The Intelligence Takeoff

The core mistake in most “AI bubble” framing is category error. Investors are treating the current capex cycle like a speculative demand spike layered onto a mature technology stack. It is better understood as the construction phase of a new industrial substrate. When railroads were overbuilt, people still needed rail. When fiber was overbuilt, the world still digitized. In AI, the relevant question is not whether spend has surged; it has. The relevant question is whether intelligence production is becoming a foundational input across software, enterprise operations, search, media, design, coding, security, and eventually physical systems. On current evidence, it is.

NVIDIA’s fiscal 2025 data center revenue exceeded $115 billion, a scale that would have seemed absurd two years ago. Microsoft, Alphabet, Amazon, and Meta together are now committing well over $150 billion annually in capex, with a growing share tied directly or indirectly to AI infrastructure. Skeptics point to the speed of the build-out as proof of excess. That reading misses the fact that these buyers are not promotional startups levering cheap credit to chase vanity metrics; they are some of the highest quality balance sheets in the world, allocating capital against internally observed demand curves. They can see utilization, query growth, training roadmaps, inference margin profiles, enterprise contract pipelines, and user retention in a way the market cannot.

The second misunderstanding is that AI economics are still being judged through a training-centric lens. Training remains expensive and strategically important, but the revenue engine is increasingly inference. Once models are deployed into search, copilots, customer service, coding tools, security workflows, drug discovery, or industrial design, the spend profile changes. Compute becomes a recurring cost of goods sold for intelligence delivered at scale. That matters because recurring intelligence consumption can justify recurring infrastructure replacement, optimization, and vertical specialization.

This is where the stack is starting to look less like a bubble and more like an arms race. HBM supply remains constrained. Advanced packaging remains scarce. Power delivery, cooling, networking, and interconnects are all bottlenecks. Companies such as SK Hynix are not riding a fad; they are supplying a chokepoint component in a constrained production system. New interconnect architectures, including optical and photonic approaches from companies like Celestial AI, are being funded because conventional electrical scaling is running into energy and bandwidth limits. These are not the hallmarks of a demand mirage. They are the hallmarks of an ecosystem straining against real throughput constraints.

The proper investment conclusion is nuanced. There will absolutely be pockets of overearning, misplaced capital, and painful repricing. Not every application company deserves today’s multiple. Not every infrastructure vendor will retain bargaining power. But the top-down claim that “AI capex is a bubble” is too blunt to be analytically useful. A better framework is to separate speculative revenue assumptions from infrastructure necessity. By that standard, the highest-conviction opportunities remain in bottleneck suppliers, enabling architectures, and platforms that turn model quality into durable enterprise workflows.

The current capex cycle is historically large, but it is not historically unprecedented when framed against the size of the platforms deploying it. Alphabet, Microsoft, Amazon, and Meta each sit atop multi-hundred-billion-dollar revenue bases and have business models in which incremental infrastructure can reinforce already dominant distribution. In prior cycles, telecom operators often borrowed heavily to build capacity ahead of uncertain monetization. In the present cycle, hyperscalers are funding AI investment from operating cash flow generated by cloud, ads, software, and commerce franchises with entrenched market positions.

That distinction matters. A dollar of AI capex from Microsoft is not the same economic object as a dollar of speculative broadband rollout from a leveraged carrier in 1999. Microsoft can absorb lower near-term returns because AI strengthens Azure, Office, GitHub, security products, and its broader enterprise account control. Alphabet can fund Gemini and TPU deployment not just to sell AI services directly, but to defend search economics, improve ad targeting, expand Google Cloud, and reduce dependence on third-party accelerators over time. The investment is strategic, not purely financial.

Another contextual error is ignoring the denominator. The world’s largest compute buyers are digitizing a widening set of activities. Search queries are becoming multi-step reasoning sessions. Coding assistants are moving from autocomplete toward agentic workflow execution. Enterprise software suites are being rebuilt around copilots. Consumers increasingly expect AI-generated media, summarization, translation, and conversational interfaces as default features. Each of those shifts turns a low-compute or moderate-compute activity into a higher-compute one.

