Brian Manning

AI Investment Thesis

Last Updated: September 15, 2026

Summary

AI is the broadest general-purpose technology since electricity, and the first whose own output feeds back into its rate of progress. The models already help write the training code, generate the data, and assist the chip design for the generation that replaces them. That feedback loop is why I underwrite this as a multi-cycle buildout rather than a single capex wave.

The United States will build most of the frontier stack, but the pace of that build is now set by physical inputs rather than algorithms: generation and grid interconnection, leading-edge wafers and advanced packaging, high-bandwidth memory, optics, and the electrical and thermal equipment inside each data hall. My objective is not to pick the winning lab. It is to own the inputs that stay scarce regardless of which model leads, and the incumbents whose problems, proprietary data, and distribution convert tokens into earnings.

The bear arguments are serious and I track them explicitly. Capex can overshoot, a meaningful share of reported demand is financed by the suppliers themselves, and measured productivity has not yet caught up with spend. None of that, in my view, makes this a repeat of the 1999–2001 fiber overbuild. The better historical comparison is electrification: capacity is contracted before it is energized, and the economic return shows up years later, mostly at the firms that redesign their operations around the new input.

1. A technology that shortens its own development cycle

Every prior general-purpose technology multiplied human output without accelerating its own R&D. Steam engines did not design better steam engines. AI does, in narrow but expanding ways: code generation inside the labs' own training infrastructure, synthetic data and reinforcement learning environments that reduce dependence on scarce human-labeled data, and model-assisted placement, verification, and layout in chip design. Each of these compresses the time between model generations.

The investment consequence is that the capability curve and the spend curve are coupled. A better model expands the set of tasks worth automating, which raises inference demand, which funds the next training run, which requires more accelerators, more memory, and more megawatts. Analyzing the trade as a one-time software upgrade cycle misses that loop. Markets have historically been poor at pricing sustained changes in the rate of improvement; they tend to extrapolate the current level instead.

This does not imply that every lab earns an adequate return, or that model-layer valuations are disciplined. I use the frontier labs' public statements as inputs for scenario ranges and adoption timing, not as endorsements of their equity or their policy positions. My base case is that a large share of the economic surplus accrues outside the labs, to the owners of scarce inputs and to incumbents with the workflows and data to apply the models.

2. Electricity, not dark fiber

The bear analogy that carries the most weight in investment committees is the telecom overbuild: capacity built on forecast demand, utilization that never arrived, and equity that went to zero. I think the analogy breaks on three points.

How demand clears. Fiber in 1999–2000 was laid against traffic projections, and much of it stayed unlit for years. AI capacity is largely contracted before it is delivered. Hyperscaler remaining performance obligations, multi-year take-or-pay agreements with neoclouds, and lab commitments to cloud capacity are signed ahead of the buildings being energized. Lead times for accelerators, HBM, optical modules, gas turbines, and large power transformers run from several quarters to several years. The queue consists of paying customers waiting on physical delivery, not idle capacity waiting on customers.

How supply expands. The fiber glut was made far worse by wavelength-division multiplexing, which multiplied the capacity of strands already in the ground at close to zero incremental cost. There is no equivalent in compute. More tokens require more wafers, more packaging capacity, more memory, and more power, each of which has its own long-cycle supply chain and its own disciplined oligopoly.

How the asset ages. Fiber lasts decades, so overbuilt capacity overhung pricing for a decade. Accelerators face the opposite risk: rapid economic obsolescence as each generation lowers cost per token. That creates a real depreciation question (addressed below), but it also means excess capacity is worked off faster, because older fleets are repurposed to inference and eventually retired rather than sitting as a permanent overhang.

Electrification is the closer precedent. Power plants were built ahead of measurable productivity gains, and the gains arrived only after factories were redesigned around distributed electric motors rather than a central line shaft. Most of the value went to the manufacturers who redesigned, not the utilities.

The analogy does not remove overbuild risk. The more useful signal than headline capex is supply behavior at the chokepoints. As long as the leading foundry, the HBM producers, and the turbine and transformer OEMs add capacity against contracted demand, a broad glut is unlikely. Prior semiconductor downturns began when second-source capacity was added on price rather than on orders. That is the signal I watch.

3. The two bear cases that deserve the most weight

Circular financing

A growing share of AI demand is financed by the supply chain itself. Chip vendors take equity stakes in customers who then buy their chips. Clouds invest in labs that commit the proceeds back to those clouds. Neoclouds borrow against GPU collateral whose value depends on residual assumptions. Large campuses are increasingly funded off balance sheet through joint ventures and private credit.

