Happy Thursday To The Greatest Living Americans! Welcome back to the Infinite Innovation Newsletter, where my mission is to keep you “In The Know” with what really going on with the stock market and its best longterm innovation investments. Today let’s take a deeper look at the chip market and the 5 companies that we think will lead the build out of the AI / Robotics Boom.
The AI Compute Supercycle: Five Chip Companies We Love Long Term
The AI buildout is no longer a single-stock story—and it is not simply about selling the most GPUs. It is becoming a full-stack infrastructure arms race spanning training compute, inference acceleration, custom silicon, high-speed networking, memory, and power-efficient data-center architecture.
At G.L.A. INVE$T RE$EARCH, we remain committed to owning the businesses enabling exponential innovation. The five names we love for the long-term AI silicon era are NVIDIA (NVDA), Cerebras (CBRS), Marvell (MRVL), Micron (MU), and AMD (AMD).
This is not a call to chase any stock at any price. It is a long-duration thesis: dollar-cost average exceptional innovation companies through volatility, stay focused on the underlying demand curve, and let compounding do the heavy lifting.
Training vs. inference
To understand where the semiconductor opportunity is going, investors need to separate two related—but economically different—forms of AI compute.
Workload | What it does | Primary need | Economic implication |
|---|---|---|---|
Training | Builds and improves the model using enormous data sets | Maximum compute, networking, memory bandwidth, and large clusters | Capital-intensive, concentrated among hyperscalers and leading AI labs |
Inference | Runs a trained model to answer a query, generate content, take an action, or power an AI agent | Low latency, efficiency, cost per token/query, broad deployment | Potentially much larger in volume as AI reaches billions of users and enterprise workflows |
Training is where frontier models are created. Think of teaching a student for years: processing vast quantities of information, testing, correcting, and reinforcing knowledge. NVIDIA has dominated this layer because its hardware, networking, CUDA software ecosystem, developer base, and system-level execution have become the default AI factory.
Inference is where the trained model is put to work. Every ChatGPT-style response, AI search result, autonomous-agent task, recommendation, coding assistant request, medical-image analysis, and robotics decision requires inference. Think of this as the student applying what they learned in the real world—millions or billions of times per day.
The major investment point: training created the first wave of AI infrastructure spending; inference could create the broader, more persistent, and more diverse second wave. That is why we want a portfolio that owns the dominant incumbent, the inference disruptor, the custom-silicon/networking enabler, the memory bottleneck, and the credible GPU/CPU challenger.
The demand-supply reality
The demand side remains powerful because every major platform company has the same strategic problem: it cannot afford to be under-compute in an AI economy.
Microsoft, Amazon, Alphabet, Meta, Oracle, sovereign cloud initiatives, enterprise software companies, and AI-native startups are all competing for capacity. Dan Ives has characterized NVIDIA as the “cornerstone” of the AI revolution, pointing to continued enterprise AI demand and the importance of hyperscaler spending.
The supply side is more complex than simply “who can make chips”:
Leading-edge wafer capacity is limited.
Advanced packaging remains critical for AI accelerators.
High-bandwidth memory is a key bottleneck.
Optical connectivity and high-speed networking become harder as clusters scale.
Power availability, cooling, and data-center construction can constrain deployment even when chip demand is present.
Customers increasingly want multiple sources of compute rather than a single-vendor dependency.
NVIDIA’s Blackwell products have shown the tightness of the current environment: the company said Blackwell sales were “off the charts,” while Blackwell Ultra set inference benchmarks and training benchmarks in MLPerf v5.1. NVIDIA also reported nearly $95.2 billion of supply-related purchase commitments, illustrating how aggressively it is securing capacity to serve demand.
Cathie Wood and ARK have made an important complementary point: the market will not belong to only one architecture forever. ARK expects custom silicon to exceed one-third of compute by 2030, while its research argues AMD can offer better performance per dollar on certain workloads.
That is the core of our thesis: AI demand is large enough to support multiple winners. But the winners will not all win in the same part of the stack.
