchers across both OpenAI and Anthropic are hinting at imminent RSI milestones.
* Achieving functional RSI will create an unprecedented compound demand for compute, as models continuously utilize massive server clusters to self-develop.
[13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) Hardware Demand Signals: SK Hynix, AMD, and Hyperscaler Margins
Kim counters market skepticism by laying out concrete data points from major semiconductor players and cloud infrastructure providers:
* **SK Hynix:** Executives stated during their IPO process that client demand is 5x to 6x higher than current supply capacity, prompting plans to double high-bandwidth memory (HBM) capacity over the next five years.
* **AMD:** CEO Lisa Su aggressively raised her forecast for the AI data center chip total addressable market (TAM) from $120 billion to $220 billion in just three months—a rare move for a conservative public CEO.
* **Inferencing Profitability:** Morgan Stanley estimates that inferencing workloads yield 60% to 80% profit margins for hyperscalers, disproving comparisons to unprofitable dot-com era infrastructure.
* **OpenAI Spending Regrets:** Sam Altman publicly acknowledged on a recent podcast that OpenAI made a mistake when it temporarily slowed compute acquisitions, and is now working to accelerate capacity purchases.
* **Amazon's Cash Flow Strategy:** CEO Andy Jassy’s shareholder letter clearly outlined that Amazon’s $200 billion capex commitment is backed by concrete customer demand and high long-term free cash flow visibility.
[16:36](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=996s) The $6 Trillion AI Opportunity and the "Return on Revenue" Argument
The conversation opens with a discussion on the sustainability of massive AI capital expenditures. While critics warn of an eventual cap ex plateau, Tae Kim argues that this static view ignores dramatic revenue growth across cloud hyperscalers—such as Microsoft Azure growing at 40% and Google Cloud hitting 80%. He emphasizes that this surge in top-line growth generates immense operating cash flow to fund continued data center expansion. The true catalyst over the next 12 to 18 months is agentic AI, which is actively rearchitecting enterprise workflows and R&D processes. Kim points to AMD CEO Lisa Su’s strategy as a prime example, noting that AMD has put roughly 100 generative AI models into production and is burning through a trillion tokens a month at a double-digit growth rate to drive internal product development and customer support.
Beyond standard ROI, Kim introduces the concept of "Return on Revenue" (or return on fear): if a company fails to integrate agentic AI, competitors like Verizon or AT&T will leverage it to disrupt them and siphon off their market share. While enterprise adoption is currently uneven—concentrated heavily in a San Francisco startup bubble where usage is high—broader corporate adoption remains in its infancy. Citing observations from industry analysts like Jared Sleeper and data from Ara, Kim notes that total enterprise AI usage could easily jump 100x if average corporations adopted AI to the degree of leading tech firms. With the corporate IT and knowledge management market sitting at approximately $6 trillion annually, Kim sees vast room for growth beyond the $120 billion in combined annual recurring revenue (ARR) currently generated by leading frontier model developers like OpenAI and Anthropic.
* Enterprise IT and knowledge management represent a $6 trillion total addressable market (TAM).
* Top frontier model creators currently generate roughly $120 billion in combined ARR, leaving massive headroom to scale toward $200B–$400B.
* AMD is actively running over 100 production GenAI models internally, consuming over 1 trillion tokens monthly.
> "Everyone's worried about ROI. ROI is important, but you also need return on revenue because if you don't use AI, your rival is going to use AI to beat you in the market."
---
[20:56](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1256s) Debunking Sensational Headlines: Financial Times and Wall Street Journal
The hosts and Kim address recent media headlines that spooked the market, starting with a *Financial Times* report claiming Nvidia was backstopping a $50 billion data center lease in Texas. Kim unpacks the story, pointing out that the initial headline omitted critical context: the arrangement is actually a 15-year lease commitment valued at around $20 billion, with renewal options extending it to $50 billion over 30 years. On a per-year basis, it represents a minor $1 billion to $2 billion operational commitment—a total "rounding error" for a company operating at a $320 billion annual run rate and pacing toward $400 billion to $500 billion. Rather than a sign of Nvidia buying its own chips, it represents standard tenant leasing that can be subleased for profit.
The group also touches on a *Wall Street Journal* report alleging that Nvidia is in talks to backstop SoftBank up to $250 billion for OpenAI-related ventures. While both companies declined to comment at the time, Kim urges calm, pointing out how figures like CNBC's Jim Cramer hyperventilated by advising viewers to "sell everything at the open" over data center panic. Kim emphasizes looking at underlying industry fundamentals: AMD expanding its total market forecast, researchers prioritizing recursive self-improvement (RSI), and global leaders prioritizing AI hardware supply. A recent visit by the South Korean president to Silicon Valley—which brought together top executives including Jensen Huang, Haw Tan, Dario Amodei, and Sam Altman—underscores the intense, ongoing global shortage of High Bandwidth Memory (HBM).
* The FT's $50 billion Texas data center headline breaks down to roughly $20B over a 15-year lease (with options up to 30 years via Hut 8), representing ~$1B–$2B annually.
* Nvidia's revenue run rate is hovering near $320 billion, with projections aiming for $400B–$500B in the near term.
> "They just wanted to say the biggest number... It's a rounding error. They're doing a $320 billion run rate a year... This is how people run with the sensationalized headlines and people panic and freak out."
---
[25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) Nvidia’s CUDA Moat and Dominant Supply Chain Control
The discussion pivots to whether Nvidia's software dominance via CUDA is eroding, given claims from competitors like AMD and model developers like Moonshot AI (Kimi) that AI agents can automatically write GPU kernels to run models on alternative silicon. Kim remains skeptical of the "death of CUDA" narrative, explaining that CUDA’s true moat lies in over a decade of deep reliability, scale, and battle-tested software optimization that has been run millions of times. Beyond software, Nvidia’s most underrated moat is its balance sheet power and supply chain dominance. By prepaying suppliers, Nvidia has locked up the vast majority of critical component inventory—from TSMC wafers and HBM memory to optical interconnects from startups and established players like Lumentum and Coherent.
Addressing persistent energy and power bottleneck fears, Kim highlights Nvidia CEO Jensen Huang's recent comments on Bloomberg. While Huang acknowledged real-world constraints in data center shells and power, he stated that the semiconductor supply chain has enough capacity to double Nvidia's revenue year-over-year. This capability far exceeds current Wall Street consensus estimates.
* Nvidia's co-design approach tightly integrates networking, CPUs, GPUs, and custom optics into a unified stack.
* Aggressive prepayment strategies have allowed Nvidia to corner the market on TSMC wafer allocations, HBM memory, and optical components.
> "Jensen said this last week... 'I think the chip industry has enough supply to double their revenue every year.' Basically implying Nvidia has enough supply for energy and all that stuff. No one is pricing that in."
---
[28:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1705s) Strategic Investments and Ecosystem Ownership
In the final stretch, the hosts reflect on Nvidia’s expanding balance sheet investments across the entire AI ecosystem, drawing parallels to its early, highly profitable stake in CoreWeave as well as recent investments in optical hardware leaders Lumentum and Coherent. Rather than predatory "vendor financing," Kim views these equity investments as Jensen Huang aggressively catalyzing the supply chain. By capitalizing startups, optical component makers, and specialized GPU cloud providers, Nvidia ensures these key partners can build out the infrastructure required to meet incoming demand. In two to three years, Nvidia's corporate venture portfolio could comfortably exceed a trillion dollars in value.
The episode concludes on a lighter note as the hosts joke about Jensen Huang's iconic image—highlighting a charity auction where one of his leather jackets fetched a massive sum. The hosts tease about open-sourcing a "Jensen jacket dupe" before thanking Kim for bringing clarity to the market volatility.
* Nvidia continues to leverage its massive cash balance to invest in key ecosystem partners, including CoreWeave, Lumentum, Coherent, and SSI (Safe Superintelligence).
* Host joke: *"The man can buy a leather jacket for like five grand and sell it for a million dollars. I don't know what else you need to see."*
> "In a year, two, three years, Nvidia is going to have all these stakes in these companies and it's going to look like he was a good investor... because he has been in the past."
## Notable quotes
> [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) When you have more capable models that come out, people find uses for them... the next six to nine months are going to be bigger than anyone believes. — Tae Kim
> [16:36](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=996s) Everyone's worried about ROI. ROI is important, but you also need return on revenue because if you don't use AI, your rival is going to use AI to beat you in the market. — Tae Kim
> [20:56](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1256s) They just wanted to say the biggest number... It's a rounding error. They're doing a $320 billion run rate a year... This is how people run with the sensationalized headlines and people panic and freak out. — Tae Kim
> [25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) Jensen said this last week... 'I think the chip industry has enough supply to double their revenue every year.' Basically implying Nvidia has enough supply for energy and all that stuff. No one is pricing that in. — Tae Kim
> [28:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1705s) In a year, two, three years, Nvidia is going to have all these stakes in these companies and it's going to look like he was a good investor... because he has been in the past. — Tae Kim
> [28:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1705s) The man can buy a leather jacket for like five grand and sell it for a million dollars. I don't know what else you need to see. — Host
## Mentioned
Tae Kim, TBPN, Meta, NVIDIA, AMD, Kimi, DeepSeek, Mark Zuckerberg, Andrew Bosworth, Alexandr Wang, Scale AI, Y Combinator, Semianalysis, Anthropic, OpenAI, Google, Microsoft, Amazon, Apple, Jensen Huang, Dario Amodei, Sam Altman, SK Hynix, Morgan Stanley, Lisa Su, Andy Jassy, Microsoft Azure, Google Cloud, Verizon, AT&T, Jared Sleeper, Ara, Financial Times, Wall Street Journal, Hut 8, SoftBank, CNBC, Jim Cramer, Haw Tan, TSMC, Moonshot AI, Bloomberg, Lumentum, Coherent, CoreWeave, Safe Superintelligence
Financial analyst Tae Kim argues that public market fears of an AI spending peak are unfounded, contending that looming recursive self-improvement, agentic workflows, and persistent hardware bottlenecks will drive a compounding surge in compute demand.
Refuting Capex Cut Fears and Media Narratives (1:45)
Kim and the hosts critique market overreactions to recent news reports alleging that hyperscalers are scaling back artificial intelligence investments. Addressing a July 2nd Reuters report on Mark Zuckerberg's comments regarding agentic AI, Kim notes that executives such as Meta CTO Andrew Bosworth and Scale AI CEO Alexandr Wang quickly clarified that Meta intends to aggressively increase capex both this year and next. According to Kim, Meta holds an advantage because AI improvements integrate directly into its core advertising engine to deliver immediate return on investment, bypassing the need for standalone API monetization.
Kim also refutes reporting from the Financial Times concerning an alleged $50 billion Nvidia data center lease in Texas. He clarifies that the deal is a 15-to-30-year operational lease valued at $1 billion to $2 billion annually. Kim characterizes this expense as a minor operational commitment relative to Nvidia's approximately $320 billion annual revenue run rate, rather than an indicator of artificial demand or self-dealing.
Compute Demand Surges from Agents and RSI (12:04)
Addressing arguments that algorithmic efficiency and open-weight models will diminish semiconductor demand, Kim compares recent sentiment to prior reactions surrounding DeepSeek. He notes that China's Kimi is a massive 2.8-trillion-parameter model that requires at least 64 GPUs per cluster, which overwhelmed servers at release. In Kim's view, as model capabilities expand, practical deployment expands alongside them rather than reducing compute needs.
