2025 AI Infrastructure Stock Scorecard for Supply Chain

A supply chain AI contract does not fail because a semiconductor stock misses a quarter. It fails when the platform cannot scale, the cloud bill changes the economics, the vendor loses access to compute, or the board starts asking whether the company bought into an infrastructure bubble. That is why the 2025 AI infrastructure stock price target conversation matters to procurement teams, even if nobody in the room is trying to trade NVIDIA, Micron, Broadcom, or AMD.

The better question is narrower than the market usually makes it: did 2025 validate the infrastructure stack that supply chain AI platforms depend on, or did it expose that stack as speculative? On the evidence available through July 2026, the buildout looks financially durable. The caution is that infrastructure durability says almost nothing by itself about whether a demand planning model, inventory optimization engine, routing assistant, or warehouse AI pilot will produce measurable operating value.

Start With CapEx, Not The Ticker Tape

Stock performance is a useful confirmation signal, but hyperscaler capital spending is the foundation. Microsoft, Google, Amazon, and Meta increased AI CapEx from about $100 billion in 2023 to about $258 billion in 2024, then to an estimated $405 billion-plus in 2025, a 58% year-over-year increase. For 2026, Big Tech AI CapEx was projected above $500 billion.[1]

That spending is not abstract enthusiasm. It buys data centers, GPUs, custom accelerators, high-bandwidth memory, switching, optical interconnects, power systems, and cloud capacity. Those are the layers underneath the AI features now being sold into supply chain planning suites, transportation management systems, procurement analytics tools, and warehouse automation platforms.

Longer-term projections add weight, with Morgan Stanley projecting $2.9 trillion in additional AI infrastructure spending through 2028 across data centers, chips, networking, and power infrastructure.[2] A buyer signing a multiyear AI platform contract still has plenty to diligence, but the infrastructure side of the stack is not being funded like a short-lived experiment.

The 2025 AI Infrastructure Stock Scorecard

The scorecard is most useful when each company is read as an infrastructure layer, not as a stock recommendation. The point is not whether a procurement leader should own the shares. The point is whether capital markets, analysts, and customers continued to support the suppliers behind the compute, memory, networking, and custom silicon that AI platforms consume.

The cutoff issue matters. NVIDIA’s roughly 42% figure is a calendar 2025 return, while AMD’s roughly 155% figure comes from trailing-12-month data available in July 2026, so those two numbers do not describe the same measurement window.[3][4] That does not make either number useless. It does mean the scorecard should be read as infrastructure validation, not a clean league table.

NVIDIA: The GPU Backbone Still Had Institutional Support

NVIDIA remains the easiest company to overuse in an AI argument. A rising NVIDIA share price does not prove that a supply chain control tower will forecast a stockout correctly. It does, however, show that the GPU layer behind training and inference retained broad financial support after the first wave of AI enthusiasm should have faced harder scrutiny.

In calendar 2025, NVIDIA was up about 42%. By July 2026, reporting cited Q1 FY27 revenue of $81.61 billion and 95% analyst bullishness.[3] For supply chain buyers, the useful reading is not “buy the leader.” It is that the core compute supplier for many AI platforms was still producing earnings momentum and analyst confidence at a scale that makes sudden ecosystem abandonment unlikely.

That matters when a vendor’s roadmap depends on GPU availability for model retraining, scenario simulation, synthetic data generation, or inference-heavy exception management. The buyer still needs service-level terms, usage economics, and model governance. But the underlying GPU ecosystem was not starved of capital in 2025.

Micron: The Memory Bottleneck Became A Procurement Issue

Micron deserves more attention from supply chain teams than it usually gets. High-bandwidth memory is not a decorative component in AI infrastructure. It influences throughput, latency, and the practical economics of running larger models at production scale.

Micron was reported up about 145% year to date in 2025, with EPS projected to double to $16.68.[5] That is not a generic semiconductor rebound story. It points to a constraint layer in AI infrastructure attracting capital because production inference requires more than raw GPU count.