Our five long-term AI chip names
Company | Primary AI role | Why we love it | Key risk |
|---|---|---|---|
NVIDIA | Full-stack training and inference leader | Hardware, CUDA ecosystem, networking, systems, developer moat | Valuation, competition, export restrictions, execution through rapid product cycles |
Cerebras | High-performance inference disruptor | Wafer-scale architecture and low-latency inference potential | Customer concentration, manufacturing complexity, early public-market volatility |
Marvell | Custom AI silicon, optical interconnect, networking | Picks-and-shovels exposure to hyperscaler diversification | Custom-program timing, concentration, execution on integration and ramps |
Micron | High-bandwidth memory and AI memory cycle | AI accelerators need massive memory bandwidth and capacity | Memory remains cyclical; pricing can reverse |
AMD | GPU, CPU, and heterogeneous compute challenger | Expanding AI accelerator opportunity from a smaller base | NVIDIA ecosystem advantage and customer qualification timelines |
1. NVIDIA — The AI factory
NVIDIA remains one of our favorite long-term core AI holding’s. This is not merely a GPU company. It is an integrated AI infrastructure company with a deep software moat, systems expertise, networking assets, developer loyalty, and an increasingly complete roadmap across training and inference.
The market often focuses on whether a competitor can produce a chip with competitive benchmark performance. That question matters—but the more important question is whether the competitor can replicate the surrounding ecosystem: CUDA, enterprise tooling, libraries, networking, systems integration, support, developer familiarity, and the accumulated knowledge base of the global AI community.
NVIDIA’s data-center revenue reached $62.3 billion in its fiscal fourth quarter, with full-year revenue of $215.9 billion; its Q1 fiscal 2027 revenue outlook was $78 billion, while the company noted that it did not assume China data-center revenue in that guidance.
Dan Ives’s framing is directionally correct: NVIDIA is still the central infrastructure provider for the AI spending wave. Tom Lee’s bullish AI framework likewise rests on the idea that compute scarcity remains strategically valuable as corporations and nations race to build AI capability.
Coach Kyle’s view: NVDA is still a foundational compounder. We do not need it to own 100% of AI compute to win. It simply needs to maintain leadership in the highest-value workloads while expanding its share of system-level economics.
2. Cerebras — The inference wild card
Cerebras represents a different bet: not that it will “replace NVIDIA,” but that inference becomes so important that radically different architecture can earn meaningful share.
Its wafer-scale approach is designed for high-speed, low-latency model serving. That matters when the product is not simply a chatbot response but an AI agent, real-time enterprise workflow, search product, robotics decision, customer-service interaction, or coding environment where latency affects user adoption and unit economics.
Cerebras raised its 2026 adjusted revenue target to $880 million to $890 million, citing demand for AI chips. It has also been expanding data-center capacity in Europe to address demand for regionally hosted, low-latency inference.
Coach Kyle’s view: CBRS is the higher-beta, asymmetric part of this basket. The upside comes from the possibility that inference becomes a separate and immense compute market, where speed-per-query and cost-per-token matter more than traditional GPU orthodoxy. The risks are real: customer concentration, execution, capital intensity, and valuation volatility, but I am comfortable with both the risks and the volatility.
3. Marvell — The custom silicon and connectivity toll road
Marvell is the underappreciated “picks-and-shovels” play for a more heterogeneous AI world. As hyperscalers build custom XPUs and deploy increasingly giant AI clusters, the opportunity expands beyond accelerators into custom ASIC design, electro-optics, storage, switching, SerDes, and the data movement required to prevent expensive compute from sitting idle.
This matters because AI systems are not just chips. The most advanced clusters require enormous bandwidth to move data between accelerators, memory, storage, and servers. The faster the compute gets, the more valuable the interconnect becomes.
Marvell’s data-center business has become its central growth engine, with reporting indicating data-center revenue represented 73% of sales, supported by optical interconnect and custom-ASIC demand. Management has pointed to accelerating demand for high-bandwidth connectivity in AI clusters.
ARK’s custom-silicon outlook reinforces the long-term setup. If custom silicon becomes a larger portion of total compute, companies that help hyperscalers architect and connect those systems are positioned to benefit.
Coach Kyle’s view: MRVL gives us exposure to AI infrastructure diversification. It is not a pure training-GPU trade. It is a bet that the winners in AI will require specialized chips and increasingly sophisticated data movement.
