AI Infrastructure Stocks: Best Picks for GPU Cloud Investors

Contents

• AI Infrastructure Stocks: NVIDIA: GPU Monopoly
• AMD: The Challenger
• Broadcom: The Hidden Play
• TSMC: The Foundry
• CoreWeave: The Cloud GPU Disruptor
• Valuation & Growth
• Capex Trends
• ETFs & Index Exposure
• Investment Thesis
• FAQ
• Related Resources
• Sources

Where the money flows: GPUs, chips, memory, cloud services. March 2026: NVIDIA (NVDA), AMD, Broadcom (AVGO), TSMC (TSM), CoreWeave (CRWV). This is data. Not investment advice. Consult a qualified financial advisor before making investment decisions.

Real demand. Data centers overbuilding. Capex stays elevated 2-3 years. Capacity wins. No capacity loses.

AI Infrastructure Stocks: NVIDIA: GPU Monopoly

Market Position

NVIDIA controls ~95% of the AI GPU market (as of Q4 2025 / Q1 2026). H100 and newer GPUs (H200, B200) are the only chips that matter for large model training and inference.

Why the monopoly?

• 10 years of CUDA dominance. Every ML framework (PyTorch, TensorFlow) assumes CUDA. Switching costs are enormous.
• Software moat. CUDA libraries (cuDNN, CUTLASS, TensorRT) are unmatched. Training on AMD GPUs requires rewriting optimization code.
• Manufacturing excellence. TSMC (partner) gives NVIDIA first-priority wafer allocation. AMD can't compete on supply.

Revenue & Earnings (FY2025, ending January 2026)

NVIDIA's earnings are exponential. Data center revenue was $60.9B in FY2025, up from $18.4B in FY2024 ($26.1B in FY2023). The growth rate is decelerating (217% vs 280% vs infinity % in prior years), suggesting market maturation.

Key Products

H100 (Hopper, 2023): $15K-40K per GPU. 80GB HBM3. Cloud rental: $2.69/hour (RunPod SXM) to $3.78/hour (Lambda SXM).

H200 (Hopper-refresh, late 2025): $20K per GPU (estimate). 141GB HBM3e. Cloud rental: $3.59/hour.

B200 (Blackwell, early 2026): $30K-50K per GPU (estimate, not officially priced). 192GB HBM3e. Cloud rental: $5.98-$6.08/hour.

Valuation (March 2026 estimates)

• P/E Ratio: ~35-40x (elevated, but justified by growth)
• Forward P/E: ~25-30x (more reasonable if growth continues)
• Price-to-Sales: ~15-20x (very high; implies 5-7 year payback at current margins)

NVIDIA trades at premium valuations because:

1. Earnings growth is exceptional (280% YoY).
2. Market share is near-monopolistic.
3. Switching costs lock in customers.

Risk: Growth will eventually decelerate. When? Probably 2027-2028, when capex cycles normalize.

Capex Exposure

Every $1 spent on capex in data centers translates to roughly $0.10-0.15 of NVIDIA revenue (GPU is 10-15% of total data center capex). As capex rises, NVIDIA revenue rises proportionally.

AMD: The Challenger

Market Position

AMD is in second place with ~5% of the AI GPU market. Their MI300X (192GB HBM3) is the closest competitor to H100 but lags on software maturity and supply.

Revenue & Earnings (Q4 2025 estimates)

AMD's growth is real but slower than NVIDIA's. Data center was ~$1.8B annually (2025), up from ~$1.3B in 2024. That's 38% growth, impressive but 6x slower than NVIDIA.

Key Products

MI300X (CDNA 3, 2023): $12K per GPU (list price, often discounted). 192GB HBM3. Cloud rental: $1.99/hour (DigitalOcean) to $3.45/hour (Crusoe), March 2026.

MI325X (CDNA 3 refresh, late 2025/early 2026): TBD specs. Marketing claims 1.3x MI300X performance. Price and cloud availability TBD.

Challenges

1. CUDA moat is real. Recompiling PyTorch code for AMD's ROCm is not trivial. Libraries are incomplete.
2. Supply is constrained. AMD relies on TSMC's 5nm capacity, competing with NVIDIA. NVIDIA gets priority.
3. Customers are risk-averse. Big cloud providers (AWS, Azure) prefer NVIDIA because it's the safe choice (no one gets fired for choosing NVIDIA).

Upside

If AMD solves the software moat (ROCm maturity), they could steal 10-15% market share. Price competition would follow. But this is a 2-3 year play, not immediate.

Valuation: P/E ~20-25x. More reasonable than NVIDIA, but reflects slower growth.

Broadcom: The Hidden Play

Market Position

Broadcom manufactures:

• Networking chips (Ethernet switches, optical transceivers) for data center interconnect
• Memory controllers for GPUs and CPUs
• Custom silicon for cloud providers (AWS, Meta, Google)

Broadcom is the infrastructure play. Every GPU cluster needs Broadcom networking. Every data center upgrade touches Broadcom.

Revenue & Earnings (FY2025, fiscal year ends September)

Broadcom's segment breakdown is complex. Their "Networking & Cloud Services" division saw 30%+ growth in 2025 due to AI data center demand.