Kim points to recursive self-improvement (RSI)—where models autonomously design, test, and optimize subsequent AI systems—as the primary demand driver over the coming 6 to 18 months. Citing statements by OpenAI CEO Sam Altman indicating that forthcoming model advances will outpace the prior two years combined, Kim emphasizes that researchers at OpenAI and Anthropic are targeting RSI milestones. Achieving autonomous self-improvement would establish a continuous, compounding workload on data center clusters.
Supply Chain Signals and Expanding Enterprise TAM (13:08)
Kim contrasts market skepticism with quantitative data from semiconductor suppliers and cloud providers. Memory supplier SK Hynix reported during its IPO process that customer demand outpaces capacity by five to six times, spurring plans to double high-bandwidth memory capacity over five years. Concurrently, AMD CEO Lisa Su raised her AI data center total addressable market projection from $120 billion to $220 billion in three months, while AMD runs over 100 internal generative models processing a trillion tokens monthly. Financially, Morgan Stanley estimates hyperscaler inferencing profit margins between 60% and 80%, countering comparisons to unprofitable dot-com infrastructure.
Beyond direct return on investment, Kim outlines a competitive dynamic of "Return on Revenue," where enterprises adopt AI defensively to protect their core business from rivals. With corporate IT and knowledge management representing a $6 trillion addressable market, Kim sees significant headroom beyond the estimated $120 billion in combined annual recurring revenue generated by leading frontier model developers.
Nvidia's Supply Prepayments and Ecosystem Dominance (25:05)
Addressing claims that AI agents generating custom GPU kernels will erode Nvidia's CUDA software moat, Kim argues that CUDA's real advantage lies in more than a decade of reliability, scale, and operational testing. Beyond software, Kim identifies Nvidia's primary competitive moat as its balance sheet and supply-chain dominance. By deploying aggressive cash prepayments, Nvidia has secured critical allocations of TSMC wafers, advanced packaging, high-bandwidth memory, and optical interconnects from suppliers like Lumentum and Coherent.
Kim also highlights Nvidia's coalition of tech companies representing $18 trillion in market capitalization—including Microsoft, Google, OpenAI, and Amazon—formed to protect domestic open-weight models from restrictive regulations. Citing statements from CEO Jensen Huang that the supply chain possesses enough capacity to support doubling Nvidia's revenue year-over-year, Kim views Nvidia's equity investments in partners like CoreWeave and optical component makers as proactive efforts to capitalize the infrastructure required to meet expanding demand.
While Kim's outlook anticipates rapid hardware expansion driven by agentic computing and self-improving models, he acknowledges that current enterprise adoption remains largely concentrated among tech-forward firms and Silicon Valley startups. Whether mainstream corporate adoption can scale by the estimated 100x required to absorb long-term hyperscaler capital commitments remains an ongoing test for the broader market.
## Background
该视频由科技播客频道 TBPN 于 2026 年 8 月 3 日在 YouTube 平台发布。节目简介显示，本期为对知名财经与半导体分析师 Tae Kim 的独家专访，重点拆解市场对 AI 交易的广泛误读，并深入探讨 Kimi 等高参数开源模型为何会反向刺激算力需求、Meta 的资本开支真相、递归自我改进（RSI）的技术飞跃、英伟达的供应链护城河与生态投资，以及数据中心和能源危机背后的真实产业逻辑。
## TL;DR
Tae Kim 结合半导体供应链硬核数据全面反驳了近期弥漫市场的“AI 算力过剩”论调，指出随着 2.8 万亿巨型参数模型、智能体（Agentic AI）以及递归自我改进（RSI）的加速逼近，算力需求正面临多重指数级爆发。SK 海力士订单超载 5 至 6 倍、AMD 翻倍上修市场规模预估，加之云厂商在 GPU 推理端高达 60% 至 80% 的丰厚利润率，印证了企业为了业务生存必须全力扩大 AI 资本支出的必然性。英伟达凭借全栈协同设计、巨额预付款对核心稀缺元器件的独占锁定，以及贯穿生态的战略资本网络，构筑起极高的壁垒，媒体对所谓 500 亿美元得州数据中心租赁担保等负面头条纯属夸大其词。
## Atmosphere & style
整场访谈节奏明快且极富戏剧张力，主持团队与 Tae Kim 默契十足，在一连串关于“老黄皮夹克拍卖与蒸馏”以及“建立 Tae Kim 救市战略储备”的机智行业爆梗中笑声不断。对话在高密度的产业链硬核数据推演与轻松幽默的硅谷圈内调侃之间自如切换，以极其冷静且理性的基本面视角逐一粉碎散布在社交平台与媒体头条上的恐慌情绪（FUD）。整期内容散发着强烈的科技乐观主义与坚定定力，在金融市场剧烈动荡的背景下，为听众提供了一场既干货淋漓又兼具解压情绪价值的深度对谈。
## Bulletin
- [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) 市场误解了高效开源模型的算力影响，如月之暗面 Kimi 参数量高达 2.8 万亿且需 64 卡 GPU 集群部署，开源模型与智能体演进不仅不会削减算力需求，反而在加速算力吞噬。
- [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) 前沿 AI 实验室暗示递归自我改进（RSI）的到来远比公众预期更快，叠加推理扩展与智能体落地，算力需求正面临三重指数级爆发。
- [13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) 供应链硬核数据印证算力依然极端紧缺：SK 海力士 HBM 订单需求为其供应极限的 5 到 6 倍，AMD 亦在 3 个月内将 2030 年 AI 加速器/CPU 潜在市场规模从 1200 亿美元大幅上调至 2200 亿美元。
- [20:55](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1255s) 英伟达得州 500 亿美元数据中心租赁传闻遭媒体夸大，实为 15 年期约 200 亿美元基础租约，年均仅 10 亿至 20 亿美元且具备转租盈利空间，相对其超 3000 亿美元的年化营收仅为极小比例。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=0s) 宏观情绪因地缘政治冲击、资金解杠杆及媒体散布 FUD 而处于极度悲观的超跌挤泡沫状态，但底层技术指标与产业基本面依然向好。
- [02:45](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=165s) 路透社对扎克伯格内部表态的断章取义引发 Meta 削减 CapEx 恐慌，但 Meta 随后澄清其今年与明年将大幅增加资本开支，坚定推进前沿模型竞争。
- [05:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=350s) Meta 的变现逻辑不仅限于核心广告业务的 AI 赋能，更是承担重组研发团队、争夺通用前沿大模型主导权的双重战略使命。
- [07:41](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=461s) 英伟达在短短数日内联合超 18 万亿美元市值的科技巨头签署开放权重支持信，旨在先发制人抗衡华盛顿借“反蒸馏”之名施加的前置行政监管。
- [09:54](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=594s) 以 Anthropic 的 Dario Amodei 为代表的阵营对完全无约束的开源前沿模型保持警惕，担忧一旦发生网络或生物安全事故将对科技巨头根基造成毁灭性打击。
- [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) 高参数开源模型的涌现与 Agentic AI 的爆发打破了“算力过剩论”，随着模型利用海量算力进行自我迭代，算力消耗将呈飞轮式剧增。
- [15:01](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=901s) 超大规模云厂商的 GPU 推理业务利润率高达 60% 至 80%，强劲的云业务营收增速与经营现金流足以支撑当下庞大的资本开支。
- [15:01](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=901s) 企业在 AI 采用上面临“营收生存危机”，不重构流程提高效率的企业将迅速被对手颠覆，这驱动着 AT&T 等传统企业每月消耗多达 1 万亿 token。
- [18:40](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1120s) 目前绝大多数企业对 AI 的使用仍停留在报销 ChatGPT 账号的浅层阶段，全行业若能对齐硅谷头部深度，AI 调用量将面临 100 倍的增长空间。
- [19:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1190s) 企业 IT 与知识管理年市场规模达 6 万亿美元，目前 OpenAI 与 Anthropic 合计年经常性收入仅百亿美元规模，未来具备数千亿美元的指数级成长空间。
- [22:51](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1371s) 韩国总统访美引得黄仁勋、陈福阳、Sam Altman 与 Dario Amodei 齐聚，凸显 HBM 内存已成为制约各大算力霸主的极度短缺战略物资。
- [25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) 即便 AI 智能体能辅助编写 GPU 内核，英伟达的护城河依然稳固：全栈系统协同设计及其雄厚资产负债表对台积电产能、HBM 和光学组件的独占锁定难以被轻易打破。
- [27:30](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1650s) 针对能源与机房瓶颈的担忧，黄仁勋强调供应链能力足以支撑芯片行业营收每年翻倍，当前华尔街的增长预期明显偏低。
- [29:59](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1799s) 英伟达积极通过战略资本注资 CoreWeave、光学厂商及前沿 AI 团队，既打通了下游算力设施的扩产瓶颈，又为其资产负债表构建起庞大的股权生态版图。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=0s) Tae Kim 再次做客：市场情绪重置与 FUD 蔓延
节目开场，主持人热情迎接嘉宾 Tae Kim 进入直播间，并调侃 Tae Kim 上一次上节目时精准命中了市场周期的绝对底部（"bottom ticked it"）。Tae Kim 笑着回顾了他当时对 CPU、存储芯片以及 NVIDIA 的看好判断：虽然经历了一波深度回调，但 NVIDIA 依然上涨了 5% 到 10%，CPU 标的翻倍，高带宽内存（HBM）相关股票更是大涨了 100%。主持人开玩笑称，应该为市场建立一个“Tae Kim 战略应急储备”，一旦股市下跌就立刻请他连线救市。
