AI Infrastructure Economics Model

Key Dynamics

Anthropic: Fastest Growing, Cleanest Path

$1B to $14B ARR in 14 months. Claude Code alone at $2.5B+ run rate. Multi-cloud strategy (AWS Trainium + Google TPUs + Azure) avoids single-vendor lock-in. Profitable by 2028 — years ahead of OpenAI.

OpenAI: Scale at a Cost

$20B ARR but burning $9B+ annually. Inference spend with Microsoft: $12.4B through Q3 2025. Projected $115B cumulative losses through 2029. Not profitable until ~2030. The revenue is real; the margins aren't.

Google: Structural Advantage

Proprietary TPUs cut serving costs 78% in one year. 10B+ tokens/minute through Gemini API. $175-185B 2026 CapEx. Least GPU-constrained of any player because they own the silicon.

Microsoft: GPUs in Boxes They Can't Plug In

CEO confirmed GPUs literally sitting in inventory due to power/space shortages. "Short for many quarters." Azure AI adding ~16-20 percentage points to cloud growth but physically bottlenecked.

Nvidia: The Arms Dealer Wins

$115.2B data center revenue (FY2025). Blackwell selling faster than any product in company history. Controls 70%+ of TSMC CoWoS-L packaging capacity. Entire 2025 Blackwell production sold out through mid-2026.

Supply Chain Bottlenecks

CoWoS Packaging: The Real Chokepoint

TSMC's Chip-on-Wafer-on-Substrate (CoWoS) advanced packaging is required for all HBM-equipped AI chips. Nvidia holds 70%+ of CoWoS-L capacity. Even doubling capacity in 2025 couldn't keep up with 113% YoY demand surge. This is the single biggest constraint on GPU supply.

HBM Memory: Sold Out Through 2026

All three suppliers (SK Hynix 62%, Samsung 35%, Micron 11%) are sold out. HBM3e transition absorbing 96% of output. New fabs (SK Hynix M15X, Samsung P5) don't come online until 2027-2028. HBM4 mass production begins Feb 2026 but at limited volume.

The Training-to-Inference Shift

Inference is moving from ~50% to ~70% of all AI compute by 2027. Agentic workloads are the primary driver — each user action triggers 10-50+ internal model calls. This changes the hardware mix: more memory bandwidth, more CPUs alongside GPUs, and persistent always-on deployment vs. bursty training campaigns.

The Power Problem

This is the section that matters most. Forget GPUs for a second — the binding constraint on AI scaling is electricity, not silicon.

The Numbers

• US total generation capacity: ~1,300 GW (1.3 TW)
• Current datacenter draw: ~25-42 GW (3-4% of US electricity)
• Projected 2030: ~134 GW (7-12% of US electricity)
• Projected shortfall by 2028: ~10 GW (enough to power 7.5 million homes)

Why It's Hard

• Grid interconnection queue: 10,300 projects totaling 1,400 GW waiting. Median wait: 5 years.
• Transformer lead times: 12-48 months for power transformers
• Switchgear: Up to 3 years
• Northern Virginia (largest DC market): 4-7 year grid connection wait

How They're Working Around It

• Onsite gas turbines: xAI, Meta, and others bypass the grid entirely with dedicated natural gas plants. US proposals for new gas-fired generation tripled in 2025.
• Nuclear PPAs: Microsoft (Three Mile Island, 835 MW by 2028), Amazon (Susquehanna, $20B+), Google (Kairos SMR, 500 MW), Meta (TerraPower+Oklo+Vistra, up to 6.6 GW)
• Modular datacenters: 3-6 month builds vs 2-3 years traditional
• Energy company acquisitions: Google bought Intersect Power for $4.75B

GPU Rack Power Trajectory

Per-rack power is heading toward 1 MW. Every new GPU generation makes the power problem worse, not better. Efficiency gains in compute-per-watt get eaten by density increases.

Executive Summary

No, the current buildout is not enough. But it's not as simple as "build more." The constraints are layered: silicon supply, power infrastructure, and construction timelines each impose different bottlenecks at different time horizons.

The Token Economy: How Much Compute Do AI Agents Actually Need?

The Agentic Multiplier

This is the piece most analyses miss. Traditional AI chat (ask a question, get an answer) consumes relatively modest tokens — maybe 2,000-5,000 tokens per interaction. But agentic AI tools like Claude Code, Cursor, and Devin work fundamentally differently:

• A single user prompt triggers 10-50+ internal model calls (tool use, reasoning, code generation, validation)
• Average coding prompts are 20,000+ input tokens (3-4x general prompts)
• A heavy Claude Code user can consume 2.4 billion tokens in a single month
• Agent teams use approximately 7x more tokens than single-agent sessions
• The industry average prompt length has tripled in under 2 years (2K to 5.4K tokens)

The OpenRouter 100-trillion-token study found that programming workloads grew from 11% to over 50% of all tokens by late 2025. Claude alone handles ~60% of coding workloads on that platform, with average prompts over 20,000 tokens.

