## Background
This episode of Invest Like the Best was published on May 20, 2026, by Colossus | Investing & Business Podcasts on RSS / web. It features host Patrick O'Shaughnessy in conversation with Gavin Baker, Founding Partner and CIO of Atreides Management, discussing the physical constraints defining the future of AI.
## TL;DR
Gavin Baker explores 'watts and wafers' as the primary physical bottlenecks determining the next phase of the AI revolution. He analyzes how near-term power shortages on Earth will eventually give way to orbital compute in space, alongside the pivotal role TSMC plays in semiconductor supply discipline. Baker also provides deep commentary on valuation discrepancies across tech, the disaggregation of pre-fill and decode GPU architectures, and how frontier AI labs and hyperscalers are positioning themselves.
## Atmosphere & style
The conversation carries an intellectually intense, highly analytical, and forward-looking tone between two seasoned market observers. Gavin Baker speaks with energetic conviction, blending granular hardware and macroeconomic technicality with broad historical and geopolitical analogies. The dialogue feels like an intimate, high-stakes briefing from the frontlines of technology investing, oscillating between hard-headed financial realism and profound awe at the speed of technological disruption.
## Key takeaways
- `14:00` The central physical bottlenecks dictating the next phase of AI expansion are 'Watts' (energy/power infrastructure) and 'Wafers' (semiconductor manufacturing capacity).
- `03:32` Anthropic demonstrated unprecedented growth by adding $11 billion in ARR practically overnight, burning roughly 80% less capital than OpenAI to reach comparable revenue scale.
- `05:00` DeepSeek's release showed that reasoning models exponentially increase compute demand during inference, accelerating hardware demand rather than destroying it.
- `15:45` Near-term terrestrial power shortages are expected to ease by 2027–2028, with orbital compute (space-based data center racks) serving as the long-term solution beyond 2028.
- `25:00` TSMC's strict supply discipline acts as the primary buffer against a total AI market overbuild, holding the line to prevent excessive GPU capacity overshoot.
- `25:00` Elon Musk's 'TerraFab' initiative aims to establish massive U.S. semiconductor manufacturing by partnering with Intel and drawing top talent from overseas foundries.
- `30:30` Economic returns in AI remain concentrated at the frontier along the Pareto frontier, with capabilities following Richard Sutton's 'Bitter Lesson' where compute beats specialized algorithmic design.
- `35:33` The AI business model is shifting from flat-rate consumer subscriptions to usage-based enterprise pricing ('pay by the drink'), unlocking massive ARR growth.
- `42:49` Disaggregating pre-fill (context memory-bound) and decode (token bandwidth-bound) workloads extends older GPU operational lifespans and improves private credit financing structures.
- `51:27` To survive competition from frontier model providers, application-layer startups must sit directly in the 'token path' or achieve non-obvious operational scale.
- `57:43` A severe valuation disconnect exists in public markets, where semi-cap equipment trades at roughly 40x annualized earnings while DRAM memory trades at single-digit multiples.
- `01:04:12` Major hyperscalers exhibit varying strategies, with Microsoft pivoting to hold compute internally, while Amazon and NVIDIA lead in deep engagement with chip startups.
## Detailed walkthrough
### `00:00` Introduction and Sponsor Announcements
The episode opens with sponsor messages from **Ramp**, **Rogo (Felix)**, **WorkOS**, and **Vanta**, highlighting their offerings for finance teams, AI-native automation, and enterprise infrastructure. Patrick O'Shaughnessy then introduces the episode as his sixth conversation with **Gavin Baker**, CIO and founding partner of **Atreides Management**. This milestone ties Gavin with Bill Gurley for the most appearances on *Invest Like the Best*.
Patrick sets the stage for the discussion, noting that Gavin's central thesis centers on **"Watts and Wafers"**—the twin physical constraints of energy and semiconductor manufacturing capacity that will define the next phase of the artificial intelligence revolution. Gavin will outline how near-term power shortages could ease around 2027–2028 before **orbital compute** (space-based data centers) scales up as a long-term solution, alongside why TSMC’s capacity decisions and GPU architecture disaggregation are crucial variables to track.
`02:29` Reflecting on the Extraordinary Market Setup of Early 2026
Patrick asks Gavin to reflect on living through the unique market and technology environment of March and April 2026. Gavin distinguishes between two types of market drawdowns: those where your thesis is invalidated and you must take your medicine, and those where stock price action diverges wildly from underlying fundamentals, allowing disciplined investors to lean in and build "pent-up alpha." He describes the spring of 2026 as one of the most extraordinary moments in the history of capitalism and American business.
While tech stocks and the NASDAQ were selling off, the fundamental growth in AI was breaking every historical benchmark. Gavin highlights **Anthropic**, which added an astounding **$11 billion in Annual Recurring Revenue (ARR)** in a tiny window. To put this in perspective, he contrasts it with three of the most successful SaaS and cloud companies of the last decade—Palantir, Snowflake, and Databricks—which collectively employ tens of thousands of people and took ten years to build their combined revenue base. Anthropic added that equivalent volume of business practically overnight.
- **Anthropic ARR Growth:** Added $11 billion in ARR, outstripping the decade-long cumulative growth of major SaaS giants in a single month.
- **Historical Context:** The SaaS and cloud revolution created $5 to $10 trillion in market value, yet no company in that era exhibited an S-curve or exponential growth trajectory like current frontier AI labs.
`05:00` DeepSeek Retrospective, Geopolitics, and Tech Valuations
Gavin connects the March 2026 dynamics to the earlier market panic following the release of **DeepSeek** in early 2025. He recalls how DeepSeek published its paper on a U.S. holiday Monday, triggering an AI market sell-off a week later on "DeepSeek Monday." However, sharp observers quickly realized DeepSeek's reasoning models actually exponentially increased demand for compute during inference. Immediately following the paper, AWS availability zone prices in Asia doubled, GPU rental prices spiked, and DRAM prices went vertical.
A similar disconnect occurred in early 2026. Investors were panicked by macroeconomic and geopolitical events—specifically tension surrounding the **Strait of Hormuz**. Gavin points out that a closed Strait of Hormuz actually boosted American relative manufacturing competitiveness: U.S. natural gas (NG1 on Bloomberg)—a primary input for domestic electricity—dropped by 20%, whereas European and Asian energy prices doubled or tripled. Combined with the U.S. being the world's largest producer and exporter of oil and gas, the American economy was far less vulnerable to energy shocks than during the 1970s. This macro backdrop allowed tech stocks to trade at their cheapest valuations relative to the broader market in ten years, right at the peak of AI operational momentum.
- **DeepSeek Market Misconception:** Initial market fears of compute efficiency actually proved to be massively bullish for hardware, as reasoning models consumed far more tokens during inference.
- **US Energy Advantage:** Natural gas price divergence gave American industrial and compute infrastructure a distinct cost advantage over international competitors during the Middle East crisis.
`08:40` Valuing Frontier AI Labs: Anthropic vs. OpenAI
Patrick and Gavin analyze the market valuations of **OpenAI** and **Anthropic**, the two reference assets for frontier AI models. Gavin stresses that despite their similar top-line numbers, the two companies exhibit vastly different structural Returns on Invested Capital (ROIC). Anthropic operates with a dramatically lower cost per token and has burned approximately **80% less capital** than OpenAI to reach a comparable operational scale. Gavin credits Anthropic's efficiency and capital management, referencing leadership under figures like CFO Sarah Friar at OpenAI and Krishna Rao at Anthropic.
Gavin argues that Anthropic is severely supply-constrained by available compute. If given unlimited compute access, Anthropic's unconstrained run-rate revenue (URR) could easily be **$100 billion to $200 billion** today. To stretch their compute supplies, Anthropic even deprecated the token output of Claude (with Claude Opus generating roughly 70% fewer tokens for identical queries), proving that token density and reasoning efficiency matter deeply.
- **Anthropic Capital Efficiency:** Achieved comparable revenue scale to OpenAI while burning roughly 80% less capital.
- **Compute Bottleneck:** Anthropic trades at a implied conservative multiple relative to its unconstrained potential, as lack of hardware limits its true revenue potential.
- **Elon Musk’s Investor Trust:** Gavin contrasts frontier lab fundraising with Elon Musk’s decade-long track record of compounding SpaceX at a low-30s percentage rate annually by deliberately not overpricing rounds, thus preserving an "investor trust superpower" that grants infinite capital access.
`14:00` Watts and Wafers: Navigating Near-Term Power Shortages
Transitioning to the episode's core theme, Gavin dives into the twin physical bottlenecks facing AI: **Watts (power)** and **Wafers (semiconductor manufacturing)**. On the energy front, private equity heavyweights (Blackstone, Apollo, KKR) note that the core bottleneck has shifted from raw power generation and silicon supply to local **zoning and regulatory approvals**. Companies are treading cautiously ahead of political midterms to avoid becoming targets over local power grid strains or corporate restructuring.
On the supply side, the industrial manufacturing of power infrastructure remains tight—such as the specialized casting of massive turbine blades, where Western manufacturing capabilities have lagged for decades. Nevertheless, Gavin remains confident that free-market capitalism will resolve terrestrial power shortages by **2027 or 2028** as new energy projects come online. Beyond 2028, however, the ultimate solution for power constraints will transition off-planet to **orbital compute**.
- **Primary Gatekeepers:** Local zoning, permitting, and regulatory approvals have surpassed chip availability as the top bottleneck for ground data centers.
- **Timeline for Power Relief:** Terrestrial grid capacity and new energy installations are projected to ease the power crunch by 2027–2028.
`15:45` Redefining Orbital Compute: Racks in Space
Gavin clarifies common misconceptions about orbital compute. Critics often mock space data centers by imagining massive, Pentagon-sized structures floating in orbit. In reality, Gavin explains that orbital data centers consist of standard, modular server racks—similar in dimensions to an 8-foot-tall, 3,000-pound **NVIDIA Blackwell rack**—deployed individually as satellites into sun-synchronous orbits.
Equipped with 500-foot solar array wings that constantly face the sun and trailing radiators designed to dissipate heat into the vacuum of space, these satellites bypass ground-based power constraints entirely. SpaceX engineers at Starbase express extreme confidence in this architecture. Space-based compute nodes eliminate complex terrestrial wiring by linking individual racks across vacuum using high-speed optical lasers—a technology already battle-tested across thousands of **Starlink** satellites.
- **Physical Form Factor:** Orbital compute uses standardized racks (comparable to a 100 kW Blackwell rack) rather than mega-structures.
- **Power Scale:** Starlink V3 satellites already operate at 20 kilowatts; scaling this up to 60 kW or 100 kW per satellite is a natural technical progression.
- **Thermal Management & Maintenance:** Cooling is managed via passive directional radiators pointed away from the sun. Until autonomous humanoid robotics (like Optimus) can perform in-space repairs, failed racks will simply be replaced by launching new units via low-cost Starship payloads.
- **Inter-Satellite Links:** Racks communicate via laser mesh networks in a vacuum, forming a distributed "virtual data center" in low Earth orbit.
`20:00` Terrestrial Data Centers, the Silicon Shield, and Bubble Risks
The economic viability of space-based matrix multiplication raises immediate questions about the fate of terrestrial infrastructure. America historically consumes every available energy source as aggressively as possible, and compute will follow the exact same dynamic. This makes the bear case for edge AI less convincing. Inference is well-suited for orbital compute, whereas training will remain on Earth for the foreseeable future. Consequently, terrestrial data centers will retain high value for decades. However, companies ramping up terrestrial power production and cooling infrastructure face structural timing risks as these long-lead capacity additions arrive just as the market recognizes the reality of orbital compute. Meanwhile, interim solutions are advancing rapidly, such as Boomerang Space repurposing jet engines for power generation.