At the component layer, the capex surge is inseparable from the memory and networking surge. HBM demand has become one of the cleanest indicators that the market is dealing with a genuine systems transition rather than a short-lived spending fever. Modern accelerators are no longer compute-bound in isolation; they are system-bound. If the GPU cannot be fed, the expensive silicon is underutilized. That is why suppliers such as SK Hynix have captured extraordinary pricing power. It is also why advanced packaging houses and substrate suppliers have become strategically relevant to the entire AI roadmap.

The network effect of this capex is physical. A data center designed for conventional cloud workloads is not automatically fit for high-density AI clusters. Operators need different rack power envelopes, thermal management, liquid cooling adoption paths, networking topologies, storage architectures, and software orchestration layers. Rebuilding that footprint takes time and money. It is one reason the spend looks so dramatic in reported numbers: companies are not merely buying chips; they are reconfiguring the physical fabric of computation.

Seen this way, the capex surge is less analogous to a customer fad and more analogous to a manufacturing sector tool-up ahead of a new production regime. The product being manufactured is machine intelligence. The hyperscalers are the first industrializers. The supply chain is still tight, the architecture is still evolving, and the profit pools are still sorting themselves out. But the scale of the build is consistent with the scale of the opportunity.

The central bearish argument against AI infrastructure has been simple: where is the revenue? That question was fair in the early training wave, when large models consumed enormous compute but monetization lagged. It is now incomplete. Revenue conversion is happening, but unevenly, and often in forms that classical software investors underweight. Some of it appears directly as API revenue or premium subscription ARPU. More of it appears as retention improvement, seat expansion, task substitution, support cost reduction, cloud wallet capture, or search defense.

Take Microsoft. The company does not need Copilot alone to justify infrastructure deployment. If AI features make Office stickier, increase E5 penetration, pull enterprise workloads into Azure, and reinforce GitHub’s position with developers, then compute is monetizing across multiple lines. Alphabet faces a similar dynamic. AI overviews in search may pressure some legacy ad placements, but they also protect the search habit, keep users inside Google’s surface area, and create new higher-intent commercial moments. The revenue conversion path can therefore be indirect and still be economically rational.

NVIDIA’s own numbers are a useful clue. A business does not grow data center revenue past $115 billion in fiscal 2025 unless customers believe the hardware can generate real returns or preserve critical strategic ground. Some of those returns are clearly pre-positioning. But the breadth of the buyer base — hyperscalers, sovereigns, enterprise service providers, model labs, and increasingly vertical software companies — suggests more than pure experimentation. The installed base is feeding actual workloads.

There is also a timing mismatch that creates confusion. Compute is procured in chunks. Revenue arrives in streams. A company might buy a cluster in quarter one, spend six months integrating it into products, and only begin to show the resulting revenue uplift over the following year. Public markets often compare a lumpy capital outlay against a still-forming revenue line and declare the ratio absurd. That misses the operational cadence of infrastructure businesses.

Inference further improves the revenue picture because it converts models from research artifacts into continuously consumed services. Training happens episodically at the frontier. Inference happens every time a model answers a query, rewrites a contract, ranks a sales lead, debugs code, or interprets a radiology image. As throughput improves and unit costs fall, providers can widen the set of use cases that clear economic thresholds. That is what turns a capex cycle into a durable demand cycle.

The proper analytical frame is therefore compute-to-economic-output conversion, not merely compute-to-immediate-line-item-revenue conversion. When AI lowers labor intensity, speeds internal throughput, improves user engagement, raises product quality, or defends an existing cash-generating franchise, infrastructure is being monetized even if the accounting presentation lags investor expectations. That does not mean every dollar is well spent. It means the revenue test has to be applied with more sophistication than a simple one-quarter payback screen.