For a supply chain AI buyer, memory constraints show up indirectly. A planning platform may become slower as more users run scenarios. A logistics model may cost more to serve when exception volume spikes. A warehouse AI application may perform well in a pilot and then struggle when multiple sites, SKUs, and event streams are added. HBM strength does not solve those problems by itself, but it confirms that the market is funding one of the bottlenecks that would otherwise cap production AI.

Broadcom: Custom Silicon Is Not A Side Story

Broadcom’s role is different from NVIDIA’s. It is not mainly the visible GPU brand attached to AI demos. Its relevance is the hyperscaler move toward custom silicon, especially when inference cost becomes the limiting factor for production workloads.

Broadcom reported AI semiconductor revenue of $10.8 billion in Q2 FY26, up 143% year over year, and 92% of analysts were bullish.[3] That combination matters because hyperscalers are not only buying general-purpose AI capacity; they are also trying to control unit economics through custom accelerators.

Supply chain AI will need that efficiency if it becomes embedded in daily operations rather than reserved for occasional planning runs. Continuous ETA prediction, dynamic inventory positioning, labor planning, supplier risk monitoring, and automated exception triage all become more expensive when inference is always on. Custom silicon gives the cloud providers another path to reduce cost per workload.

AMD: Diversification Changes The Buyer’s Risk Profile

AMD is best read as a competitive pressure signal. Supply chain buyers do not need AMD to displace NVIDIA for AMD to matter. They need credible alternatives that improve supply availability, pricing discipline, and vendor negotiating leverage.

As of July 2026, AMD was reported up about 155% over the trailing 12 months, and 82% of analysts were bullish.[4][3] Because that return blends late 2025 and part of 2026, it should not be described as a clean calendar 2025 performance number. Its procurement meaning is still clear enough: capital markets were rewarding a second major GPU supplier at the exact point when AI buyers needed the compute market to become less brittle.

That does not remove concentration risk. It does make it easier to believe that cloud providers and AI platform vendors will have more than one path for scaling workloads over a multiyear contract term.

Marvell And Arista: The Less Glamorous Layers Still Count

Networking is where some AI infrastructure scorecards become too neat. Marvell and Arista are not interchangeable with the compute and memory names, and they do not need to be forced into the same role. Their relevance is the data movement layer: switches, interconnect, and data center fabrics that let AI workloads operate across large cloud environments.[1][6]

For supply chain applications, that layer becomes more important as AI spreads across facilities, suppliers, carriers, planning systems, and execution systems. Distributed inference is not just a model problem. It is a data center and network problem, especially when users expect recommendations fast enough to change a shipment, rebalance inventory, or trigger a warehouse action.

What The Scorecard Actually Validates

The 2025 scorecard validates financial durability at the infrastructure layer. It says hyperscalers were still spending heavily, chip and memory suppliers were still receiving market support, and analysts still saw enough earnings momentum to remain broadly constructive on key names. For a supply chain executive, that is useful because the AI vendor ecosystem depends on these companies whether the invoice comes through a SaaS provider, a cloud marketplace, or a systems integrator.

It does not validate a vendor’s feature claims. A transportation AI module can run on world-class infrastructure and still fail because carrier data is late, exception codes are inconsistent, operating teams do not trust recommendations, or the workflow never reaches the person who can act. A forecasting engine can use excellent compute and still miss value if planners override it outside any measurable governance loop.

This is where the MIT finding is clarifying rather than contradictory: 95% of generative AI pilots failed to achieve measurable P&L impact, according to research reported in July 2025.[7] That figure does not disprove the infrastructure buildout. It says application execution is a separate problem.

Infrastructure Momentum And Pilot Failure Can Both Be True

The most expensive mistake in an AI procurement review is treating infrastructure validation as application proof. The second most expensive mistake is treating failed pilots as proof that the infrastructure buildout is fake. They answer different questions.

