The ledger remembers what the hype forgets.
Raymond James just dropped a bombshell that sent ripples through the semiconductor world: AMD upgraded to Strong Buy with a $641 price target. That's not a casual pat on the back—it's a bet that Advanced Micro Devices, the perennial second-place finisher, is about to become something else entirely.
Here's the part nobody's talking about: this isn't just a financial rating. It's a technical verdict on AMD's MI300 series, its Chiplet architecture, and its ability to survive the NVIDIA storm.
The Context: Why This Matters Now
AMD has been the "second" for so long it's become a running joke. Second in GPUs. Second in server CPUs. Second in AI accelerators. But second place is starting to look a lot more interesting.
The AI boom didn't just lift NVIDIA's boat. It created a structural shortage across the entire AI supply chain—from HBM memory to CoWoS packaging—and AMD positioned itself exactly at the intersection of that bottleneck.
The MI300X is a monster. 192GB of HBM3. 2.4x the memory of NVIDIA's H100. Built on TSMC's 4nm/5nm process with a Chiplet architecture that's turning out to be ahead of its time. When NVIDIA's Blackwell platform also pivoted to Chiplet design, it was a quiet vindication of AMD's bet.
But here's the uncomfortable truth: AMD's real bottleneck isn't its silicon. It's TSMC's CoWoS packaging capacity. The AI wars are actually a war for TSMC's production allocation, and whoever wins that allocation wins the AI market.
Core: The Hidden Signals in the $641 Target
Let's decode the technicals, because the street is missing the real story.
The Chiplet Advantage: It's not just about making chips smaller
AMD's Chiplet architecture is the industry's most aggressive bet. Since Zen 2, AMD has been putting multiple smaller dies into a single package. It's like building a distributed system within one physical chip—each die speaks to the others through high-speed interconnects.
This is different from NVIDIA's monolithic approach. The MI300X has 13 chiplets. Each one is smaller, easier to manufacture, and easier to optimize. The result: AMD can push HBM capacity to levels that monolithic chips can't match.
The market reads this as "just packaging." It's actually an architectural edge.
The 192GB HBM3 memory in the MI300X isn't just bigger. It's more valuable in inference workloads—where you're running large language models in production, not training them. Inference is memory-bandwidth-bound, and the MI300X has memory bandwidth in spades.
The Yield Question: Why the Strong Buy is a Technical Bet
Raymond James doesn't hand out Strong Buy ratings without talking to the supply chain. This rating implicitly signals something critical: AMD's yield rates are improving faster than expected.
The MI300X is a big die with 2.5D packaging and HBM integration. These chips are hard to make. Yield rates are typically 70-85% during ramp-up. But the rating suggests the street believes AMD is getting closer to 85%+.
That's not a small thing. Yield rate is the difference between 50% gross margin and 55%. Between meeting guidance and beating it. The bank's confidence is a signal that AMD's production ramp is on track.
The Real Bottleneck: CoWoS
The elephant in the room—and the one the rating implicitly bets on—is TSMC's CoWoS packaging capacity. AMD is 100% dependent on TSMC for advanced packaging. So is NVIDIA. The two are fighting for the same pool of CoWoS capacity.
TSMC is doubling its CoWoS capacity in 2024, but the demand is still outstripping supply. The $641 target price suggests the bank is betting that AMD's allocation will be enough to support its growth story.
If TSMC's expansion slips, AMD's growth story starts to crack. That's the hidden risk nobody's pricing in.
The Contrarian Angle: The Second-Place Advantage
While everyone's focused on NVIDIA's 80%+ market share, there's a structural shift happening that works in AMD's favor.
The cloud giants don't want a single supplier. Microsoft, Meta, Oracle, Google—they're all strategically building a "second supplier" relationship with AMD. It's not just about price. It's about bargaining power against NVIDIA, about negotiating better terms, about not being held hostage by CUDA.
This isn't a theory. Microsoft is already AMD's largest AI chip customer, accounting for an estimated 30-40% of MI300X revenue. And Microsoft is simultaneously NVIDIA's largest customer. That's not a contradiction. That's a hedge.
The AI market is so big that second place is actually first place in a different game.
The market is pricing AMD as a second-tier player. But in the race for AI inference—which is growing faster than training—AMD's architecture gives it a structural advantage. The MI300X's 192GB HBM capacity is the kind of spec that matters when you're running large language models in production.