The most important shift underway is from a training-defined narrative to an inference-defined one. Training has dominated headlines because frontier model runs are spectacularly expensive and technically dramatic. But training is only the R&D and model-creation side of the equation. The economic scale comes when models are deployed into live workflows. That is where infrastructure demand becomes broader, steadier, and more difficult to reverse.

Inference has several properties that change capital allocation. First, it is persistent. Once an enterprise embeds a model into customer service, software development, compliance review, or internal knowledge retrieval, the company is committing to an ongoing compute bill. Second, inference is latency-sensitive and often user-facing, which increases the value of optimized serving hardware, memory bandwidth, interconnect efficiency, and software stack integration. Third, inference volume can grow much faster than training volume because each successful product can generate billions of requests.

This matters for the whole supply chain. HBM demand is not only a frontier training story; high-throughput inference systems also benefit from memory bandwidth and efficient data movement. Networking architectures become central because multi-accelerator systems must serve at scale without wasting power in communication overhead. Storage hierarchies and caching strategies start to determine gross margins. The infrastructure conversation broadens from “who can train the largest model” to “who can serve intelligence cheaply, quickly, and reliably.”

That is where new technologies such as photonic interconnects become relevant. Electrical interconnects face growing penalties in power and distance as cluster sizes expand. Celestial AI and related efforts are trying to reduce the energy tax of moving data between processors and memory-rich systems. If they succeed, they will not merely shave costs; they will alter the feasible architecture of large inference clusters and potentially increase utilization of the most expensive silicon in the stack.

The inference inflection also changes competitive positioning among model providers and hyperscalers. In a training-centric world, access to massive capital and research talent dominates. In an inference-centric world, distribution, customer integration, software tooling, and system-level optimization matter more. That favors incumbents with large installed user bases and enterprise trust. It also means infrastructure investment can defend platform power even if the frontier model lead rotates.

From an investor standpoint, inference is the bridge between today’s heavy spending and tomorrow’s durable cash flows. It is the mechanism by which AI stops being a science project and becomes an operating layer. Once that transition is accepted, the current build-out looks less like speculative overcapacity and more like the inevitable overshoot-and-correct pattern that accompanies the early industrialization of a general-purpose technology.

Traditional bubble diagnostics are poorly suited to this phase of AI because they focus on valuation excess, promotional fervor, and weak-fundamentals capital formation. Some of those conditions exist in pockets of the market, especially among application names with thin moats and inflated TAM narratives. But they are a poor fit for the capex decisions of the major infrastructure buyers.

One failed metric is the idea that rapid spending growth itself proves irrationality. In industries experiencing architecture shifts, capex often comes in discontinuous waves. Data centers designed for CPU-centric cloud workloads are not adequate substitutes for AI-optimized clusters. Power density, cooling, networking, and memory requirements all jump. If the input requirements change discontinuously, rational investment can also change discontinuously.

Another failed metric is short-term utilization anxiety. Bears often assume that any temporary underutilization means the cycle is broken. In reality, large infrastructure systems are intentionally built ahead of fully realized demand because product launches, enterprise rollouts, and developer ecosystems cannot scale against a perpetually constrained base. Temporary slack can be economically rational if it accelerates ecosystem capture or prevents competitors from locking in scarce supply first.

A third failed metric is the comparison of AI hardware spend to immediate software monetization. This misses the fact that AI is not one product category. It is a capability layer that will be embedded across search, productivity, software development, security, healthcare, design, logistics, and industrial systems. Asking whether current capex is justified by present standalone AI revenue is like asking whether early cloud capex was justified only by infrastructure-as-a-service revenue while ignoring SaaS, digital media, and mobile app ecosystems that later sat on top of it.