随后，对话切入当下半导体与 AI 板块的宏观情绪。Tae Kim 直言，当前市场情绪极其负面。过去几个月市场经历了抛物线式的狂热暴涨，大量散户与对冲基金涌入，而现在正处于快速平仓解套的挤泡沫阶段。导致抛售的触发因素接踵而至：先是地缘政治方面，中东局势及轰炸事件重创了市场预期；紧接着，媒体与社交平台上的短视“暴论”（viral hot takes）持续散布 FUD（恐惧、不确定与怀疑）。从路透社断章取义报道扎克伯格对智能体（Agentic AI）进展缓慢的内部言论，到《华尔街日报》针对供应商融资（Vendor Financing）的质疑报道、中国长鑫存储（CXMT）相关动向引发的恐慌，再到 The Information 针对 ASML 的报道以及 Kimi 模型的讨论，一系列杂音令投资者神经高度紧张。
- **过往预测表现**：上次连线后芯片股触底反弹走出指数级行情，HBM 概念股涨幅达 100%，CPU 相关标的实现翻倍。
- **近期负面压制因素**：地缘政治震荡加剧、短线资金与散户追高后回撤解杠杆、主流财经媒体对巨头资本支出及技术的片面解读。
---
### [02:45](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=165s) 扎克伯格内部讲话被断章取义与 Meta 的资本开支之争
针对 7 月初路透社关于扎克伯格言论的报道，主持人与 Tae Kim 进行了详细梳理。路透社当时发文称扎克伯格在内部全员大会（Town Hall）上表示“AI 智能体技术的进展慢于预期”，且内部裁员与架构重组不够干脆利落。外界一度将此解读为 Meta 投资回报不及预期、可能削减资本支出（CapEx）的危险信号。然而，Meta 的反击极其迅速：高管 Boz 随后做客播客澄清，Alexandre Wang 也在社交媒体和 Y Combinator 线下活动中密集披露模型与数据方面的实质性进展，SemiAnalysis 亦撰写深度长文唱多 Meta，指出其算力充沛且组织架构高度协同，敢于在 AI 时代全面“All-in”（"properly yolo in the AI era"）。
Tae Kim 指出，路透社随后也修正报道称 Meta 实际上将在今年和明年大幅增加资本开支。所谓“智能体进展受阻导致削减开支”的恐慌完全是虚惊一场。主持人提到，Meta 的公关口径此前给人一种“AI 将替代员工”的被动紧缩感，若能主动向市场传递作为超大规模云厂商（Hyperscaler）大举进攻的姿态，效果会好得多；Tae Kim 则解释，那原本只是一场并未打算公开的内部例会，却被媒体摘取单句做成了耸人听闻的头条。
```mermaid
flowchart TD
accTitle: Meta 舆情发酵与反转链条
accDescr: 展示 Meta 内部讲话被路透社报道引发市场恐慌，随后管理层与机构多方辟谣反转的完整脉络
subgraph SG1["舆论发酵阶段"]
A["扎克伯格内部全员会表态：<br/>Agent 进展不及预期、重组未达理想"] --> B["路透社断章取义单句做标题：<br/>引发削减 CapEx 恐慌"]
B --> C["媒体与散户大肆传播 FUD：<br/>质疑 Meta AI 投资回报率 (ROI)"]
end
subgraph SG2["官方澄清与事实反转"]
C --> D["高管与分析机构密集发声：<br/>Boz 播客、Alex Wang 披露数据、SemiAnalysis 唱多"]
D --> E["真实战略浮出水面：<br/>Meta 计划今年与明年大幅提升 CapEx，全面进军前沿模型"]
end
```
---
### [05:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=350s) Meta 的变现闭环：广告赋能还是前沿大模型？
对话深入到投资者最为关心的核心问题：Meta 的 AI 收入到底何时能够兑现？主持人指出，虽然 Meta 依靠 AI 优化广告推荐系统（"AI has accelerated everything"）带来了显而易见的 ROI，但如果管理层不断扩大 CapEx 规模，宣称全力押注 Agentic 智能体框架并推行开源路线，市场必然会追问：除了让现有的广告业务变得更强之外，纯 AI 产品收入或 API 调用收入何时能跨越 10 亿甚至 100 亿美元大关？
Tae Kim 分析认为，Meta 现阶段的底层逻辑仍然是重返前沿 AI 模型的军备竞赛。虽然此前经历了核心人才流失的阵痛期，但随着 Alexandre Wang 大举招兵买马，整个研发团队已经完成重组。尽管构建下一代前沿模型需要 6 到 12 个月的时间才能显现质的飞跃，但近期发布的过渡模型表现已经显著超出外界预期。资本市场虽然对巨额 CapEx 的短期回报感到焦虑，但 Meta 显然不打算在前沿阵地上退缩。
- **商业化分歧**：市场希望 Meta 明确表态“买 GPU 只是为了把核心广告业务做到极致以保证 ROI”，但 Meta 实际在承担探索通用前沿模型的双重支出。
- **团队重组现状**：经历人员更迭后，新引入的人才正在全力推进自研模型架构，短期成果已展现韧性。
---
### [07:41](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=461s) 开源阵营大集结：NVIDIA 联手硅谷围剿监管与封闭路线
针对 NVIDIA 近期牵头推动的一封关于开源/开放权重（Open-weight）模型的行业联合信，Tae Kim 与主持人展开了重点探讨。Tae Kim 赞叹 NVIDIA 的行业号召力：在短短三四天内，NVIDIA 几乎联合了整个科技行业，累计市值超过 18 万亿美元的企业签署了联名信，在开源立场上直接形成了对 Anthropic 等封闭阵营的“围剿之势”。包括 OpenAI、后来跟进的亚马逊以及谷歌均加入了这一阵营，唯独苹果依旧按兵不动——主持人对此感到不解，因为理论上苹果原本应该从开放权重模型的繁荣中获益最多。
对于 NVIDIA 如此激进举动背后的动机，Tae Kim 指出，黄仁勋在接受彭博社采访时已明确表达了对华盛顿监管动向的担忧。白宫和国会内部近期至少有四位官员表态称“我们不反对开放权重，但我们反对蒸馏（Distillation）”，这被科技界普遍解读为针对开放权重模型实施前置监管的先兆。包括微软和 NVIDIA 在内的巨头担心政府打着“安全”旗号出台苛刻的行政限制，因此必须先发制人凝聚行业共识。主持人补充道，监管针对海外蒸馏行为的担忧尚可理解，但对于像 Gemma 或 Llama 这样由美国本土巨头主导的开源模型，巨头自身具备强大的安全团队与防事故机制，一旦出现侵权或蒸馏也可以直接通过法律诉讼解决，根本无需动用一刀切式的严苛行政限制。
- **联名阵营规模**：涵盖微软、NVIDIA、OpenAI、亚马逊、谷歌等，对应总市值超 18 万亿美元；苹果当时仍未表态。
- **核心政策博弈**：华盛顿监管层试图以“遏制非法模型蒸馏”为切入点限制开源生态，NVIDIA 则联合产业盟友抢先筑起防御墙。
### [09:54](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=594s) 开源模型的安全隐患与 Dario 的担忧
访谈承接上文关于前沿模型开源的争论。主持人与嘉宾探讨了将顶尖前沿模型直接开源给公众所带来的潜在风险。一旦大模型产生安全事故（Safety Incident），无论是网络安全漏洞还是生物安全危害，对于 Meta、Google 这类拥有庞大核心业务（如社交网络平台或 Google 搜索）的科技巨头而言，对其品牌形象与商业根基都将是毁灭性的打击。以 Google 的 Gemma 模型为例，若其被恶意利用，后果不堪设想。
相比之下，海外或外国开发团队在开源模型时往往缺乏同等顾虑，更容易抱持“将模型抛向市场、由外界自行承担后果”的姿态。嘉宾指出，这正是 Anthropic 首席执行官 Dario Amodei 一直以来高度警惕的核心原因：当缺乏约束的前沿能力被随意释出，安全防线将面临严重冲击。
---
### [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) Kimi 破除“算力过剩论”与即将到来的 RSI 飞跃
主持人提问：市场近期似乎在重演此前的“DeepSeek 时刻”，许多人认为开源高效模型的出现将大幅拉低推演与训练成本，进而导致 AI 算力投资降温（AI Trade 回调），这一逻辑是否站得住脚？Tae Kim 直言，这种担忧几乎完全是对现实的误判。
外界对于月之暗面（Moonshot AI）推出的 Kimi 等模型抱有“算力需求骤减”的恐慌，但深入研读技术白皮书和官方博客便会发现，Kimi 绝非某种轻量级微型模型，而是拥有高达 2.8 万亿参数（2.8 Trillion Parameters）的庞然大物。其推理和服务消耗极其惊人，上线首日服务器便被海量请求直接冲垮。官方博客明确建议该模型部署在配备 64 张 GPU 的互联服务器集群上，这完美适配英伟达的算力集群架构。正如一年前 DeepSeek 的推理能力爆发并没有引发算力过剩，反而刺激了全行业对推理算力的贪婪渴求一样，更强的模型只会让人们发掘出更多应用场景。
```mermaid
flowchart TD
accTitle: 算力需求与能力迭代的正向飞轮
accDescr: 展示从模型能力提升到 Agentic AI 落地、再到 RSI 带来的算力消耗爆发式增长机制。
subgraph Capability ["能力突破阶段"]
A["高参数开源/前沿模型发布<br/>(如 2.8T Kimi、推理模型)"] --> B["企业探索全新应用场景"]
B --> C["智能体 Agentic AI 全面爆发<br/>(代码编写与工作流重构)"]
end
subgraph Acceleration ["飞轮升级阶段"]
C --> D["模型自我迭代<br/>(递归自我改进 RSI)"]
D --> E["自我开发与能力演进<br/>消耗天量训练与推理算力"]
end
subgraph Demand ["市场结果"]
E --> F["算力需求呈指数级剧增<br/>打破算力过剩论"]
end
```
嘉宾进一步指出，当前智能体（Agentic AI）正步入爆发期，未来 6 到 9 个月的发展斜率将远超外界想象。Sam Altman 近日在 Y Combinator 活动及播客中已公开表态：未来 6 个月 AI 能力的飞跃程度，将戏剧性地超越过去两年的总和。这种技术跃迁很大程度上源于递归自我改进（Recursive Self-Improvement, RSI）的临近。无论是 OpenAI 还是 Anthropic 内部，研发团队都暗示 RSI 距离现实比公众想象的要近得多。当模型开始利用海量算力进行自我迭代、代码演化与自我提升时，它对底层硬件的算力吞噬将呈指数级爆炸，市场对算力过剩的悲观论调完全忽略了这一飞轮效应。
- Kimi 模型的参数量高达 2.8 万亿，推荐运行在 64 卡 GPU 的超大集群环境。
- 当前 AI 算力正由“推理扩展（Reasoning Ramp）”叠加“智能体扩展（Agentic Ramp）”，若叠加 RSI 机制，算力需求将呈三重指数级增长。
- Tae Kim 提到自己在社交平台 X 上发表有关 RSI 的观点时，频繁获得前沿 AI 实验室核心研究人员的点赞认同。
---
### [13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) 巨头囤积算力？来自 SK 海力士与 AMD 的真实需求印证
针对主持人关于“巨头囤积算力（Compute Hoarding）以及 Meta 等公司未对外转售算力是否会造成资源浪费”的疑问，Tae Kim 引用半导体供应链的硬核数据驳斥了这一顾虑。
在近期 SK 海力士（SK Hynix）的沟通中，管理层明确表示，尽管他们计划在未来 5 年内将产能翻倍，但当前来自核心客户（极大概率包括英伟达黄仁勋）的采购订单需求，仍然是其供应极限的 5 到 6 倍之多。供需失衡依然极度悬殊。此外，在 AMD 的 Advancing AI 大会上，首席执行官苏姿丰（Lisa Su）在短短 3 个月内，就将公司对 2030 年数据中心 AI 加速器/CPU 潜在市场规模（TAM）的预期从 1200 亿美元大幅上调至 2200 亿美元。
| 指标 / 事件 | 先前预测 / 现状 | 调整后预测 / 实际需求 | 信号解读 |
|:--- |:--- |:--- |:--- |
| **SK 海力士 HBM 产能需求** | 当前满负荷供给 | 客户订单为供应能力的 **5 至 6 倍** | 供应链下游（如英伟达）面临长期供不应求 |
| **AMD 2030 年 AI 市场预估 (TAM)** | 1200 亿美元（3 个月前） | 上调至 **2200 亿美元** | 严肃公众上市公司 CEO 罕见在短周期内巨幅上修 TAM |
| **云厂商 GPU 推理业务利润率** | 传统云服务水平 | 摩根士丹利测算利润率达 **60% - 80%** | GPU 云并非烧钱无底洞，而是高盈利性业务 |
现场主持人惊叹于苏姿丰的调整幅度：“像她这样严谨、执掌非 meme 类硬科技上市公司数十年的资深商业领袖，绝不可能凭空将预期翻倍，除非她亲眼见证了来自客户极其疯狂的确定性订单。”
---
### [15:01](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=901s) 资本开支恐慌与“收入回报率”的生存之战
面对金融市场上“当前 AI 狂潮宛如 2000 年互联网泡沫重演”的论调，Tae Kim 提供了截然不同的基本面视角。据摩根士丹利估算，超大规模云厂商（Hyperscalers）提供的 GPU 推理业务利润率高达 60% 至 80%。即使算力采购看似庞大，但随着云厂商自身收入维持高速增长——Azure 增速达到 40%、Google Cloud 增速高达 80%、AWS 亦保持强劲双位数扩张——这些由营收转化的巨额经营现金流足以支撑新一轮资本开支（CapEx）。正如亚马逊 CEO Andy Jassy 在 4 月致股东信中所强调的自由现金流逻辑：“我们不会凭空押注 2000 亿美元，我们看到了真实的需求，这项投资在中长期内将创造无可匹敌的自由现金流。”
更关键的驱动力在于企业在智能化浪潮下的生存危机。在 AMD 的大会上，AT&T 高管分享了其实践：他们已经将 100 多个生成式 AI 模型部署到生产线，每月消耗多达 1 万亿个 token，且该规模仍在双位数递增。企业之所以大举采用编码智能体和自动化流程重塑研发，是因为如果一家企业不在研发端进行上百次迭代以提升体验，其竞争对手（如 Verizon 之于 AT&T）就会利用 AI 全面降维打击，夺走全部市场份额。
Tae Kim 总结道：“所有人都在紧盯投资回报率（ROI），但大家更需要关注‘营收回报率（Return on Revenue）’或‘营收生存率’——若你不使用 AI，竞争对手就会用 AI 将你置于死地并抢走所有客户。”
---
### [18:40](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1120s) 企业渗透率鸿沟：从个别试水到百倍爆发
针对 AI 的普及程度，主持人和嘉宾指出，虽然科技行业已经处于高强度的“Token 极限狂欢（Token Maxing）”状态，但全社会企业层面的 AI 渗透其实才刚刚揭开序幕。
