How to keep up with the rapid pace of AI

To keep up with the constant stream of developments in AI, the first step is to use the technology yourself. Gavin Baker emphasizes that many investors draw definitive conclusions about AI based on free tiers, which is a mistake. He believes you must pay for the highest-tier models from the leading labs.

The free tier is like you're dealing with a 10 year old and you're making conclusions about the 10 year old's capabilities as an adult... the 200 per month tiers, those are like a fully fledged 30, 35 year old. It's really hard to extrapolate from an 8 or a 10 year old to the 35 year old.

The core of the AI conversation happens on X (formerly Twitter). An Insider post even noted that OpenAI largely runs on what is discussed there. The platform hosts public discussions and even disputes between major research teams, such as a notable fight between Meta's PyTorch team and Google's Jax team. The leaders of each lab had to publicly step in to resolve it. Following the 500 to 1,000 people at the cutting edge of AI, like Andrej Karpathy, is crucial because everything in the field is downstream of their work.

Additionally, it's important to listen to any podcast featuring someone from one of the four leading labs: OpenAI, Gemini, Anthropic, and XAI. To process this wealth of information, one of the best use cases for AI is to help you keep up. Gavin suggests listening to a podcast and then using an AI tool with low friction, like Grok, to discuss the interesting parts. This modern capability is remarkable.

It's like somebody said on X, we imbued these rocks with crazy spells and now we can summon super intelligent genies on our phones over the air. It's crazy.

Hardware delays and the two new scaling laws that saved AI

Gavin Baker explains that the release of Gemini 3 was very important because it confirmed that scaling laws for pre-training are intact. This is significant because nobody fully understands how or why these scaling laws work. They are an empirical observation, not a formal law. Gavin compares our current understanding to that of ancient civilizations and the sun. They could measure the sun's movements with incredible precision, like aligning the pyramids with the equinoxes, but they had no concept of orbital mechanics. They didn't know why it worked. Confirming this empirical observation is crucial for the industry.

Based on pre-training scaling laws alone, AI progress should have stalled in 2024 and 2025. This is because the industry hit a hardware limit with Nvidia's Hopper chips. It's not feasible to make a coherent cluster of more than 200,000 Hopper GPUs, where each GPU essentially knows what the others are doing. However, immense progress occurred thanks to two new scaling laws in post-training: reinforcement learning with verified rewards and test-time compute. This highlights a key concept from Andrej Karpathy.

With AI, anything you can verify you can automate.

These post-training advancements, often referred to as reasoning, effectively saved AI. They bridged an 18-month hardware gap created by the delay of Nvidia's next-generation Blackwell chips. The transition from Hopper to Blackwell was the most complex product transition in tech history, involving shifts from air-cooling to liquid-cooling and massive increases in weight and power consumption per rack. Had reasoning not emerged, the lack of new hardware would have meant no AI progress, which would have had major implications for the market.

The Blackwell delay gave Google a temporary advantage. They trained Gemini 3 on their 2024 and 2025-era TPUs, which were generations ahead of the Hopper chips everyone else was using. Gavin uses an analogy: Hopper is a World War II P-51 Mustang, Google's new TPUs are F-4 Phantoms, and the delayed Blackwell is an F-35. Google has also leveraged its position as the lowest-cost producer of tokens to put pressure on the rest of the AI ecosystem, a rare strategy in big tech where market leaders are not typically the low-cost option.

Looking ahead, the first models trained on Blackwell are expected in early 2026, likely from xAI. This is because, according to Nvidia's CEO Jensen Huang, Elon Musk's team builds data centers faster than anyone. This speed allows xAI to deploy Blackwells at scale first, helping Nvidia work out the bugs in the new, complex system for the benefit of all subsequent customers. Since we know pre-training scaling laws are intact, these future Blackwell-trained models are expected to be amazing.

