MRTNZ Investment Office · Theme 360
Own the bottlenecks. Trade the applications selectively.
Compute, networking, power, and cooling remain the cleanest exposures. The next leg depends on software proving that AI usage can become durable margin.
The 360 view
Being right about AI adoption is not enough. The investable question is who captures the rent, who funds the build, and which valuations already assume a perfect handoff from infrastructure to applications.
The bottleneck map
AI value must pass through a physical system with long lead times. The narrowest point—not the loudest product—sets the economics of the current trade.
Supply improves, but power and commissioning remain slow-moving constraints.
The slowest physical dependency sets the commissioning date, even when every component has already been ordered.
Accelerators, HBM, networking, power, and cooling monetize the build before downstream returns are fully visible.
MU · NVDA · ANET · VRT · ETN · GEV · CEGStay while: lead times, power access, and thermal capacity remain constrained.Cloud and data platforms move forward when AI revenue begins outrunning depreciation and capital intensity.
ORCL · MSFT · AMZN · GOOGL · METARotate when: utilization, backlog conversion, and AI revenue growth improve together.Applications earn larger weights company by company when agents produce durable growth, retention, or margin.
PLTR · NOW · CRM · CRWDPromote when: monetization survives bundling pressure and shows measurable customer ROI.Capacity scores are illustrative MRTNZ scenario indices from 0–100. They demonstrate bottleneck mechanics; they are not measured industry utilization, supply forecasts, or security price targets.
Investor perspective · July 28, 2026
Baker argues that widening hyperscaler credit spreads overstate funding risk because contracted GPU capacity is priced below current spot economics and operating cash flow may accelerate as older contracts reprice.
Read the original perspective →This is an attributed investor thesis, not an MRTNZ fact. The attached chart cites Wells Fargo Securities, Bloomberg, and FactSet and shows wider hyperscaler CDS spreads alongside comparatively low debt-to-assets and high interest coverage.
MSFT, AMZN, GOOGL, ORCL, and META gain if utilization and contract repricing lift cloud cash generation faster than depreciation and capital intensity.
What would prove it: Cloud growth, AI backlog conversion, operating cash flow, and interest coverage improve together.AI-native software companies that secured 2024–2025 capacity may enjoy favorable unit economics, but renewals can move that rent back to the platforms.
What would prove it: Gross margin remains durable after compute contracts renew rather than only before repricing.The thesis still breaks if grids, substations, cooling, permits, and commissioning cannot turn financed equipment into usable capacity.
What would prove it: Energized megawatts and utilization rise on schedule without material project deferrals.What else stood out
Baker highlighted a joint Micron/Meta workload example in which spilling shuffle data from DRAM to SSD caused a severe slowdown. The investable read-through is that memory capacity and bandwidth can remain scarce after accelerator supply improves.
A compute-moat thesis he amplifiedBaker amplified the view that algorithms diffuse faster than the capital, power, networking, cooling, and operating capability required to compound large-scale compute.
Open ecosystems are strategically valuableBaker favors a plural model ecosystem and views Nvidia's broad software and open-model support as strategically important, not merely a hardware sales tactic.
What breaks the view: The view weakens if cloud operating cash flow decelerates, repricing fails to lift revenue, credit spreads correctly anticipate project losses, or physical commissioning delays strand funded capacity.
Capital stack
Four layers, four different return profiles. Counts reflect this curated 23-name universe.
Silicon, memory, foundry capacity, lithography, and networking remain the first rent collectors.
Cycle, policy, and customer concentration.Hyperscalers fund the build and own distribution. The question is whether revenue outruns depreciation.
Capital intensity and price competition.Power, cooling, grid equipment, and servers monetize physical scarcity one layer below the model race.
Long-cycle execution and theme premiums.The largest long-run pool, but today’s most selective trade: usage must become durable revenue and margin.
Bundling, weak ROI, and multiple compression.Opportunity board
The score combines 35% direct AI exposure, 25% business quality, 20% Street setup, and 20% price dislocation.
23 names · ranked by MRTNZ opportunity score
Top five now
Rankings are a research queue, not an instruction to buy. Each needs price discipline and an earnings check.
High-bandwidth memory. Latest annual revenue growth: +48.9%.
Operational AI software. Latest annual revenue growth: +56.2%.
AI cloud capacity. Latest annual revenue growth: +17.3%.
Accelerators and AI systems. Latest annual revenue growth: +65.5%.
Cloud, models, and copilots. Latest annual revenue growth: +14.9%.
Rotation playbook
The key transition is from scarcity economics to measurable enterprise returns.
Accelerators, memory, networking, power, and cooling capture the immediate buildout.
Cloud and model revenue must widen faster than depreciation and capital intensity.
Software earns a larger weight only when agents lift growth, retention, or margin.
Lower compute prices help adopters but can compress the first wave of infrastructure winners.
12–24 month framework
These are MRTNZ scenarios, not Street forecasts or price targets.
Infrastructure demand stays constrained while platforms and applications convert usage into expanding revenue and cash flow.
Compute and power keep earning; hyperscaler returns improve slowly; application winners remain company-specific.
Utilization disappoints, accelerator pricing falls, and the market reprices suppliers before end-demand catches up.
Method
Market facts come from the current Nasdaq snapshot and reported annual financials.
Street setup uses published target and rating aggregation, not MRTNZ valuation.
AI exposure and opportunity scores are MRTNZ judgments. They rank the research queue; they do not forecast returns.
Investor perspectives are attributed third-party theses. They are included to sharpen the debate, not treated as verified facts or MRTNZ forecasts.
Refresh cadence follows the Market 360 data pipeline at open, midday, and close.
Read the AI infrastructure capital-cycle field note →