Some of this is normal behavior in a shortage: suppliers with strong balance sheets extending credit to secure volume. Some of it inflates reported demand relative to end-customer cash. I separate the two by watching:

• Cash conversion at hyperscalers and neoclouds, and the gap between reported AI revenue and operating cash flow.
• Depreciation lives. Most hyperscalers now depreciate servers over five to six years. If the economic life of an accelerator generation is closer to three or four, reported earnings overstate the return on the fleet, and GPU-backed lenders are over-advanced against the collateral.
• Counterparty concentration. Whether contracted backlog is diversified across unrelated buyers, or depends on a handful of labs that are themselves funded by their suppliers.

Scarcity of wafers and power can hide circularity for a long time. When supply loosens, it surfaces quickly, because the marginal buyer disappears at the same moment residual values fall.

Missing productivity

Enterprise AI spend has run well ahead of measured productivity in official statistics. That lag is consistent with every prior general-purpose technology; the dynamo took roughly four decades to show up in manufacturing productivity. The real mistake is to treat a lag as proof that demand is illusory.

Returns become measurable when workflows are rebuilt so that agents act directly on systems of record and proprietary data, not when a chat interface is placed on top of an existing process. The largest gains are likely to be small percentage improvements on very large operating bases: a point of recovery at a mine, a point of efficiency across a generating fleet, lower loss rates on a loan book, fewer false positives in a security operations center, better routing across a logistics network. Each is worth billions, none requires a breakthrough, and the firms that own those assets already hold the domain expertise, operating data, and capital to buy compute.

I size positions on the assumption that productivity gains arrive unevenly and later than the optimists expect, and that the market will overreact in both directions along the way.

4. Binding constraints: domestic power, foundry concentration, and Taiwan

The U.S. leads in frontier models, chip design, hyperscale cloud, and capital formation. It is also attempting to rebuild energy and manufacturing capacity it allowed to atrophy over three decades. Intelligence at scale is a physical product, and the schedule of the buildout is now set by physical supply chains.

1. Power. The question is no longer whether AI requires more electricity but when megawatts can be delivered to a specific site. Interconnection queues in most U.S. ISOs run for years, large power transformers and high-voltage breakers carry multi-year lead times, and heavy-duty gas turbine slots are booked well into the next decade. Firm generation (existing nuclear, uprates and restarts, gas) and behind-the-meter solutions are being valued accordingly. Zoning, transmission rights, and interconnection position often matter more than the generating equipment itself.
2. Foundry and packaging. Leading-edge logic and most CoWoS-class advanced packaging are concentrated in Taiwan. U.S. and allied capacity is growing, but process leadership, yield learning, and packaging scale take multiple product cycles to replicate. Reshoring is a decade-long industrial program, not a single policy announcement.
3. Taiwan. A single geography sits under most of the world's advanced AI silicon. That is not a reason to avoid the chip layer; it is a reason to incorporate sovereignty risk, inventory buffers, and second-source optionality (domestic packaging, allied memory, trailing-edge and specialty nodes) into sizing.

Policy that speeds permitting, domestic generation, transmission, and credible fab and packaging capacity supports the thesis. Policy that confuses exporting the American technology stack to allies with unrestricted transfer to adversaries works against it. Capacity located in the U.S., or under allied control, is what keeps the buildout financeable and insurable.

5. The five-layer map

I map the portfolio to NVIDIA's five-layer framework (Applications, Models, Infrastructure, Chips, Energy) as a way to locate where capital is spent and where the constraints sit, not as a forecast of which model company wins. Each theme below corresponds to positions on the STOCK Watchlist. I list the layers bottom-up, from Energy to Applications, so the binding constraint comes first.

Energy

Power is the gating input for the entire stack. An accelerator without energized capacity is a depreciating asset earning nothing, so delivered megawatts, not installed silicon, determine how quickly contracted demand converts to revenue.

• Nuclear. The existing U.S. fleet is the only large source of firm, carbon-free baseload that can be contracted today, which is why hyperscalers are signing long-dated PPAs, funding uprates, and backing restarts. Factory-built small modular reactors are a longer-dated option on scalable baseload; licensing, first-of-a-kind cost, and fuel supply keep meaningful volume at the end of the decade or later.
• Solar. The fastest new generation to permit and build, and the lowest-cost incremental energy in most U.S. markets. Paired with battery storage, it firms output enough to relieve peak load, and distributed residential systems reduce strain on the distribution grid that data centers compete for.
• Grid. Transformers, switchgear, high-voltage breakers, and transmission are the equipment standing between generation and every campus. Lead times are measured in years, and interconnection position has become a strategic asset in its own right.
• 800V DC. As rack densities approach one megawatt, distributing 54V DC inside the rack requires impractical amounts of copper and loses too much power in conversion. The architecture is moving to roughly 800V DC across the hall, with solid-state transformers, higher-voltage power supplies, busways, and new protection gear, eliminating several AC-to-DC conversion stages between the utility and the chip.
• BTM (behind-the-meter). On-site generation (reciprocating engines, aeroderivative and industrial turbines, fuel cells) that energizes a campus before the utility interconnection is complete, and often remains as backup or peaking capacity afterward. Reciprocating engines and fuel cells scale in modular increments and avoid the heavy-duty turbine backlog.

