Gavin Baker is an experienced technology-sector investor who specializes in artificial intelligence (AI) and related technologies. He is known for his in-depth analyses of chips, data centers, and AI economics. In an interview on the Invest Like The Best podcast with Patrick O'Shaughnessy, he shared his insights into the future of AI, including a prediction that the next major data centers could be located in space. Baker has many years of experience investing in companies such as Nvidia and Google and often discusses complex topics such as scaling laws in AI and the economic impact of new chips. His past included plans to become a ski instructor or river guide, but he ultimately got into investing after reading books by Peter Lynch and Warren Buffett, which changed the course of his life.
AI economics and the chip battle: Nvidia versus Google
Baker describes in detail the current battle between Nvidia and Google in the field of AI chips. Nvidia is developing GPUs with the Hopper and Blackwell architectures, while Google has its TPUs (Tensor Processing Units). According to Baker, the transition from Hopper to Blackwell is the most complex product transition in the history of technology. Racks with Blackwell chips weigh around 1,360 kg (compared with 454 kg for Hopper) and consume up to 130 kW of power, equivalent to the consumption of 130 American households. This transition involves switching to liquid cooling and requires massive changes to data centers, such as reinforcing floors or installing new power sources.
Baker emphasizes that Google has a temporary advantage thanks to its TPU v6 and v7 chips, which are cheaper for generating tokens (units of computation in AI). According to him, Google is "sucking the economic oxygen" out of the market by offering low-margin services, making it difficult for smaller companies to compete. However, with the arrival of Nvidia's Blackwell chips in 2026, Baker expects the situation to reverse. The first model trained on Blackwell will probably come from Elon Musk's xAI because the company is building data centers the fastest. Baker compares Blackwell to the modern F-35 aircraft, while older TPUs are like the older F-4 Phantom.
Scaling laws and progress in AI
Baker explains that progress in AI is driven by scaling laws, which state that more chips and data lead to smarter models. Google's Gemini 3 confirmed that these laws still apply to pre-training. However, without new chips, progress would have stagnated in 2024 and 2025, but "reasoning models" such as OpenAI's o1 saved the day. These models allow AI to "think" longer before answering, increasing intelligence without requiring better hardware.
According to Baker, AI is now shifting from pure intelligence to usefulness. Models such as Gemini Ultra or Super Grok are already so smart that the average person cannot tell the difference between them unless they ask about complex topics such as PCI Express versus Ethernet protocols. The key elements of usefulness are: enormous context (for example, knowing your preferences, such as an east-facing balcony for morning sunlight, according to Andrew Huberman), reliability (no hallucinations), and task duration (from booking a table to planning an entire family vacation). Baker predicts that AI will soon master tasks with verifiable outcomes, such as accounting (whether the books balance) or sales (whether a deal was closed).
Space-based data centers
One of Baker's most interesting predictions is that we will see data centers in space within the next 3-4 years. From a physics perspective, space is ideal: solar energy is 30% more intense and available 24/7 without batteries. Cooling is free – simply point a radiator into the darkness of space, where the temperature is close to absolute zero, although according to experts, it is not quite as simple as it may seem. Laser communication in a vacuum is faster than communication through fiber-optic cables on Earth. Baker connects this with technologies such as Starlink, where a query travels directly from a phone to a satellite, reducing latency compared with terrestrial networks.
Baker sees a connection between Elon Musk's companies: Tesla (Optimus robot bodies), xAI (intelligence), and SpaceX (rockets and space computing). Space-based data centers could overcome limitations on Earth, such as shortages of power or cooling, which is the most expensive part of terrestrial centers.
Economic traps for software companies
Baker warns of a "death trap" for traditional software companies such as Salesforce or Adobe, which have margins of 80-90%. AI requires massive computing resources with margins of only 35-40%, so these companies must reduce their margins to offer AI agents or risk being steamrolled by startups. Microsoft is the exception because it is doing this correctly.
On geopolitics, Baker says that Chinese open models such as DeepSeek are a "godsend" for Meta, which uses them to improve its models. However, China is falling behind due to a shortage of powerful Nvidia chips, which the US banned from export and China subsequently banned from purchase as well. Baker considers China's restrictions on rare-earth exports a mistake because the US is developing new processing methods, which will resolve the problem more quickly.
Evidence of returns on AI investments
Baker demonstrates that AI investments pay off based on publicly traded companies. Major buyers of specialized AI GPUs have a higher return on invested capital than before the increase in AI spending. One example is C.H. Robinson, a freight logistics company. Previously, it took 15-45 minutes to provide a quote, and the company responded to only 60% of requests. With AI, it now responds to 100% within seconds, which led its shares to rise by 20% following its quarterly results. Baker emphasizes that AI helps with verifiable tasks such as customer support or sales.
Source: theneuron.ai