旧金山湾区的初创公司圈子充斥着将 AI 融入每一道工作流的“硅谷信息茧房”，但在美国绝大多数传统企业中，大部分员工对 AI 的认知依旧停留在“听说过、有空我也得去瞧瞧”的阶段。嘉宾援引 Jared Sleeper 在 X 平台上的分析以及 Coatue 的研报数据：当前企业间的 AI 采用鸿沟依然极其巨大。如果全行业普通企业能达到头部前沿公司同等的 AI 采用深度，全市场的 AI 调用量将当场爆发 100 倍。
主持人补充道，在 Ramp 的企业消费数据中也可以观察到这种浅层现象：不少公司口中的“采用了 AI”，实际仅仅是为员工报销了一个月费 20 美元的 ChatGPT Pro 账号而已。这与真正运用底层代码智能体（Coding Agents）、深度推理、全自动化研发和工作流重构相比，完全不可同日而语。当前市场仅仅触及了极其浅层的功能，深层生产力的重构才刚开始。
### [19:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1190s) 6万亿美元企业市场与前沿AI模型的指数级增长
在讨论AI如何实际加速工作中的重复性流程以及技术扩散刚刚开启时，对话转向了企业IT市场的庞大规模与前沿模型公司的变现潜力。企业在IT与知识管理方面的年市场规模（TAM）高达约6万亿美元。相比之下，当前两家最核心的前沿大模型公司——OpenAI与Anthropic——合计的年化经常性收入（ARR）即便在经历爆发式增长后，目前也仅在100亿至200亿美元的区间内（转录中提及合计120亿美元水平）。
面对如此庞大的企业市场蛋糕，这两家公司正在以指数级的速度狂飙。从基本的逻辑推导来看，随着技术加速渗透，它们的合并收入在未来一到两年内增长到2000亿、3000亿甚至4000亿美元并非天方夜谭，而是具备坚实的商业支撑。然而，大众与市场往往忽视底层强劲的扩散趋势，反而极易被耸人听闻的媒体头条所误导。
- **企业IT与知识管理年市场规模**：约6万亿美元（$6 trillion）。
- **前沿模型增长轨迹**：OpenAI与Anthropic当前合计ARR仅占TAM极小一部分，未来一至两年具备向数千亿美元体量冲刺的指数级成长空间。
---
### [20:55](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1255s) 破除媒体恐慌：剖析英伟达“500亿美元得州数据中心租赁”真相
针对媒体上引发恐慌的报道，对话重点剖析了《金融时报》（Financial Times）关于英伟达的一则头条新闻。该报道声称英伟达将为得克萨斯州一个耗资500亿美元、使用其芯片的数据中心提供租赁担保（backstop a lease）。这种标题给市场的直观冲击极为负面，甚至让人觉得英伟达在“自买自卖芯片”（进行可疑的内部循环交易）。
然而，只要真正通读文章就会发现事实大相径庭：这是一笔为期15年的租约（根据Hut 8披露，15年期承诺价值约为200亿美元）；只有在英伟达选择在15年后继续续约的情况下，整个30年周期的总价值才会达到500亿美元。折算下来，如果续约，每年涉及的金额仅为10亿到20亿美元左右。更重要的是，英伟达作为租户还可以对外转租并从中赚取利润，绝非每年白白承担数十亿美元亏损。
```mermaid
flowchart TD
accTitle: NVIDIA Texas Data Center Lease Breakdown
accDescr: Diagram illustrating the sensational headline versus the 30-year actual lease breakdown for NVIDIA.
subgraph Media["媒体报道 vs 实际结构"]
H["头条耸人听闻：<br/>500亿美元得州数据中心担保"] --> R["真实条款拆解：<br/>15年基础租约（约200亿美元）"]
R --> E["可选续约：<br/>满30年累计达500亿美元"]
end
subgraph Impact["年度财务影响"]
E --> A["年均开支：仅10亿 - 20亿美元"]
A --> B["商业逻辑：作为租户对外转租盈利"]
A --> C["对比营收：年化3200亿冲刺5000亿<br/>（纯属四舍五入的零头误差）"]
end
```
对于一家当前年化营收运行率（run rate）已达3200亿美元、并有望在下一年迈向4000亿至5000亿美元的巨头而言，每年10亿至20亿美元不过是“四舍五入的零头误差”（rounding error）。媒体为了追求眼球效应，刻意提炼出极其庞大的数字做成标题，从而引发了市场不必要的惊慌。
- **媒体标题与真实条款**：宣称“500亿美元”，实为15年期约200亿美元租赁，仅在满期续约后30年总额才可能达500亿美元。
- **财务占比**：分摊至每年仅10亿至20亿美元，相对英伟达3000亿至5000亿美元的营收体量几乎微不足道。
---
### [22:51](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1371s) 恐慌抛售潮与真实基本面：HBM极度紧缺下的巨头争夺战
随后，两人讨论了《华尔街日报》等媒体关于英伟达正与OpenAI洽谈、可能为软银（SoftBank）高达2500亿美元交易提供担保的传言。尽管两家公司未予置评，且具体细节尚未公开，但市场前一日已产生极大的负面情绪，知名财经主持人吉姆·克莱默（Jim Cramer）甚至在节目中对观众大喊“开盘卖掉一切”，将恐慌情绪推向高潮。
Tae Kim指出，这种情绪化的恐慌脱离了真实的产业基本面：
1. **TAM持续上修**：AMD首席执行官苏姿丰（Lisa Su）在过去短短三个月内大幅调高了其AI潜在市场规模预期。
2. **RSI赛跑**：递归自我改进（RSI, Recursive Self-Improvement）的曙光已经显现，所有顶级AI研究人员都意识到技术突破即将发生，正在争分夺秒地抢占先机。
3. **工作流重构**：智能体AI（Agentic AI）正迫使全球每家企业重塑内部业务流。
最直观的例证莫过于韩国总统近期造访旧金山湾区硅谷时，英伟达CEO黄仁勋（Jensen Huang）、博通CEO陈福阳（Hock Tan）、Anthropic创始人Dario Amodei以及OpenAI的Sam Altman全员火速到场。巨头齐聚的根本动力非常纯粹——他们迫切需要高带宽内存（HBM），而HBM正面临极其严峻的产能短缺与难以置信的旺盛需求。
- **市场情绪反差**：克莱默等名嘴喊出“卖掉一切”，但产业核心指标全线向好。
- **关键瓶颈**：HBM内存处于严重供不应求状态，成为制约各大算力霸主的核心战略物资。
---
### [25:05](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1505s) CUDA护城河真会消亡吗？全栈系统协同与供应链垄断力
围绕AMD在AI发布会上的表态以及月之暗面（Kimi）在技术博客中提到通过大模型编写GPU内核，主持人询问在Agentic AI时代，当AI智能体能够代写底层驱动和内核代码时，CUDA的软件壁垒是否会就此崩塌。
Tae Kim对此给出了冷静的分析。AMD出于商业竞争动机，自然倾向于宣称“CUDA不再是障碍”。然而，CUDA真正的威力在于其经历了“数百万次生产环境调用锤炼出的极端可靠性”（super reliable），所有隐藏的边界缺陷（bug）都已被深度优化。AI生成的代码在理论和局部测试中表现尚可，但在真实、复杂的大规模生产环境中能否稳定运行，仍需打上一个巨大的问号。
更为关键的是，英伟达的真正护城河早已超越了单纯的CUDA软件层：
1. **全栈系统级协同设计（Co-design）**：网络互联、CPU、GPU以及软件栈之间的无缝紧密咬合与超大规模集群能力。
2. **供应链的绝对统治力**：依托强大的资产负债表与充裕的现金储备，英伟达充当行业“大猩猩”，通过巨额预付款（prepay）直接锁死了关键稀缺元器件。从台积电（TSMC）的晶圆代工产能、海量HBM内存，到令光学初创公司倍感绝望的顶尖光学组件供应，英伟达几乎全面完成了独占锁定。这种供应链压制力正是其未来一到两年持续兑现巨额营收的关键底层支撑。
---
### [27:30](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1650s) 能源瓶颈疑云与英伟达隐秘的“万亿美元AI版图”
针对市场上热议的“能源瓶颈会先于芯片瓶颈爆发”的担忧（Energy FUD），黄仁勋在接受彭博社采访时做出了极为耐人寻味的回应。他虽然坦承存在数据中心外壳机房（shell）、电力供应及各类元器件短缺等挑战，但紧接着明确强调：“芯片行业拥有足够的供应链能力，支撑其营收每年翻倍。”这一表态实际上暗示英伟达已经做好了供应链层面的充分储备以对冲能源与物理制约，而当前华尔街分析师对英伟达下一年度的营收预期，远远低于“每年翻倍”的增长速度。
对话最后探讨了英伟达庞大的对外投资生态。英伟达不仅向生态客户供应硬件，还悄然入股了各个垂直领域的关键核心力量——早期投资CoreWeave带来了巨额账面回报，近期又入股了Lumentum和Coherent等光学巨头，并参投了由伊利亚（Ilya Sutskever）创办的SSI（Safe Superintelligence）。
虽然部分市场评论家质疑这是为了拉升出货的“供应商融资”（vendor financing），但其本质逻辑在于：黄仁勋预见到了如海啸般倾泻而来的算力需求，必须利用英伟达的资金实力协助上下游供应链扩充产能、锁定配件。在两到三年后，当市场回看英伟达的资产负债表时，其持有的顶级AI企业股权很可能会膨胀至上万亿美元，进一步巩固早期规模优势构建起的巨大财富滚雪球效应。
### [29:59](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1799s) 英伟达的生态投资战略与黄仁勋的前瞻布局
GPU 云计算业务利润极其丰厚，但承接这类业务的初创及算力服务公司往往面临沉重的资本开支压力，必须筹措巨资来搭建基础设施与组件供应链，才能在未来一到两年内提供充足的 GPU 云服务。黄仁勋敏锐地洞察到了行业供需的这一断层，就像他此前深度支持 CoreWeave 等算力新贵一样。英伟达直接向这些产业链关键企业注资，帮助它们扩充产能以制造行业急需的关键组件与算力设施。再过两三年回头看，英伟达在整个 AI 与云生态中所持有的庞大股权网络，必将再次证明黄仁勋卓越的投资眼光——一如他在过去历次技术周期中所展现出的远见。
```mermaid
flowchart LR
accTitle: 英伟达算力生态投资循环
accDescr: 展示英伟达通过资本投资扶持产业链、扩大云算力供给并最终获得投资与生态双重回报的正向闭环。
subgraph Capital ["资本注入与赋能"]
A["英伟达战略资本"] --> B["新兴基础设施与 GPU 云服务商<br/>(如 CoreWeave 等)"]
end
subgraph Expansion ["产能扩建与交付"]
B --> C["扩充硬件与供应链组件"]
C --> D["满足未来 1-2 年激增的<br/>GPU 云计算需求"]
end
subgraph Return ["生态回报"]
D --> E["持有大量高价值股权<br/>巩固产业护城河"]
end
```
- **资本与供给瓶颈**：AI 云端需求爆发之际，基础设施服务商必须拥有充足资本快速铺设硬件，而英伟达的介入恰好为上下游打通了资金与芯片的供给通道。
- **投资收益闭环**：通过提前布局 CoreWeave 等一系列配套伙伴，英伟达不仅锁定了下游算力的消耗场景，更在未来数年中坐享数倍的股权增值红利。
---
### [30:38](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1838s) 皮夹克拍卖梗、模型“蒸馏”调侃与节目尾声
访谈随即在主持人的捧腹笑声中转向轻松的调侃。主持人拿黄仁勋标志性的行头打趣道：“那个男人能花 5000 美元买一件皮夹克，转头就能卖出 100 万美元，你还要看什么证据证明他的能力？”（"the man can buy a leather jacket for like five grand and sell it for a million dollars"）。现场甚至开玩笑逼问 Tae Kim 是不是那位拍下夹克的神秘竞拍者，Tae 笑着否认后，主持人顺着 AI 行业的技术黑话接梗：“现在真正的问题是，还要多久才会有人把老黄的皮夹克给‘蒸馏’（distill）并开源出来，好让我们只花两块钱就能买到一件高仿皮衣？”
在一片欢声笑语中，主持团队向 Tae Kim 表达了由衷的感谢，感谢他包容全场飞舞的各种段子与玩笑。随着大盘近期的剧烈波动，大家纷纷祈愿这期干货满满又充满欢乐的对谈能成为金融市场的“幸运符”（"lucky charm for the markets"），并在“市场大底已经出现”（"Bottom is in"）的共鸣与道别声中圆满收尾。
## Notable quotes
> [02:45](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=165s) properly yolo in the AI era — SemiAnalysis / 主持人引用
> [05:50](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=350s) AI has accelerated everything — 主持人
> [10:25](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=625s) 未来 6 个月 AI 能力的飞跃程度，将戏剧性地超越过去两年的总和。 — Sam Altman / Tae Kim 引用
> [13:08](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=788s) 像她这样严谨、执掌非 meme 类硬科技上市公司数十年的资深商业领袖，绝不可能凭空将预期翻倍，除非她亲眼见证了来自客户极其疯狂的确定性订单。 — 主持人
> [15:01](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=901s) 我们不会凭空押注 2000 亿美元，我们看到了真实的需求，这项投资在中长期内将创造无可匹敌的自由现金流。 — Andy Jassy / Tae Kim 引用
> [15:01](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=901s) 所有人都在紧盯投资回报率（ROI），但大家更需要关注‘营收回报率（Return on Revenue）’或‘营收生存率’——若你不使用 AI，竞争对手就会用 AI 将你置于死地并抢走所有客户。 — Tae Kim
> [22:51](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1371s) 开盘卖掉一切 — Jim Cramer / Tae Kim 引用
> [27:30](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1650s) 芯片行业拥有足够的供应链能力，支撑其营收每年翻倍。 — Jensen Huang / Tae Kim 引用