Who's Using This and How Much?

5.6 quadrillion tokens/month = ~2.2 billion tokens/second average. At a peak-to-average ratio of 3.5x, that's ~7.5 billion TPS at peak.

The Supply Side: Can We Serve 7.5 Billion TPS?

Current Global Inference Fleet (2026 estimate)

Average demand: 2.2B TPS. Effective supply: 1.2B TPS. That's a ~45% shortfall.

At peak (7.5B TPS), the gap is enormous. But peaks are managed through queuing, rate limiting, degraded service (smaller/faster models), and geographic load balancing. The real question is whether average throughput can be sustainably served.

Revenue Impact: Who Wins, Who Loses?

The $650B+ CapEx Wave (2026 alone)

The Big Four hyperscalers (Amazon, Google, Microsoft, Meta) are spending $610-665B in 2026 on infrastructure, with ~75% directly AI-tied. This money flows to:

The AI Model Companies

The Investment Case: Key Numbers for Each Sector

GPUs/Silicon ($200-300B annual market by 2027)

• Nvidia: Near-monopoly on AI training/inference GPUs. $115B DC revenue in FY2025, tracking to $200B+ in FY2026. Rubin in H2 2026 extends the lead. Risk: customer ASICs (Trainium, TPU) slowly erode share long-term.
• TSMC: Sole manufacturer of all cutting-edge AI chips. CoWoS packaging is THE bottleneck. Revenue grows in lockstep with AI demand.
• AMD: Distant #2. MI300X relevant for memory-heavy workloads. ROCm software gap limits adoption. ~6% of Nvidia's DC revenue.
• SK Hynix: HBM leader (62% share). Sold out through 2026. $8B+ annualized HBM revenue growing rapidly.

Cloud/Infrastructure ($400B+ annual market)

• AWS: $128.7B revenue (2025), $200B CapEx in 2026. Trainium custom silicon is a multibillion-dollar business growing 150% QoQ.
• Google Cloud: $43B revenue (2025), growing 48% YoY. TPU advantage enables 78% cost reduction. $175-185B CapEx in 2026.
• Azure: $75B+ revenue (2025), growing 34%. Power-constrained but spending $120-145B in 2026.

Power/Energy ($50-100B+ annual AI-related investment)

• Nuclear: Multiple GW of deals signed. TMI restart (2028), SMR fleets (2030+). Long timeline but massive TAM.
• Natural gas: 250+ GW of new gas plants proposed in US, 1/3 datacenter-linked. Fastest path to new power.
• Utilities: Grid operators and IPPs see unprecedented demand growth after decades of flat load.

How Much More Is Needed?

The model (adjustable above) suggests:

• GPUs: The current fleet (~4.4M inference chips) needs to roughly double to ~8-10M by end of 2027 to keep pace. That's ~$300-500B in GPU purchases.
• Power: ~10 GW additional capacity needed by 2028. At $10-15B per GW for datacenter-grade power, that's $100-150B in power infrastructure.
• Datacenters: ~50-80 new large-scale AI datacenters (100+ MW each) needed globally. The announced pipeline covers most of this, but construction timelines mean 2026-2027 will be tight.
• Total additional investment needed (2026-2028): $500B-1T beyond what's already committed. The committed CapEx (~$650B/year) covers maybe 60-70% of what's needed.

Risks to This Model

• Efficiency breakthroughs: DeepSeek showed that algorithmic improvements can dramatically reduce compute needs. If models get 5-10x more efficient, the supply gap narrows fast.
• Demand disappointment: If enterprise AI agent adoption is slower than projected (trust issues, integration complexity, regulatory friction), the demand side could be lower.
• Price competition: Token prices fell 10-50x since 2023. If this continues, revenue per token drops even as volume grows.
• Geopolitical disruption: Taiwan risk (TSMC), export controls, tariffs could all disrupt supply chains.
• The "good enough" plateau: If current model capabilities are "good enough" for most enterprise use cases, the push for ever-larger models (and their compute hunger) may slow.