Key Products

Custom silicon for hyperscalers: AWS Trainium (training chip), Inferentia (inference chip). Broadcom provides the interconnect and memory controllers.

Ethernet switches (25G-400G): Required for GPU cluster networking. Every multi-GPU system needs these.

PCIe adapters and memory modules: Interface between GPUs and CPUs. High-margin business.

Valuation

• P/E Ratio: ~18-22x
• Forward P/E: ~15-18x

Broadcom is cheaper than NVIDIA and trades more reasonably than semiconductors. The business is defensive: hyperscalers must buy networking and memory regardless of market conditions.

Why It's a Hidden Play

Broadcom's AI exposure is indirect. Investors focusing on "GPU stocks" miss Broadcom. But Broadcom benefits from the same capex wave. Without Broadcom's networking, GPU clusters can't communicate. Revenue is reliable and growing 25-30% annually.

TSMC: The Foundry

Market Position

TSMC manufactures NVIDIA's and AMD's GPUs using 5nm (and newer 3nm) process. TSMC is capacity-constrained. They have limited 5nm wafer capacity and must prioritize customers.

NVIDIA gets first pick. AMD gets second. Everyone else waits.

Revenue & Earnings (FY2024, fiscal year ends June 2025)

TSMC is benefiting enormously from AI. GPU orders pushed 2024 into unexpected growth. 2025 growth will moderate (demand is still strong but supply isn't the constraint anymore).

Capex Exposure

TSMC is spending $30-40B annually on new fab construction (2025-2027). Massive capex to expand 3nm and 5nm capacity. This capex is driven by AI demand.

Higher capex = lower free cash flow in the near term, but future capacity = future revenue growth.

Valuation

• P/E Ratio: ~22-26x
• Forward P/E: ~18-20x

TSMC is cheaper than NVIDIA but carries geopolitical risk (Taiwan, US-China tensions). Also, if AI capex cycles soften, TSMC's growth moderates faster than NVIDIA's.

CoreWeave: The Cloud GPU Disruptor

Market Position (Post-IPO)

CoreWeave went public in March 2025 (IPO price $40.00; ticker: CRWV). CoreWeave is the largest dedicated cloud GPU provider (vs AWS/Azure, which also offer CPUs).

Business model: Lease GPUs on 24-hour notice. Cheaper than buying, faster than hyperscaler waiting lists.

Revenue & Earnings (FY2024 reported / FY2025E)

CoreWeave is growing rapidly but not yet net-profitable due to heavy capex. Margins are strong on a gross basis because they own their GPU infrastructure rather than reselling wholesale capacity.

Key Products

GPU rentals by model:

• H100: $1.99-$3.78/hour (1x)
• B200: $5.98-$6.08/hour (1x), $68.80/hour (8x cluster)
• H200: $3.59/hour (1x)
• Custom clusters: NVLink interconnect, high-speed networking

Investment Thesis

CoreWeave solves a critical problem: ML engineers can't wait 6 months for GPU delivery. CoreWeave offers 24-hour provisioning. They're taking market share from private cloud (teams buying own servers) and waiting for hyperscalers.

Upside: If CoreWeave reaches $5B+ revenue (from ~$1.9B in 2024), significant multiple expansion is possible as margins improve. That requires continued GPU cloud market growth through 2027-2028.

Downside: If hyperscalers (AWS, Azure) offer better GPU deals, CoreWeave margins compress. Price wars are likely.

Valuation

• Price/Sales: ~5-8x (reasonable for a high-growth cloud company at ~$1.9B revenue)
• Operating losses are significant due to capex-heavy GPU infrastructure buildout

CoreWeave is a higher-risk play than NVIDIA or AMD given pre-profitability status, but carries significant upside if GPU cloud demand remains strong. High-risk, high-reward.

Valuation & Growth

Comparative Valuation (March 2026)

Observations:

• NVIDIA is the most expensive, justified by dominance and growth.
• CoreWeave is the cheapest (but also riskiest, pre-profitable).
• Broadcom is undervalued relative to growth (hidden play).
• TSMC is middle-of-the-road; geopolitical risk keeps multiple low.

Capex Trends

Data Center Capex by Year (Industry Estimate)

GPU capex is the fastest-growing segment. As % of total data center capex, GPU went from 3% (2023) to 6-7% (2026) to estimated 8-10% by 2028.

Revenue Implications

For every $10B in GPU capex:

• NVIDIA captures ~$8-9B (80-90% share)
• AMD captures ~$0.5-1B (5-10% share)
• Others capture ~$0.5B

This is why NVIDIA's growth is exponential (capex is accelerating) and why that growth will decelerate (capex eventually normalizes).

Deceleration timeline: Likely 2027-2028. When? When cloud providers stop expanding data centers to squeeze utilization out of existing capacity.

ETFs & Index Exposure

Investors can gain AI infrastructure exposure via index funds and ETFs without picking individual stocks.

