The Real AI Infrastructure Investing Lesson Behind Intel Intel was supposed to be the loser of the AI cycle. GPUs took the spotlight. The market wrote off CPUs…

The Real AI Infrastructure Investing Lesson Behind Intel

Intel was supposed to be the loser of the AI cycle.

GPUs took the spotlight. The market wrote off CPUs as legacy, and Intel as structurally behind.

Then Intel reported Q1 revenue of $13.6 billion, beating expectations for the sixth consecutive quarter. Its Data Center and AI division printed $5.1 billion, up 22% year over year. The stock pushed through levels last seen in the dot-com bubble.

If AI was a one-way GPU trade, that shouldn’t have happened.

This is the real lesson for AI infrastructure investing: narratives peak around the visible tip of the stack, but cash flows compound in the layers everyone ignores—until they don’t.

---

Why Intel Became the "AI Loser"—Then Broke Dot-Com Highs

The market’s original AI verdict on Intel

The early AI trade was simple:

• AI = large models
• Large models = training
• Training = GPUs
• GPUs = one ticker

Intel failed that story.

It wasn’t the GPU leader. It wasn’t the brand taped to AI conference keynotes. It was the incumbent CPU manufacturer that had supposedly missed the wave.

So the narrative hardened:

"Intel is structurally late to AI."

Capital rotated into the pure-play GPU story. The market priced in a future where GPUs captured the lion’s share of AI economics and CPUs became plumbing.

What actually showed up in Intel’s numbers

While the narrative drifted, the P&L started to tell a different story:

• Q1 revenue: $13.6 billion
• Sixth consecutive quarter beating expectations
• Data Center and AI division: $5.1 billion, +22% year over year
• Stock: broke above dot-com-era highs

Those are not the numbers of a company structurally excluded from AI.

What changed? Not a marketing slogan.

The AI workload did.

---

The AI Stack Is Shifting: From Training to Inference and Agents

The first chapter of AI was training headline models. The next chapters are more mundane—and much larger.

Training: where GPUs dominated the first chapter

Training is capital-intensive, bursty, and dominated by a few large players:

• You raise billions.
• You rent or buy dense GPU clusters.
• You train a frontier model.

This phase created the impression that AI was a single-asset, GPU-only trade. For a while, it was directionally correct. Training is embarrassingly parallel; GPUs excel at it.

But training is not where most of the world’s compute will live.

Inference: the compounding, less glamorous layer

Inference is what happens after training:

• Models are deployed into products.
• Users query them in real time.
• Enterprises embed them into workflows.

Inference has different characteristics:

• Recurring rather than episodic
• Operational rather than experimental
• Latency- and cost-sensitive, not just "maximum horsepower"-driven

It’s also where:

• CPUs matter more.
• Efficiency, density, and orchestration start to dominate.
• The market severely underappreciated who would benefit.

As workloads shift from training to inference, the stack rebases on infrastructure that can do more than just peak parallelism. It needs to integrate, schedule, secure, and serve at scale.

Agentic AI and edge workloads: where scale really lives

Now layer on agentic AI and edge computing:

• Agentic AI moves from simple prompts to autonomous task sequences.
• Edge computing moves from centralized data centers to distributed devices and local nodes.

Both trends share a common feature:

They drive persistent, distributed, real-world compute—not occasional lab runs.

That means:

• More orchestration across heterogeneous hardware
• Tighter coupling with existing CPU-based systems
• Heavy emphasis on cost, power, and integration at the edge

This is precisely the territory where a scaled CPU incumbent with deep data center and edge reach can reassert itself.

---

CPU vs GPU in AI: What Most Investors Missed

GPUs as the poster child of the AI trade

GPUs earned the early narrative:

• They enabled the jump in model size and capability.
• They became synonymous with "owning AI."
• They delivered outsized equity performance.

Most allocators stopped the analysis there.

The implicit assumption was binary:

• GPUs win; CPUs are commoditized.

Reality is more layered.

The CPU’s quiet reassertion in the AI era

The CPU did not disappear in AI. It shifted roles:

• From starring in the narrative
• To running the control plane, coordination logic, and a significant share of inference

As AI workloads mature, CPUs are:

• Managing data pipelines and pre/post-processing
• Orchestrating GPU fleets in data centers
• Powering a large portion of inference workloads, especially where latency, cost, and integration matter more than raw peak FLOPs

The transcript’s core point is blunt:

"The CPU is reasserting itself as the foundation of the AI era."

That doesn’t mean GPUs are irrelevant. It means AI infrastructure is not a single-chip story. It’s a system story.

Why Intel’s position matters for future AI cash flows

Intel sits at several key junctions of that system:

• Data centers: entrenched relationships, tooling, and platforms
• Edge: CPUs already embedded in endpoints and gateways
• Process technology: a roadmap that major ecosystem players are now explicitly underwriting

As AI activity normalizes from "experiments" to "infrastructure," the cash flows migrate toward:

• Providers that can deliver at scale
• Architectures that balance performance, power, and cost
• Incumbents that already own integration into enterprise stacks

That’s what you see reflected in Intel’s Data Center and AI division growth and the stock’s repricing.

---

What Google, Nvidia, and Musk Are Signaling About AI Infrastructure

When operators with real P&Ls make hardware decisions, investors should listen.

Google’s AI workloads on Intel

Google is running AI workloads on Intel chips.