Detailed Stock Analysis

NVDA — Strong Buy — PT $280-300

The arms dealer always wins in a gold rush, and this is the biggest gold rush in tech history. Nvidia just reported $68.1B for Q4 (crushing estimates) and guided Q1 to $78B (vs $72B consensus). Our demand model says analysts are still conservative. The structural deficit we identified — demand outpacing supply by 20-40% — means every GPU Nvidia can ship gets absorbed immediately. They control 70%+ of TSMC's CoWoS-L capacity, Blackwell is sold out through mid-2026, and Rubin (5x inference perf) arrives H2 2026. At ~25x forward P/E for a company growing 48-65%, this is reasonably priced. The only risk is a demand shock (efficiency breakthroughs a la DeepSeek) — but even DeepSeek's efficiency gains got eaten by volume increases.

MSFT — Strong Buy — PT $550-600

The analyst-to-price gap here is extraordinary: $403 current vs $596 mean target (48% implied upside). Azure is growing 39% with AI contributing 16-20 percentage points. Yes, they have GPUs sitting in boxes they can't plug in — but that's a temporary constraint, not a demand problem. As power comes online through 2026-2027, Microsoft has the inventory ready to deploy immediately. 15M paid Copilot seats (up 160% YoY) across 90% of Fortune 500 is a recurring revenue engine. At ~23x forward P/E — the cheapest of the mega-caps relative to growth — the market is overweighting the near-term power constraint and underweighting the demand runway.

DELL — Strong Buy — PT $160-180

The most undervalued stock in the AI infrastructure chain. Forward P/E of ~10x for a company that just guided FY2027 revenue of $138-142B (12% above Street consensus of $125.5B). $64B in AI server orders, $43B backlog, ISG revenue up 73% YoY. Dell is the #1 AI server ODM and every datacenter buildout needs their hardware. The stock got unfairly punished by the Morgan Stanley underweight call at $101, but the Q4 beat and massive forward guide should force re-ratings. At 10x forward earnings with 19-23% revenue growth, this is mispriced.

MU — Buy — PT $480-520

Here's a stock where our model disagrees with analyst targets. MU is trading at $416 above the analyst mean of $340-358, yet the forward P/E is just ~10x. Why? Because analysts are slow to update HBM projections. HBM is sold out through 2026 (all three suppliers confirmed), the TAM is growing from $35B to $100B by 2028, and Micron just guided Q2 revenue of $18.7B with $8.42 EPS. The stock price is ahead of consensus but the fundamentals justify even higher. Every AI GPU needs HBM; Micron is one of only three companies on Earth that can make it.

VRT — Buy — PT $300-320

Our research identified power/cooling as the binding constraint on AI scaling. VRT is the purest play on that thesis. Q4 organic orders up 252%, backlog doubled to $15B, and they're guiding 28% organic revenue growth with 43% EPS growth for 2026. Every datacenter being built needs Vertiv's power management and thermal systems. At ATH but the order growth says it's not done.

TSM — Buy — PT $450-480

The monopoly. Every cutting-edge AI chip — Nvidia, AMD, Apple, Amazon, Google — goes through TSMC. CoWoS packaging is THE bottleneck in our supply model. Guiding ~30% revenue growth for 2026. At ~22x forward P/E, the Taiwan geopolitical discount is baked in, and the Arizona fabs de-risk it over time. You can't build AI infrastructure without TSMC.

Nuclear Power Trio: CEG, VST, TLN

Our model shows a ~10 GW power shortfall by 2028. Nuclear is the only zero-carbon baseload that can fill it. CEG just became the nation's largest electricity producer (Calpine acquisition). VST has the Meta deal. TLN has the Amazon Susquehanna deal. All three are trading below analyst targets with forward guidance catalysts ahead. VST at ~14x forward P/E on the Q4 earnings miss pullback looks particularly attractive as a buy-the-dip opportunity.

ORCL — Speculative Buy — PT $200-250

The widest bull-bear spread on the board. $523B RPO backlog from Stargate is either transformative or vaporware. Stock is -57% from its high of $345. If Stargate materializes at even 30% of announced scale, ORCL is drastically undervalued. If disputes stall the project, there's more downside. Our model says the demand for what Stargate would provide is real — the question is whether Oracle can execute. Position size accordingly.

AMD — Hold — No Price Target

Not a sell, but not where we'd put new money. AMD is a fine company doing ~$35B revenue, but our model shows Nvidia's structural advantages (CoWoS capacity, CUDA ecosystem, Blackwell performance) are widening, not narrowing. AMD's DC revenue is ~6% of Nvidia's. The Meta-AMD deal announced Feb 24 is positive but won't close the gap. At ~31x forward P/E, you're paying a premium for a #2 player. In this space, we'd rather own NVDA, TSM, or even MU.

Model Portfolio: AI Infrastructure

If building a concentrated AI infrastructure portfolio, here's how we'd weight it:

Key Catalysts Timeline (2026)