The AI Super Cycle: Why This Veteran Investor Is Betting $4B on Hardware
Tobiloba OdejinmiBy Tobiloba Odejinmi
Education
Jun 1, 2026 • 11:30 AM
10m10 min read
Verified
Source: Pexels
The Core Insight
Gavin Baker, a 20-year veteran investor and founder of Atrades Management, argues that the AI industry is not in a bubble but a 'super cycle' driven by physical infrastructure constraints. By focusing on 'watts, wafers, and tokens,' Baker identifies the true bottlenecks, electricity, silicon fabrication, and connectivity, rather than software-as-a-service (SaaS) or consumer chatbots. His strategy prioritizes companies solving the physical limitations of scaling AI, while hedging against broader market volatility.
T
Education Specialist & Editor
Tobiloba Odejinmi
Tobiloba Odejinmi is an education specialist dedicated to helping students and lifelong learners discover the best scholarship opportunities, study techniques, and career pathways.
The Kodawire Editorial Team consists of experienced journalists and subject matter experts dedicated to delivering accurate, well-researched, and engaging content.
Infrastructure Over Software: The most sustainable returns in AI are found in the "picks and shovels", semiconductors, power, and connectivity, rather than consumer-facing chatbots.
The Physical Governor: Unlike the debt-fueled dotcom era, current AI spending is backed by free cash flow from profitable tech giants, and supply chain bottlenecks (TSMC, energy grids) prevent market over-supply.
The Inference Shift: Revenue potential for inference (running models) is estimated to be 5–10x larger than pre-training, shifting the investment focus toward specialized inference chips.
Barbell Strategy: Balance exposure to established giants like Nvidia and Micron with niche infrastructure players like Astera Labs and Cerebras.
For two decades, Gavin Baker of Atrades Management has navigated the volatile currents of the technology sector. While the current market discourse is dominated by fears of an "AI bubble," Baker’s $4.1 billion portfolio suggests a different reality: we are in the midst of a "super cycle." This is not a speculative frenzy built on cheap debt, but a capital-intensive build-out of physical infrastructure. The distinction is critical. In the late 1990s, the dotcom bubble was fueled by companies with no revenue and massive debt loads. Today, the AI build-out is funded by the free cash flow of the world’s most profitable companies, Microsoft, Google, Amazon, and Meta. Much like the boring habits that build wealth, this long-term infrastructure play requires patience and a focus on fundamentals rather than hype.
The most compelling argument against a bubble is the existence of physical constraints. We are not in a world of infinite supply. The industry is governed by "watts, wafers, and tokens." Because chip fabrication capacity at firms like TSMC is finite and energy grids are struggling to keep pace with data center demand, the market is effectively prevented from over-supplying itself. This scarcity acts as a natural stabilizer, ensuring that capital expenditure remains tethered to real-world capacity.
The physical constraints of wafer production act as a stabilizer for the AI market. (Credit: Nic Wood via Pexels)
How I Researched This
To provide this analysis, I have examined the recent 13F filings for Atrades Management and cross-referenced Baker’s public commentary on the AI infrastructure stack. My research process involved isolating the specific bottlenecks Baker identifies, connectivity, memory, and power, and verifying these against current industry reports from semiconductor leaders like SK Hynix and TSMC. I have stripped away the speculative noise to focus on the structural, long-term thesis that defines Baker’s two-decade track record.
The Four Pillars of AI Infrastructure Investment
Baker’s investment strategy rests on four distinct pillars that prioritize physical utility over software hype:
Verticalized Small Language Models (SLMs): As enterprises seek to leverage proprietary data without compromising privacy, the focus is shifting toward models optimized for local devices. These SLMs allow for high-performance reasoning without the need to send sensitive data to the cloud.
Sovereign Infrastructure: The "moat" in 2026 is no longer just code; it is the speed of physical deployment. Companies that can compress the timeline of building data centers from years to months hold a massive competitive advantage.
Performance per Watt: With hyperscalers spending billions on compute, the primary driver for AI labs is cost-efficiency. The ability to generate more tokens per watt of electricity is the metric that will determine which hardware providers win the market.
Energy and Space: Terrestrial grids are reaching capacity. The future of compute may lie in portable energy solutions and orbital compute, with companies like SpaceX serving as the essential "highway" to deploy infrastructure where it is needed most.