This isn’t a sentimental decision. It’s:

• A cost–performance trade-off
• A bet on ecosystem maturity
• A recognition that large-scale AI services are not GPU-only forever

For allocators, this is a tell: hyperscalers are optimizing the full stack, not worshiping a single component.

Nvidia selecting Intel for next-gen systems

The most revealing signal is from the apparent "competitor":

Nvidia selected Intel for its next-generation systems.

This underscores an often-missed point:

• The AI boom is systems-driven, not monolithic.
• Even the GPU leader is partnering with CPU and process incumbents where it makes economic and engineering sense.

If Nvidia is underwriting Intel’s role in its future systems, the idea that Intel is "outside" the AI infrastructure story becomes difficult to defend.

Terafab and Intel’s advanced process node

Elon Musk’s Terafab is building on Intel’s most advanced process node.

Again, strip out the personalities and focus on capital allocation:

• Sophisticated operators are locking in capacity on Intel’s cutting-edge process.
• That is a forward statement of belief in Intel’s infrastructure relevance.

The through-line:

• Google isn’t nostalgic.
• Nvidia isn’t charitable.
• Musk isn’t sentimental.

They’re making hard-nosed decisions about where AI infrastructure will actually sit.

---

A Playbook for AI Infrastructure Investing Beyond the Hype

The Intel story is not about one ticker. It’s an allocation lesson.

Separate AI narratives from AI cash flows

Most investors bought the AI billboard.

Serious capital needs to own the plumbing:

• Narratives live in presentations and headlines.
• Cash flows live in:
  • Contracts
  • Capacity reservations
  • Long-lived infrastructure

When evaluating AI infrastructure investing, ask:

• Who is getting pre-committed capacity?
• Whose utilization is rising as AI services go live, not just get announced?
• Where are hyperscalers, model labs, and large operators quietly building dependencies?

Map the full stack: training, inference, agents, edge

A basic stack map for AI infrastructure:

• Training layer: GPU-dense, capital-intensive, episodic
• Inference layer: recurring, often CPU-heavy, sensitive to latency and cost
• Agentic layer: orchestrates tasks, tools, and workflows, driving persistent compute
• Edge layer: pushes AI into devices, local nodes, and real-world environments

Each layer has distinct characteristics:

• Different utilization patterns
• Different hardware mixes
• Different supplier power dynamics

The early trade captured training. The durable trade will capture inference, agents, and edge.

Why overlooked infrastructure can be the best risk-adjusted trade

The infrastructure play that most investors ignored is now printing results.

This is a recurring pattern:

• The market writes off an incumbent too early.
• A new technology wave appears to bypass them.
• Quietly, the value chain shifts in their direction.
• By the time the numbers are undeniable, the easy multiple has moved.

For allocators across public and private markets, the Intel example suggests a simple discipline:

Look for assets the narrative has declared "finished" while the cash flows are just getting started.

In AI, that often means:

• CPUs alongside GPUs
• Data centers alongside model labs
• Edge systems alongside headline cloud platforms

---

FAQ: AI Infrastructure Investing for Institutional and Private Capital

What is AI infrastructure investing?

AI infrastructure investing focuses on the compute, networking, data center, and edge systems that power AI workloads, rather than on consumer-facing apps or headline AI brands. It means underwriting semiconductors, data centers, connectivity, and capital-intensive platforms that turn AI demand into durable, contracted cash flows.

How is AI inference different from AI training for investors?

Training is the upfront, compute-heavy process of building AI models; inference is the ongoing process of running those models in production. Training is episodic and GPU-heavy. Inference is recurring, scale-driven, and often more CPU-anchored. For investors, inference can represent a steadier, larger base of infrastructure demand over time.

Are CPUs really relevant in the AI era dominated by GPUs?

Yes. GPUs are critical for parallelizable training workloads, but CPUs remain the control plane and backbone for orchestrating AI systems, running many inference tasks, and enabling edge and agentic workloads. As AI moves from experiments to embedded, always-on services, CPU-centric infrastructure has room to compound value.

Why did the market misprice Intel’s AI exposure?

The market anchored on the early AI narrative: training equals GPUs, and GPUs equal AI. That left little room for an incumbent CPU player that appeared late to the training wave. As the value shifted toward inference, agents, and the edge, Intel’s positioning started to show up in revenue before it fully registered in the narrative.

How can private capital participate in AI infrastructure beyond public equities?

Private capital can underwrite data center build-outs, structured capital for semiconductor and equipment ecosystems, edge infrastructure platforms, and cash-flowing service businesses tied to AI workloads. The principle is the same: focus on contracted, infrastructure-like exposure to AI demand rather than purely speculative thesis beta.

What is agentic AI, and why does it matter for infrastructure?

Agentic AI refers to systems that can autonomously sequence tasks, call tools, and interact with other software without constant human prompts. That shifts compute demand from occasional queries to persistent, workflow-level activity—intensifying requirements for scalable, efficient infrastructure across CPUs, memory, networking, and edge devices.

---

Manhattan’s View: Where Serious Capital Goes in the AI Era

The Intel episode is not about being contrarian for its own sake. It’s about being early to where the AI value chain is actually migrating.

Most investors chased the obvious GPU story. The operators who matter—Google, Nvidia, Musk—quietly wired their future into a broader infrastructure base that includes Intel.

That’s the environment Manhattan Private Credit operates in:

• Event-driven, capital-structure-aware
• Focused on infrastructure-like cash flows
• Comfortable in the parts of the market the narrative has already dismissed

The biggest opportunities rarely look obvious at the start.

Learn more at manhattanprivatecredit.com.