4. Micron — The memory bottleneck
AI accelerators are only as useful as the memory systems feeding them. Training and inference both demand immense bandwidth, and high-bandwidth memory has become one of the key enablers—and constraints—of modern AI infrastructure.
This is why memory deserves a seat at the AI table. The AI stack needs logic, but it also needs memory. Bigger models, longer context windows, more concurrent users, agentic workflows, retrieval systems, and multi-modal applications all intensify the memory requirement.
Dan Ives has described memory chips as the “golden child” of the AI revolution and cited what he viewed as a sharp demand-to-supply imbalance. We would not treat a single supply-demand ratio as a permanent fact—the memory market has always been cyclical—but the broader point is right: AI has materially altered the quality of demand for leading memory products.
Coach Kyle’s view: MU is our memory-cycle expression of the AI buildout. The right mindset is not to forget cyclicality; it is to recognize that the AI cycle may create a structurally higher-value mix and a more durable demand engine than past commodity-memory upcycles.
5. AMD — The credible challenger
AMD belongs in the long-term basket because AI customers want alternatives. They want negotiating leverage, supply flexibility, workload-specific optimization, and a second major ecosystem for accelerators and CPUs.
The AMD opportunity is not dependent on immediately unseating NVIDIA. It is dependent on winning enough deployments, enough cloud instances, enough enterprise workloads, and enough cost-sensitive or customized use cases to compound from a much smaller AI-accelerator base.
ARK has argued that AMD can already deliver better performance per dollar than NVIDIA in certain workloads, and it sees agentic AI as a tailwind for AMD’s broader CPU and GPU exposure. ARK also projects AI-system annual sales could reach $1.5 trillion by 2030.
Coach Kyle’s view: AMD offers the clearest high-quality challenger thesis. It brings CPU relevance, GPU ambition, a broad customer base, and the strategic importance of being a viable second source. Its execution must be judged by real customer adoption and software progress—not only by benchmark headlines.
The Musk Compute Empire: Securing Chips for Tesla, xAI, and SpaceX
Elon Musk is not treating AI chips as a commodity purchase. Across Tesla, xAI, and SpaceX, he is pursuing a multi-pronged compute strategy: buy leading-edge NVIDIA capacity now, design proprietary inference silicon for deployment at massive scale, rebuild Tesla’s Dojo training effort, and ultimately explore orbital data centers powered by solar energy.
This is exactly why the semiconductor opportunity is broader than one company: Musk’s companies need GPUs, custom inference chips, memory, networking, power, data-center capacity, and potentially space-qualified silicon.
1. xAI: Buy compute now
For xAI and Grok, Musk has pursued the immediate route: acquire enormous quantities of state-of-the-art AI infrastructure. xAI has relied heavily on NVIDIA hardware, including H100, H200, and GB200-class systems, while expanding the Colossus supercomputer footprint. Reporting indicates xAI is targeting a one-million-chip cluster and is building a third Mississippi data center, alongside more than $20 billion of planned investment in the state.
That matters for NVIDIA today. The largest model builders cannot wait several years for every internally designed chip program to mature; they need leading training capability immediately. NVIDIA remains the most proven way to deploy frontier-scale AI compute at speed.
Coach Kyle’s takeaway: Musk’s xAI strategy validates the near-term NVIDIA thesis. When time-to-compute matters, the dominant AI platform is still the default answer.
2. Tesla: Build proprietary inference silicon
Tesla’s AI5, AI6, and future generations are designed primarily to run AI models efficiently in Tesla vehicles, Cybercab, Optimus, and potentially Tesla data centers. This is fundamentally an inference-first strategy: Tesla wants lower cost, lower latency, greater control, and massive deployment scale for real-world AI decisions.
Musk has said Tesla’s AI5 and AI6 chips should be excellent at inference and “pretty good” for training, with internal effort concentrated on those chips. The rationale is straightforward: a vehicle fleet and robotics fleet may eventually require millions of always-on inference systems. Buying a premium external accelerator for every endpoint is not the optimal long-term unit-economics model.
This supports our Cerebras and AMD discussion as well. As inference expands, the addressable market becomes more architecture-specific. NVIDIA will remain a major force, but specialized chips can win where power efficiency, latency, cost per query, and tightly controlled workloads matter most.