Even valuation multiples can mislead. NVIDIA’s multiple at various points has looked stretched on trailing frameworks and oddly reasonable on forward consensus, depending on where one catches the cycle. But the more important issue is not whether every hardware leader is perfectly priced today. It is whether the market is misclassifying a structural demand curve as a temporary mania. A company can be both strategically indispensable and tactically overvalued. Those are separate questions.

Finally, bubble metrics tend to underweight supply chain friction. In true speculative booms, supply can often rush in and commoditize returns quickly. In AI infrastructure, returns are protected by technical complexity, manufacturing bottlenecks, packaging scarcity, memory concentration, and integration know-how. HBM capacity cannot be doubled overnight. Advanced lithography cannot be conjured without ASML tools and long lead times. Next-generation interconnects require years of ecosystem validation. Those frictions dampen the classic boom-bust oversupply dynamic.

None of this means risk is low. It means the right risk lens is execution, timing, and margin migration — not simplistic analogy. The market will probably suffer periodic panics when spending outruns visible revenue, but those panics should be read as cycle volatility inside a real build-out, not automatic proof that the build-out itself is illusory.

The hyperscalers are spending because AI threatens to reorder both offense and defense across their franchises. For Microsoft, the prize is deeper enterprise entrenchment and Azure share gain. For Alphabet, it is search defense, cloud growth, and maintaining control of the consumer information interface. For Amazon, it is preserving AWS leadership while turning AI into a higher-value layer over its infrastructure base. For Meta, it is using AI to improve ad performance, content recommendation, and developer leverage. Each company can justify infrastructure spend even under conservative direct monetization assumptions because the strategic downside of underinvesting is severe.

This is what many external observers miss: the return hurdle is not a textbook corporate finance hurdle detached from competitive dynamics. The relevant comparison is often between spending heavily and risking mediocre returns, versus spending too little and losing distribution, developers, enterprise accounts, or search habit. In that context, overspending can be rational insurance.

The calculus is reinforced by procurement realities. NVIDIA supply has historically been constrained at the highest-end accelerator tiers. HBM availability has been tight. Advanced packaging capacity has been limited. That means capex timing is also supply reservation. If a hyperscaler waits for perfect visibility before ordering, it risks losing place in the queue. Rational buyers in constrained markets often pull spend forward.

There is a software angle as well. Owning the infrastructure stack creates optionality in pricing, model deployment, and vertical integration. Alphabet’s TPU program is partly about cost and partly about strategic independence. Microsoft’s deep integration with OpenAI and its own tooling stack strengthens control over how developers consume AI. AWS is building choice into Bedrock while still needing the physical backbone to capture workload demand. Infrastructure is not just capacity; it is bargaining power.

Importantly, hyperscalers do not need all AI demand to normalize at today’s feverish pace for the investments to work. They need enough enduring demand to support high utilization on core fleets, plus continued migration of enterprise and consumer workflows toward AI-assisted operation. Given the number of surfaces now being rebuilt around language, vision, and multimodal reasoning, that threshold may be lower than skeptics assume.

In practical terms, the hyperscaler calculus says the following: if AI becomes the primary interface to digital work, you cannot afford to be short compute, short memory bandwidth, short network throughput, or short developer trust. If AI becomes merely a large product category rather than a universal interface, the investments still likely reinforce cloud and platform franchises. Either way, the incentive to build remains strong. The disagreement is about magnitude and timing of returns, not about whether the strategic imperative exists.

The deepest reason AI infrastructure spend is not a bubble is that infrastructure itself is becoming a moat. In software, investors became accustomed to asset-light business models in which code scaled cheaply and capital intensity often signaled weakness. AI changes that equation. At the frontier and increasingly in production, performance, latency, reliability, and cost are all functions of physical system design. The companies that control scarce infrastructure, or the bottlenecks that enable it, can hold strategic advantage for longer than software investors expect.

NVIDIA is the clearest example. Its moat is not just superior chips. It is an integrated stack of accelerators, networking, software libraries, developer trust, and a release cadence that keeps customers near the frontier. SK Hynix’s role in HBM adds another layer: even the best accelerator design is constrained if memory supply lags. The system is only as strong as its narrowest chokepoint. That creates a chain of moats across the stack rather than a single-winner market.