On the silicon side, global compute output remains heavily reliant on a small group of senior engineers at TSMC in Taiwan. Accounting for a massive share of the nation's GDP, water, and electricity, these engineers view themselves as guardians of Morris Chang’s legacy and Taiwan's "silicon shield." Looking back at visiting the Hsinchu Science Park over two decades ago, asking if TSMC could ever catch Intel was met with the response that it was a "beautiful dream" for their grandchildren. TSMC ultimately achieved it, accelerated by Intel's self-inflicted mistakes. TSMC operates on an unusual culture of trust—highlighted by the fact that Nvidia CEO Jensen Huang has never signed a formal contract with TSMC, relying instead on verbal handshakes and long-term partnership fairness.
Every foundational technology shift—from canals and railroads to the South Sea Bubble and the dot-com era—has historically triggered an investment bubble. Markets correctly identify a transformative tech shift, leading to what Michael Mauboussin terms a "breakdown in diversity," where market consensus becomes universally bullish. This feeds a bubble that funds the buildout, eventually causing supply to overshoot demand and triggering a crash. However, the current AI buildout differs fundamentally from the 2000 telecom crash:
* **Operating Cash Flow Funding:** Current capital expenditures are overwhelmingly funded by corporate cash flows rather than speculative debt.
* **Asset Utilization:** Current GPUs are running at **100%** capacity utilization, unlike the **99%** unutilized dark fiber during the dot-com bubble.
* **Valuation Discrepancies:** Equity valuations today remain far more anchored to underlying earnings than during prior technology frenzies.
`25:00` The Legacy of George Vanderheiden and the "TerraFab" Vision
Fighting a technology bubble prematurely can be career-ending for fund managers. The late Fidelity portfolio manager George Vanderheiden resisted the 1999 tech bubble, eventually retiring in early 2000 under immense client pressure for underperforming the market while holding **40%** of his portfolio in tobacco and **40%** in homebuilders. Over the subsequent three years, his value-oriented allocation outperformed the NASDAQ by roughly **20x to 30x**, proving his core adage that "being early is the same thing as being wrong."
Today, TSMC's strict supply discipline acts as the primary buffer preventing an catastrophic AI overbuild. If TSMC bowed to Jensen Huang's demands, Nvidia could potentially sell **$2 trillion to $3 trillion** in GPUs by 2026 or 2027, eventually forcing an oversupply crisis. If the industry avoids a bubble, TSMC deserves the credit for holding the line. While competitors like Intel or Samsung may eventually break discipline to capture unmet demand, TSMC's capacity expansion pacing remains the single most important variable for tracking bubble risk. Its **9 to 15 month** node lead maintains a delicate "Goldilocks zone" that keeps silicon supply tight enough to prevent an overbuild while starving competitors of scale.
Concurrently, Elon Musk's "TerraFab"—a joint venture involving SpaceX and Tesla aimed at building the world's largest semiconductor fab in the US—presents a structural shift:
* **Intel Partnership:** TerraFab leverages Intel's 50-year institutional knowledge base, which sits only **3 to 5 quarters** behind the leading-edge nodes.
* **Equipment Access:** Major semiconductor equipment suppliers (ASML, KLA, Lam Research, Applied Materials) are sending their top "A-teams" to support Musk's team, eager to prevent a TSMC monopsony.
* **Talent Acquisition:** Musk's hardware engineering reputation draws elite global talent. Plans include building localized cultural hubs ("Taiwan Town," "Korea Town," "Japan Town") next to the Texas fab to relocate whole teams and top talent from overseas foundries.
* **Execution Speed:** While traditional fabs take three years to construct, Musk's team previously built a massive data center in **122 days**, prompting Samsung to allocate him dedicated office space inside their Texas facility to manage their buildout pace.
`30:00` The Pareto Frontier, DeepSeek, and TurboQuant
Contrary to predictions that open-source models (like DeepSeek) would allow enterprise users to avoid heavy token spend, economic returns in AI remain overwhelmingly concentrated at the frontier. The distribution of intelligence versus cost—the Pareto frontier—has shifted rapidly over the past year:
* **Google's Shift:** Nine months ago, Google dominated the Pareto frontier. However, conservative design choices on TPUv8 (meant to reduce dependence on Broadcom and Nvidia) cost them their leadership position.
* **Current Leaders:** The Pareto frontier is currently held by Anthropic, OpenAI, and xAI's Grok 4.3 (which leads in cost efficiency for 500-billion parameter models), while Gemini 3.1 trails slightly behind.
* **Richard Sutton’s Bitter Lesson:** The principle that general methods leveraging more compute always beat human domain-specific algorithmic ingenuity remains the core thesis of AI scaling.
A notable market scare in March stemmed from "TurboQuant"—a Google memory optimization paper detailing DRAM efficiency gains published during long-term agreement negotiations with Micron, Samsung, and SK Hynix. While financial markets panicked over potential DRAM demand destruction, practically zero AI hardware engineers believed TurboQuant would reduce net DRAM consumption.
The ultimate risk to the scaling narrative is a temporary violation of the Bitter Lesson caused by Artificial Superintelligence (ASI). If a **400 to 600 IQ** model emerges, its primary goal will be acquiring efficiency to maximize its own intelligence and resources, potentially inventing algorithmic breakthroughs that temporarily reduce raw compute demand.
`35:00` Usage-Based Pricing and Continual Learning
While harness engineering and runtimes (providing models with context, tools, and memory states) significantly boost performance, the market's business model is undergoing a structural shift. The standard **$250 to $300 per month** flat-rate subscription tier now subjects power users to heavy rate-limiting and degraded model performance, as newer models produce significantly fewer redundant tokens. To access full model reasoning capabilities, users must migrate to enterprise usage-based pricing ("pay by the drink").
This transition mirrors the mobile telecom industry between 2005 and 2007, where growth exploded as users moved beyond capped plans to hybrid usage models. Usage-based pricing will enable companies like OpenAI and Anthropic to easily surpass **$200 billion in ARR**, as enterprises deploy hundreds of concurrent autonomous agents per employee. However, this dynamic creates a growing divide between enterprise users with unlimited API access and consumer tier users running rate-limited models.
The next critical technical leap involves **continual learning**—a model's ability to dynamically update its weights in real time:
* **Current Mechanics:** Current models rely on reinforcement learning (RL) during mid-training or post-training RL gyms, requiring millions of repetitive trials to ingest simple rules.
* **Human Efficiency:** Human learning is hyper sample-efficient (e.g., touching a hot stove once teaches permanent avoidance).
* **Future Takeoff:** Achieving true real-time weight updating would trigger an extremely fast takeoff trajectory, representing one of the final major hurdles alongside solving context memory limits.
`40:00` Startup Chip Architectures, the "Iron Triangle," and Disaggregation
Gavin Baker weighs in on the thousands of startup "flowers blooming" in the semiconductor space to bypass compute bottlenecks. He views this explosion of competition as healthy for the tech ecosystem—and ultimately beneficial even for NVIDIA CEO Jensen Huang, as it lowers potential regulatory antitrust scrutiny. Drawing an analogy to the military's "iron triangle" of tank design (where engineers must make trade-offs between attack, defense, and mobility), Baker explains that chip architects face physical constraints dictated by TSMC's manufacturing rules. Competitors like Google's TPU, Amazon's Trainium, and AMD's MI series are essentially attempting to build a better GPU. Currently, Baker notes that Trainium 3 is "tugging on Superman’s cape" as it ramps into production with a switch scale-up network needed to economically run Mixture-of-Experts (MoE) models, while AMD continues to fly from behind.
For semiconductor startups, Baker’s rule of thumb is that capturing even a 1% market share will be worth $100 billion. However, trying to beat NVIDIA head-on at making a standard GPU is a losing game. NVIDIA gets better pricing from TSMC than any 200-person startup, receives advanced insights into TSMC’s roadmap, and enjoys early visibility across every major frontier model builder. Instead, successful startups must build something fundamentally *different* and inherently *hard* to replicate.
The emerging architectural split between pre-fill (understanding the prompt/context, which is memory-capacity bound) and decode (generating output tokens, which is memory-bandwidth bound) provides a broader canvas for startup innovation:
* **The Naval Analogy:** Baker cites his colleague Andrew Fox, who likens pre-fill to "loading the cannon" on an 18th-century British warship, while decode is "firing" the cannon.
* **Cerebras’ Wafer-Scale Approach:** Atreides was a venture investor in Cerebras, which took the difficult route of building wafer-scale computers. Despite facing constraints where the chip's internal compute/memory far exceeds its shoreline I/O bandwidth, Cerebras is tackling this by attempting to bond optical wafers directly on top.
* **Resilience:** Building custom silicon requires extreme grit; it took Cerebras three generations of chips to refine their architecture after early setbacks.
`45:00` Extending GPU Lifespans and the Impact on Private Credit
The disaggregation of pre-fill and decode is poised to transform the financial economics of AI infrastructure, effectively saving the private credit market. Private credit lenders are currently taking severe losses on SaaS software loans. Many investors previously underwrote GPU debt assuming a short three-to-four-year lifespan, while AI skeptics claimed rapid tech cycles meant GPUs were obsolete in one or two years. Baker strongly pushes back against this narrative.
By placing specialized inference hardware (such as Cerebras systems or Groq LPUs, which NVIDIA effectively acquired) in front of older generation GPUs like Hopper or even Ampere, companies can offload pre-fill tasks to the older chips. This extends the functional life of GPUs to 10 or 15 years, running them until they literally melt. Lowering the depreciation risk mathematically alters the cost of financing the AI buildout:
* **Cheaper Financing:** Instead of high-yield debt terms (like CoreWeave's lowest financing rates in the low sevens), GPU fleets can be financed at 5% to 6% interest rates.
* **Hyperscaler Fleets:** Citing Cotu investor Jamin Ball, Baker notes that while "sellers of shortage" (like NVIDIA) have outperformed "buyers of shortage" (the hyperscalers), hyperscalers possess massive existing CPU fleets. Because agentic workflows rely heavily on CPUs for orchestration and tool calls, these older assets will retain immense economic value.
`50:00` The Application Layer, Token Path, and Open Source Game Theory
Shifting to AI-native founders outside of hardware, Baker addresses the struggle of creating lasting software moats. Under Jensen Huang’s "five-layer cake of AI," profits currently accrue to energy, data centers, chips, and frontier models rather than the application layer. Successful early breakout applications like Cursor, Cognition, and Anthropic's coding tools succeeded by focusing aggressively on code. Baker references Replit founder Amjad Masad’s observation that coding may be the shortest path to Artificial Superintelligence (ASI), as an AI that codes can program itself to perform any subsequent task.
To survive, application-layer startups must either operate within the "token path" (a concept popularized by Jamin Ball, exemplified by data platforms like Databricks) or build hard, non-obvious operational scale before model companies subsume their functionality:
* **Data Moat Vulnerability:** Baker cautions against believing a narrow vertical data moat will protect startups. If that data is generated by humans, frontier labs can likely train cheaper, broader models that render vertical niches obsolete.
* **Jensen's Countermove & Anti-Distillation:** Baker believes NVIDIA could develop a frontier-level model if pushed, using open-source models as a strategic counterweight to frontier model labs trying to move down the stack.
* **Chinese Open Source & Distillation:** Current Chinese open-source models (e.g., DeepSeek) heavily rely on distilling American frontier tokens—sometimes built on as few as 150,000 original reasoning traces scraped via APIs. American labs are now fighting back with anti-distillation techniques.
* **The New Prisoner's Dilemma:** Frontier labs face a game theory dilemma over whether to release their top-tier models via API. If a lab holds back to prevent distillation, a defecting competitor can release their model, capture massive revenue, and convert those financial resources directly into model intelligence.
Here is the detailed walkthrough for Part 4 of the conversation.