Nvidia's accelerating roadmap challenges Google's custom chips

Nvidia's new Blackwell chip is a significant step forward. The GB300 model, in particular, is drop-in compatible with existing data center racks, making adoption easier. Companies that use the GB300 will likely become the low-cost producers of AI tokens. This shift has profound strategic implications, especially for a company like Google.

Previously, it might have been a rational decision for Google to run its AI services at a negative 30% margin. By being the low-cost provider, they could remove the economic oxygen from the environment, making it difficult for competitors to raise capital. However, this entire calculus changes if Google is no longer the low-cost producer. Continuing to run at a loss while being a higher-cost producer could become painful and even impact their stock price. The competitive dynamics are shifting as Nvidia's Blackwell chips are used for training and then inference.

The performance gap between Nvidia's GPUs and Google's TPUs seems to be widening. This is partly due to Google making more conservative design choices. Another factor is Google's partnership with Broadcom for its TPUs. In this arrangement, Google handles the 'front-end' design, like an architect designing a house. Broadcom handles the 'back-end' manufacturing and relationship with Taiwan Semiconductor, acting like the builder who constructs the house. For this service, Broadcom earns a very high gross margin of 50-55%.

This financial arrangement is becoming a point of friction. With TPUs potentially becoming a $30 billion business, Google would be paying Broadcom around $15 billion. At this scale, it becomes economically rational for Google to bring the entire semiconductor program in-house, much like Apple does. Google's recent move to bring in MediaTek can be seen as a warning shot to Broadcom over these high costs.

The broader challenge is that creating a custom accelerator is incredibly complex. It's not just about the chip. Nvidia is accelerating its own product cycle, releasing a new GPU every year, to make it harder for anyone to keep up.

What's the NIC going to be? What's the CPU going to be? What's the scale up switch going to be? What's the scale up protocol? What's the scale out switch? What kind of optics are you to use? What's the software that's going to make all this work together? And then it's like, oh shit, I made this tiny little chip.

It takes multiple generations to develop a competitive chip. The first TPU was an achievement simply because it was made, but it wasn't truly competitive until its third or fourth version. This learning curve is steep, and very few have mastered it. The Amazon ASIC team, which developed the Gravitron CPU, is often cited as the best example of a successful in-house chip team.

How hardware advancements unlock practical AI applications

While AWS's Trainium and Google's TPU are becoming formidable ASICs for AI, the larger story is the impact of the next generation of hardware. The upcoming Blackwell models are expected to be amazing, primarily because they will cause a dramatic reduction in the per-token cost of AI. This cost reduction is significant because it allows models to "think" for much longer, which in turn enables them to perform more complex, real-world tasks.

A recent example of this progress is Gemini 3 making a restaurant reservation. This is a step towards a true personal assistant that can handle hotel, flight, and Uber bookings. This capability extends to the enterprise. Some tech-forward companies already use AI for over 50% of their customer support, a $400 billion industry. AI's proficiency at persuasion makes it well-suited for both sales and customer support.

A useful framework for identifying where AI will have an impact comes from Andrej Karpathy: AI can automate anything that can be verified. This applies to any function with a clear right or wrong outcome.

Does the model balance? They'll be really good at making models. Do all the books globally reconcile? They'll be really good at accounting... Did you make the sale or not? That's just like AlphaGo. Did you win or you lose? Did the guy convert or not?

If these applications show a return on investment by 2026, the massive spending on hardware like Blackwell will be justified and the cycle will continue. Currently, companies are in a prisoner's dilemma, terrified of slowing down. However, as the spending numbers become astronomical with Blackwell and the subsequent Rubin generation, economic realities and ROI will start to dominate the decision-making process over the fear of falling behind.

The ROI on AI has been unambiguously positive

The return on investment (ROI) for AI has been empirically, factually, and unambiguously positive. The largest spenders on GPUs are public companies that report audited quarterly financials. A calculation of their return on invested capital (ROIC) shows it is higher now than it was before they increased spending on AI infrastructure.