> [30:38](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1838s) the man can buy a leather jacket for like five grand and sell it for a million dollars — 主持人
> [30:38](https://www.youtube.com/watch?v=PAUXPWUDxb4&t=1838s) 现在真正的问题是，还要多久才会有人把老黄的皮夹克给‘蒸馏’（distill）并开源出来，好让我们只花两块钱就能买到一件高仿皮衣？ — 主持人
## Mentioned
Tae Kim, NVIDIA, Mark Zuckerberg, Meta, Andrew Bosworth (Boz), Alexandre Wang, SemiAnalysis, OpenAI, Anthropic, Dario Amodei, Google, Gemma, Llama, Moonshot AI (月之暗面), Kimi, DeepSeek, Sam Altman, SK Hynix (SK 海力士), AMD, Lisa Su (苏姿丰), Microsoft Azure, Google Cloud, AWS, Andy Jassy, AT&T, Verizon, Jared Sleeper, Coatue, Ramp, ChatGPT Pro, Financial Times, Hut 8, Jim Cramer, Broadcom (博通), Hock Tan (陈福阳), TSMC (台积电), CoreWeave, Lumentum, Coherent, Safe Superintelligence (SSI), Ilya Sutskever, CUDA
## Video description
This is our full interview with Tae Kim.
We discussed why the AI trade is being misunderstood, why Kimi and better open-weight models could increase compute demand instead of reducing it, Meta's AI spending, NVIDIA's open-weight strategy, the race toward RSI, why data center fears are overblown, AMD's expanding AI market forecast, the future of agentic AI, NVIDIA's CUDA moat, and much more.
TBPN is made possible by:
Ramp - https://ramp.com
Public - https://public.com
Cisco - https://www.cisco.com
Console - https://www.console.com
CrowdStrike - https://www.crowdstrike.com
Figma - https://www.figma.com
MongoDB - https://www.mongodb.com
NYSE - https://www.nyse.com
Railway - https://railway.com
Shopify - https://www.shopify.com/
Codex - http://openAI.com/codex
Sign up for TBPN’s daily newsletter at TBPN.com
Follow TBPN:
https://TBPN.com
https://x.com/tbpn
https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231
https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235
https://www.youtube.com/@TBPNLive
[Source](https://www.youtube.com/watch?v=PAUXPWUDxb4)
Source quotes
[2:45] SemiAnalysis / 主持人引用: properly yolo in the AI era
[5:50] 主持人: AI has accelerated everything
[10:25] Sam Altman / Tae Kim 引用: 未来 6 个月 AI 能力的飞跃程度，将戏剧性地超越过去两年的总和。
[13:08] 主持人: 像她这样严谨、执掌非 meme 类硬科技上市公司数十年的资深商业领袖，绝不可能凭空将预期翻倍，除非她亲眼见证了来自客户极其疯狂的确定性订单。
[15:01] Andy Jassy / Tae Kim 引用: 我们不会凭空押注 2000 亿美元，我们看到了真实的需求，这项投资在中长期内将创造无可匹敌的自由现金流。
[15:01] Tae Kim: 所有人都在紧盯投资回报率（ROI），但大家更需要关注‘营收回报率（Return on Revenue）’或‘营收生存率’——若你不使用 AI，竞争对手就会用 AI 将你置于死地并抢走所有客户。
[22:51] Jim Cramer / Tae Kim 引用: 开盘卖掉一切
[27:30] Jensen Huang / Tae Kim 引用: 芯片行业拥有足够的供应链能力，支撑其营收每年翻倍。
[30:38] 主持人: the man can buy a leather jacket for like five grand and sell it for a million dollars
[30:38] 主持人: 现在真正的问题是，还要多久才会有人把老黄的皮夹克给‘蒸馏’（distill）并开源出来，好让我们只花两块钱就能买到一件高仿皮衣？
Source quotes
[10:25] Tae Kim: When you have more capable models that come out, people find uses for them... the next six to nine months are going to be bigger than anyone believes.
[16:36] Tae Kim: Everyone's worried about ROI. ROI is important, but you also need return on revenue because if you don't use AI, your rival is going to use AI to beat you in the market.
[20:56] Tae Kim: They just wanted to say the biggest number... It's a rounding error. They're doing a $320 billion run rate a year... This is how people run with the sensationalized headlines and people panic and freak out.
[25:05] Tae Kim: Jensen said this last week... 'I think the chip industry has enough supply to double their revenue every year.' Basically implying Nvidia has enough supply for energy and all that stuff. No one is pricing that in.
[28:25] Tae Kim: In a year, two, three years, Nvidia is going to have all these stakes in these companies and it's going to look like he was a good investor... because he has been in the past.
[28:25] Host: The man can buy a leather jacket for like five grand and sell it for a million dollars. I don't know what else you need to see.
Original transcript
[0:00] We have Tay Kim in the waiting room.
[0:01] Let's bring him in to the TVP Ultra
[0:04] Dome. Tay, how you doing?
[0:06] >> Hey guys, doing great. What's going on?
[0:09] >> So, so tell me, last time you were on
[0:10] the show, you bottom ticked it. What's
[0:12] going on?
[0:13] >> Uh, I think I made the bullish call on
[0:16] CPUs, memory, and Nvidia. Nvidia is up
[0:20] like 5 10%, but nice.
[0:21] >> Uh, the CPU names have still doubled
[0:23] even after this big draw down, and the
[0:25] HPM names are up 100%. So, I I'm hoping
[0:29] that, you know, it's the same thing
[0:31] again. I come on here and stocks go up
[0:34] again.
[0:34] >> Yeah. Ideally, we could have like an
[0:36] emergency reserve of take appearances.
[0:38] So, if the market is ever down, reserve,
[0:41] the strategic reserve, we call you up,
[0:42] you jump on and then it's
[0:44] >> it was funny because it was literally
[0:46] the exact bottom and you know, it went
[0:48] exponential after that.
[0:50] >> Take Kim effect. So where are we right
[0:53] now with the level of FUD, the level of
[0:56] uh downward pressure on the AI trade
[0:59] broadly, the chips, the semi-de like
[1:01] reset for us on like where sentiment is
[1:03] and then we can work through the
[1:05] different pieces of counter examples.
[1:08] So, I think sentiment's very negative.
[1:10] We kind of had this
[1:12] >> huge up parabolic up move the last few
[1:14] months and likely a lot of retail and
[1:17] hedge funds piled in
[1:19] >> and we're seeing this unwind down. I I
[1:21] think the first a big part of it was the
[1:23] Iran war getting worse. uh every time
[1:26] you we had the
[1:28] >> the first ceasefire negotiations stock
[1:31] started taking off right right after
[1:32] that and then when you we had the actual
[1:34] ceasefire uh we had a follow through and
[1:37] then as soon as Trump started bombing
[1:39] Iran again you know chip stocks have
[1:42] kind of plummeted in the last two three
[1:43] weeks uh and then now we we're seeing
[1:45] just you know back to the old uh pattern
[1:48] of uh media and the viral hot takes uh
[1:52] spreading a lot of FUD uh I think we saw
[1:54] earlier earlier this month. Um I think
[1:57] uh Reuters quoted like Zuckerberg's
[2:00] about a genetic AI. They took it out of
[2:02] context and then every media person was
[2:06] with a hot take that this met Meta was
[2:08] seeing bad returns and they're going to
[2:10] cut capex and then we had leaks right
[2:12] after that saying that it looks like
[2:14] Meta is going to raise capex. So we're
[2:16] seeing a lot of this hot take FUD. Uh
[2:19] yesterday I think we had a flurry of
[2:21] stuff that scared people. um the Wall
[2:23] Street Journal vendor financing uh
[2:25] article that we'll see we'll see what
[2:27] happens with that. We had the CMXT IPO
[2:31] in China and everyone freaked out over
[2:33] that. We had the information article on
[2:35] ASML. Uh we could go through each one
[2:37] and then the Kim thing it's obviously a
[2:39] big thing.
