The 13F filings landed like a cold splash of water on a market drunk on green candles. Bridgewater Associates — the macro fund that built its name on reading the global tides before they turn — cut its NVIDIA stake by 27% while quietly building a larger position in AMD.
The immediate reaction was predictable. NVIDIA bulls called it noise. AMD bears shrugged it off. But the fund's reputation for early signal detection warrants a closer look. If Bridgewater's move signals what I suspect — an institutional read on the AI chip war's next phase — then the market hasn't priced it in yet.
Volatility isn't always a storm; sometimes it's a map. And this map points to a tectonic shift in the semiconductor landscape that goes far beyond quarterly earnings.
The Context: More Than a Portfolio Rebalance
Bridgewater's positioning is never random. The firm manages roughly $100 billion in assets, and its 13F filings are scoured by every institutional investor looking for a roadmap. Their 27% reduction of NVIDIA shares alongside a strategic increase in AMD suggests a deliberate macro-level bet.
The obvious narrative is valuation. NVIDIA's stock has gone parabolic since ChatGPT ignited the AI arms race. At a PE ratio near 55x, it trades at a significant premium to both its own historical averages and the broader market. AMD, by contrast, sits at around 40x, with a more compelling risk-reward profile.
But I don't think it's that simple. Based on my years of watching these cycles — from the 2017 ICO mania to the 2025 institutional convergence — this has all the hallmarks of a thesis shift.
Here's what I'm seeing beneath the surface.
The Core: NVIDIA's Moat Is Narrowing
Let's start with the technology itself. NVIDIA's Blackwell architecture is genuinely impressive. The B200 chip is a dual-die design with 208 billion transistors, connected via TSMC's CoWoS-L advanced packaging. It's the fastest AI training chip on the planet. The software ecosystem — CUDA — is the moat that's been both NVIDIA's greatest strength and its most effective barrier to entry.
CUDA has been the standard for over a decade. Developers write in it, frameworks are built on it, and the entire AI stack runs through it. That's a 3-5 year lead over AMD's ROCm ecosystem, which is still playing catch-up.
But the most important details are in the packaging and the process, not the specs.
AMD's MI300X is built on a Chiplet architecture. Instead of trying to manufacture one massive die, it breaks the design into 13 smaller chiplets. This isn't just a different engineering choice; it's a fundamentally different approach to manufacturing flexibility. It means AMD can allocate production across multiple process nodes more easily and has more negotiating leverage with TSMC.
The MI350 is coming in 2025 on TSMC's 3nm node — the same generation as NVIDIA's Rubin in 2026. The technology gap is closing. On the interconnect front, NVIDIA's NVLink 5.0 offers 1.8TB/s bandwidth, but AMD's Infinity Fabric has improved significantly, too.
The real story isn't raw performance. It's the shrinking gap in cost efficiency and supply chain flexibility.
The Hidden Variable: TSMC's Capacity Balancing Act
Here's where the analysis gets really interesting. Both companies are fabless. Both depend on TSMC for the most advanced nodes and CoWoS packaging. And right now, CoWoS is the most contested real estate in the tech world.
NVIDIA has been the whale in the room, getting priority access to TSMC's most advanced capacity. In 2024, they basically had a lock on the available CoWoS production.
But here's the trend I'm watching: TSMC is planning to double its CoWoS capacity in 2025 to roughly 60,000-80,000 wafers per month. When that capacity comes online, a lot of things shift. AMD will get access to significantly more production capacity, potentially doubling its MI300 shipments from around 500,000 units in 2024 to over 1 million in 2025.
This capacity expansion turns the market from a seller's market to a buyer's market. AMD will be able to scale its production to meet demand, and NVIDIA's ability to command premium prices — which is currently built on scarcity — will face its first real challenge.
That's the supply-side signal that might be the real reason behind Bridgewater's move. If you're a macro fund, you're looking at the moment when the supply constraint breaks and competition actually begins. From my analysis, that moment is 2025-2026.
The Contrarian Angle: The Inference Market's Silent Shift
The market narrative has been fixated on AI training chips. But the next chapter is the inference market.
Inference is the process of running a trained AI model to make predictions. Think of it this way: Training is the initial sprint of the marathon, but inference is the actual race that runs every day. As AI applications get deployed at scale — in search, healthcare, customer service, and code generation — the inference workload will explode.
This market is growing at an even faster rate than training, and it's much more price-sensitive. Enterprises deploying AI at scale don't need the absolute best performance. They need the best performance per dollar.
AMD's MI300 series has a significant advantage in this space. It offers about 80-90% of NVIDIA's performance for about 60-70% of the price. In a cost-sensitive inference market, that's the kind of value proposition that wins contracts.
NVIDIA will retain the high-end training crown. But AMD is positioning itself to be the cost-effective standard for inference. In the current market cycle, I believe the market is still underpricing AMD's ability to capture this specific segment.
The Geopolitical Wildcard
The US export controls are another piece of the puzzle that favors AMD.
NVIDIA's China revenue has dropped from around 25% of total revenue to roughly 10-15% due to the export controls. AMD's exposure is lower — about 15-20% of revenue — and the impact is less severe because its MI300 series performance gap with NVIDIA was already more significant.
This means AMD has less geopolitical risk in its portfolio. For a fund like Bridgewater, which prioritizes risk-adjusted returns, that's a meaningful factor.
China's domestic AI chip efforts — Huawei's Ascend series, for example — are also more directly targeting NVIDIA's high-end products than AMD's mid-range offerings. This means the long-term threat from Chinese competition is more of a concern for NVIDIA than for AMD.
The Valuation and the Numbers
The financial data paints a clear picture.
NVIDIA's gross margins sit at around 75%. AMD's are at approximately 50%. NVIDIA's ROE is around 70%, and AMD's is 15%. NVIDIA is a cash machine, generating around $40 billion in operating cash flow in 2024, versus AMD's $5 billion.
But here's the rub: NVIDIA's PE is 55x, and its PS is 30x. AMD's PE is 40x, and its PS is 10x. NVIDIA's price has run ahead of even its extremely strong fundamentals.
AMD's revenue is growing in line with NVIDIA's, and if its market share increases from 10% to 20-25% over the next two years — as the supply chain opens up and the inference market expands — the stock has much more room to run. That's the "value" proposition that Bridgewater is likely buying into.
The Takeaway: What to Watch Now
The Bridgewater shift isn't a panic move or a rejection of the AI thesis. It's a rebalancing within the AI thesis.
The question to ask now isn't "will AI chips be in demand?" — that's answered. The question is "which company is best positioned to benefit from the next phase?" The answer, at this point, is increasingly complicated.
I don't expect NVIDIA to lose its crown. The CUDA ecosystem is a genuinely sticky moat that takes years to erode. But I do expect to see more aggressive price competition and more strategic wins for AMD.
In the coming 12 months, the key data points to watch will be:
- AMD's MI350 launch — if it ships on time and with strong performance, that's a bullish signal
- TSMC's CoWoS capacity ramp — if the expansion goes smoothly, AMD gets the oxygen it needs
- The first earnings reports of 2026 — where we'll see if AMD's share gains are real, or just noise
Bridgewater may not always get the timing perfect, but their structural reads on market shifts have a way of being right.
The next cycle of the AI chip war will be about efficiency, not just performance. And in that race, the gap is closing.