---
`55:00` Overinvesting in Cybersecurity, "The Last Samurai," and AI-Assisted Workflows
Gavin Baker explains how Atreides is preparing for a world dominated by ultra-capable AI systems like "Mythos 3" or "Mythos 4." Defensively, his main focus is aggressively overinvesting in cybersecurity. He strongly advises everyone to set up a family or company safe word that cannot be socially engineered, leaving all digital devices behind—literally walking down to the ocean—to establish it. This is designed to combat sophisticated voice and video cloning scams where an AI-generated simulation of a relative or colleague calls via FaceTime asking for funds. On the analytical side, Baker reflects on his own role using an analogy from the movie *The Last Samurai*. He describes Tom Cruise’s character as a Civil War veteran hired during the Meiji Restoration to train peasant armies, who ultimately realizes that traditional methods are rendered obsolete by modern weaponry. The machine gun represents AI: *"If we do not all become masters of the machine gun, we're going to get mastered."* Baker remains optimistic that his decades of investing experience will allow him to integrate these "machine guns" into his workflow and continue contributing human value for a long time.
Within Atreides, AI agents are actively running to optimize research and daily operations. Baker notes that his most useful personal agent generates curated summaries of the key insights from industry podcasts and speeches. Given that tech leaders—including Jensen Huang, Elon Musk, Dario Amodei, and executives from OpenAI, xAI, Google, Cursor, Fireworks, and Base10—produce roughly six hours of content daily, AI is essential for finding needles in haystack commentary. Another critical internal tool analyzes corporate proxy statements to evaluate executive compensation structures. It automatically breaks down whether management is rewarded via basic RSUs or performance stock units (PSUs) and tracks changes over time, saving analysts countless hours of labor-intensive work.
- **Defensive Security Protocol:** Establish offline, non-socially-engineerable safe words with family and colleagues to verify identity against hyper-realistic AI video and voice impersonations.
- **Key External Sources Tracked:** OpenAI, xAI, Google, Cursor, Fireworks, Base10, as well as commentary from Jensen Huang, Elon Musk, and Dario Amodei.
- **High-Value Analytical Agent:** Parsing proxy filings to evaluate executive alignment through PSU (Performance Stock Unit) metrics versus standard RSUs.
---
`57:43` Cross-Sectional Valuation Distortions and Commodity-Cycle Risks
Baker voices growing concern over extreme valuation dislocations within the technology ecosystem. He points out that semi-cap equipment manufacturers trade around 40 times next quarter's annualized earnings, whereas DRAM memory providers trade at mid-single-digit multiples. In previous peak cycles, this valuation gap was far narrower (e.g., 5x vs. 12x, or 3x vs. 45x). While semiconductor equipment business models have improved relative to memory—partly due to recurring maintenance revenues—a 1,000% multiple gap makes no financial sense, especially given the margin expansion driven by High Bandwidth Memory (HBM). Similarly, NVIDIA remains near historic lows relative to the broader market, effectively pricing in an unfathomable amount of market share loss, while low-quality or adjacent industrial plays command massive premiums.
This dynamics mirrors a classic commodity cycle: during severe supply shortages, high-cost, low-quality suppliers experience the largest percentage gains because they transition from the brink of insolvency to gushing cash flow. Hyperscalers are forced to buy from unreliable, high-cost vendors whose components suffer high failure rates simply because no other capacity exists. Retail investors on social media (such as X/Twitter and Reddit) often pump these low-quality names to extreme valuations, even though institutional investors know these stocks will plummet once supply normalizes. Baker finds comfort in holding high-quality compounders with reasonable valuations, contrasting them with speculative bubbles in areas like nuclear power, quantum computing, and small-cap retail favorites.
- **Valuation Disconnect:** Semi-cap equipment trading at ~40x annualized earnings vs. DRAM companies trading in the mid-single digits.
- **Historical Comparison:** Historical peak-cycle multiples for memory were around 5x vs. 12x for equipment.
- **Commodity Cycle Dynamic:** High-cost, low-quality component manufacturers outperform early in supply shortages due to extreme cash-flow swings, but carry massive downside risk once capacity catches up.
---
`01:01:45` Market Baskets, Disaggregation, and the Mis-Categorization Opportunity
Baker highlights how market dynamics shifted dramatically in January. During 2024 and 2025, the AI trade moved in tight correlation—investors could hold long GPU compute or scale-up networking while hedging with short power or memory positions. However, those cross-sectional correlations broke down completely. Scale-up networking began rallying while scale-out networking fell, and DRAM underperformed NAND and HDDs (hard disk drives). This shift occurred as market participants became educated on specific subsectors, leading to rigid factor baskets.
This breakdown created significant mispricing opportunities in miscategorized companies. Baker highlights Astera Labs as a prime example, having first invested during their Series C round. The broader market mistakenly placed Astera into "copper loser" baskets under the assumption that optical connections would rapidly replace copper interconnections. However, because Astera’s core product lineup includes switches—where both copper and optical interconnects are used to link switches to accelerators—being a switch provider means they intrinsically benefit from accelerator deployment regardless of the physical medium.
- **Market Shift:** The tight correlation among AI subsectors dissolved in January, requiring fine-grained stock selection rather than sector-wide basket trades.
- **Astera Labs Case Study:** Misgrouped into "copper loser" baskets despite switch architectures utilizing both copper and optical interconnects to accelerators.
---
`01:04:12` Hyperscaler Review: Google, Meta, Amazon, and Microsoft
Baker provides a rapid-fire analysis of the major public hyperscalers navigating the AI landscape:
- **Google:** While Google lost its initial exclusive TPU advantage as the broader market scaled GPU compute, its sheer volume of installed compute capacity remains unmatched. Google holds immense structural advantages through search, YouTube video data (which is uniquely valuable for robotics training), and accelerating Google Cloud Platform (GCP) growth.
- **Meta:** Mark Zuckerberg receives immense credit for turning Meta into an AI-first organization faster and more aggressively than its peers. Meta secured massive talent and compute contracts early, and its MSL (Meta AI Synthetic Lab) released Muse—a major positive surprise that, while not quite on the absolute Pareto frontier with OpenAI or Anthropic, demonstrated rapid rate-of-change progress.
- **Amazon:** AWS occupies a powerful position due to its custom Trainium chips and the impending P&L productivity gains from automation and robotics across its retail footprint. Internal Nova models are also performing better than public perception suggests.
- **Microsoft:** CEO Satya Nadella made bold, courageous decisions to pivot strategy. After a brief hesitation in early 2025 where standard CapEx return algorithms faltered, Microsoft shifted focus from merely reselling GPU capacity to OpenAI toward reserving compute internally for its own products and proprietary model development. While short-term Azure growth took a hit—and the market punished the stock—this strategy ensures Microsoft is not wholly dependent on external API-accessible frontier models.
- **Hyperscaler Engagement with Startups:** Amazon and NVIDIA lead the industry in deep startup engagement, followed by Google. Broadcom remains the premier ASIC partner for custom chips, whereas AMD displays virtually zero engagement with early-stage startups.
<ElicitationsGroup message="Where would you like to go next?">
{/* Reason: Offers clear logical next steps following Gavin Baker's insights on semiconductor valuations and hyperscaler positioning. */}
<Elicitation label="Explore Part 5 of the transcript walkthrough" query="Provide the next part of the transcript walkthrough."/>
<Elicitation label="Analyze the 'Watts vs. Wafers' physical constraints" query="Explain Gavin Baker's theory on watts and wafers as the main physical constraints for AI expansion."/>
</ElicitationsGroup>
`01:10:00` Missing Engagement from Hyperscalers and Value Destruction at the Application Layer
Gavin Baker expresses surprise and concern over the fact that certain major hyperscalers—specifically Microsoft and Meta—have minimal to zero direct engagement with the emerging ecosystem of specialized AI chip startups. In contrast, NVIDIA, AMD, and Google maintain active relationships with these nimble teams. Baker notes that many of the world's best engineering teams are leaving large public corporations to join agile startups, making this lack of engagement a potentially significant oversight and competitive disadvantage for the passive hyperscalers.
Reflecting on the broader impact of AI, Baker and host Patrick O'Shaughnessy discuss the severe value destruction taking place at the application layer. While high-profile AI-native successes like Cursor and Cognition garner massive attention, AI has net destroyed trillions of dollars in application-layer economic value. Baker highlights a stark, defining metric of the current era: the companies generating the most economic value today are those with the highest effective ratio of utilized GPUs per human.
* **GPU Allocation:** Current value generation heavily favors companies optimizing high GPU-to-employee operational ratios.
* **Hyperscaler Strategy:** NVIDIA, AMD, and Google actively engage with specialized hardware startups, whereas Meta and Microsoft have historically shown far less involvement in this specific early-stage hardware layer.
---
`01:12:00` Personal Safety Risks and Geopolitical Instability
As AI becomes deeply politicized, Baker voices growing concern regarding personal safety for prominent leaders in the space. Citing the alarming incident where Molotov cocktails were thrown at OpenAI CEO Sam Altman’s home, Baker warns that society is entering a high-beta, high-variance environment. He worries that rising political violence in the United States will increasingly target key figures steering the AI revolution.
On the geopolitical stage, AI dominance is already reshaping global power dynamics. Pointing to the war in Ukraine, Baker argues that Ukrainian military successes stem not merely from hardware capabilities like drones, but from wielding the best battlefield AI outside of the United States and Israel. This emerging reality presents a double-edged sword: while an American AI lead benefits the West, it risks destabilizing foreign adversaries who must figure out how to respond to an unprecedented technological asymmetry.
* **Targeted Violence:** Growing concerns over the physical security of prominent AI figures following attacks on industry leaders.
* **Battlefield AI:** Autonomous and AI-driven combat systems in Ukraine represent a practical manifestation of AI reshaping real-world military doctrine.
---
`01:13:30` Pax Americana, Social Equity, and Optimism for the AI Future
To frame the geopolitical risks, Baker reflects on historical precedents, drawing a parallel to the post-1945 "Pax Americana." Following World War II, the United States held an absolute nuclear monopoly yet chose to rebuild former adversaries like Germany and Japan rather than conquer the globe. Baker hopes American AI supremacy leads to a similar era of global stability rather than conflict. He underscores his deep personal optimism by sharing the story of a father who utilized laboratory compute and AI agents to discover a life-saving drug for his daughter's rare, incurable disease, dramatically changing her life.
Despite his stance as an "AI maximalist," Baker cautions that AI acts as a societal event horizon and discontinuity. He stresses that while Luddite critiques of technological progress will ultimately prove wrong, leaders must humbly and thoughtfully address public anxieties. He highlights the dystopian risk of top-tier AI being accessible only to the ultra-wealthy, arguing that democratic access is essential to ensure technology benefits all of humanity.
* **Historical Context:** Compares AI dominance to the post-WWII nuclear monopoly and the subsequent Pax Americana.
* **Medical Breakthroughs:** AI agents and compute allocations are actively accelerating drug discovery for rare genetic diseases.
* **Equitable Access:** AI leaders must ensure elite capabilities do not remain locked behind immense capital barriers.
## Notable quotes
> `25:00` being early is the same thing as being wrong. — George Vanderheiden
> `55:00` If we do not all become masters of the machine gun, we're going to get mastered. — Gavin Baker
## Mentioned
Gavin Baker, Patrick O'Shaughnessy, Atreides Management, Anthropic, OpenAI, SpaceX, Tesla, Elon Musk, TSMC, NVIDIA, Jensen Huang, DeepSeek, Google, Meta, Amazon, AWS, Microsoft, Satya Nadella, Mark Zuckerberg, Cerebras, Astera Labs, Intel, Samsung, ASML, KLA, Lam Research, Applied Materials, Palantir, Snowflake, Databricks, Sarah Friar, Krishna Rao, Morris Chang, George Vanderheiden, Michael Mauboussin, Richard Sutton, Andrew Fox, Jamin Ball, Amjad Masad, Sam Altman, Dario Amodei, Ramp, Rogo, WorkOS, Vanta, Ridgeline, Starlink, Blackwell, Grok, Gemini, TPU, Trainium, Cursor, Cognition, Replit
Gavin Baker argues that artificial intelligence scaling is governed by the physical constraints of 'Watts and Wafers'—terrestrial power grids and semiconductor manufacturing capacity. While these bottlenecks temporarily cap growth and dictate infrastructure architecture, Baker contends they also protect the industry from a debt-fueled bubble by enforcing supply discipline.