[2:39] >> Yeah, we'll definitely get there and I
[2:41] want to talk about open source and
[2:43] Nvidia's strategy there obviously. Uh
[2:45] starting with the Mark Zuckerberg news
[2:47] in Reuters. This was July 2nd. Meta's
[2:50] Zuckerberg says, "AI agent tech
[2:52] progressing slower than expected." Um,
[2:55] Zuckerberg added that the company's
[2:56] reorganization that included major job
[2:58] cuts was not as clean as it could have
[3:00] been. Zuckerberg and other meta
[3:02] executives have been seeking uh to
[3:04] moderate some of the c of the
[3:05] organizational changes introduced this
[3:07] year. And they said that uh the
[3:10] trajectory of agentic development over
[3:12] the last four months hasn't really
[3:15] accelerated in the way we expected. uh
[3:17] the company's bets on new structure
[3:19] haven't come to fruition yet and so
[3:22] people were sort of reading this as
[3:24] maybe Meta is going to pull back but
[3:26] then it felt like the response was
[3:29] extremely quick with with Bos going on a
[3:32] podcast and Alex Wang sharing a whole
[3:35] bunch of progress across um a few
[3:38] different uh models and data points and
[3:41] then semi analysis wrote a whole
[3:42] bullcase for MSL talking about how they
[3:46] have compute and also So they have uh
[3:48] more of like the internal structural
[3:50] alignment to sort of properly yolo in
[3:53] the AI era if I'm boiling it down as
[3:56] brutally as possible just because with
[3:59] Google there's always this debate
[4:01] between oh do you sell the TPUs or do
[4:02] you sell the cloud comp do you have
[4:04] bended in the product whereas Mark
[4:06] Zuckerberg's able to sort of like go all
[4:08] in on this new idea and so maybe there's
[4:11] more more glimmers of hope there but uh
[4:14] what else have you been tracking
[4:15] downstream of meta those ambitions.
[4:18] >> Well, I mean, they've been very upfront
[4:20] that they're investing heavily in AI.
[4:22] Alexander Wang is tweeting multiple
[4:24] times every few weeks that they're
[4:26] they're going full force. They're going
[4:28] to redo open source AI models. They I
[4:30] think he said that the YC event over the
[4:33] weekend
[4:34] >> and it's I mean, if you actually look at
[4:36] and then Reuters came out, I think with
[4:38] an article saying that they're actually
[4:39] going to raise capex dramatically uh
[4:42] this year and next year. So all that
[4:44] kind of fear that that quote about
[4:47] authentic AI from the town hall um that
[4:49] kind of like spooked the market for a
[4:51] few days, it kind of it was completely
[4:53] false. um the stuff like it's a yeah it
[4:56] feels like it's a comm's air because the
[4:59] the language that's been coming out of
[5:01] meta has been a little bit like AI is
[5:04] going to replace our employees and it
[5:06] feels like it'd be much better for them
[5:08] to to come to the market with a message
[5:11] of we're going on the offensive like
[5:13] we're a hyperscaler
[5:14] >> to be to be fair that that was an
[5:16] internal town hall they didn't mean to
[5:18] leak it and Reuters leaked that one
[5:20] quote and put out put out the headline
[5:22] before the article it's interesting like
[5:24] Meta did Meta basically go through like
[5:26] an 8-year period where like internal
[5:28] town halls didn't instantly leak.
[5:30] >> I think everything leaked always. I
[5:31] think I think everything's been
[5:33] >> I know. But there was there was a period
[5:34] where like the the the the sort of
[5:36] attention of the media was way way way
[5:39] less on like what Meta was doing
[5:41] internally relative to the 2010s
[5:45] >> and all that attention just went to the
[5:47] labs, right?
[5:48] >> Yeah. Yeah. Yeah. No, that makes sense.
[5:50] Um yeah I I guess the question is like
[5:54] the question that I keep coming back to
[5:57] is like where is their revenue ramp?
[5:59] Where is their AI revenue going to ramp
[6:02] and when? Right? Because as you
[6:05] >> super would say ads like the ads like
[6:07] the AI has
[6:08] >> they're accelerating everything.
[6:10] >> That's always been my view too. But when
[6:12] you're
[6:13] >> when you're continuing to ramp capex
[6:16] >> Yeah.
[6:16] >> with and saying like we're going all in
[6:18] on Agentic and we're building a harness
[6:21] and we're also going to do open source
[6:23] and it's like well what is the strategy?
[6:25] >> Sure.
[6:26] >> Like what's the breakthrough? What is
[6:28] going to take you to a billion dollars
[6:30] of like pure AI product revenue or or or
[6:34] just API revenue? What's going to and
[6:36] then to five and 10
[6:39] >> and what's going to allow you to like
[6:40] justify the spend other than I think the
[6:43] market would love if they just said,
[6:44] "Yeah, we actually need all these GPUs
[6:46] because we can actually be 10 time we're
[6:48] already good at ads. We could be 10
[6:49] times better and that's where we're
[6:51] going to get the the ROI on all this
[6:53] capex."
[6:54] >> Well, well, they're definitely getting
[6:55] ROI on that. the market is worried
[6:57] about, you know, all this extra capbacks
[6:59] on the they're going for the frontier AI
[7:01] model race again
[7:03] >> and they had to reset like they had to a
[7:06] lot of people left and now Wang hired a
[7:08] ton of people and uh we'll see what
[7:10] happens over the next it's going to take
[7:12] time. It's going to take six to 12
[7:13] months before we see any more progress.
[7:15] But that model that came out a few weeks
[7:18] ago um was a lot better than people
[7:19] expected. It wasn't
[7:21] >> you know the frontier but it was much
[7:23] better than what people expected.
[7:24] >> Yeah. Yeah. Uh so uh how have you been
[7:28] processing the NVIDIA letter around uh
[7:33] open source and all the back and forth
[7:36] all the people jumping on the companies
[7:39] that have been staying back how you work
[7:41] through that?
[7:41] >> It's been very impressive what they've
[7:43] been they basically united the entire
[7:46] tech industry against Enthropic in in
[7:48] the last like three four days. 18
[7:50] trillion in market cap has signed on
[7:52] last time I checked across uh
[7:55] >> Google took a little time. Amazon signed
[7:57] on eventually. I mean the only
[7:59] >> Yeah, they signed on yesterday. They
[8:00] tweeted out. Interesting.
[8:02] >> I think Apple is still the hold out
[8:05] which is kind of strange because they're
[8:06] the one that would most benefit from
[8:09] open source open weight models being you
[8:12] know more available I would think. But I
[8:14] would I don't know what Apple,
[8:16] >> but I mean they pretty much got the
[8:18] whole tech industry to uh kind of corner
[8:21] anthropic in their position. Yeah.
[8:23] >> Um
[8:24] >> Open AI signed on.
[8:25] >> Yeah. What did you think of?
[8:27] >> Obviously, Nvidia is afraid.
[8:28] >> Yeah. I don't know. I don't know how how
[8:31] if I don't know if they're really
[8:36] >> I don't read I don't read it as being
[8:37] like cornered by any means.
[8:40] >> Right. Well, Jensen is on the record
[8:42] that, you know, he said, I think to
[8:44] Bloomberg that there was rising
[8:46] sentiment that something was going to
[8:47] happen on the on the regulation front at
[8:51] White House or whatever. So, this we saw
[8:53] that this was this was last week. You
[8:55] had at least four people in the admin
[8:59] >> say, "We're not against open weights.
[9:01] We're against distillation." And at
[9:03] least I was reading into that of some
[9:05] type of regulatory action around open
[9:07] weights and and then positioning it as
[9:10] we're targeting this is this is like
[9:13] wrote yesterday. Yeah. Yeah.
[9:15] >> Um about you know push back and
[9:17] restrictions and he's done doing it
[9:19] under the safety umbrella but u
[9:22] definitely Microsoft and video are
[9:24] worried that uh what the white house or
[9:27] congress is going to do something on
[9:28] this front and that's why
[9:30] >> they uh took it seems very reasonable
[9:32] that he would know have no problem with
[9:33] like Gemma or Llama or any of the open
[9:36] source from like American hyperscalers
[9:38] where if you find out that they're
[9:39] distilling you just walk across the
[9:41] street and sue them. Uh, and also these
[9:44] big companies have huge huge I mean they
[9:47] have safety teams but also just like
[9:48] huge incentives to not have a safety
[9:50] incident happen on their watch because
[9:52] you're trying to like catch up to the
[9:54] frontier and then all of a sudden you
[9:55] have a safety incident that's going to
[9:57] be really bad for your overall brand uh
[9:59] and you have a different business to
[10:01] protect whether it's social networking
[10:04] or uh Google search if all of a sudden
[10:06] the Gemma model winds up being a thorn
[10:09] in someone's side for a cyber security
[10:11] reason or a bio reason uh that would be
[10:13] uh really really bad. But uh a foreign
[10:16] company that is just like hurling it
[10:18] over here can kind of just be like you
[10:20] guys deal with the consequences
[10:21] potentially. So I think that's what what
[10:22] what Dario is worried about. Uh what
[10:25] about um the overall idea of like we're
[10:29] we're it feels like we're sort of
[10:31] replaying the deepseek moment. Open
[10:33] source is going to reduce cost and so
[10:36] that's a reason to pull back on the AI
[10:38] trade overall. How have you processed
[10:39] that? It's almost it's almost a perfect
[10:41] catalog. Uh people are worried about
[10:43] Kimmy. Yeah. But when you actually read
[10:45] the technical paper and their blog
[10:47] posts, uh this is not a tiny efficient
[10:50] model. This is 2.8 trillion parameters.
[10:52] It's going to require a ton of compute
[10:54] to serve. I mean, we saw the first day
[10:56] they put it out that their servers got
[10:58] slammed.
[10:59] >> Um even in the blog post, they say it's
[11:01] best run on a a kind of a a server with
[11:05] 64 GPUs. So big super clusters that
[11:08] areworked well and and that's perfectly
[11:10] runs great on Nvidia. And if you
[11:12] remember, you know, during the whole
[11:13] Deep Seek thing about a year a year or
[11:15] so ago, uh the market freaked out that,
[11:18] you know, Deepseek was going to was so
[11:20] efficient that it will lead to a compute
[11:21] glut. But Deepseek was an example of the
[11:25] reasoning model that actually it was the
[11:26] opposite. It created a ton of demand.