This positive ROI comes from various sources, including OPEX savings and efficiency gains. For example, moving large recommender systems from CPUs to GPUs has led to massive efficiencies and accelerated revenue growth for these companies. According to Gavin Baker, this is a valid part of the return calculation. The ROI is clearly there.

This has created internal tension at every major internet company. Those responsible for revenue are often annoyed by the number of GPUs allocated to researchers. For the revenue teams, the equation is simple and linear.

If you give me more GPUs, I will drive more revenue. Give me those GPUs, we'll have more revenue, more gross profit, and then we can spend money.

This creates a constant fight for resources. Part of what drives the research side is a near-religious belief that they will achieve Artificial Superintelligence (ASI). For many, the ultimate goal is to live forever, and they believe ASI is the path to achieving it. While that would be a good return, the ultimate economic benefits are unknown. If humans have already pushed the boundaries of physics, biology, and chemistry, the economic returns to ASI may not be as high as some expect.

The biggest bear case for AI is the rise of on-device models

One of the most significant bear cases for the continued explosive demand for AI compute is the rise of edge AI. This scenario involves powerful, pruned-down versions of models like Gemini 5 or Grok 4 running directly on devices like phones. These on-device models could operate for free at 30 to 60 tokens per second. Apple's strategy seems to align with this, positioning the iPhone as a privacy-safe AI distributor that can call on larger cloud-based models only when necessary.

If a model with a '115 IQ' is considered 'good enough' for the average user, it could significantly dampen the demand for massive, centralized computing resources. This is considered the most plausible bear case, aside from scaling laws breaking or low economic returns from advanced AI. The bullish case relies on the continuation of scaling laws, not just for pre-training, but also for post-training and test-time compute. Advances like extremely long context windows, which allow a model to hold and process vast amounts of specific information like a company's entire Slack history, could unlock new capabilities.

When comparing the evolution of AI models to an S-curve, like that of the iPhone, it's becoming difficult for non-experts to see the differences between top-tier models. To spot the progress, one has to ask highly specialized questions in a field of deep expertise. For most people, the models are already so intelligent that progress is hard to perceive. This suggests a necessary shift in focus from making models more intelligent to making them more useful. Ultimately, this increase in usefulness will need to translate into major scientific breakthroughs that create entirely new industries.

The enterprise adoption and ROI of AI

The foundation of usefulness in AI is consistency and reliability, which heavily depends on its ability to retain context. For complex tasks like planning a personalized trip, an AI needs to remember numerous specific preferences, such as a desire for an east-facing balcony for morning sun, in-flight Starlink, and historical hotel choices. The ability to manage this context is crucial for expanding the length and complexity of tasks an AI can handle.

While booking a restaurant is useful, planning an entire vacation that considers the preferences of multiple family members is a much harder, more economically valuable problem. Beyond personal assistance, AI is expected to become proficient in sales and customer support soon. It is already accelerating product development, with engineers using it to create better products faster. This trend is expected to appear in every industry vertical.

The adoption of AI follows a familiar pattern seen with previous technologies like the cloud. Startups are the first to embrace it, while larger Fortune 500 companies are typically last due to their conservative nature and regulatory hurdles. VCs are broadly bullish on AI because they witness real productivity gains in their portfolio companies. Today's startups can achieve a given level of revenue with significantly fewer employees compared to two years ago, as AI handles aspects of sales, support, and product creation.

A significant shift occurred recently when non-tech Fortune 500 companies began reporting specific, quantitative benefits from AI. For example, the freight forwarder C.H. Robinson used AI to improve its quoting process for shipping. Previously, it took them 15 to 45 minutes to quote a price, and they only responded to 60% of requests. With AI, they now quote 100% of inbound requests in seconds. This led to a strong quarter and a 20% stock increase, demonstrating AI's impact on both revenue and costs.

