Hook: The Quiet Rotation That Speaks Volumes
Bridgewater Associates cut its NVIDIA position by 27% in Q4 2024. Simultaneously, the fund increased its stake in AMD. The market read this as routine portfolio rebalancing. It is not. This is a structural signal about the end of the AI chip monopoly. The data points are stark. NVIDIA's PE ratio sits near 55x. AMD trades at roughly 35x. The technological gap between them is narrowing. The liquidity narrative is shifting from training to inference. Bridgewater, the ultimate macro liquidity observer, does not make moves based on noise. They position for structural shifts. This is one of those shifts. The question is not whether NVIDIA is a good company. It is. The question is whether the current valuation already prices in the next three years of perfection. The answer, based on the numbers, is yes. And Bridgewater is betting that AMD's trajectory offers a better risk-reward asymmetry over the next 24 months.
Context: The Macro Liquidity Map and the AI Capex Cycle
Markets lie, but liquidity tells the truth. The AI chip market is currently the largest absorber of capital expenditure in the tech sector. Hyperscalers—Microsoft, Meta, Amazon, Alphabet—are committing billions to AI infrastructure. This creates a massive liquidity pool that flows directly to NVIDIA. But the second derivative of that flow is changing. The market is shifting from a scarcity-driven phase to an expansion phase. TSMC's CoWoS capacity is slated to double in 2025. That is the key variable. The bottleneck that gave NVIDIA its pricing power is being resolved. When supply expands, pricing power erodes. The current cycle position suggests we are in a late-stage expansion. NVIDIA's B200 is supply-constrained today, but the production ramp is on schedule. By Q3 2025, the market should see balance. By 2026, there will be oversupply risk. This is the classic semiconductor cycle. The question is whether the AI demand curve remains steep enough to absorb the supply. The current data suggests yes, but the rate of growth will decelerate. This is where the inference market becomes critical.
Core: The Technical Convergence and the Valuation Divergence
Let me break down the technical reality. NVIDIA's current advantage is not the silicon. It is the software ecosystem. CUDA has a 3-to-5-year lead over AMD's ROCm. That is a fact. But hardware is converging. NVIDIA's Blackwell B200 uses a dual-die design with CoWoS-L packaging, hitting 208 billion transistors. AMD's MI300X uses a chiplet approach with 13 small dies on TSMC's 4nm node. The process nodes are identical. The architectural philosophies differ, but the performance gap is closing. NVIDIA leads in interconnect with NVLink 5.0 delivering 1.8TB/s versus AMD's Infinity Fabric at 1.2TB/s. That is about one generation of advantage. In terms of energy efficiency, NVIDIA is roughly 2.5x better per unit of performance. But that gap narrows with the MI350 series expected in 2025 on TSMC's 3nm node.
Here is the critical insight that most retail investors miss. The market is now bifurcating. Training demand is becoming commoditized. Inference demand is exploding. The training market, where NVIDIA holds an 85% share, is growing at 40-60% annually. The inference market, where NVIDIA holds about 70%, is growing at 60-80%. AMD is stronger in inference relative to its overall share. The MI300 series offers better price-to-performance for inference workloads. As AI applications move from development to deployment, the inference share of total compute will exceed training by 2026. This is a structural tailwind for AMD. The data supports this. AMD's data center revenue mix is already shifting, with inference accounting for about 20% of AI revenue versus 10% for NVIDIA. This divergence will widen.
The Supply Chain and Geopolitical Arbitrage
Let's examine the geopolitical dimension. Both companies are fabless. Both depend on TSMC. But their exposure to export controls differs significantly. NVIDIA's China revenue has dropped from 25% of total to roughly 10-15% due to export restrictions on H100, H200, and B200. AMD's China exposure is smaller, around 15-20%, and the impact is muted because its chips are already positioned as second-tier alternatives. Bridgewater's rotation reduces geopolitical tail risk. This is not speculation. It is risk management. The US-China tech decoupling is not a single event. It is a process. Each round of restrictions pushes NVIDIA further out of the Chinese market while AMD's relative position remains stable. Additionally, the threat from domestic Chinese AI chips like Huawei's Ascend series is directed primarily at NVIDIA's high-end products. AMD's mid-tier positioning is less exposed. This is a subtle but critical distinction.