[11:29] And I think the same thing is going to
[11:30] happen with Kimmy where uh when you have
[11:32] more capable models that come out,
[11:35] people find uses for them. And right
[11:37] now, just like last year when reasoning
[11:39] models took off, Agentic AI and agents
[11:42] are taking off right now. And the market
[11:45] is kind of like not realizing that
[11:47] because right now, just like last year
[11:49] when reasoning models were taking off,
[11:50] right now Agentic AI is taking off and
[11:53] the next six nine months are going to be
[11:55] bigger than anyone believes. Sam is on
[11:58] the record. Sam is on the record over
[12:01] the weekend saying
[12:04] >> at the YC thing again like people don't
[12:06] I don't know why people don't like
[12:07] listen to the it's on YouTube. Yeah.
[12:09] >> That the next 6 months it's going to be
[12:11] much more dramatically better for AI
[12:14] than the last two years in terms of
[12:15] advances and capabilities. And I I heard
[12:17] you say RSI before. I think it's going
[12:19] to be RSI. Uh people inside OpenAI and
[12:23] definitely Anthropic. anthropic put on
[12:26] this RSI I think is a lot closer than
[12:28] people think and if RSI actually happens
[12:30] in the next 369 months that's going to
[12:33] soak up insane amount of I mean we have
[12:36] this exponential ramp for reasoning
[12:38] exponential ramp for aentic and then if
[12:40] RSI actually happens and I think it it
[12:43] sounds like both frontier labs think
[12:45] it's going to happen very soon that's
[12:46] going to soak up an unbelievable amount
[12:48] of compute as the AI models you know use
[12:52] more compute to to self-develop and
[12:54] improve and I think that's one thing
[12:56] that missing
[12:58] >> that both anthropic and openi are kind
[13:00] of winking that oh it's happening
[13:01] anytime I tweet something on RSI all
[13:04] these frontier AI speakers like my tweet
[13:06] so I think that's good
[13:08] >> uh what what is your what is your sort
[13:10] of framework around compute hoarding
[13:14] because certainly certainly it is it has
[13:17] been happening when you look at when you
[13:19] look at you know uh like going back to
[13:22] the meta example Right? They're not
[13:24] selling compute yet. They're maybe
[13:27] curious about it or or exploring some
[13:29] deals, but they have all this computing
[13:32] on their own ability to create the
[13:34] capability that will have enough demand
[13:36] to uh justify justify that. Do you just
[13:41] think it there there's so much demand
[13:43] overall that it just, you know, even if
[13:45] there's there's hoarding, it just will
[13:46] leak out
[13:48] >> and it's okay.
[13:49] >> Well, there's so much demand overall. I
[13:50] mean, the SKH Highex executives said uh
[13:54] during their IPO run um that their
[13:57] customers are asking five to six times
[13:59] more than they're able to serve and
[14:01] they're going to double capacity over
[14:02] the next five years. They said, and
[14:04] their customers, and I'm going to assume
[14:06] it sounded like Jensen, um are asking
[14:09] for five to six times more than they're
[14:11] they're able to build. So, there's
[14:13] overwhelming demand. You guys were at
[14:15] the advanced AI uh AMD event. Lisa Sue
[14:19] raised her CPU, a Genentic CPU forecast.
[14:23] Just three months ago, it was 120
[14:25] billion for 2030. Three months later,
[14:27] they raised it to 220 billion.
[14:29] >> Yeah.
[14:30] >> Like, she doesn't do that. She doesn't
[14:32] do that.
[14:32] >> You have that just on the
[14:34] >> I've got that ready. I can do whatever.
[14:36] >> I mean, like,
[14:38] >> well, I just love this chart because he
[14:40] called it perfectly.
[14:42] >> He actually did. It's crazy.
[14:44] CEOs don't don't you know raise their
[14:47] tams by like these multiples in a few
[14:50] months if they're not seeing insane
[14:52] demand coming in
[14:53] >> especially not public CEOs who are
[14:55] serious business leaders who've been
[14:57] running like non-me stocks for decades
[15:00] and are like serious
[15:01] >> everyone's freaking out that this is
[15:02] like the dotcom bubble all over again
[15:05] finance but what if these hyperscaler
[15:08] GPU cloud businesses are amazing
[15:10] businesses like Morgan Stanley says if
[15:13] you inference, it's 60 to 80% profit
[15:16] margins, right? These are amazingly
[15:18] profitable businesses as long as we keep
[15:20] growing the next few years. And and
[15:22] again, just like last year, we're we're
[15:24] on this exponential run right now over
[15:27] the next two quarters. And the market
[15:29] isn't seeing that. Everyone's freaking
[15:31] out that oh no, we're spending too much.
[15:33] And even uh Sam Alman uh podcast came
[15:36] out today and another podcast.
[15:39] >> Sam is out there.
[15:40] >> He said that he regretted you know,
[15:43] pulling back on the compute purchases.
[15:45] They made a mistake by uh not putting
[15:47] the pedal to the metal because now
[15:49] things are taking off again. So like I
[15:53] like Amazon the CEO in April if you
[15:57] everyone read his annual letter Andy
[15:59] Jessi wrote he talks about how free cash
[16:01] flow works. We're not betting $200
[16:03] billion on a hunch. We see the demand.
[16:06] We know it's going to be insanely
[16:07] profitable and free cash flow positive
[16:10] in the medium to long term. So that's
[16:12] why you're investing $200 billion now
[16:14] and in a year or two we're going to see
[16:16] insane amounts of free cash flow. The
[16:18] thing that people are worried about
[16:20] right now, it takes time to build out
[16:22] these data centers and fabs. And you bet
[16:24] now hold
[16:27] like if you if you see uh if you see
[16:30] free cash flow that that that assumes
[16:32] that uh like like the revenues have to
[16:36] catch up and then the capex can't grow
[16:38] more exponentially. And so that means
[16:40] you have to see some sort of plateauing.
[16:43] Maybe it's at the end of the chart.
[16:44] Maybe it's this 2030 range. But there is
[16:46] a different world of just like continued
[16:48] growth forever and then we sort of like
[16:50] run out of money.
[16:51] >> The push back I have there is that's a
[16:53] static view, right? If they don't grow
[16:55] revenue
[16:57] >> for the next three years, yes, you can't
[16:59] do that. But they're growing Azure is
[17:02] growing 40%. Google Cloud is growing
[17:04] 80%.
[17:05] >> Yeah.
[17:05] >> You know, Amazon's growing
[17:07] >> high uh double digits. So if if if
[17:11] revenue is growing 40 you to 80% this
[17:15] year or next year and the year after
[17:17] that's more revenue you have that's more
[17:19] operating cash flow you have to invest
[17:21] >> right
[17:22] >> so that's that's what people are missing
[17:25] and if this stuff if if the data center
[17:27] you're building now you're spending all
[17:29] this now
[17:30] >> generates unbelievable free cash flow in
[17:33] 12 to 18 months because you know this
[17:35] agent AI is actually changing and
[17:38] rearchitecting ing all the workflows
[17:40] inside companies and you need to do the
[17:42] agentic AI coding agents to make your
[17:45] product better because if you don't if
[17:47] you don't iterate 100 different
[17:49] iterations of your product in R&D
[17:51] >> if you don't do AI just like AT&T is
[17:54] doing at the Gent uh advancing AI at AMD
[17:57] he's talked about they're putting a 100
[17:59] Gen AI models into production they're
[18:01] burning a trillion tokens a month and
[18:04] then that's growing double digits the
[18:06] reason why they're doing that is because
[18:08] you you By using aentic AI, you're
[18:10] providing a better customer service, a
[18:12] better you have a better product R&D,
[18:14] and you're helping your companies make
[18:16] better products and services. And if you
[18:18] don't incorporate AI into your company,
[18:22] >> Verizon, your other company is going to
[18:24] do is going to incorporate AI and then
[18:26] disrupt you and then you lose all your
[18:28] revenue. Yeah.
[18:29] >> So, everyone's worried about ROI. ROI is
[18:31] important, but you also need return on
[18:33] revenue because uh if you don't use AI,
[18:37] your your rival is going to use AI to
[18:39] beat you in the market.
[18:40] >> Yeah. Yeah. No, I I think the diffusion
[18:41] story is still even though we got like
[18:44] sort of jitters by the token maxing
[18:46] thing, just the actual usage of AI
[18:49] across companies is still pretty limited
[18:51] in terms of the amount of people that
[18:53] are using it, the time that those people
[18:55] are using it. Like there definitely is a
[18:57] San Francisco bubble of startups where
[18:59] everyone is using AI a lot. But if you
[19:02] just walk into a normal business, a lot
[19:05] of people are like, "Yeah, I got to
[19:06] check that out." Which is a
[19:08] >> Let me give you some some context here.
[19:10] Some
[19:10] >> Arazian uh Jared Sleeper over on X was
[19:13] saying uh enterprise adoption disparity
[19:16] remains enormous. and he cited Ara
[19:20] saying uh usage would 100x if every
[19:22] company adopted AI to the degree of the
[19:25] most advanced companies. There's like
[19:26] this there's this small group of
[19:28] companies that are
[19:28] >> people forget in the ramp in the ramp
[19:31] data like adopting AI can mean like
[19:33] having a chatbt pro account for someone
[19:35] which is like not exactly the same as
[19:38] like using codecs and like coding agents
[19:40] and stuff like it's important I think
[19:41] that you know if I have someone on my
[19:43] team I want them to be able to go and do
[19:45] a deep research report but uh that's
[19:48] like table stakes the question is like
[19:50] are you actually speeding up anything
[19:52] that's repetitive in your job and uh
[19:54] that diffusion is just starting to take
[19:56] hold.
[19:57] >> So, so the total market size in terms of
[19:59] IT and knowledge management in
[20:02] corporations, it's about $6 trillion,
[20:05] right, a year.
[20:06] >> Uh the two main frontier AI model
[20:08] companies, Open AI and anthropic, I'm
[20:11] going to say I think this is roughly
[20:12] accurate are doing $120 billion combined
[20:15] in ARR.
[20:16] >> Yeah.
[20:16] >> You know, why can't that go to 200, 300,
[20:20] 400 billion in the next year or two? I
[20:22] mean, they're they're growing at
[20:23] exponential rates. Yeah. when we're
[20:25] taking off and the if the market is $6
[20:27] trillion, right? Why can't they grow to
[20:30] 200, 300, 400 billion in the next couple
[20:32] years? I mean, it's like just do a
[20:35] little logic and rational deduction.
[20:37] >> Yeah.
[20:38] >> Um this is definitely possible and it's
[20:40] happening right now and it's
[20:42] accelerating and people aren't, you
[20:44] know, they just taking, you know, these
[20:46] big headlines where we had this uh you
[20:48] know, $50 billion from Financial Times
[20:51] and we find out it's over 30 years. It's
[20:53] like on the homepage.
[20:55] >> Yeah. Yeah. Okay. I wanted to ask you
[20:56] about this. Nvidia revealed this tenant
[20:58] for $50 billion data center that will
[21:00] use its chips. Explain what is actually
[21:02] going on here.
[21:03] >> So, the Financial Times, you know, put
[21:05] on their homepage today that Nvidia is
[21:08] going to backs stop a lease for a data
[21:10] center in Texas for $50 billion. And I
[21:13] saw that I was like, "Oh my gosh." Oh,
[21:15] that doesn't sound good.
[21:17] >> No, it literally sounds like they're
[21:19] buying their own chips. It's like it
[21:21] sounds like the most bad thing you could
[21:23] do.