These positive results help alleviate concerns about a potential "ROI air gap." The fear was that massive capital expenditure on new chips like Nvidia's Blackwell, used primarily for training models, would not generate immediate returns. Since the ROI comes from inference (using the model), not training, there was a risk that return on invested capital would drop significantly. However, examples like C.H. Robinson suggest companies are finding ways to navigate this period of heavy investment.

Private equity is well-positioned to apply AI to traditional businesses

While the market is captivated by the top ten companies, attention may shift to the other 490 in the S&P 500, especially as more companies face challenging quarters. Historically, well-run companies have a long track record of success because they effectively use technology. This suggests that companies with an internal culture of experimentation and innovation will likely do well with AI. For example, the best investment banks are probably going to be earlier and better adopters of AI than their competitors.

Gavin Baker has a strong opinion on venture capitalists setting up holding companies to improve traditional businesses with AI. He believes they are entering a field that private equity has dominated for 50 years.

You're just not going to beat private equity at their game. This is what Vista did in the early days.

While private equity has faced challenges recently due to rising multiples, more expensive assets, and higher financing costs, these firms are well-positioned for the future. Gavin thinks that private equity firms will be very effective at systematically applying AI across their portfolios.

Why building frontier AI models is harder than it looks

Gavin Baker explains that while foundation models were once called the fastest-appreciating assets in history, a key element was missing. The classic internet company flywheel, where users generate data that improves the product, did not apply to early AI models. However, the advent of "reasoning" capabilities has fundamentally changed this dynamic.

That flywheel has started to spin and that is really profound for these Frontier Labs...if a lot of people are asking a similar question, they're consistently either liking or not liking the answer, then you can kind of use that. Like that has a verifiable reward, that's a good outcome and then you can kind of feed those good answers back into the model.

Building these frontier models is much harder than many assumed. Gavin points to major companies like Meta, Microsoft, and Amazon, who have invested heavily but failed to produce top-tier models. Mark Zuckerberg's prediction that Meta would have the best AI in 2025 was, according to Gavin, "as wrong as it was possible to be." This widespread failure highlights the immense difficulty of the task.

Several factors contribute to this difficulty. First, the technical complexity of managing infrastructure is a huge hurdle. Keeping a large cluster of GPUs running at high utilization is incredibly challenging, and performance varies wildly between companies. A lab with 90% uptime has an insurmountable advantage over one with 30%. Second, top AI researchers possess what they call "taste"—an intuition for choosing the right experiments. As models grow, these experiments become prohibitively expensive, requiring tens of thousands of GPUs for days at a time, making the right choices critical.

These challenges create significant barriers to entry. Moreover, the leading labs like OpenAI, Anthropic, Gemini, and xAI have a compounding advantage: they use a more advanced, internal version of their model, known as a checkpoint, to train the next one.

They have a more advanced checkpoint internally of the model. And they're using that model to train the next model. And if you do not have that latest checkpoint, it's getting really hard to catch up.

This creates a cycle where the leaders continuously pull further ahead, making it nearly impossible for others to catch up without a competitive model to start with. For companies like Meta, Chinese open-source models offer a potential starting point to try and bootstrap their way back into the race.

Nvidia's Blackwell chip is reshaping the geopolitical AI landscape

China has made a potentially terrible mistake by trying to force its domestic open-source AI development onto Chinese-designed chips instead of using Nvidia's upcoming Blackwell chips. This decision is poised to widen the performance gap between American AI labs and Chinese open-source projects. In a recent technical paper, the Chinese lab DeepMind even alluded to this problem.

One of the reasons we struggle to compete with the American Frontier Labs is we don't have enough compute. That was their very politically correct, still a little bit risky way of saying, because China said, we don't want the Blackwells, and they're saying, won't you please give us the Blackwells?

By the time China realizes its need for these advanced chips, likely in late 2026, the US may have already solved the rare earth mineral supply issue. These minerals are not actually rare, just messy and difficult to refine. With DARPA programs and deposits in friendly nations, the US is likely to solve this problem faster than anticipated, reducing China's geopolitical leverage.