Broad tech ETFs:

• QQQ (Invesco QQQ Trust): Holds NVIDIA, AMD, TSMC, Broadcom. Exposure to AI infrastructure as 40-50% of holdings (estimated).
• XLK (Technology Select Sector SPDR): S&P 500 tech. Smaller AI infrastructure exposure (~30%).
• SPY (S&P 500): Diversified. AI infrastructure is 10-15% of holdings.

AI/Semiconductor-specific ETFs:

• SOXL (Direxion Daily Semiconductor Bull 3X Shares): Leveraged 3x exposure to semiconductor index. Risky, volatile, only for short-term traders.
• XSD (SPDR S&P Semiconductor ETF): Equal-weight semiconductor stocks. Balances NVIDIA dominance (SOXL is NVIDIA-heavy).
• ROBO (ROBO Global Robotics and Automation ETF): Robotics and automation focused, includes NVIDIA, AMD, and other AI hardware and software companies. Trades on NYSE Arca.

Dividend considerations:

NVIDIA, Broadcom, and TSMC all pay dividends (2-3% yield). AMD and CoreWeave do not (growth-focused). For dividend-seeking investors, TSMC and Broadcom are better choices.

Investment Thesis

Bull Case

1. Capex is structural, not cyclical. AI model training requires ever-larger clusters. Every 18 months, models double in size. Each doubling requires new infrastructure.

2. NVIDIA's moat is unbreakable (5+ years). CUDA lock-in is real. Even if AMD GPUs were cheaper, rewriting optimization code costs more than the GPU savings.

3. Broadcom is a hidden beneficiary. Network capex grows faster than GPU capex. Undervalued.

4. CoreWeave is a true disruptor. They're capturing demand that hyperscalers can't or won't supply.

5. Capex will remain elevated through 2027-2028. Data center operators are still under-supplied. Expect 2-3 more years of 25-30% capex growth.

Bear Case

1. Valuations are stretched. NVIDIA at 35-40x P/E assumes perfect execution for 5+ years. One bad quarter = 20% correction.

2. Capex cycles are finite. Capex growth will decelerate in 2027-2028 when utilization climbs. At that point, revenue growth falls from 200% to 20%.

3. Competition is coming. AMD's MI325X, Intel's Gaudi 3, Google's TPU 5 are improving. NVIDIA's share could fall from 95% to 80% by 2028.

4. Geopolitical risk. US-China tensions could impact TSMC, AMD supply chains. Export controls on advanced chips are tightening.

5. CoreWeave margins are unsustainable. Competition from hyperscalers will force price cuts. Path to profitability is unclear.

FAQ

Should I buy NVIDIA stock?

This is not investment advice. NVIDIA is the highest-quality company in the space but trades at premium valuations. Appropriate for long-term portfolios if you believe in AI capex extending through 2028-2030. If you're uncomfortable with 30x+ P/E, wait for a correction or consider diversifying into AMD or Broadcom.

Is AMD a good value play?

AMD is cheaper than NVIDIA, but the software moat is real. Unless AMD significantly improves ROCm, they'll cap out at 5-10% market share. Value investors might like the price, but the competitive position is weak.

Should I invest in chip makers or cloud GPU providers?

Chip makers (NVIDIA, TSMC) have higher margins and longer growth visibility. Cloud GPU providers (CoreWeave) are riskier but could deliver higher returns if they dominate the market. Diversify.

What's the biggest risk?

Capex deceleration. If cloud providers stop building in 2027 and instead optimize existing capacity, GPU demand crashes. Watch cloud provider earnings reports for capex guidance.

Are AI infrastructure stocks overvalued?

NVIDIA is probably 15-25% overvalued. AMD is fairly valued. Broadcom is undervalued. CoreWeave is speculative (could be overvalued or undervalued depending on execution). Overall, the sector is fair-to-rich, not in bubble territory (yet).

Which stock will outperform over the next 2 years?

CoreWeave (if they execute and profitability emerges). Broadcom (if networking capex accelerates). NVIDIA (if capex remains strong). AMD (if they steal share from NVIDIA).

What's a good entry point?

NVIDIA: $100-120 per share (assumes 15% correction). AMD: $90-110. Broadcom: $180-200. TSMC: $120-140. CoreWeave: context-dependent on post-IPO trading range; monitor earnings reports for profitability progress. (These are not target prices, just illustrative reference ranges based on historical valuations. Not financial advice.)

Should I dollar-cost average or lump-sum invest?

Given the momentum, lump-sum investing in NVIDIA/Broadcom is riskier than DCA over 6 months. For speculative plays (CoreWeave), DCA is safer.

Related Resources

• NVIDIA GPU Pricing Comparison
• Top AI Stocks & Core Infrastructure Tools
• AI Infrastructure Companies Overview
• MLOps Tools & Platform Comparison

Sources

• NVIDIA FY2025 Earnings Report
• AMD Q4 2025 Earnings (estimate)
• Broadcom FY2025 Earnings Report
• TSMC FY2024 Earnings Report
• CoreWeave IPO Prospectus & Investor Relations
• IDC Global DataSphere Capex Analysis
• DeployBase Market Data & GPU Pricing (tracked March 22, 2026)