Power Grid Bottlenecks and the Orbital Compute Horizon (14:00)
Baker identifies regulatory permitting and local zoning approvals, rather than raw generation or silicon availability, as the primary immediate barriers to terrestrial data center expansion. Private equity sponsors report that ground-based power shortages should ease between 2027 and 2028 as new domestic utility capacity comes online. In the interim, developers are deploying stopgap solutions, such as repurposing jet engines for localized power generation.
Beyond 2028, Baker expects incremental compute scaling to transition off-planet into sun-synchronous orbital data centers. Rather than monolithic structures, this architecture relies on modular, standardized satellite racks—comparable in scale to a 100-kilowatt NVIDIA Blackwell unit—equipped with 500-foot solar arrays and directional radiators that dissipate heat into space. Drawing on SpaceX Starlink technology, these distributed orbital nodes would link across a vacuum using optical laser mesh networks, bypassing terrestrial land and power constraints entirely.
Silicon Supply Discipline and Technology Bubble Dynamics (20:00)
Unlike the 2000 telecom crash, which was financed with speculative debt and resulted in 99% unutilized dark fiber, Baker observes that current AI hardware runs at 100% capacity utilization and is funded primarily by corporate operating cash flows. TSMC's strict capacity discipline acts as the critical barrier preventing an catastrophic industry overbuild; Baker asserts that if TSMC expanded indiscriminately to meet unconstrained demand, NVIDIA could sell $2 trillion to $3 trillion in hardware, culminating in an oversupply crash.
Baker also notes severe valuation distortions across public markets. Semi-cap equipment manufacturers trade around 40 times annualized earnings, whereas DRAM providers trade at mid-single-digit multiples despite margin expansion from High Bandwidth Memory. Baker compares this discrepancy to classic commodity cycle behavior, where low-quality, high-cost suppliers experience outsized price gains during severe supply crunches but carry substantial downside once capacity normalizes.
Hardware Disaggregation and Extended GPU Lifespans (45:00)
AI compute is structurally bifurcating between prompt context ingestion ('pre-fill,' which is memory-capacity bound) and output token generation ('decode,' which is memory-bandwidth bound). Baker explains that placing specialized inference hardware, such as Cerebras wafer-scale systems or Groq LPUs, in front of older-generation GPUs allows operators to route pre-fill tasks to older Hopper or Ampere chips.
This architectural disaggregation counteracts claims that GPUs face rapid obsolescence within one or two years. By extending functional chip lifespans to 10 or 15 years, data center operators significantly lower depreciation risk. In Baker's view, this adjustment alters AI infrastructure economics, enabling operators to secure lower financing rates of 5% to 6% rather than high-yield debt terms in the low sevens.
Frontier Economics, Token Pricing, and Application Vulnerability (50:00)
Economic value in AI remains concentrated at the capability frontier—led by Anthropic, OpenAI, and xAI's Grok 4.3—and within organizations boasting high ratios of utilized GPUs per employee. Baker contrasts Anthropic's capital efficiency with OpenAI, noting Anthropic burned approximately 80% less capital to reach a comparable operational scale. Facing compute shortages, labs are shifting away from $250-to-$300 monthly flat subscriptions toward enterprise usage-based pricing, a structural transition Baker compares to mobile telecom's expansion between 2005 and 2007.
Conversely, Baker cautions that AI has net destroyed economic value across traditional application-layer software. Startups relying on narrow, human-generated vertical data moats remain vulnerable, as frontier labs can train broader models that subsume vertical tools. Simultaneously, frontier labs confront a prisoner's dilemma regarding API access, as competitors and foreign open-source models scrape reasoning traces—sometimes as few as 150,000 tokens—to distill frontier capabilities.
Baker views the AI transition as an irreversible technological discontinuity akin to the post-1945 nuclear monopoly, balancing profound optimism for scientific and medical breakthroughs against warnings of geopolitical friction, executive safety risks, and the danger of capability access becoming restricted to elite institutions.
[25:00] George Vanderheiden: being early is the same thing as being wrong.
[55:00] Gavin Baker: If we do not all become masters of the machine gun, we're going to get mastered.
Original transcript
[0:00] Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all run on Ramp. Mine does too, and so should yours. Learn more at ramp.com slash invest. Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, context, and standards. Send Felix an email like, take these comments and turn them for me, or update my tracker with the context of these emails, or run the ability to pay math on this buyer, and Felix sends back finished PowerPoint decks, Excel models, and sourced research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai slash Felix. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Hello and welcome everyone I Patrick O and this is Invest Like the Best This show is an open exploration of markets ideas stories and strategies that will help you better invest both your time and your money If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus along with all of our podcasts at colossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum, All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Positive Sum may maintain positions in the securities discussed in this podcast. To learn more, visit PSUM.VC. My guest today is Gavin Baker, the founding partner and CIO of Atreides Management. and this is our sixth conversation. The central theme is watts and wafers, the two physical constraints that in Gavin's view will dictate the next phase of AI. On power, he thinks the near-term shortage starts to ease in 2027 and 28 as new sources of energy come online and that orbital compute helps solve this problem in the long term. On wafers he explains what is different this time from the dot bubble and why TSMC capacity decisions may be the single most important variable to watch We also discuss Elon Carafab the disaggregation of GPUs the role of new chip companies and whether economic value of AI will keep accruing to the frontier models. Please enjoy this awesome sixth conversation with Gavin Baker. All right, so this is our sixth time doing this, if you can believe it, which puts you back into first place, or at least tied first place with Gurley, back into steam territory. Always my favorite conversation about markets and everything going on. Even since last time when we did this, which was so exciting and spectacular, I think we're in an even more interesting time now. Maybe just start by riffing on how it felt for you living through March and April of this year, which felt to me just like a completely unique economic technology and market environment. And you're the biggest student of history and of these times. So what did it feel like? I would say broadly speaking, there are two kinds of drawdowns. There are drawdowns where you're wrong, company missed estimates, your hypothesis was invalidated, and you have to take your medicine and you crystallize that loss. And then there are drawdowns or periods of underperformance where you underperforming because of companies you know really really well and where you profoundly disagree with the price action and you can lean in And instead of crystallizing negative performance, you can kind of build pent-up alpha, pent-up future performance. And for me, that is what March felt like. The NASDAQ was selling off. At the same time, what was happening in AI was, I think, the most extraordinary moment in the history of capitalism, the history of American business. What I just mean by that is that Anthropic, they added $11 billion of ARR. And what is astonishing to me about this is that the SaaS and cloud revolution, it created, we'll call it between $5 and $10 trillion of value. I would say arguably the three highest profile SaaS companies in the last 10, 12 years are Palantir, Snowflake, and Databricks. and these three companies employ thousands of people, tens of thousands collectively. They've all spent 10 years building their businesses. And Anthropic added their combined businesses in one month. Nothing like that has ever happened in the history of capitalism. Forget my career, just the flat out history of capitalism, the history of business. It's wild. And then Krishna comes on this show and shares with
[5:00] That's 500% in DR. Yeah, you do the math on that for three years. So there's just no precedent for this. And we tech investors hear a lot of discussions about S curves and investing in exponentials. I've just never seen an exponential like this. It felt even more extreme than DeepSeek, which was a very similar setup. They happened at about the same time. If we go back to 25, there's a huge sell-off on DeepSeek, which was very strange because the paper gets published seven days before DeepSeek Monday. It got published, I believe, on a Monday that was a holiday in America. I read it. I thought, hmm, this feels like it might not be that positively for the AI trade. I took action. And then we had DeepSeek Monday, where AI really imploded a week later. That was really strange because by DeepSeek Monday, it was super clear that this was going to be the most positive thing that ever happened to compute demand. Prices of the AWS availability zones in Asia had already doubled You were seeing GPU availability go down And this was just the first time we saw how much more compute hungry reasoning models are during inference than non models And so that was a similar setup. You had to do some work to see that. I mean, it's not that hard to say, oh, wow, stocks are selling off. The price of DRAM is going vertical. The price of GPUs in measure going vertical, GPU availabilities going down. And then like two or three days later, GPU prices in America started going up, GPU rental prices. All you had to do in March was simply observe what was happening to Anthropic. And there's all these people who seem to regret not buying during 22, not buying during COVID, not buying during DeepSeek. You had the same valuation set up at the beginning of April and an even clearer AI inflection. So there have been all these chances to buy into AI. And then, of course, what complicated it was the straight and foremost. I became a believer and am a believer that I think maybe one thing that the market was mispricing I no macro expert I do do a lot of pro security investing So I do have access to people who are experts that are excited to share their thoughts and opinions with me And that the Strait of Hormuz being closed is actually relatively awesome for America. Why? Particularly for the goals of the current administration. So electricity is a very important industrial or manufacturing input. The key input into American electricity prices, which feeds into AI, is in G1, natural gas one on Bloomberg. That was down 20%. And natural gas in Asia, Europe, everywhere else doubled or tripled. Our relative manufacturing competitiveness improved overnight. And for better or worse, that is what the Trump administration seems to care about. They are very focused on America's relative position. And I think a lot of people had memories of the 1970s. What made the 70s so traumatic was it wasn't just that prices went up. It's that there were actual gas shortages. Then you go through, okay, well, the US economy is dramatically less energy intensive than it was. The United States is now the world largest producer of oil and gas And we become now the world largest exporter of oil and gas And on top of that there this relative manufacturing advantage that made it easier to stay focused on AI fundamentals, stay focused on what were historically attractive valuations. I think on a relative basis, tech essentially got as cheap as it's been versus the rest of the market has at any point over the last 10 years. And just think about that in the context of market efficiency. We have the most extraordinary moment in the history of capitalism that's wildly bullish for AI, and you get a chance to buy AI at really attractive valuation. What do you make of the multiples that specifically Anthropic and OpenAI, which in my mind are like the reference assets that are the most pure play takes on this trend, really being not that crazy? Like if you just look at the sales multiple and compare it to maybe what Databricks and Snowflake and these companies traded at their peak, how do you process it? How do you make sense of it? I do think OpenAI and Anthropic are pretty different animals from a capital efficiency perspective. And Anthropic clearly has a dramatically lower cost per token than OpenAI. They just do. And you could just see that in the amount of money that they have burned to get to a roughly similar range.