[21:23] >> Yeah. Then they actually read the
[21:25] article, like halfway down the article,
[21:26] it's like a 15-year lease. Uh, and it's
[21:30] only $50 billion if they renew the lease
[21:33] after 15 years. So, it's like over 30
[21:36] years if they renew it. Then, if you
[21:38] think about that, you're like, "Wait a
[21:39] minute. 50 billion divide by 30 if they
[21:41] renew it." That's
[21:43] >> okay. Yeah. Nvidia's 15year lease
[21:45] commitment for the Texas site is worth
[21:47] basically 20 billion. and renewal
[21:49] options would take the total value to 50
[21:52] billion over 30 years according to HUT
[21:55] 8.
[21:55] >> Okay. But what do you think their what
[21:58] what are their plans for the site? Is
[22:00] this they are going to have some like
[22:04] what do you expect them
[22:05] >> so so my my point is this is a billion
[22:08] you know whatever a billion or$2 billion
[22:10] a year right it's a non-story but it's a
[22:13] it's a big headline sensational headline
[22:15] on the homepage.
[22:16] >> Yeah. Yeah. And also and also it's not
[22:18] that it's it's not like you're taking a
[22:20] $2 billion loss every year. It's you are
[22:24] the tenant and then you are also renting
[22:25] that out. So hopefully you're making
[22:26] profit.
[22:27] >> Yeah. It it's a rounding error. It's
[22:29] like, you know, they're doing 320
[22:32] billion run rate a year. Now that's
[22:34] going to go to 400 500 billion next
[22:36] year. And we're talking about something
[22:37] that might be a billion. You know, like
[22:40] this is not a story, but you this is how
[22:43] people run with the sensationalized
[22:45] headlines and people panic and freak
[22:48] out.
[22:48] >> I think they just wanted to say the
[22:49] biggest number.
[22:51] >> That's exactly the point. And and we're
[22:53] going to see what happens with this Wall
[22:54] Street Journal article. Um both OpenAI
[22:57] and Nvidia are not commenting so far.
[23:00] >> Uh we'll see. But take us through the
[23:01] rumor.
[23:02] >> Have to wait.
[23:02] >> Rumor will tell.
[23:03] >> Well, it's not a rumor. It's the Wall
[23:04] Street Journal and other uh people
[23:06] reporting that Yeah. Nvidia is in talks
[23:09] with open AAI to backs stop soft bank up
[23:12] to 250 billion you know billion we don't
[23:15] know the details and I don't want to
[23:17] speculate and comment but let's actually
[23:19] see the details before we I I think the
[23:21] market had a really big negative
[23:23] reaction yesterday to the story uh
[23:25] because everyone I mean Jim Kramer was
[23:27] telling his audience like sell
[23:29] everything at the open today because AI
[23:32] and data fenders.com
[23:35] you know it was insane it's just let Let
[23:37] let's see the actual deal and the
[23:40] metrics and the numbers before we panic
[23:43] and freak out.
[23:44] >> Yeah. Yeah. Yeah. Yeah. That makes
[23:46] sense.
[23:47] >> So, honestly, when you say freak out and
[23:50] sell everything, sell your dollars, sell
[23:52] your h sell your house, sell your
[23:54] stocks, then I'll freak out.
[23:56] >> But until then, Taye, I feel I feel
[23:59] okay.
[23:59] >> I I mean, I just see the fundamentals. I
[24:01] see the CEO of AMD expanding her TAM you
[24:05] know dramatically in just over the last
[24:07] 3 months. I see RSI under horizon like
[24:10] every AI researchers like oh my god this
[24:12] is going to happen we have to get there
[24:14] sooner and then I see you know the
[24:17] obvious use case of uh Aentic AI where
[24:21] you have to rearchitect your workflows
[24:23] internally. Every company has to do
[24:25] this. So everything is taking off. You
[24:28] see like um you see uh when the
[24:31] president of Korea came to San Francisco
[24:33] area last week, you know, they had like
[24:35] a day in the valley instantly. Nvidia
[24:40] CEO Jensen Huang, Broadcom CEO Hawkan,
[24:43] Dario, you know, Sam Alman are there,
[24:46] right?
[24:48] >> Yeah. Do a little logic deduction. Why?
[24:49] Why are they there like crazy? Because
[24:52] they need HBM memory and they're dying
[24:54] to have it. So if you think about that,
[24:57] that means there's insane demand and HPM
[24:59] memory is in in
[25:01] >> in shortage and there's tremendous
[25:03] demand for it, right?
[25:05] >> Talk about the Oh, sorry. Talk about the
[25:07] Nvidia CUDA mode. It feels like uh a big
[25:11] piece of AMD's advanced AI event was uh
[25:16] maybe the CUDA mode isn't as much of an
[25:18] issue anymore in the age of agentic AI.
[25:21] you can have an AI agent write you the
[25:24] software that you need to use any chip
[25:27] and that creates uh less pricing power
[25:29] for Nvidia, but there's another world
[25:32] where you're not really
[25:35] like Nvidia doesn't necessarily need a
[25:37] moat because everything's just growing
[25:38] so fast that they're still growing. But
[25:40] how have you interpreted the processing
[25:41] of like the potential death of the CUDA
[25:44] mode?
[25:44] >> So AMD is on it. Kimmy wrote like a
[25:48] couple paragraphs in their blog post
[25:49] about how they created a GPU kernel, all
[25:52] that. So, everyone, you know,
[25:53] >> sure,
[25:54] >> get scared or whatever.
[25:55] >> It it's like we'll see what it's like in
[25:58] real life. You know, this is just, you
[26:00] know, AMD is incentivized to say, "Oh,
[26:02] Kudo is not a problem anymore."
[26:04] >> CUDA's been a tremendous mo and I think
[26:07] it continues to be a moat. And the
[26:08] reason why is uh it's super reliable.
[26:12] all the bugs have been optimized and
[26:14] fixed and that comes from hitting the
[26:16] software you know millions of times and
[26:19] millions of times right like you don't
[26:20] know if you know if you use cloud code
[26:23] or Kimmy that what they figure out using
[26:26] their training data is going to work in
[26:29] the real world right they could talk
[26:30] about one little piece that does well
[26:32] let let's see how it actually works yeah
[26:34] but
[26:36] big mode is uh its scale it's co-design
[26:40] of actually working through the
[26:41] networking, the CPU, the GPU, and how
[26:44] everything works together. And the other
[26:47] big thing is their balance sheet and
[26:49] their ability to uh get supply
[26:52] commitments from, you know, I like I
[26:54] think I said this before, um optical
[26:56] startups are like upset because Nvidia
[26:59] secured all the supply for all the
[27:01] optical components. Same thing with TSMC
[27:04] wafer. Same thing with HPM memory. So
[27:06] the Nvidia is using their size and
[27:09] gorilla and be able to prepay and get
[27:12] components that are in shortage. So
[27:13] they've become the dominant, you know,
[27:15] over the next year or two, you're going
[27:16] to see Nvidia able to add tons of
[27:18] revenue because they were able to lock
[27:20] up all the supply components. That's
[27:22] another thing that people don't really
[27:23] talk about is their supply chain and
[27:25] their ability to work with partners and
[27:27] secure secure component inventory. Is
[27:30] there still energy FUD that we would run
[27:33] into an energy bottleneck before we run
[27:35] into a chip bottleneck? So,
[27:37] >> so Jensen said this last week on the
[27:39] Bloomberg interview that there are a lot
[27:41] of bottlenecks including um data center
[27:44] shell power and all those things uh
[27:47] components um energy whatever. So all
[27:50] those things it sounds really bad right
[27:51] and then right after that he said I
[27:54] think we have the chip industry has
[27:56] enough supply to double their revenue
[27:58] every year. basically implying Nvidia
[28:00] has enough supply uh for for energy and
[28:02] all that stuff. No one is pricing that
[28:04] in. So, everyone talks about
[28:06] bottlenecks. Nvidia CEO just basically
[28:09] told you on Friday that they have enough
[28:12] uh enough supply chain and all the
[28:15] bottleneck stuff to double revenue every
[28:17] year. No one, you know, Nvidia's revenue
[28:20] uh estimates for next year are a lot
[28:22] lower than double. I'll tell you that.
[28:25] Do you think the market uh prices in
[28:28] just how much of almost every important
[28:32] AI company in every category Nvidia
[28:34] actually owns? like it it it feels like
[28:37] every single like we're constantly
[28:39] focused on who's going to raise capex
[28:42] next and and and and where's this
[28:44] quarter coming in and it feels like in
[28:46] two or three years people will look at
[28:48] Nvidia's balance sheet and be like wait
[28:50] they have what what I imagine then will
[28:53] be you know could could end up being a
[28:56] trillion dollar plus of just like
[28:58] ownership and all of these great
[29:01] companies which again just goes back to
[29:02] the advantages of that early scale while
[29:06] they're, you know, while while all these
[29:07] companies are trying to compete away
[29:09] Nvidia's margins and all these different
[29:11] things, they've been able to accumulate
[29:14] again positions in in all of these
[29:16] incredible companies. I mean, we saw the
[29:18] SSI news uh yesterday is a great example
[29:22] of that. Um, but but how do you look how
[29:24] do you see it?
[29:26] >> So, I I think look at Jenison's history
[29:28] in investing in these companies and
[29:30] coreweave and see how much money they
[29:32] made. uh they just bought stake in uh
[29:34] the optical companies Lumenum and Co
[29:36] coherent. Um Jensen is enabling the
[29:40] future because he sees this overwhelming
[29:42] title of demand and he needs these
[29:44] companies to be able to build up their
[29:46] supply chain uh and to give uh supplies
[29:50] and chips to to Nvidia. So they actually
[29:52] ramp very hard. Um you know everyone's
[29:55] freaking out that this is vendor
[29:56] financing. What if hypers scale GPU
[29:59] cloud is so profitable and these
[30:02] companies need capital to build up that
[30:04] supply so they can uh serve the GPU
[30:08] cloud uh services over the next year or
[30:10] two. Maybe Jensen sees that coming like
[30:12] he did with all these other companies
[30:14] like Cororeweave and uh that's why he's
[30:17] investing in these companies to be able
[30:19] to expand their ability to make the
[30:21] components uh the industry needs. So I I
[30:24] think you're exactly right. In in a
[30:26] year, two, three years, uh Nvidia is
[30:28] going to have like all these stakes in
[30:30] these companies and it's going to look
[30:32] like uh he was a good investor. Um cuz
[30:35] he has been in the past.
[30:37] I mean,
[30:38] >> the man can buy a leather jacket for
[30:40] like five grand and sell it for a
[30:42] million dollars. I don't know what else
[30:43] you need to see.
[30:44] >> I mean, think think about
[30:45] >> Are you a secret bidder? Did you win
[30:47] that?
[30:47] >> No.
[30:48] >> You got to get you a jacket. Uh, the
[30:50] real the real question is
[30:52] >> how long until someone distills a jacket
[30:55] and open sources it you can get a dupe
[30:58] of a Jensen jacket for two bucks. That's
[31:00] what I want. Anyway, thank you so much
[31:02] for coming on the show. Jordy, you got
[31:03] anything else?
[31:04] >> Uh, this was great.
[31:05] >> Yeah, this was great.
[31:07] >> Thanks for thanks for putting up with
[31:09] all of our jokes.
[31:11] >> Hopefully this becomes the lucky charm
[31:13] for the markets.
[31:14] >> Yes, I agree.
[31:16] >> I I agree. I agree.
[31:18] >> Bottom is in. Great to see you, Tay.
[31:19] Have a good rest of the