The same logic is opening room for emergent infrastructure players. Companies working on optical connectivity, advanced packaging, power delivery, thermal management, and memory-centric architectures are not peripheral beneficiaries. They are candidates to become essential. Celestial AI is relevant precisely because moving data efficiently is becoming one of the defining technical and economic problems of large-scale AI. If photonic fabrics materially lower energy per bit and expand bandwidth, they could unlock new cluster designs and shift where value accrues.

Infrastructure moats also propagate upward. A hyperscaler with lower inference cost can price more aggressively, ship richer product experiences, and tolerate longer customer adoption cycles. A cloud vendor with better AI network architecture can win enterprise workloads that care about latency and reliability. A model provider with privileged access to optimized infrastructure can iterate faster and defend margins. The moat compounds across hardware, software, and distribution.

This does not mean every infrastructure layer has permanent pricing power. Over time, some components will commoditize, margins will normalize, and standards will diffuse. But commoditization tends to arrive later in systems with manufacturing complexity, qualification cycles, and ecosystem dependency. The market is still early enough that bottleneck control matters more than eventual end-state commoditization.

For capital allocators, this reframes the opportunity set. Rather than asking which AI application has the prettiest demo, it may be more profitable to ask which system constraint must be relieved for the next order-of-magnitude scale increase to occur. The winners in such periods are often the picks-and-shovels businesses solving the stubborn physical problems the market cannot wish away.

The investment takeaway is not to buy “AI” indiscriminately. It is to separate structural beneficiaries from narrative passengers. The highest-confidence beneficiaries remain companies controlling bottlenecks in compute, memory, lithography, packaging, network throughput, and power-efficient data movement. NVIDIA remains the flagship, but the broader list includes memory leaders such as SK Hynix, enabling equipment providers, advanced interconnect and optical fabric companies, and selected hyperscalers whose installed distributions allow them to monetize intelligence at scale.

The second implication is to be disciplined about where the market is underestimating duration. If inference becomes a long-lived consumption layer, then suppliers positioned around recurring deployment economics deserve more credit than those tied purely to one-off training cycles. This creates a premium for businesses aligned with memory bandwidth, serving efficiency, rack-level optimization, and interconnect innovation.

Third, investors should expect value migration within the stack. Early in a new infrastructure cycle, the obvious silicon winner captures outsized economics. Over time, constraints move. Memory suppliers gain leverage. Networking becomes harder. Cooling and power architecture matter more. Software orchestration determines utilization. Optical and photonic solutions may step from experimental to essential if electrical approaches hit diminishing returns. The market opportunity is therefore dynamic rather than static.

Fourth, volatility should be used analytically rather than emotionally. The sector will continue to suffer from impossible expectations, periodic digestion phases, and sharp repricing around capex disclosures. Those drawdowns will be real. But if the underlying thesis is correct — that machine intelligence is becoming a core production input — then cyclical disappointment should create windows into structurally advantaged assets rather than invalidate the build-out wholesale.

Finally, portfolio construction should reflect that this is a systems transition, not a single-company story. Exposure can be layered: dominant accelerators, memory chokepoints, lithography monopolies, hyperscaler platform winners, and selective emergent architecture bets. What should be avoided is lazy bubble language that treats all spend as equal and all valuation risk as thesis risk. The world may well overbuild some AI capacity. It is still very likely to need far more intelligence infrastructure than it has today.

That is the heart of the intelligence takeoff thesis. The spending is large because the prize is large. Some capital will be wasted, some narratives will collapse, and some equities are already discounting perfection. But the build itself is grounded in genuine demand, real bottlenecks, and powerful strategic incentives. That is not what a bubble looks like at its core. It is what the early industrialization of a new general-purpose capability looks like.