The Elephant in the Room: Software
I've been watching this space since 2017, and there's one thing that's been AMD's Achilles' heel for years: software.
CUDA has a decade of development behind it. AMD's ROCm stack has been playing catch-up. But this year, something shifted. ROCm 6.0 came out, mainstream AI frameworks started adding better AMD support, and the gap is narrowing.
The bank's Strong Buy rating is a bet that this trend continues. That's not certain. But it's a bet that's being made with full knowledge of the risks.
Market Demand: The Real Driver
The AI chip market is expected to hit $50-60 billion in 2024. NVIDIA holds 80%+. AMD is targeting 10-15% of the pie. That's $5-9 billion in AI revenue—significant growth from the ~$1 billion AMD had before.
The MI300X's pricing helps: $15,000-20,000 per unit, which is 30-40% cheaper than NVIDIA's H100. In a market where every cloud provider is trying to reduce costs, that price advantage matters.
AI inference is where the real growth is. The inference market is growing faster than training, with projected CAGR of 80%+ through 2025. And inference is where AMD's architecture has the clearest advantage.
The Geopolitical Wildcard
The US export controls have cut AMD off from China, which represents 20-30% of global AI chip demand. That's a real loss.
But here's the counterintuitive part: the export controls actually help AMD in the long run. By taking both NVIDIA and AMD out of the Chinese market, the controls have created a forced "friendshoring" effect. The AI supply chain is consolidating around North American demand, and AMD is the second supplier for all those American and European cloud providers.
The loss of China is a near-term headwind. But the strategic positioning in the "free world" AI supply chain is a bigger tailwind.
The Financial Reality: What the Numbers Say
AMD's gross margin is about 50-52%, up from 45% in 2020. That's driven by the mix shift toward AI chips. The company's target margin for 2025 is 52-55%, which is achievable if the MI300 series continues to ramp.
The company has positive operating cash flow of about $4 billion, and free cash flow of $2.5 billion. For a fabless company, the capital expenditure is light—about 5-8% of revenue. That's a high-quality earnings stream.
The valuation story is simple: AMD is trading at about 40x trailing earnings, while NVIDIA is at 60x. The $641 target implies a PE of about 50x on 2025 earnings. That's a premium to the current multiple, but it's still a discount to NVIDIA.
The question is: does AMD deserve NVIDIA-like multiples? The market is betting that it does not, because it's pricing in the software ecosystem gap, the CoWoS risk, and the customer concentration.
The Risks We're Not Talking About
Risk 1: NVIDIA's Blackwell is a beast.
The Blackwell platform is already shipping. It's more powerful than the H100, and NVIDIA's CUDA ecosystem is still the default for AI developers. If Blackwell hits the market as hard as expected, AMD's growth story could stall.
Risk 2: CoWoS is the bottleneck.
If TSMC's CoWoS expansion slips, AMD gets the short end of the allocation. That's a risk no bank can hedge against.
Risk 3: Software, software, software.
ROCm is improving, but it's still not CUDA. If developers don't switch, AMD's hardware advantages won't translate into share.
The Takeaway: Watch the Signals
This is the key thing to watch over the next 12-18 months:
- AMD's Q3 earnings (October) — will show MI300 revenue and 2025 guidance.
- TSMC's CoWoS capacity updates — the real bottleneck for AMD's growth.
- Microsoft's AI infrastructure spending — the biggest AMD customer's behavior is a leading indicator.
- MI400 series progress — the next-gen chip will determine if AMD can close the gap to NVIDIA.
The ledger remembers what the hype forgets. AMD's story isn't just about beating NVIDIA. It's about being the alternative that cloud providers need—a second source for AI compute that has real architectural advantages in the fastest-growing segment of the market.
The $641 price target is a bet that AMD will pull it off. Not a guarantee. A bet.
Caught in the current of real-time value—the AI race is just beginning, and the market is still pricing AMD as a follower.
The question isn't whether AMD can beat NVIDIA. It's whether the "second place" is becoming the better place to be.
Final Thoughts
AMD's $641 price target isn't just a number. It's a statement about the future of the AI compute market. The AI infrastructure buildout is too big for NVIDIA to serve alone. And AMD is the only other player with the architecture, the supply chain, and the customer relationships to take advantage of that reality.
The ghost of Ethereum isn't in the ledger. But the ghost of AMD's past—the second place, the also-ran, the cheaper alternative—is now the story of a company that's finally in the right place at the right time.
Watch the signals. The verdict will be in by 2025.
