Today I checked the prices of memory sticks on JD.com, and the price of a 64GB memory stick is almost catching up to the price of a MacBook M4 with 16GB of RAM! The price increase is outrageous! 🥸
Micron Technology (#MU), which I had been strongly recommending, has also hit a new high. When I recommended it, Micron's stock price was $160, and today it is $286, an increase of 78%. Today, there was another major event in the U.S. stock market that could continue to ignite interest in the SRAM market. Some promising U.S. stocks are worth keeping an eye on! 🧐
Today, the news of NVIDIA's $20 billion collaboration with #Groq shook the entire chip industry. Honestly, every time Jensen Huang is one step ahead, winning in a game that others haven't yet understood. This time, it may rewrite the competitive landscape of the TPU market and the SRAM market.
You may have heard: NVIDIA did not directly "buy" Groq but instead arranged a "non-exclusive license," acquiring Groq's chip technology and poaching their entire core team, including the CEO and CTO. You should know that Groq's founder, Jonathan Ross, was the leader behind Google's TPU chip. Three months ago, Groq had just completed a $750 million financing round, with a valuation of about $6.9 billion, and now they've taken away all their key personnel. This move is quite something. Although Groq continues to operate, the soul of the company has entered NVIDIA's doors. To put it simply, NVIDIA is saying, "I don't want your shell; I want your people and brains."
Why spend such a high price? Many people's first reaction is: "Aren't AI chips booming? Why do they need to acquire someone else's technology?" But if you've been paying attention to memory prices recently, you'll understand that Jensen Huang is playing a bigger game.
Because HBM is becoming a "luxury item," our team reviews financial reports and disassembles BOMs (Bill of Materials) daily, and one thing is clear: in an H100 GPU, just the 8 HBM3e memory chips cost over $4,000, accounting for nearly half of the entire card! And these HBM chips are solely supplied by Samsung, SK Hynix, and Micron, with production capacity tightly constrained. By the end of 2024 to 2025, HBM prices are expected to double, and even having money may not guarantee availability.
In this situation, NVIDIA will be in a very passive position, which undoubtedly poses a fatal threat to its profit margins:
→ NVIDIA's gross margin may be compressed;
→ Cloud vendors (like Microsoft and Meta) may slow down AI server purchases due to high costs;
→ Worse, if geopolitical tensions rise and HBM supply is cut off, the entire AI computing power expansion could "lose power."
If you were an NVIDIA shareholder, you might also be worried that although NVIDIA is "technologically invincible," it could ultimately be choked by a memory chip.
Thus, Groq's solution becomes particularly important; it does not use HBM and relies entirely on on-chip SRAM. Their chips do not connect to external HBM but integrate 220MB of SRAM directly on the chip, essentially building a "small warehouse" in the factory workshop, eliminating the need to run far to a large warehouse (HBM) to move goods. The result is: extremely low latency, fast inference speed, and low power consumption.
Although this design cannot handle large model training (due to limited SRAM capacity), it excels at inference. For example, running Llama2-70B, Groq can achieve over 500 tokens/second, faster than many GPUs, and without competing for HBM.
So, seeing this, it becomes clear why NVIDIA is investing $20 billion to acquire talent and technology. NVIDIA is not trying to replicate Groq but rather to turn this "non-HBM-dependent" technology route into its Plan B. In case HBM does encounter issues, it will have a backup plan and may even launch a new generation of low-HBM-dependent inference chips, specifically targeting high-value scenarios like edge AI, real-time services, and financial risk control.
More importantly, Groq has a "compiler dream team." No matter how strong the hardware is, it needs software to run effectively. The most impressive aspect of Groq is not the chip itself but its compiler. It can "perfectly translate" AI models into instructions that the chip can execute efficiently, achieving nearly 100% hardware utilization. In contrast, traditional GPUs often have a utilization rate of only about 70% due to their overly general architecture and complex scheduling.
NVIDIA has the CUDA moat, but on the inference side, tools like TensorRT and Triton are not yet "extreme" enough. By bringing Groq's compiler team on board, it is like directly parachuting in a special forces team to help elevate its software stack. In the future, using NVIDIA's inference chips may be faster, more power-efficient, and cost-effective, which would create a true ecological barrier.
The best part is that NVIDIA did not swallow Groq but allowed it to continue operating independently. This approach has two benefits: first, Groq can continue to sell services externally (like GroqCloud), validating its technology; second, it avoids antitrust scrutiny and prevents deterring other potential partners.
For NVIDIA shareholders, this is like taking a calming pill, as it prevents Groq from becoming a future competitor (it already has a good reputation among cloud vendors) while absorbing its core capabilities. Jensen Huang's move can be described as "borrowing the shell to practice skills, leaving no future troubles."
From this $20 billion investment collaboration, we can see Jensen Huang's sharp vision, which greatly enhances my confidence in NVIDIA's "ability to withstand supply chain risks." Our investment in #NVDA is never just a bet on GPU sales but a bet on its ability to continuously define the future of AI computing power. Now, it is reinforcing its position across the board in memory walls, software stacks, and talent reserves, indicating that the management team is thinking further ahead than the market.
In the short to medium term, HBM prices are likely to continue rising for a while (as shown in 👇 Figure 3), as even NVIDIA feels a sense of urgency that cannot be underestimated. Currently, the price increases for memory and hard drives are ongoing, benefiting three U.S. companies: Micron Technology (#MU), Seagate Technology (#STX), and Western Digital (#WDC), all of which are available on @MSX_CN (as shown in 👇 Figure 4).
However, in the long term, it is important to note ⚠️ that if NVIDIA can indeed cut the HBM usage of inference chips in half using SRAM and a new architecture, the demand structure for AI memory may change dramatically, which could be a negative for Micron, Hynix, and Samsung.
In summary, NVIDIA's $20 billion investment is not about "buying technology" but about buying future options. While the whole world is scrambling for HBM, it is quietly preparing a path to win without HBM. This is the moat of a top tech company; while you are still building walls, it has already built a spaceship. 🧐
Currently, these companies are all on #MSX. When trading U.S. stocks, I choose to use the #RWA tokenized platform #MSX to invest and participate in the U.S. stock market: http://msx.com/?code=Vu2v44
Early U.S. stock investment fans and partners can message me privately. After filling out the form, you can enter the U.S. stock exchange and discussion community for free (currently limited to 10 people per week, assistant review may take some time, thank you 🙏)!
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