Data center expansion is limited by physical energy grid capacity. (Credit: Curioso Photography via Pexels)
The Risks You Need to Know
Investors must recognize that this thesis is highly sensitive to supply chain chokepoints. If TSMC were to suddenly triple its capacity, the current supply-demand balance would collapse, potentially forcing companies to take on debt to fund massive, unoptimized capital expenditures. Furthermore, the reliance on specific hardware manufacturers creates a single point of failure. If a competitor were to emerge that could replicate the precision of ASML’s lithography machines, the current valuation of the entire semiconductor stack would face significant downward pressure.
Unpacking the Portfolio: Where the Smart Money is Flowing
Baker’s portfolio reflects a "barbell" approach. On one end, he holds established giants like Nvidia and Micron, which provide the foundational compute and memory necessary for the industry. On the other, he targets niche infrastructure players:
Astera Labs (ALAB): Acting as the "plumbing" of the data center, Astera solves the connectivity bottleneck that occurs when clusters scale to hundreds of thousands of GPUs.
Unity Software: While known for gaming, Unity is a critical player in building "world models." By simulating physics and 3D environments, it provides the training ground for robotics and AGI.
Cerebras, Positron, and Sci-Fi: These firms represent the forward-looking bet on inference-specific chip architectures.
What the Numbers Really Mean
The shift from pre-training to inference is a mathematical imperative. Pre-training is a one-time cost, but inference is a recurring revenue stream. Estimates suggest the revenue opportunity for inference is 5–10x larger than pre-training compute. When you look at the operating margins of memory leaders like SK Hynix, which have reached 70%, it becomes clear why capital is flowing into these specific hardware components. The math is simple: if you can reduce the cost of a token by even a fraction of a cent, the scale of enterprise adoption increases exponentially. Understanding these metrics is as vital as learning tax-saving strategies to protect your long-term gains.
The Other Side of the Story
Most market analysts argue that the AI sector is a bubble because of the sheer volume of capital expenditure. However, this perspective ignores the "speed of atoms." The contrarian view is that we are not spending too much; we are spending exactly what is required to overcome the physical limitations of our current infrastructure. The "bubble" narrative assumes that demand is static, but the demand for compute is currently outpacing the physical ability to manufacture the chips required to meet it.
The Decision Matrix
If you are evaluating your own exposure to the AI sector, consider your time horizon:
If you are a short-term trader: The volatility in the semiconductor sector may be too high, and the reliance on QQQ puts as a hedge suggests that even the experts expect broader market turbulence.
If you are a long-term investor: Focus on the "picks and shovels." Look for companies that solve physical bottlenecks, connectivity, power efficiency, and memory, rather than companies that simply build chatbots.
Long-term investors should focus on infrastructure fundamentals rather than short-term software hype. (Credit: cottonbro studio via Pexels)
The Silent Wealth Killer
The biggest trap for retail investors is the "software-first" bias. Many investors flock to the latest AI application or chatbot, ignoring the fact that these companies often have high customer acquisition costs and low moats. The silent wealth killer is ignoring the infrastructure layer. If you are invested in the software but not the silicon, you are betting on the tenant rather than the landlord. In a gold rush, the landlord always wins. This is a core principle of building a sustainable financial roadmap.
13F Filings (SEC EDGAR): The primary source for tracking institutional movement.
Semiconductor Industry Reports: Essential for monitoring wafer capacity and memory pricing trends.
Energy Grid Load Data: A critical, often overlooked indicator of where data center expansion is physically possible.
What Do You Think?
Gavin Baker’s thesis hinges on the idea that physical constraints, watts and wafers, are the only things keeping this market from overheating. Do you believe that these physical bottlenecks are enough to sustain the current valuation of the AI sector, or is the market ignoring a looming correction? I will be replying to every comment in the first 24 hours.
Baker argues that the current AI build-out is funded by the free cash flow of profitable tech giants and is constrained by physical limitations like energy and chip fabrication, unlike the debt-fueled dotcom bubble.
The industry is governed by 'watts, wafers, and tokens,' representing energy, chip fabrication capacity, and compute output.
Pre-training is a one-time cost, whereas inference is a recurring revenue stream with a market opportunity estimated to be 5–10x larger.
Active Engagement
Was this information helpful?
Join Discussions
0 Thoughts
Editorial Team • Question of the Day
"If you had to choose between investing in the "landlords" (infrastructure/chips) or the "tenants" (AI software/chatbots) for the next decade, which would you pick and why?"