This dynamic significantly impacts the competition between AI companies. XAI will likely be the first to release a model using Blackwell, giving them an important advantage. Meanwhile, OpenAI is facing a "Code Red" because they are a high-cost producer of tokens, partly from paying a margin on compute. This financial pressure is evident in their need to raise massive amounts of capital. In contrast, Anthropic is more efficient, burning less cash and growing faster, thanks to its relationships with Google for TPUs and Amazon for Trainium chips. Recognizing the shifting landscape, Anthropic's CEO Dario Amodei also signed a $5 billion deal with Nvidia, making Anthropic the third major AI lab, alongside XAI and OpenAI, to align with Nvidia in its battle against Google.

The first-principles argument for data centers in space

One of the most important developments in the next three to four years will be the rise of data centers in space. From a first principles perspective, data centers are fundamentally superior in space compared to on Earth. The primary inputs for a data center are power, cooling, and chips. Space offers significant advantages for the first two.

For power, a satellite can be in the sun 24 hours a day. The sun is also 30% more intense in space, resulting in six times more irradiance than on Earth. This provides abundant solar energy and eliminates the need for batteries, which are a huge cost. For cooling, which represents a majority of the mass and complexity in a terrestrial data center, space offers a simple and free solution. A radiator can be placed on the dark side of the satellite, which is close to absolute zero. This removes the need for complex HVAC and liquid cooling systems.

Connectivity between racks is also faster in space. While data centers on Earth use lasers through fiber optic cables, satellites can link to each other using lasers through a vacuum, which is even faster. This creates a more coherent network.

The only thing faster than a laser going through a fiber optic cable is a laser going through absolute vacuum. So if you can link these satellites in space together using lasers, you actually have a faster and more coherent network than in any data center on Earth.

For inference tasks, this offers a much better user experience. A request can go directly from a phone to a satellite and back, rather than a long journey through cell towers and fiber optic cables on the ground. The main obstacle to this vision is launch cost and capacity. Economically viable data centers in space will require many successful launches of rockets like SpaceX's Starship.

This development highlights the convergence of Elon Musk's companies: Tesla, SpaceX, and xAI. SpaceX can provide the data centers in space that power the AI developed by xAI. This AI, in turn, can serve as the intelligence for the Optimus robot from Tesla, which uses Tesla Vision as its perception system. Each company creates a competitive advantage for the others, forming a powerful ecosystem.

The historical cycle of shortages and gluts applies to AI compute

A historical economic pattern shows that shortages are always followed by gluts in capital cycles. This concept can be applied to the current shortage of AI compute. Companies like Mark Chen's have indicated they would consume ten times as much compute if it were available, highlighting the massive current demand.

While a glut in compute will eventually occur, AI is fundamentally different from traditional software. Every time AI is used, it consumes compute, a constant demand that did not exist with older software models. If companies had access to ten times more compute, the quality of their services would simply improve. For example, the premium $200 tier would get much better, and the free tier would become as powerful as the current premium tier.

To generate a return on this investment in compute, companies are exploring new monetization strategies. Google has started to introduce ads into its AI mode, which will likely give other companies permission to do the same for their free tiers. Another avenue for revenue is through taking commissions on actions facilitated by the AI, such as booking a vacation suggested by the model.

Taiwan Semi's caution is a governor on the semiconductor industry

Semiconductor inventory cycles are inevitable due to the iron law that customer buffer inventories must equal lead times. However, a true capacity cycle hasn't occurred since the late 90s, largely because Taiwan Semi has been so effective at aggregating and smoothing supply. The current problem is that Taiwan Semi is not expanding capacity as fast as its customers want. This is a mistake, according to Gavin Baker, because competitors like Intel have empty fabs that will eventually be filled due to the shortage of compute. Gavin notes that while Intel's fabs are not as good, the company is now benefiting from the strategy set by former CEO Patrick Gelsinger, whose firing he considers shameful.