Contrarian: The Decoupling Thesis and the CUDA Delusion
Here is the contrarian angle that the market is ignoring. The consensus view is that NVIDIA's CUDA moat is unassailable. I disagree. The moat is real, but it is not permanent. The transition to inference changes the software requirement. Training requires the flexibility of CUDA's low-level programming. Inference is more about optimization and throughput. This is where ROCm's improvements matter less than raw price-to-performance. Moreover, the hyperscalers are actively developing their own custom silicon. Google's TPU, AWS's Trainium, and Microsoft's Maia are all designed to reduce dependency on NVIDIA. These in-house chips will not replace NVIDIA for frontier training, but they will take share in inference workloads. This pressure comes from above and below. NVIDIA is caught between custom silicon at the top and AMD's price aggression at the bottom.
Let's talk about valuation. This is where the empirical evidence is most compelling. NVIDIA's PE of 55x prices in flawless execution for the next three years. The market is paying for certainty. But the AI hardware cycle is inherently volatile. The history of semiconductors is a history of boom and bust. AMD's PE of 35x offers a margin of safety. Its AI revenue is growing at a comparable rate, but the valuation does not reflect the same level of optimism. Bridgewater is not predicting NVIDIA's decline. They are predicting a compression in the valuation gap. The ROIC data supports this. NVIDIA's ROIC is a stunning 60%. AMD's is 12%. But the direction of travel matters. NVIDIA's ROIC is peaking. AMD's is improving. The trajectory is more important than the absolute level.
The Capacity Expansion Signal
TSMC's CoWoS capacity is the single most important variable for 2025-2026. The current capacity is approximately 30-40k wafers per month. This is expected to double to 60-80k per month by the end of 2025. This expansion changes the supply-demand dynamics. NVIDIA will get more wafers, but so will AMD. The current allocation priority favors NVIDIA, but TSMC has an incentive to cultivate a second major customer. This reduces TSMC's own concentration risk. The expansion means AMD's MI300 shipments could double from 500k units in 2024 to over 1 million in 2025. This is a material change in market share dynamics. The market is still pricing AMD as a perpetual second-place player. The capacity data suggests the gap will narrow faster than expected.
The Inference Demand Curve
The training market is becoming saturated. The number of frontier models being trained is limited. The inference market, however, is expanding with every AI application deployment. Every chatbot, every code assistant, every image generator requires inference compute. The cost of inference is declining, which drives more usage. This is a classic Jevons paradox—as the cost of a resource decreases, demand increases. AMD is better positioned for this price-sensitive market. The MI300's 80-90% of NVIDIA's performance at a lower price point is attractive for inference workloads where cost per token is the key metric. This is where AMD will gain share. The data supports a shift from the current 70-15 NVIDIA-AMD split in inference to a 60-25 split by 2027.
Takeaway: Positioning for the Structural Shift
Alpha is found where others see only noise. The Bridgewater rotation is a signal that the AI chip market is entering a new phase. The era of NVIDIA's absolute dominance is ending. This does not mean NVIDIA will fail. It means the investment case has changed. The asymmetry has shifted. AMD offers a better risk-reward profile with lower valuation, improving technology, and a structural tailwind from the inference market. Survival is the first metric of success. In the semiconductor cycle, the companies that survive are those that adapt to the changing demand structure. The next 24 months will be defined by the inference transition. The question is not whether you believe in AI. The question is whether you are positioned for the next phase of the cycle. The data says the rotation from training to inference is underway. Bridgewater has positioned accordingly. We do not predict; we position. The question now is whether you will follow the liquidity or chase the narrative. Structure emerges from the chaos of contraction. The contraction in NVIDIA's valuation will create the structure for AMD's expansion. The only question is timing. The data suggests the timing is now.