[10:00] I think they burned maybe 80% less than OpenAI. So as businesses, they clearly have very different structural ROICs. I think Sarah Fryer is one of the most exceptional CFOs. I think they're doing a lot of things to try to improve this. And they've secured a lot of compute. They've secured a lot of compute. That's another big difference. It turns out being aggressive really paid. Anthropic at $900 billion for $50 billion in ARR, growing at ridiculous rates. And I think maybe a true statement is that if Anthropic could just wave a magic wand and get all the compute they wanted, they'd probably be doing well north of $100 billion today, maybe $150. They have clearly deprecated the intelligence of Claude. There's an analysis. Claude is even on Opus is generating 70% less tokens for the exact same question. As we talked about last time, token quantity equals quality of answer and quality of thinking at some level. And there is an intelligence density per token that also matters. I felt that as a user. So I think they would be doing materially more, 100, 150, maybe 200 billion. So you might be buying it at more like five times unconstrained I going to make up a new number URR unconstrained revenue Why do you think they don raise billion at a trillion valuation or something like this? If you were the Anthropic CFO, Krishna's awesome, we just had him on. Or if you're Sarah, it seems to me like if the inbound I received following the Krishna episode is any indication everyone I've ever met is trying to invest in both these companies. I think it's wise. The future is uncertain. You are clearly in a very capital intensive game. Even if you are Anthropic, I'm sure is at very positive gross margins on inference today. I think Anthropic probably starts generating cash this year if they are not already generating cash, which I think is probably the case. But still, you probably want to be able to raise more capital, access more compute. The world is uncertain. Ukraine is starting to really, really win. How is Russia going to respond? I think there's still a lot of uncertainty in Iran. All this uncertainty, I think, probably amplifies geopolitical uncertainty over time. So it's an uncertain world If I think about Elon Elon has always made investors money He treats it like a sacred covenant And as a result because he made people money for now 20 years he has a superpower That is he can essentially raise as much capital as he wants whenever he wants. I do think being focused on making investors money is wise and creates benefits that don't just last for like a year or They can last for the next 20 to 30 years. And the way Elon did this was systematically underpricing SpaceX or whatever else. Like, what is the actual method? Just never being greedy on valuation, never pushing valuation. Just that simple. My friend Antonio pointed out SpaceX compounded it low 30% per year for a decade. And that was just because Elon was, I think, focused on preserving the superpower and having trying to strike a fair balance between investors and employees. I think it's wise, but could Anthropic raise money at probably at least a 100% premium to this rumored latest mark? Of course. Let's get to the Watson wafers part of the discussion. Always my favorite thing to talk about with you. The importance of this infrastructure build out Every time I feel like it getting overheated and then the next time I talk to you it seems like we should have done way more than we did You studied S and the steepness of those S a lot And you know a lot about history Talk us through how you're thinking about Watts and Wafers today as the key to inputs into this whole thing. I think capitalism is going to solve the Watts shortage absent big regulatory political blowback, which I think is a real possibility. The head of data center infra investing at one of the big PE firms, Blackstone Apollo, KKR said it used to be energy and chips were our biggest gating factors. Now it's zoning and approval, much more important. I think a lot of companies are waiting till after the midterms to take action in terms of maybe workforce reductions. Nobody wants to be a pinata during the midterms. You've seen a lot of companies that make turbines, announce a plans to significantly increase capacity. There's like two of these machines that can cast these big blades. We haven't made one in 80 years in the West. We don't know how to make them anymore. All of that is true. By no means am I minimizing the industrial engineering magic and artistry that goes into those, but capitalism is very good. It's
[15:00] problems like these over time. There's other sources of energy besides these turbines with a longer time frame. So I think the watts shortage will probably begin to alleviate 27, 28. And then I think orbital compute will really solve that. I do want to reframe orbital compute because I think when people hear data centers in space, which we discussed our last episode, They picture a Pentagon-sized building in space. They're like, well, we can't do that. That's not what it is. A Blackwell rack weighs 3,000 pounds. It's eight feet high. It's four feet deep, three feet wide. It's racks in space. And SpaceX has showed you an illustration. And it's a rack. That's the satellite. But it's probably about the size of a Blackwell rack. It has these solar wings that are probably 500 feet long on each side. You keep it in a sun-secretous orbit, so those solar panels are always in the sun. Because it in an exactly sun orbit the radiator which extends behind it for hundreds of feet is in the shape This is a common criticism yeah How are you going to go over here I spent a lot of time at Starbase over the years and I talked to a lot of SpaceX engineers and I do think it is the most talented group of engineers on planet Earth. And they're very confident they have solved this. And they're not always confident. There's some engineering that needs to happen to turn the Starship into a Mars colonial transporter. Will they do that? Absolutely. What are they more focused on? I'd say probably the repair and maintenance. These are the two big responses, the radiator and how do you repair whatever issue goes wrong in the rack? And the answer is, until you have probably floating optimuses, you don't. Starship is going to change the space economy in ways we cannot imagine. And particularly if regulation becomes a constraint to data centers, none of it's going to matter. You're going to sell as much orbital compute as you can make. And then obviously you link these racks using lasers traveling through vacuum, which are already on every Starlink. And it's just mind blowing to me that SpaceX operates the world largest satellite fleet which is 98 or 99 of all satellites in orbit Every Starlink they cooling it today I think Starlink V3 is going to operate at 20 kilowatts A Blackwell rack is only 100 kilowatts And people talk a lot about density Well if you're connecting the racks with lasers through vacuum, you can make the rack bigger. Physically, you're focused on weight, not size. In a data center on Earth, where you're trying to connect racks, ideally using copper, minimize lengths. Cabling is a big cost. You do want that rack to be small. Copper when you can, optics when you must. But in space, there's all sorts of things that SpaceX can do that I think maybe some of these naysayers are not contemplating. They operate more satellites than anyone. They have a 20 kilowatt satellite today. So maybe you just scale that up to 60 kilowatts to start. They seem very confident they're going to go right to 100 to 120. The same company now also operates the largest data center on Earth. They have the world's best hardware engineers and all sorts of people, almost all of whom are not smart enough or practical enough to work at SpaceX, are these armchair skeptics. I don want to quote Larry Ellison but somebody was being skeptical And Larry was just like listen he out there landing rockets I don see anybody else landing rockets And the reality is 10 years later no other company is consistently capable of landing and fully reusing an orbital rocket. None of this makes sense without reusability. That means you have to land it. I would like to redefine orbital compute has racks in space, not giant floating pentagon sized data centers in space. That's silly. What makes a data center is you're connecting these racks with lasers. So it'll be racks in space that are connected with lasers into a virtual data center. Vanta automates security and compliance for over 16,000 fast moving companies like Ramp, Cursor and Harvey, keeping them audit ready around the clock. It's the number one agentic trust platform, and it now helps companies like yours watch for the risks that show up between audits across your vendors, your AI tools, and your whole environment. Every new tool your team signs up for, every vendor that turns on AI features is an opportunity for something to go wrong, and most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24-7 GRC engineer in the background, finding issues, drafting fixes for you, and cutting vendor assessment time by up to 50%. Whether you're a fast-growing startup or a global enterprise, Vanta helps you earn
[20:00] improve trust. Invest like the best listeners get a special offer for $1,000 off at vanta.com slash invest. Ridgeline is the first end-to-end system of record with embedded AI for investment management firms running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared to anything else in investment management software, which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline's unified platform. If you're serious about your firm's AI strategy, Ridgeline should be part of that conversation. You can request a demo at ridgeline.ai. And if you think about that state of the world, let's say that all happens and we're really good at getting these things up economically and running matrix multiplication all over space. What does that mean for terrestrial data centers? Someone once said America was going to suck as hard as it can on every energy source it can get. And I just think the same is true of compute. It why I probably less worried about like an edge AI bear case than I was We going to consume as much compute as we can Inference, I think, is very sensible for Orvo compute. Training will be done on Earth for a long time. So I don't think that this is super bearish for terrestrial data centers. I think those are going to be valuable for my lifetime. But I do think if you're in this ecosystem of power production and cooling, and you are massively ramping capacity, a lot of these capacity ramps are going to be hitting, just as I think all of the silly skeptics start to understand that orbital compute is very real. I think it's worth thinking long and hard about that if you're one of those companies. And then all sorts of cool stuff is happening in the interim. We're getting really good at repurposing jet engines. There's that boomerow space that is doing this. Capitalism is hard at work on watts. On wafers, though, it's just this group of plenty older humans in Taiwan who are the most important humans in Taiwan, whatever they are. The overwhelming fraction of the country GDP water usage electricity usage They talk about the silicon shield They all view themselves as inheritors of Morris Chang sacred legacy Now, I vividly remember visiting Science Park more than 20 years ago and talking to them. Do you think you could catch Intel? And they said, this is such a beautiful dream, but it's a dream for our grandchildren. room. And they did it, partly because of Intel's self-inflicted wounds. They think very differently. One reason Jensen flies over there so much is he wants them to expand capacity. I do think it's wild that Jensen has never had a contract with Taiwan Semi. They do business on what seems fair in handshakes. Just fascinating. No contract. It's going to be fair over time. We're partners, we're going to be fair to each other. The truth is, based on every prior market precedent for a foundational new technology like AI, you've always had a bubble. Carlotta Perez wrote this great book about this. Markets are efficient. They correctly understand that this is a foundational new technology. There's what Mobison calls a breakdown in diversity. Everyone becomes bullish on this new technology. And I am beginning to worry a little bit about a diversity breakdown And then you get a bubble That bubble funds the build out of this new technology but supply gets ahead of demand and you get a crash. And it's a particularly severe crash if it's a debt-fueled build out like the year 2000. And one thing that's really good about the current build out is it's still overwhelmingly funded out of operating cash flows, which is a really important fundamental difference versus the year 2000. has is valuation, has is the fact that every GPU is running at 100% utilization when 99% of fiber was unutilized. So there's all these fundamental differences. History doesn't repeat, but it rhymes. And as an investor, we have to be very cognizant of it and recognize that based on the last two or 300 years, forget the internet bubble. We had a railroad bubble, a canal bubble, every kind of bubble, South Sea bubble. We should expect a bubble. That's terrifying. Nobody wants a bubble. The reason it's terrible is if you're valuation sensitive, you like massively underperform. You get fired by probably all your clients. George Vander Heiden, who is no longer with us, Great Fidelity Portfolio Manager, he fought the bubble in 99 and he retired in early 2000 because I think he just couldn't take it. He knew it was wrong. His clients were deeply skeptical. George, you're out of