Taiwan Semi's reluctance to expand is rooted in a deep-seated paranoia about an overbuild. They are highly skeptical of demand forecasts and famously dismissed Sam Altman's projections.

They're the guys who met with Sam Altman and laughed and said he's a podcast bro, he has no idea what he's talking about. They're terrified of an overbuild.

This caution from Taiwan Semi, along with power constraints, acts as a natural governor on the industry. These governors are beneficial because they create a smoother and longer growth cycle. If both constraints were removed simultaneously, for instance if Taiwan Semi opened up capacity just as space-based data centers solved power issues in five or six years, an overbuild could happen very quickly. For now, these governors ensure stability.

Power as a constraint benefits the most advanced AI technologies

After a long period of being an uninteresting topic, power has suddenly become a critical constraint, particularly for AI. This constraint, however, is beneficial for the most advanced compute players. When watts are the limiting factor, the total cost of compute hardware becomes irrelevant. The primary goal shifts to maximizing output, or tokens, per watt. This means a company can generate significantly more revenue from more efficient, albeit more expensive, hardware.

Gavin Baker illustrates this with an example: a $50 billion data center that produces $25 billion in revenue is far superior to a cheaper $35 billion data center that only generates $8 billion. This dynamic ensures that the best technologies will win regardless of price, granting them incredible pricing power. As an investor, this is a significant implication.

The solutions to this power demand are not what some might expect. Building new nuclear power plants in the United States is simply not feasible in the required timeframe due to heavy regulation. The process is fraught with delays over minor issues.

You know, one ant... that is America. Yeah, it's crazy. Actually, humans need to come first. We need to have a human centric view of the world.

Instead, the viable solutions are natural gas and solar. AI data centers, especially those not used for inference, can be located anywhere. This allows them to be built in places like Abilene, Texas, which is situated in a large natural gas basin. Thanks to fracking, the U.S. has a plentiful and accessible supply of natural gas. The power industry is already starting to adapt. For instance, Caterpillar recently announced plans to increase its capacity for turbine manufacturing by 75% in the coming years, signaling that the system is beginning to respond to the new demand.

Young AI-native entrepreneurs are getting polished faster

A new generation of young, AI-native entrepreneurs is emerging, and they are remarkably impressive. They become polished much faster than previous generations, primarily because they use AI as a constant advisor.

These young CEOs consult AI for a wide range of business challenges, from preparing for investor meetings to navigating difficult HR situations or refining sales strategies. The advice AI provides is often very effective, which explains why venture capitalists are seeing massive productivity gains across their portfolio companies.

How should I deal with pitching this investor I'm meeting with Patrick o'. Shaughnessy? What do you think the best ways I should pitch him are? ... I have this difficult HR situation. How would you handle it? ... We're struggling to sell our product. What changes would you make? And it's really good at all of that today.

This trend extends to other fields like investment. Young talent now enters the workforce with a level of knowledge that previously might have taken until their early 30s to acquire. The accessibility of specialized knowledge through podcasts and the internet, combined with a native ability to use AI, is creating a generation of professionals who advance at an accelerated pace.

Semiconductor venture capital makes a comeback to fuel AI

The semiconductor venture capital space is experiencing a significant resurgence, largely ignited by the success of Nvidia. A unique characteristic of this trend is that the average founder is often around 50 years old—a seasoned expert in their field. For example, a top networking architect at a large public company, inspired by the market potential, might decide to start their own specialized company. This is crucial for the entire AI hardware ecosystem.

Even tech giants like Nvidia, AMD, and Google cannot innovate alone. A product like Nvidia's Blackwell rack contains thousands of parts, with Nvidia only making a few hundred of them. To maintain an aggressive annual innovation cycle, they need the entire supply chain—makers of transceivers, wires, backplanes, and lasers—to accelerate with them. Gavin Baker notes that this revitalized venture ecosystem creates a network of companies that can keep pace, and also puts pressure on existing public companies to innovate faster.