Chips

Once demand is proven, the scarcity sits in silicon: accelerators, the memory attached to them, the packaging that joins them, and the tools that manufacture all three. Several of these segments are monopolies or tight oligopolies with multi-year capacity lead times.

• GPU. The general-purpose training and inference accelerator, and still the reference platform for the rest of the stack. The moat is the full system (CUDA and its libraries, NVLink scale-up, rack-level integration) and a one-year product cadence that makes competitors chase a moving target.
• CPU. Every accelerator node still needs host processors for scheduling, data preparation, memory management, and I/O. Agentic workloads raise CPU demand further, because tool calls, code execution, and sandboxed environments run on general-purpose cores rather than on the accelerator.
• Inference. As serving volume overtakes training, cost per token and tokens per watt become the metrics that matter. That opens room for architectures optimized for memory bandwidth, low-latency decode, and throughput rather than raw training performance.
• Edge. Inference that has to run on the device (phones, PCs, vehicles, industrial equipment) because of latency, connectivity, cost, or data-privacy requirements. This moves part of the compute bill from the data center to the consumer's hardware purchase.
• Custom ASIC. Hyperscalers designing their own accelerators for their highest-volume internal workloads (Google's TPU, Amazon's Trainium and Inferentia, Meta's MTIA), usually with a merchant design partner. The motive is to reduce dependence on merchant GPU pricing and optimize for known models; it also shifts value toward the design-services partners, packaging, and memory suppliers.
• Analog. Robots, vehicles, and industrial systems depend on sensors, data converters, motor drivers, and power management that cannot be replaced by digital logic. As AI moves into physical systems, analog content per unit rises, in a segment characterized by long product lives and high switching costs.
• Power semis. Silicon carbide, gallium nitride, and high-efficiency controllers that convert and regulate power from the rack input down to sub-1V at the processor. Each point of conversion efficiency translates directly into rack density and cooling load.
• Memory. High-bandwidth memory is the practical constraint on model size and serving throughput, and HBM consumes several times the wafer area per bit of conventional DRAM, which tightens supply across the memory market. Enterprise NAND is absorbing KV-cache offload and checkpoint storage, and high-capacity nearline HDDs remain the lowest-cost medium for training data.
• Foundry. Leading-edge logic nodes and advanced packaging (CoWoS-class 2.5D and 3D integration) that every frontier accelerator depends on. This is the most concentrated chokepoint in the stack, and the foundry's capacity-add discipline is one of the more reliable indicators of how long the cycle has to run.
• EDA & IP. Design software and licensed instruction-set and interface IP without which no advanced chip reaches tape-out. Rising design complexity, more custom silicon starts, and the move to multi-die systems all increase spend per design.
• SME (semiconductor equipment). EUV lithography, where ASML holds a monopoly, along with etch, deposition, and inspection and metrology tools. Gate-all-around transistors, backside power delivery, HBM stacking, and higher-layer 3D NAND all increase process steps and tool intensity per wafer.

Infrastructure

This layer converts chips into delivered tokens. Without the facility, the power path, cooling, and the network fabric, purchased silicon is capacity on paper.

• Optical. At current SerDes speeds, copper links reach only a few meters, so any connection beyond the rack, and increasingly within it, moves to optics. Demand is shifting from pluggable 800G to 1.6T transceivers and toward co-packaged optics on switch silicon, driven by the traffic between GPUs in training and large-scale inference clusters.
• Networking. The scale-up and scale-out fabrics (NVLink, InfiniBand, AI-optimized Ethernet, and emerging standards such as UALink) that let thousands of accelerators run as a single system. Networking performance directly sets cluster utilization; a slow fabric leaves expensive accelerators waiting on each other.
• Data Center. The physical plant: shell, power distribution, liquid cooling, and racks. As densities rise, direct-to-chip liquid cooling and heavier electrical infrastructure are becoming standard, and facilities with secured power and entitlements carry scarcity value.
• NeoClouds. Specialized GPU clouds that absorb demand hyperscalers cannot deliver on the required timeline, serving labs and enterprises with dedicated capacity. The model depends on long-term contracts, GPU-backed financing, and a small number of large customers, so it carries the most leverage to the cycle and the most exposure to residual-value and counterparty risk.
• Hyperscaler. The largest buyers of AI infrastructure and the default platform for enterprise AI deployment. Their capex guidance sets volume for the entire supply chain, and their existing enterprise relationships, security certifications, and data gravity give them distribution the labs do not have.
• Quantum. A long-dated option on the next discontinuity in compute, relevant to simulation, materials, and cryptography. Error-correction progress is real but commercial scale remains uncertain, so the position is sized as an option rather than a core holding.