[25:00] app. He had white hair. He's a truly great man. I only overlapped with him briefly, but he was a very important mentor and friend to my good friend and mentor, Jennifer Urich. So I have a lot of Vander Heiden DNA through her. He was the same person who said being early is the same thing as being wrong. George retires because he can't take the underperformance and he can't take clients saying, what's wrong with you? You don't get it. And he has like 40% of his funded tobacco, So 40% did home builders. And literally, he probably outperformed the NASDAQ by like 20 or 30x over the next three years. And I have been optimistic that this fundamental shortage of wafers, which really today is controlled by Taiwan Semi, will prevent one. If Taiwan Semi did what Jensen wanted, I think NVIDIA could sell $2 trillion of GPUs in 26 or 27, maybe $2.5 trillion. maybe $3 trillion. But there is a limit where consumers would consume so much, they probably would be in an overbuild. So Taiwan Semi, if we don't get a bubble, we need to throw a party for them because they will have single prevented a bubble You are starting to see companies go to Intel and Samsung Let's just assume TSM stays super supply constrained versus blatant demand. What happened? The history of markets is, I don't know who, but one of Intel and Samsung, they're not going to stay disciplined. They will break. And then at some level, that will force everyone else to break. I think a lot of this may come down to the degree to which Taiwan Semi can maintain a lead over Intel and Samsung. You got to remember, it's whatever it is. It's 9, 12, 15 months. The leading node adjumine. Exactly. The pace at which they expand capacity. If I were to watch one thing to understand where there's a bubble, it's Taiwan Semi's capacity decisions. And I think there's a Goldilocks zone where they expand enough. They make it hard for Intel or Samsung to really truly emerge as like a at scale second source with something well north of 30% market share. And yet they also keep this fundamental constraint on wafers that helps us avoid a bubble And then obviously I think the TerraFab is going to play into this too Say more about that It a SpaceX I believe Tesla involved as well joint venture to build the world largest fab here in America. I think they're going to be successful. One, they have a partnership with Intel, which is very important because they're getting access to 50 years of institutional knowledge that's just nine months, a few quarters, 12 months, three to five quarters behind the front. That's an advantage. It's also an advantage that I believe the TerraFab is going to get attention from the A-teams, all the semi-cap equipment companies. One big reason Taiwan semi-caught up is ASML and KLA 10 core and LAM research and applied materials. They wanted them to catch up. They don't like having a monopsony. The A-teams were in Taiwan working until made some mistakes and presto. So the A-teams will be here because of Elon's reputation in hardware engineering. And then to a degree that I think is maybe hard for people to imagine in America, where politics has replaced religion, I think because Elon had his foray into politics, it makes it hard for some people in America to see him clearly which is sad because I do think he probably doing more for America than any other American He single bringing manufacturing back to America He revived defense tech I think SpaceX is in some ways the most important defense contractor in America. He's doing a Starlink. It's amazing for the world. He's creating all these blue-collar manufacturing jobs, which is like a goal, I think, of a lot of liberals and good for America. He's done more than any living human to decarbonize the world. And if you are upset about data sitters on Earth for environmental reasons, well, here you go. It's sad, but he is a living deity in China, Taiwan, South Korea, and Japan. Having watched him for a long time, what he's going to do is they're going to recruit the best people because the best engineers want to work for Elon, especially in hardware engineering. He's going to recruit incredible engineers. Next to the TerraFab, there'll be a Taiwan town. Oh, these are your favorite restaurants? I'm going to move them and their whole staff from Taiwan to Texas. And we're going to make everything the way they like it. And then we'll have Japan town. Same thing. Then we're going to have Korea town. We're going to have all these things exactly dialed to recruit the best
[30:00] engineers. And that's just not the way that the people who run Intel at Siam Sung think. So he's going to have the best talent. He's going to have the A-teams at the Wafer Fab equipment companies. He has Intel, which is important. It's so good for all of any administration's political goals. And I think it's different enough that it will not alienate Taiwan's semi. And these have long lead times, right? TerraFab is going to be pumping out whatever GPs, whatever chips, quite a long time from now. We'll see. Elon tends to do things differently. Everybody else has taken three years to build a data center. He built one in 122 days. Sam Saga had to give him an office in their fab in Texas because he was so unhappy about the pace at which they're expanding and building. Are you surprised by, you mentioned DeepSeek earlier, the simple reaction to that was, okay, these models are just going to get 95% as effective for some tiny fraction of the cost to still Chinese open source models. We'll be able to use these for most of what we want to do. Fast forwarded a little bit of time, two years from now, there's no reason I have to spend a million dollars a year in my small little firm on tokens or something But then the actual reality seems quite different than this And I curious why there that dissonance in your mind I do think it fascinating the returns to the frontier all the economic returns to AI at the model layer, not all of them, but an overwhelming amount of them have been at the frontier, which is surprising to me. And I think it's been surprising to a lot of people. This is one of the most important questions to be answered, and you need to have a hypothesis on it as an investor. Are frontier tokens going to continue capturing the overwhelming majority of economic value created at the model layer? And it is surprising. I remember when Gemini 3.1 Pro came out. It was mind-blowing to me. It was so good. Today, it's intolerable. There's probably a little bit of a dynamic where companies prototype with frontiers, then when they put something into production. You're hearing a lot of people do use Vertex or open source. But still, it is a fact today that the overwhelming majority of these economic returns come from frontier tokens. And that's surprising. And whether or not it continues, I think, is a very interesting question. And I'm much more open-minded to that, having had the experience I've had with Gemini My 3 and then Opus And then I do use Grok 4 a lot It is on the Pareto frontier The companies that are on the Pareto frontier are and this is by the way a big change in a consequence of what we talked about last time, Google losing their per-cost token leadership as a result of making very conservative design decisions with TPUB8 to try and take it away partially from Broadcom and NVIDIA continuing to make aggressive choices. But Google dominated the Pareto frontier, the Pareto frontier being intelligence versus cost. And I think this is the most important thing to look at to analyze AI labs. Google dominated that nine months ago. Every point on the Pareto frontier, OpenAI, XAI, and Anthropic were inside of them. Now, the Pareto frontier is dominated by Anthropic, OpenAI, and then Grok 4.3 is on the Pareto frontier. It's clearly the best, lowest cost, 500 billion parameter model. And then Gemini 3.1 is hanging on to the Pareto frontier. And if I were to bet or bet that they're subsidizing that out of pride, I would just say a violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade, to all of AI. The closer someone is to AI, the more skeptical they are that this will occur. One thing I think contributed to weakness in March was a much more stupid version of DeepSeek which is this thing called TurboQuant And TurboQuant is some Google memory optimization that was written up in a paper a year ago And then during the middle of an agreement, while Google was negotiating with Micron, Samsung, and Hynix to sign some LTA that would lock in really high prices for a long time, they released this. What people do is always more important than they say, and they just kind of publicize it on X. and it goes viral. Like, oh my God, DRAM is cooked. Here's this DRAM optimization. I was unable to find a single AI engineer on planet Earth who believed that TurboQuant would have any impact on DRAM demand. But nonetheless, a violation of Richard Hutton's bitter lesson, more compute will always outperform human algorithmic ingenuity, more compute and data, chinchilla optimal, beyond chinchilla optimal, I guess what people increasingly do today. it. That's a real risk, man. The people who are building these models are skeptical of that risk. The reason I am a little less skeptical is I think we're very close to ASI. And who knows if the bitter lesson holds for 400 IQ models. Maybe we get a temporary period where these, you know, if you get to ASI, the first thing it wants is probably
[35:00] to be smarter and have more resources. How does it do that? It makes itself more efficient. I think that is an actual risk. The bitter lesson literally, I believe, includes humans in it. So we're about to find out whether the bitter lesson, we'll find out if it applies to 300 IQ AI, then 400, then 500 and 600. And at some point, we may have a temporary violation of the bitter lesson based upon AI and ASI. So I'm curious how you think about some other parts of the innovation around the model, continual learning and memory being two that people seem to be most focused on as things that might create yet another new paradigm that we would enter. What do you think about the role of those two things? Yeah, well, I think we've done a lot with memory through these harnesses. And it turns out that harness engineering is not as important as the model, but it really matters. And these harnesses in these models are increasingly being co-developed. One of the big things a harness does, what you just think of is like a runtime that the model operates in And it knows where the tools are It creates context memory state has very specific prompts or instructions It a huge difference even simple versions It makes an incredible difference I think the last time I was on here one of the other times I just said like hey as an investor, it's very important that you pay for the $250 a month version to get your own intuitive sense. That's no longer possible. To understand what Frontier AI is capable of today, even for a non-coding use case, you need to have cloud code or codex, and you need to be on an enterprise plan. And the reason for this is, and this is another dynamic that's enabled by Google losing their cost leadership, is these AI models just shifted to usage-based pricing. And if you're on that $250 or $300 or $280 a month plan or whatever it is, you are getting severely rate limited. You are getting a lobotomized version of the AI. Because like we talked about, Claude now produces 70% less tokens. You want the tokens that Claude and its harness really think it needs to produce to get you a good answer? You need to be on a usage-based plan. And by the way, this is so bullish for AI. If we go back to 05 to 07 cellular had been a great growth industry really for the last 10 years And the reason was you had a combination of fixed pricing You had 900 minutes for whatever it was and then usage-based pricing over that. When did cellular stop being a great growth industry? When everybody just went to all you can eat. And by the way, long distance is the same thing. AI is just shifting from all you can eat to pay by the drink. And it turns out people really like to talk to their friends long distance. They really like to talk to their friends on the phone. and people really like to use AI. And particularly now that one person can have 100 agents working. So I think the shift to usage-based pricing is probably why you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year. Not only is MoreCompute going to become online, but they're going to be able to push frontier token pricing with these usage enterprise models. It's sad. It's sad for the world because it just means if you can't afford that, that you're not at the frontier. And I think it's gonna throw off a lot of investors' intuitive sense of the capabilities of AI. But yeah, continual learning, man, I mean, if we solve that. How do you conceptualize that? AI is constantly updating its weights. I mean, it may end up being something different. There so many mysteries about the human mind We such sample efficient learners relative to AI Many orders of magnitude Now we have a crude variant of continual learning today when something is verifiable and that's just reinforcement learning during mid-training. Continual learning is a model that dynamically adjusts its weights or adjusts in some way in real time. Like as a human, the first time I put my hand in a fire, I've learned I never put it in there before. That model today needs to put its hand in the fire a million times and then have the designers effectively put a fire in the next training run or an RL gym for it to learn. I think it has to be dynamically updating the weights, but I think people are working on really smart techniques beyond this. But if we get that, then we have a really fast takeoff and people seem confident that continual learning is kind of just around the corner. And I do think this is like the third big question. Bitter lesson violation as a result of ASI are less likely human ingenuity. Will Frontier tokens still command the premium they do? And will we get continual learning? And if so, win? What is the role of new chip companies?