You need the people who make the transceivers, you need the people who make the wires, who make the backplanes, who make the lasers. They all have to accelerate with you.

A notable aspect of AI investing today is that at every level of the technology stack, there are significant public and private competitors. This dynamic is even spurring a wave of innovation in memory, which is a critical gating factor. However, this points to a potential bottleneck. If a true DRAM (Dynamic Random-Access Memory) cycle occurs, similar to those in the late 1990s, it could act as a natural governor on the industry's growth. In the past, such cycles led to extreme shortages where prices could multiply by 10x. In recent decades, a strong cycle might mean a 30-50% price increase. A return to exponential price hikes could fundamentally change the economics of building AI infrastructure.

SaaS companies are repeating the mistakes of brick-and-mortar retail with AI

Application SaaS companies are making the same mistake with AI that brick-and-mortar retailers made with e-commerce. Retailers saw e-commerce as a low-margin business and were slow to invest, despite clear customer demand. They were attached to their existing model where customers handled transportation costs. Amazon's vision, however, was about future efficiency at scale. Ultimately, many retailers were left behind.

SaaS companies are similarly reluctant to embrace AI because it challenges their high-margin business model. Traditional software has high gross margins because you write the code once and can distribute it broadly at a low cost. Gavin Baker explains the fundamental difference with AI.

AI is the exact opposite where you have to recompute the answer every time. And so a good AI company might have gross margins of 40%.

Trying to preserve an 80% gross margin structure is a guarantee of failure in the AI space, as AI-native companies operate on much lower margins. Interestingly, these AI companies often become cash-generative sooner than traditional SaaS companies, not due to high margins, but because they employ very few people.

There is a precedent for investors accepting lower margins for long-term growth: the transition from on-premise software to the cloud. When companies like Adobe and Microsoft shifted to a SaaS model, their margins and even revenues initially dropped. Investors eventually understood that the move would be accretive to gross profit dollars over time.

Established SaaS companies like Salesforce, ServiceNow, and HubSpot have a huge advantage over AI-native startups: a cash-generating core business. They should leverage this by launching their own AI agents to automate the core functions they already provide. For a CRM company, this means building an agent that can talk to customers. The risk of inaction is immense.

What's happening right now, another agent made by someone else is accessing your systems to do this job, pulling the data into their system and then you'll eventually be turned off. It's just crazy. And it's just because, oh wow, but we want to preserve our 80% gross margins. This is a life or death decision and essentially everyone except Microsoft is failing it.

Gavin likens the situation to the famous "burning platform" memo from a Nokia executive. The SaaS platform is on fire, but there is a new, stable platform—AI—right next to it, ready to be utilized.

Balancing conviction and flexibility in investing

Since 2020, there has been a series of rolling bubbles in the market. It started with speculative stocks like EV startups, then moved to meme stocks. Now, the bubble seems to be in areas like nuclear fusion and quantum computing. Gavin Baker notes that while these technologies are transformative, the publicly traded companies are often not the best way to invest in these themes. The actual leaders in quantum, for example, are companies like Google, IBM, and Honeywell. He also clarifies that 'quantum supremacy' is often misunderstood. It doesn't mean quantum computers will be superior at everything, but rather that they can perform specific calculations that classical computers cannot.

A fascinating trend Gavin has observed relates to AI. It seems that whatever AI needs to continue its advance, it gets. He points to the remarkably fast shift in public opinion on nuclear power, which occurred right when AI's energy needs became a major concern. This pattern, where bottlenecks are quickly resolved, brings to mind Kevin Kelly's concept of the 'technium' in his book "What Technology Wants," suggesting technology itself has a drive to grow more powerful.

Reflecting on his own journey, Gavin discusses his evolution as an investor and leader. He was once a 'one-man show,' but a significant shift came when he first made money on an idea that wasn't his own. He draws an analogy to the NFL, where quarterbacks once deemed failures, like Baker Mayfield, flourished under different systems and coaching. This inspired him to focus on creating an environment where talented people can succeed.