3. Dojo: Bring training closer to Tesla’s data
Tesla has also resumed work on its Dojo supercomputer initiative after stabilizing the AI5 design. Dojo’s purpose is to process the enormous volume of video and sensor data generated by Tesla vehicles and use it to train the neural networks behind Full Self-Driving and related systems.
The key distinction is important:
AI5/AI6: Tesla-designed chips intended chiefly for deployed inference in cars and robots.
Dojo 3: Tesla’s effort to build a more internally controlled training system for autonomous-driving and robotics models.
NVIDIA infrastructure: Still strategically relevant while internal systems mature and as Tesla/xAI need immediate frontier-scale capacity.
Tesla’s ultimate goal appears to be vertical integration—owning more of the full loop from fleet data, to training, to inference, to real-world feedback. That is a powerful model if executed well, but it is technically difficult and requires sustained investment.
4. SpaceX: The orbital compute moonshot
SpaceX is pursuing the most ambitious version of Musk’s chip strategy: AI data centers in orbit. The proposed concept uses solar-powered satellites as computing nodes, with heat radiated into space rather than managed through terrestrial cooling infrastructure. Musk has said much of the enabling technology already exists within the Starlink ecosystem.
The initial proposed satellite is expected to generate roughly 150 kilowatts of peak power and 120 kilowatts of sustained compute power. SpaceX has indicated that early orbital systems would use NVIDIA GPUs, while longer-term plans could incorporate specialized radiation-hardened chips associated with Terafab, a semiconductor-fabrication initiative involving SpaceX, Tesla, and Intel.
Terafab: Musk’s Vertical AI-Chip Factory
Terafab is Elon Musk’s planned, vertically integrated semiconductor complex in Texas involving Tesla, SpaceX/xAI, and Intel. Rather than merely designing chips and relying on separate suppliers to manufacture, package, and test them, Terafab aims to bring advanced logic, memory, packaging, and testing together in one production ecosystem. Initial plans call for a $16.8 billion investment, with the facility intended to make chips for Tesla’s autonomous vehicles and Optimus robots as well as high-power processors for SpaceX’s proposed space-based data centers.
The strategic goal is compute independence. Musk wants to reduce dependence on external chip supply chains and build specialized processors for the workloads his companies care about most: low-latency inference in vehicles and robots, enormous AI training/inference systems, and eventually radiation-tolerant, power-efficient compute for satellites. Intel’s participation adds semiconductor manufacturing and packaging expertise, with the partners targeting an ambitious long-term capacity goal of 1 terawatt per year of compute.
Why it will matter
Faster innovation cycles: Owning more of design, fabrication, packaging, and testing could shorten the distance between a new Tesla or xAI chip design and large-scale deployment.
Lower unit economics at scale: Purpose-built chips can be cheaper and more power-efficient than general-purpose hardware when deployed across millions of cars, robots, or AI queries.
Supply-chain resilience: A domestic, integrated production base could reduce exposure to chip shortages, geopolitical disruption, and bottlenecks in advanced packaging or memory.
New AI deployment: Cheaper inference could accelerate autonomous transportation, robotics, scientific modeling, accessibility tools, and more capable AI services.
Space-compute optionality: Terafab’s space-oriented chips will support SpaceX’s long-term vision for orbital data centers powered by solar energy—an ambitious concept, not a near-term certainty.
For society, the upside is not simply “more chips.” It is the possibility of making intelligence cheaper to deploy in the physical world: safer autonomous systems, more productive factories and logistics, robotic assistance for labor shortages, and AI tools that can reach far more people at lower cost. The caution is equally important: semiconductor fabs are capital-intensive, technically difficult, energy-hungry, and slow to scale. Terafab’s benefits depend on execution, responsible deployment, and whether its enormous compute capacity produces real-world productivity rather than excess infrastructure.
G.L.A. conviction
Again, we are in the begging stages of the largest innovation / productivity boom in history, the companies that solve the largest problems will have the largest returns!
The long-term direction is clear: the world is moving from buying chips to building AI factories—and Elon Musk is trying to own as much of that factory as possible.
Disclosure: This newsletter is for educational and informational purposes only. It is not individualized investment advice, a recommendation to buy or sell securities, or a solicitation. Technology and semiconductor investments can be highly volatile. Do your own research and consult a qualified financial professional before making investment decisions.