[40:00] in all of this. We talked a lot about NVIDIA, their relationship with the SMC and Intel and all these sorts of things. There's a thousand flowers blooming, I think literally probably a thousand flowers blooming, trying to create a new chip to address some part of this bottleneck. I'm curious how you process this space, this opportunity, what role it will play. So I think this is good and healthy for the world. It's good for Jensen too, because a different administration might take a different view. Competition, I think, is good for everyone and seeing different architectures explored is good. And the reason is in tank design, they talk about the iron triangle. The iron triangle of tank design is that all designers of a tank, they have to make trade-offs between attack, defense, and mobility for obvious reasons. The more defense you have, which is just armor, the heavier the tank is, the less mobile it is. So you have to live in this triangle and make trade-offs. The Merkava in Israel is optimized for defense. Russian tanks, and like the Leopard, are generally more optimized for mobility. Chip design is the same. There are these fundamental constraints imposed by the laws of physics as embedded in the Taiwan semi rules that you need to live within You have TPU Tranium and AMD which are all essentially trying to be a better GPU And today I think probably Tranium is doing the best Now nobody's a better GPU, but Tranium is tugging on Superman's cape. That hadn't started yet. The Tranium 3 needs to ramp into production because it has a switch scale-up network, what you really need to economically inference MOE models. A lot of companies have a Taurus architecture. That's where Google was. Google's developing a switch scale network. And then AMD is like always kind of flying from behind. Yeah. AMD, we'll see. The MI450, we don't know yet. We'll see. We probably know more about Trinium 3 than the MI450. But that's a hard game to play. So you have to do something different. And you have to do something different that is also hard to do. So I think the best path for these startups, my rule of thumb is 1% market share is going to be worth $100 billion. $100 billion is a pretty good venture outcome. I think what Jensen would say is, okay, if somebody does something different and it gets to a 1% or 2% or 3% share, we'll make that chip. And that's coming for everyone. But if you trying to make a better GPU good luck If you are doing something different it also needs to be hard to do And you can make different trade The disaggregation of pre-fill and inference really have opened the aperture for making these different trade-offs because you can make very aggressive trade-offs for decode, aggressive trade-offs for pre-fill. Pre-fill being taking in the context, decode being, you know, write the output. Yeah. I have a great colleague named Andrew Fox who said, pre-fill, picture British naval ship from the 18th century. Pre-fill is loading the cannon, decode is firing. And what pre-fill literally is, is just the model understanding the question, the prompt, and then keeping track of its own answer. And that is fundamentally a memory capacity bound problem. Decode is the process of generating new tokens, and that is memory bandwidth constraint. So if you're a chip designer, this gives you a richer canvas to paint on. But even so, it needs to be hard because if you make different trade-offs in that iron triangle to optimize for memory capacity, and they're not hard trade-offs to make, NVIDIA is going to make those same trade-offs. They get better prices from Taiwan Semi than you're ever going to get. Good luck. And they have the advantage of working with every model company and optimizing their designs By the way another very funny thing is there this process if you a VC and you investing in a semiconductor company that is telling you they are going to have an advantage because of a Taiwan Semi process that they have special access to, I promise you, the Jensen's saw that process when it was a twinkle in Taiwan Semi's eyes. They know more about it than this little company with 200 people can imagine. Taiwan Sibi, everybody in the supply chain is showing Jensen everything. The same way they're showing Amazon everything, AMD everything, TPU everything. And that's another reason. Don't go try to make a better GPU. So you can do something different. You can paint in the pre-fill canvas. You can paint in the decode canvas. But you also have to do something hard because if it gets to scale, you're going to have those four companies has very fast followers. My firm was a venture investor in Cerebris. What Cerebris has done is something hard and fundamentally different. Wafer scale computing. It comes with a set of trade-offs, but that architectural decision they made was hard and lets them do something that no one else can do. And we'll find out how big that is. They're working on really cool things. One of the problems Cerebris
[45:00] Cerebrus has is once you start needing to glue a lot of chips together and scale up networks or scale out networks, you need a lot of IO. And IO is bound by what's called the shoreline, the sides of the chip. Cerebrus has an overwhelming ratio of on-chip computed memory relative to shoreline IO. Well, they're really smart people. They did something really hard. They're trying to see if they can put an optical wafer right on top of that. And then that solves that problem. I'm sure they're looking at hybrid bonding of DRAM to get around the much-discussed on X, these alleged limitations that are not true. A Cerebrus machine can theoretically run any size model. There are sizes of models where they're much better than other sizes. So Cerebrus, what I think is interesting is they did something different that's hard to do, really hard to do, wafer-scale computing. I do think there's a role for these. I would just encourage them all, make a different trade-off, try and do something hard. everybody's going to get funded after the Cerebrus IPO. It's not going to be a problem, but it took Cerebrus three generations of chips to get it right. Andrew Feldman, the CEO you can just see how hard it was what he did and that whole team did to get where they are today And they need to have the grit to do that the resilience This first chip is a failure. It happens. Can you come back and make a second chip? But the one last thing on this topic is this is going to be amazing for the useful lives of GPUs and may single-handedly save private credit. Say more about that. What do you mean by the private credit? Just private credit. They're in pain from these SaaS loans. And however much they're marked down, they probably need to be marked down more. Because if the public companies are struggling to adapt, how's like a debt-laden company going to adapt and invest in what is a very different margin structure business? There's a lot of private credit in GPUs too, and they were underwriting that to, I think, three or four years. The disaggregation of inference means that I think these GPUs are going to have 10 or 15-year lives. The AI skeptics are like, oh, these companies are all cooking their books. The useful life of GPU is only a year or two. The useful life of a CPU is only four years because of the rapid technological change. No. What rapid technological change has done with the disaggregation of pre-fill and inference is mean that you can put a Cerebra system or Grok LPUs that NVIDIA acquired effectively in front of a hopper or even an Ampere Use that hopper and Ampere for pre and extend the useful life of that GPU until it melts They do melt so they have a time but maybe you don have to run them as fast. This is going to be really good for the whole private credit industry. It's going to help finance the AI buildout because if you can start to finance GPUs at more like 5% or 6% instead of, I think Corby's lowest financing was like low sevens, that actually mathematically changes the cost to finance this build out. We had this technological innovation that's going to lower the cost of financing, extend the useful life of compute on earth. And then I do think the one last thing that's interesting about that is my friend Jamin from Cotu just did a podcast and Cotu had a deck and they talked about, hey, sellers of shortage are doing so much better than the buyers of shortage, buyers shortage being the hyperscalers. But if you own a giant installed base of what is currently in shortage, that's also a very good place to be. And we're hearing CPUs are way more important than they were in an agentic world. They do all these things around orchestration, tool calls. The biggest CPU fleets in the world sit at the hyperscalers. Some of these hyperscalers may catch up a little bit to the sellers of shortage. I want to talk about this idea of different and hard applied outside of the infrastructure piece of this Now you starting to interact with new founders existing CEOs and founders that have to adjust to this new world What are you seeing the most AI native founders that aren't building chips or infrastructure or models, but just people using this technology to build other stuff? How do they feel the most different to you if you've observed differences? I do think this is just for chip design. To me, it's always been a fundamental question for venture. So there are different ideas that are obvious to everyone on planet earth as soon as they hear it. And if that's where you are in venture, if it's not hard to do, if it becomes obvious to the world before you have built scale, scale is the ultimate advantage. You're in trouble. And the great thing Amazon had was, I think it was obvious to a lot of people, but it wasn't obvious to the retail CEO. Amazon, they were very smart. Any e-commerce company that VCs invested in, they would destroy. They'd be like, oh, that's so cute. We're going to take our margins of that to negative 10,000%. And that's like the guys at Wayfair, they did something hard. Amazon tried to kill them and they failed. Those were like tough, operationally, really competent CEOs. For me in venture, I always look, is this going to be obvious to the world before this company could build scale? Or is this both not obvious, different, and really hard to
[50:00] I think a lot of founders are really struggling with this in AI. I think people are becoming worried today in Jensen's five-layer cake of AI. The profits are accruing to energy, accruing to data centers, accruing to chips, they're accruing to models, not really accruing to the applications. I think cursor and cognition got to a scale. They focused on coding. 18 months ago, the people were focusing on coding. OpenAI was doing everything under the sun. The people focused on coding were cursor, cognition, anthropic. And it was really right to focus on code. Amjad Massad, the founder of Replit, tweeted something that I thought was so smart. Just something like, bitter lesson adjacent is the fact that coding might be the shortest path to ASI and useful AI. Because if you're really good at coding, you can write yourself code to do anything. So I think it was really smart of those companies to focus intensely on coding. They all probably got to a scale where they have a place. I think cognition is doing something really, really different. But I think a lot of founders are really struggling, man. I think they're trying to get confidence that in nichier areas that they can get to them and get like a data moat before the model companies get to that niche Or that it a small enough niche that the model companies won do it themselves but it can still produce a venture outcome Is this related to what you would call like the token path? I know you've used that phrase with me before. Yeah, I think it comes from a guy, an altimeter, Jamin Ball. But he just said, if you're a software company or an AI company of any kind, you have to be in the token path. So Databricks, that's in the token path. Comparable companies are in the token path. If you're not in the token path and you're not in some really niche thing, life may be hard. And even for these vertical niches, I think if you talk to the people at the model companies, they're even skeptical of some of these because all of the data that's being generated in these niches come from humans. But then you're betting that you're able to use that proprietary data in this narrow vertical to train a model that's lower cost than the frontier labs can ever get to. And maybe that's a good bet, but I just think you have to be very, very careful. Now, on the other hand, if the returns to these frontier tokens relative to other tokens come down, there going to be an explosion in value creation at the application layer And I think another really important point is I have a belief that whenever he wants Jensen can probably get pretty close to the frontier With his own model. With his own model. I don't think he wants to do that, but that is what OpenAI and Anthropic are kind of trying to do to him unsuccessfully. He's a very logical thinker. This is the logical counter move. You will see that open source frontier, which today consists of Chinese models with stolen American tokens. Somebody told me that like DeepSeek, maybe the original one was only 150,000 reasoning traces. There's many ways to launder this if you're a Chinese company. You can hit all these different APIs. You can make it hard. Now, the American labs are working really hard on anti-distillation technology. But I just think Chinese open source, they're doing really impressive things in a very resource-constrained way, but there's a lot of distillation. And this is why I think in addition to there not being enough compute to serve Mythos, they did not want it to be distilled. They wanted to use Mythos, distill it themselves, use it to RL their next model whatever it is And then I think what they and eventually I think OpenAI anyone on the frontier will do is just say there going to be some very interesting game theory because it a new kind of prisoner's dilemma. We talked about the old prisoner's dilemma being just around like, hey, you're in a prisoner's dilemma where you have to spend. The new prisoner's dilemma is going to be if you are at the frontier, do you release that model via API or not? If everyone at the frontier Tier agrees not to do that. And in Chinese open source, if one person defects, they're going to have the best model. They're going to have a lot of revenue and cash flow. And then, of course, resources equal intelligence. So they'll start to pull ahead and then that will lead to everybody else releasing it. So it's a new game theory. It's kind of the same game theory that you have with Taiwan, Simi, Samsung, and Intel. Well, the reality is if a company like NVIDIA or AMD were to ever really, really use one of these other foundries, that foundry would get better really quickly. So I do think Jensen is going to keep open source a certain time frame behind the frontier. I think that's going to be a very interesting thing to watch. And then by the way, open source gets monetized. There's this misnomer that open source is free. Open source tokens, they cost energy to produce. You need to make up on GPUs. and the open source model companies almost always get a revenue share.
[55:00] Are you preparing Atreides for the world of Mythos 3, Mythos 4? We're just trying to overinvest in cybersecurity. And I really believe everybody needs to have a safe word. Everybody needs to go leave your digital devices behind, literally go to the ocean and have a family safe word or a company safe word. And it can't be one that can be like socially engineered. And this is just to avoid cybercrime where what looks like your son or your daughter or your grandparents or your parents or whatever, FaceTimes you. It's an utterly accurate simulation of them. They know everything and can extrapolate based on what they're likely to say and says, you know, wire me a million bucks. So doing everything we can with cybersecurity. That's defensive. What about analytical or processing? What will you still be able to do that it won't be able to do, I guess, on the analytical side? So it's a good question. I just watched The Last Samurai and I asked people at my firm to watch it. And The Last Samurai if you haven seen it I highly recommend watching it It actually a movie that aged really well It a Tom Cruise movie from 20 years ago The conceit is Tom Cruise is this bitter washed up Civil War veteran who actually a very good soldier And he bitter and washed up because he feels like he participated in negative actions against the Native Americans It during the Miji Restoration And he hired by the modern elements of the Japanese government to train like an army of peasants how to fight the samurai. There's a first battle. Of course, the samurai win, even though they don't have guns. He fights valiantly. So the samurai decide not to kill him, take him to their village. He becomes a samurai. It feels like the Civil War to him. So he fights on the side of the samurai. at the end of it, he's massacred by a peasant with a machine gun. The machine gun is here. If we do not all become masters of the machine gun, we're going to get mastered. So I am trying to become a master of the machine gun. And then I'm optimistic. There's a long period of time where just like if you were a 50-year-old samurai veteran of many wars, I fought many wars, Master Dwarf, you will have advantages using the machine gun. I'm optimistic as a lifelong student of investing. I'm going to be able to master the machine gun, this new technology, integrate it into my own process, integrate it into our firm's process in ways that let me contribute value as a human being for a long time. Like everyone, I have agents running all the time now What your most useful agent My single most useful agent is a really good summary of the points that would be interesting to me from podcasts There's just six hours a day of stuff that I feel like it's in my job description to watch. Every time somebody from OpenAI, XAI, Google, Cursor, Fireworks, Base 10, I'm going to say nothing of Jensen, Elon, Dario. I feel compelled to watch and I just don't have that much time. And there's some real needles in haystacks. That is what I would say for me is the most useful. I do think there's a set of things that I always like to see. Like I'm very sensitive to management compensation. What are they insented to do? Do they just have stupid RSUs or do they have PSUs? And if they have PSUs, what are those PSUs and sent them to do? And we now have systems that do a very good first pass at that. That saves people a lot of time and frees them up for more creative work than like going through the proxy, pulling the PSU thing, looking at how it's changed versus all the proxies because there's signal in that. And that's very labor intensive and that's so good for an AI And there obviously all sorts of same things within investing Pressuring the organization in those ways I think has been helpful This is the most exciting thrilling time to be an investor I'm getting a little bit worried. The diversity breakdown thing? Say just like a little bit more about the kinds of people that are... I don't know anyone like me who's not really bullish on DRAM. There's all these interesting things happening with AI right now. One is cross-sectionally, the valuations do not make sense. They just flat out do not make sense. They cannot all be true. In other words, you have semi-cap equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid-single digit. At the peak of the last cycle, that was 5 versus 12. At one point, it was like 3 versus 45. Those can't both be true. And yes, semiconductor capex business models have improved more than the memory business models. We don't know how much HBM is going to improve memory business models yet. Yes, they have some element of recurring revenue with parts and maintenance, but it's not worth a thousand percent multiple gap. I think it's hard to square the valuation of something like NVIDIA, which is still in early April is essentially as cheap as it gets relative to the market, like in the last 10 or 12 years or whatever it is, and very cheap absolute. It's very hard to square that valuation with something like GE Verno.