A good AI procurement review therefore has to split the diligence. The infrastructure question asks whether the vendor’s cloud, model, compute, and scaling assumptions depend on a market that is still being funded. The application question asks whether the proposed workflow can reduce stockouts, improve fill rate, lower inventory, increase planner productivity, reduce expedite costs, or improve labor utilization in a way finance can measure.

The first question got a strong answer in 2025. The second question still has to be earned one deployment at a time.

How To Read Price Targets In A Supply Chain AI Buying Process

Price targets and analyst bullishness should sit in the vendor-risk section of the buying process, not in the ROI model. They help answer whether the infrastructure suppliers behind the platform are likely to keep expanding capacity, attracting talent, and serving hyperscaler demand. They do not belong in the cell where the team estimates working capital reduction or service-level improvement.

• Use CapEx as the primary durability signal: hyperscaler spending shows whether cloud capacity is being built before individual stock commentary confirms the market’s view.
• Separate calendar-year returns from trailing-12-month returns: blended windows can make a scorecard look more precise than it is.
• Map each company to a layer: NVIDIA and AMD for compute, Micron for HBM, Broadcom for custom silicon, Marvell and Arista for networking and interconnect.
• Ask the vendor where inference runs: public cloud, private cloud, edge device, custom silicon-backed service, or a mix.
• Treat infrastructure strength as a risk reducer, not a value guarantee.

The cleanest buying memo would say something like this: the infrastructure market supporting this vendor’s AI roadmap appears durable through the contract horizon, but the business case depends on our data readiness, workflow adoption, and ability to measure operational outcomes. That sentence keeps the two risks in the right boxes.

The Power Constraint Belongs In The TCO Model

Power is the constraint that prevents the infrastructure story from becoming too comfortable. Data centers could consume up to 12% of U.S. electricity by 2028, up from about 4% in 2023.[1] That does not make the buildout imaginary. It makes capacity planning and total cost of ownership more important.

For supply chain AI buyers, power exposure will often arrive through pricing rather than a direct utility bill. Cloud providers may pass higher infrastructure costs into AI services. Vendors may introduce usage tiers, inference limits, premium response-time options, or workload scheduling rules. Compute-efficient architectures will matter more as AI moves from quarterly experiments to daily execution.

This is also where custom silicon and model efficiency become procurement topics. If two vendors produce similar planning accuracy, but one requires far more inference spend to support real-time operations, the cheaper demo may become the more expensive platform. The infrastructure scorecard says capacity is being funded; the power constraint says that capacity will not be free.

The Procurement Judgment

The 2025 AI infrastructure stock scorecard gives supply chain leaders a credible durability signal. Big Tech CapEx expanded sharply, projected spending through 2028 remained enormous, NVIDIA retained its GPU leadership signal, Micron showed how valuable the memory bottleneck had become, Broadcom confirmed the custom silicon push, and AMD strengthened the case for compute diversification.

That is enough to take the infrastructure buildout seriously when evaluating multiyear AI platform contracts. It is not enough to waive through a weak use case, a vague ROI model, a vendor that cannot explain inference economics, or an implementation plan that assumes planners and operators will change behavior because the model output looks impressive.

A supply chain AI buyer can reasonably treat the 2025 stock performance and CapEx data as strong validation that the underlying technology stack has financial backing through 2028. The same buyer should still evaluate every application as if the infrastructure boom guarantees nothing about operational success.

References

1. Fidelity AI Outlook, Fidelity, Dec 2025.
2. AI infrastructure spending projection through 2028, Morgan Stanley.
3. AI Infrastructure Stocks Analyst Outlook, 24/7 Wall St., July 2026.
4. Best AI Stocks for July 2026, NerdWallet, July 2026.
5. Up 145% in 2025, This AI Infrastructure Stock Is Still Deeply Discounted, Yahoo Finance.
6. AI Networking and Data Center Infrastructure Coverage, The Motley Fool.
7. Generative AI pilot P&L impact research, MIT, July 2025.