Models

The model layer determines who controls the weights, and with them whether economic rents accrue to a small number of frontier providers or are distributed across the enterprises and governments that run their own.

• Closed. Frontier models developed by well-capitalized labs and monetized through APIs, subscriptions, and enterprise contracts. They set the capability frontier and fund the largest training runs, but carry the highest capital intensity, the most dependence on supplier financing, and the fastest price erosion as capabilities diffuse.
• Open. Open-weight models that let governments and enterprises deploy on their own infrastructure, fine-tune on proprietary data, and retain control over the weights. Adoption is driven by data sovereignty, regulatory requirements, and cost, and it shifts value toward the infrastructure and application layers rather than the model provider.

Applications

This is where tokens become revenue, cost reduction, or labor substitution. Returns accrue to the firms that already own the workflow, the customer relationship, and the proprietary data, rather than to those renting the same foundation models as everyone else.

• Robotics. Brings AI into physical labor markets. Value is spread across the hardware bill of materials (actuators, reducers, sensors, compute) and the autonomy software stack, with the nearest-term adoption in warehousing, manufacturing, and logistics, where tasks are repetitive and environments can be controlled.
• Drones. Low-cost autonomous aircraft for defense, inspection, agriculture, and logistics. The war in Ukraine demonstrated that attritable, software-defined systems change military procurement, and the same perception and autonomy stack is spreading into industrial use.
• Space. Low-earth-orbit constellations providing global broadband and direct-to-device connectivity, and longer term, orbital compute and sensing. Launch cost declines make space infrastructure a strategic layer for resilient communications and national security.
• Cyber. AI enlarges the attack surface: autonomous agents with their own credentials, faster exploit development, and more machine identities than human ones. Identity, endpoint, and data security become prerequisites for enterprise deployment, and platforms with broad telemetry are best placed to apply AI defensively.
• Devices. The installed base of phones and PCs puts on-device and cloud-assisted inference in front of billions of users, with distribution and payment relationships already established. AI features create a new replacement cycle and raise memory and silicon content per unit.
• Streaming. Large engaged audiences and proprietary content libraries that benefit from AI across recommendation, localization, advertising, and production cost. Owned catalogs and first-party engagement data become more valuable as generic content becomes cheaper to produce.
• Software. Systems of record (ERP, CRM, HR, IT service management, vertical platforms) are where agents attach to real transactions and data. Incumbents that own the workflow and the data model are positioned to capture pricing on automated work; point solutions that sit on top of someone else's system are exposed to being bypassed.

6. Portfolio construction

Early in the cycle, broad exposure to AI hardware worked. Returns are now more dispersed. The constraints are specific: HBM supply, optical capacity for scale-out networks, advanced packaging, custom silicon programs, turbine and transformer lead times, interconnection queues, and the shift of spend from training toward inference, where cost per token is the governing metric.

The book is built around four rules:

• Own inputs that remain scarce regardless of which model leads: power generation and grid equipment, lithography and process tools, leading-edge foundry and packaging, HBM, optics, rack power and cooling, and the networking that ties clusters together.
• Own distribution and workflow: hyperscalers, systems of record, device platforms, and security platforms, where agents reach paying customers.
• Size model-layer exposure as optionality, reflecting narrative risk, capital intensity, and dependence on supplier financing, rather than making it the core of the portfolio.
• Prefer incumbents with large operating bases, hard problems, and proprietary data over new applications built on the same rented foundation models available to every competitor.

As software and model capabilities commoditize, pricing power concentrates in the physical inputs further down the stack and in the incumbents that own the workflows at the top. The middle is where margins compress.

7. What would change my view

I would reduce or restructure the portfolio materially if several of the following occurred together:

• Utilization and cash flow diverge from capacity additions for several consecutive quarters across hyperscalers and the major neoclouds; a pattern, not a single miss.
• Supply discipline breaks. Leading-edge wafer, packaging, HBM, or power equipment capacity is added ahead of contracted demand, while pricing and residual-value assumptions in lease and ABS structures remain unchanged.
• Depreciation catches up with earnings. Hyperscalers shorten server useful lives, or GPU-backed lenders mark down collateral, revealing that returns on the installed fleet are below what reported earnings implied.
• Measured productivity stays flat after several years of agent deployment inside systems of record, turning the productivity bear case from an argument about timing into an observed result.
• A Taiwan disruption occurs before inventory, packaging, and second-source capacity are sufficient to absorb it.
• U.S. policy restricts domestic energy, transmission, and fab construction while still expecting American leadership in AI.

Any one of these is a reason to reassess position sizing. Several at once would indicate the cycle has changed.