This philosophy hinges on finding a balance between conviction and flexibility. He believes investing is about navigating this tension.

To quote Michael Steinhardt, it's all about finding the right balance between the arrogance to believe that you have a variant view that's going to be right in a game that is being played of tens of millions of people worldwide with the humility to recognize that at any moment there might be a new data point that is outside of your expected probability space that invalidates that variant view and to be really, really open to that.

To achieve this, he has worked to create a culture where it is safe for people to take risks, change their minds, and challenge each other—and him. He sees investing as a search for hidden truths, which are best discovered through rigorous debate and discourse. Celebrating being wrong is a key part of learning. By making it safe for people to disagree, the quality of decisions is higher.

If I'm not wrong about three things in a day, I didn't learn anything. So I want to be told I'm wrong as much as possible.

Gavin Baker on his early passions and unconventional life plan

When asked how he would pitch his career to a young person, Gavin Baker traces his passion back to his earliest interests. As a child, he loved history, poring over books about ancient civilizations like the Phoenicians and Romans. This fascination evolved into an interest in current events, which he viewed as a form of applied history. By middle school, he was eagerly reading the New York Times and the Washington Post, excited to watch history unfold in real time.

In college, his focus shifted dramatically to rock climbing, which became the most important thing in his life. He dedicated himself completely to the sport, even doing his homework at the climbing gym. His plan was not to enter the world of finance, but to build a life around his passions. He intended to be a ski bum in the winters, a river guide in the summers, and a climber in the shoulder seasons, all while trying to become a wildlife photographer and write a novel.

Gavin shares a formative experience from his time working as a housekeeper to support his skiing. This job profoundly shaped his perspective on how to treat people.

It was shocking to me how people treated me. It has permanently impacted how I treat other people. You'd be cleaning somebody's room and they'd be in it and they'd be reading the same book as you. And you'd say, 'Oh, wow, that's a great book. I'm about where you are.' They look at you like you're a space alien, like you speak. And then they get even more shocked you read. So it had a big impact on how I've just treated everyone since then. Being nice is free.

His parents, who had strict upbringings and grew up in economically disadvantaged circumstances, were surprisingly supportive of his unconventional plan. Having paid for his college education, they encouraged him to pursue his dream of being a ski bum and artist, having never pushed him in any particular direction.

Discovering investing as a game of skill and chance

Gavin Baker's journey into investing began with a sophomore summer internship at Donaldson, Lufkin, and Jeannette (DLJ). His parents had one simple request: just get one professional internship. His job was in the private wealth management division, where he was tasked with mailing company research reports to clients who owned that particular stock. He started reading the reports and was immediately hooked, finding them to be the most interesting things imaginable.

He came to conceptualize investing as a game of skill and chance, much like poker. While there is an irreducible element of chance, like a meteor hitting a company's headquarters, there is also a significant skill component that appealed to his competitive nature. Gavin believed the key to succeeding in this game was to combine historical knowledge with a deep understanding of current events to form a unique opinion on the future.

The way you got an edge in this greatest game of skill and chance imaginable was you had the most thorough knowledge possible of history. And you intersected that with the most accurate understanding of current events in the world to form a differential opinion on what was going to happen next in this game of skill and chance.

This realization sparked an intense period of self-study. During his internship, he read books by Peter Lynch and about Warren Buffett, devoured "Market Wizards," studied Buffett's shareholder letters twice, and taught himself accounting. Back at Dartmouth, he changed his majors from English and History to History and Economics. Investing became an obsession. He would read articles from The Motley Fool at the climbing gym and constantly check stock news on a computer terminal. Having never been good at anything else competitive like sports, chess, or skiing, he was drawn to investing as the one area where he felt he could excel. This passion has defined his career, and he believes it remains the only thing he's truly good at.