[60:00] is valuation because it builds in an unfathomable amount of share loss for NVIDIA. So valuations cross-sectionally are really different. Because we are in shortages, the lowest quality companies are doing the best. So if you're an oil and gas investor, a mining investor, natural resources investor, and you're well-versed in thinking of costs, this is very intuitive to you. And a real bull market for a commodity. The commodity suppliers with the highest costs go up the most because it's the most beneficial to them. They go from on the verge of bankruptcy to gushing cash. And this is, I think, one reason commodity investing is really, really hard because quality outperforms during the cycles, but you get all of the outperformance during the downturns when the high cost guys that mooned during the shortages and the commodity bull markets go bankrupt or whatever. You're seeing that happen in every industry. The lowest quality players, companies that are hated and detested by the hyperscalers and the buyers because they have high costs, they're unreliable, the parts fail at a high rate, they're sold out and raising prices And then that activity gets the interest of these retail accounts on X and these stocks get bid to the moon whereas some of the higher quality expressions have actually really underperformed As an investor it's hard because you know, within a shadow of a doubt, that that thing that's mooned 10X in three months or six months is going to go right back down, subject to what they do with all the cash. And so it worries me a little bit that people who are very skeptical a year ago are no longer skeptical. But then I just contrast that with the valuations of these high-quality companies, which are just not extended. And it makes me feel better. I always thought it was funny in 24 and 25 that anyone asked about an AI bubble or talked about it. Do you have this nuclear bubble and this quantum bubble right here, right in front of you? What are we talking about? This is so real. Some of that nuclear quantum silliness has maybe spread into more speculative, lower quality, smaller cap names, where if you have a big presence on X or Reddit, it's easy to move them. And that frightens me a little bit. But I just wish there were more AI bears. I wish there were more memory bears. Astera is a stock I've been close to a long time. There a lot of bears on that I love that Great I first invested in the Series C Good luck thinking that a copper loser And then there also you can feel the baskets in the market and the leverage baskets And what baskets you're in is really important. You know, copper, optical, DRAM, NAND. And a very interesting thing that's happened this year is in 24 and 25, the AI trade traded together. You could be long GPU compute, scale-up networking, and optical scale across, and short power or whatever it was. That trade worked from a risk management sense because I'm very factor aware. That all blew out in January of this year. Scale-up networking would go crazy while scale-out was going down, or DRAMs massively underperforming NAND and HDDs, which had not happened. And so these cross-sectional correlations within AI really fell apart and you had to get very fine grained. You couldn't hedge your memory anymore with some semi-cap equipment or NAND. Everything cross-sectionally really changed in a very interesting way in January. And I think maybe one reason for that was AI got to a quality where it was all of a really easy for a bunch of people to get really smart on these different subsectors start trading them and then they get put into baskets And those baskets- Yeah, creating price efficiency. Yeah, exactly. I think some of the biggest opportunities outside of these higher quality names that I think can compound for a long time, and they're safe, unlike these low quality names, which are terrifying, is in names that are miscategorized. Estera was in a lot of copper loser baskets. Estera, their biggest product is going to be a switch. You use both copper and optics to connect switches to accelerators. Definitionally, if you're a switch company or an accelerator company, you cannot be a copper loser because you're going to be on the other side of that connection. I wonder if you could riff just for like a sentence or two on each of the major companies, Google, Microsoft, Amazon, the major players that are public. All the conversation is centered around these exciting new companies. Maybe run through them in Riff. Google was incredible last year because they had that TPU advantage, which is now gone. The reason I think they're still in a great position is just they have the most compute of everyone. We talked about the value of installed bases being higher as a result of shortages. They have the biggest installed base of compute. Google I.O. is this week. If they don't release something
[65:00] that even slightly leapfrogs open AI and or clawed. That's interesting. And it's not a disaster for Google. It's just interesting. And it just means this NVIDIA effect we discussed is even more powerful than maybe I'd imagined. But I'm very curious to see what the Pareto frontier looks like literally in five days after Google's announced its new stuff. This is a big card for them. But Google, between the amount of data they have And the YouTube data is actually really genuinely valuable. It is valuable in a world of robotics. The amount of compute they have, the search business they have, Google's never not going to be in a good position. And then you see that with GCP going crazy. You got to give Zuckerberg a mitts credit, what he's done in terms of making Meta an AI first company internally. And I do think he is the only one of those true internet giants to have done that. I give him a lot of credit for that. I give him a lot of credit for paying up when he did for the billion dollar contracts, the talent. And Muse, I think, was a really big upside surprise. It was the first model from MSL. And it not on the Pareto frontier with XAI Google one entrant and then OpenAI and Cloud but it pretty close That was very impressive to me So I think Meta is in a better position still not as strong of an absolute position as Google, but like their better position and rates of change matter more than level, as you know, in markets, particularly over short three-year timeframes, over long timeframes, level of competitive advantages tends to dominate. But even within that changes really matter. Amazon, I think, is in a really strong position because of Tranium. I do think you're going to see real P&L efficiencies from robotics over the next 18 months in their retail business. I actually think Nova, their internal models are not where Muse is, but they're better than they get credit for. Then Microsoft, I like Satya. I admire him. I think he's an exceptional CEO. And I give him a lot of credit for the decisions he's made. But He did go from, we're going to make Google dance to being the product manager of Copilot in like three years. I would love to know during the coup attempt against OpenAI, does Satya regret his decisions? Does Satya wish that he had supported Ilya instead of Sam and that Ilya and Mira were really running OpenAI today In his heart of hearts I would love to know Because I think the Microsoft OpenAI partnership might look very different in that world I think that's a very interesting question that we'll never know the answer to. But I give him a lot of credit. What he is doing now, he's taking risk. This goes to the decisions you have to make in that cone of uncertainty, or not only how much you spend, but what you're going to spend it on. I think Microsoft flinched for like a moment in early 25. They have this algorithm. We spend this much CapEx dollars. We get this return. That algorithm was kind of off. And if you flinch, you lose position. You lose all these allocations. And it's difficult to get it back. So they flinched. And now the decision Satya is making, which the market has punished him for, but I think is the right decision. I mean, who knows how fast Azure could be growing if they're willing to just sell GPUs to open AI. We're going to use our compute internally to make our own products better. One reason Copilot is so bad or has been so bad is just one enough compute available. They're fixing that. He's the product manager of Copilot. I do think he's a great CEO. They're trying to use their compute to train their own models I am a little skeptical that they have the right team to succeed there but just like Meta they can afford to hire maybe a different team But I think he making good decisions that are risky decisions to position Microsoft for this world where frontier models are no longer API accessible. And I think it's a really courageous decision that I give him a lot of credit for. And he is foregoing. I mean, Microsoft probably be an $800 stock today if they were using their GPUs to serve solely open AI and anthropics capacity instead of using them for their own products. So I give him a lot of credit for making a great decision. I think what's really interesting is the degree to which these companies are outward facing in their decisions. The two companies who are the most deeply engaged with startups are Amazon and NVIDIA by a mile. Then there's a really intense engagement with Google, their next most intense. Broadcom is engaged in a different way. They're just everybody's favorite ASIC supplier. If you're a startup, it's considered like a level up if you get to work with Broadcom for your second-gen chip. And it's considered manna from heaven if Broadcom works with you for their first-gen chip. And then you see essentially zero engagement with startups from AMD, AMD.
[70:00] Microsoft and Meta. When I say zero, it's a little. And I just wonder about that decision because some of the best teams are no longer at big public companies. They're at these smaller startups. And I think it's going to end up being a pretty big advantage for NVIDIA, AMD, Google, right behind them to have this engagement that you just don't see from these other hyperscalers. As we wrap up, I'm curious for you to riff on any other out there knock on effects that you've started to think about for this giant trend. We've talked about the specific companies in a lot of detail that this most impacts. We talked a little bit about the application layer and what would have to happen for there to be more value occurring to that layer of the stack. I'm curious, any other just fun knock on things that you've been thinking about as this world changes so quickly? And it is wild. I mean, at the application layer, forget value accruing, just value has been destroyed. AI has net destroyed. Even if you count cursor cognition, the most successful AI natives, trillions of dollars of value has been destroyed by AI at the application layer. And just in this context the companies that are doing the best today that are seeing their values increase the most that are creating economic value are the companies with the highest effective ratio of utilized GPUs per human Maybe this just means that every human is going to get a lot of GPUs, but I think that's an interesting fact that we kind of need to be cognizant of. I will just say, and maybe this is a little dark, I am more and more worried about personal safety. And I worry about this a lot more for people who have a much bigger public presence and are much more associated with AI. I hope nothing tragic happens. There is this upsurge in political violence here in America. And as AI increasingly becomes political, I worry that's going to get directed at more and more AI political leaders. Whatever I may think or may not think of open AI, I think it is terrible that someone threw Molotov cocktails at Sam Altman's house. I am worried that we are headed into a higher variance, higher beta, higher risk world because of AI. And that's for me as an individual and then for people who are big players on the chess board. Think about what it means geopolitically. We're watching. The Ukrainians are really starting to win. And the reason they're winning, I think, is not really because they have better drones. I think they do have better drones That part of it I think the reason Ukraine is really winning is they have the best battlefield AI outside of probably America and Israel And has China has their adversaries begin to process that How do they respond? If the United States, because of its edge in AI, it's great if you're America, but it is destabilizing for the rest of the world. Something I think a lot about is creating a charity to just educate the world on how awesome the West has been. Slavery was endemic to essentially almost every civilization, and slavery was really ended by the British Empire. Tell that story. But America, after 1945, we had the nuclear bomb. No one else had it. We could have controlled the world forever. Instead, we rebuilt Germany and Japan, who were America's most reliable allies, Israel, South Korea. That's a testament to the American spirit in our country. We didn't take over the world. There were these fears that were documented at the time that the American generals, MacArthur was a little bit of an American emperor in Japan, were just going to take over the world. And they could have, and they didn't. They came home, we demilitarized. And then you had this period of great global stability between, you know, it was scary there were terrible moments Yeah you had the Pax Americana So maybe it not destabilizing Maybe it leads to another Pax Americana informed by our AI dominance And I so optimistic that AI is going to be amazing for the world. There's someone like me whose daughter was diagnosed with a very rare disease. There's no cure. He was able to assemble a lot of resources. He was able to get a lot of compute from the labs. We were made aware of what was happening. Spent up an immense amount of agents, came up using AI with a drug on the market that can actually impact his daughter's disease, and then has spun up a company to cure it. Her life is already immeasurably different because of AI. So I'm like an AI optimist maximalist, but I also just acknowledge it's like an event horizon. It for sure, I think is going to be a discontinuity we need to navigate as a society. I think the Luddites are going to be wrong, but we need to be like really thoughtful in how we address their concerns. We need to make sure that it's good for everyone. Like it is a little dystopian that now the best AI is only available to people with a lot of money. We need to solve that. We need to approach this with humility, recognize there's a lot of uncertainty and be thoughtful. When I do this with you, I tell people afterwards, I'm like, may you find something that you love as much as Gavin loves markets and companies and capitalists.
[75:00] and history on display today. As always, Gavin, thanks for your time. Thank you. Thanks, Patrick. If you enjoyed this episode, visit Colossus.com. You'll find every episode of this podcast complete with hand edited transcripts. You can also subscribe to Colossus, our quarterly print, digital, and private audio publication featuring in-depth profiles of the founders, investors, and companies that we admire most. Learn more at Colossus.com slash subscribe. Thank you. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at Vanta.com slash invest. The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months. Visit WorkOS.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit RidgelineApps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs. Check them out at Rogo.ai slash invest.
![Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473] — Colossus | Investing & Business Podcasts](https://reducer.xgoose.org/media/thumbnail/1294.jpg)