Nvidia has led the Mag 7 for two years with a return of 853% since Jan of 2023. The second closest Mag 7 return in this time frame is Meta at 495% with all other Mag 7 stocks returning less than 200%, and popular tech ETF QQQ returning 105% in that time frame. All the above is outstanding performance, yet one stock has clearly set itself apart.
Given the spectacular outperformance of more than 7X the QQQs, it goes without saying there is outsized pressure on Nvidia’s stock, whether it’s from short sellers who are still in disbelief, or from investors who are simply wanting to book gains.
Recently, a low-cost large language model (LLM) from China sent the stock down 25% in two days. In the heart of the aftermath, I spoke with Charles Payne on Fox Business News, discussing key reasons investors should stay the course, given cheaper AI development will drive more demand (not less). The stock has nearly recouped its losses, trading 6% below its high before the Deep Seek news broke. Ultimately, Deep Seek is immaterial to AI demand, or perhaps even a net add given AI is too expensive to develop for most developers, resulting in fewer consumer and enterprise applications.
When it comes to Nvidia, there’s no denying the Street gets plenty wrong. Perhaps most egregious was when the Street dropped the stock 60% based on a gaming miss, as Ethereum’s merge to proof of stake in September of 2022 was expected to be the death knell for the AI juggernaut. A month later, the AI GPUs that drove the 7X outperformance shipped. The disconnect was staggering.
The I/O Fund has a strong track record on this stock, discussing every twist and turn publicly for our free stock newsletter readers with documented gains of up to 4,100% as far back as 2018 based on a very-early AI thesis. That gaming miss that resulted in a drop of 60%? We bought that too.
However, it’s important that I pause the exuberance and discuss something stirring beneath the surface. Recently, in this quarterly only, key suppliers are providing mixed guidance on the timing of Nvidia’s Blackwell GB200 systems. The commentary is subtle, and it would require knowing this stock thoroughly to identify the change in tone.
To put it simply, if Blackwell GB200s were ramping, we would see strong sequential growth for Q1. At the very least, there would be some indication Q2 is setting up for strong growth, and yet the commentary is shifting toward a second half discussion. When you add that suppliers do not want to get into the crosshairs with Nvidia, yet are under SEC regulations on how they offer guidance, the language used is incredibly easy to miss.
Ultimately, my spidey senses are up as commentary for Q1 and Q2 should be stronger on the products and solutions that supply Nvidia’s GB200 NVL systems. Meanwhile, Nvidia’s management team is under immense pressure as the company has beat revenue estimates by $1 billion or more for six quarters. This means even a minor delay or minor adjustment in expectations could come under close scrutiny.
As a reminder, we don’t make earnings calls, as many factors can affect stock price. For example, even though I’m getting mixed signals on the timing for GB200 NVL systems, Nvidia’s B200s are ramping, and theoretically can absorb some of this demand.
Instead of making an earnings call, we present quality research so that investors are fully informed to make their own decisions. From there, we take this a step further and publish every single trade we make on our research site. In finance, full transparency is rare, yet through never-ending tenacity, my firm has offered up to 4100% gains on Nvidia alone. We continue this long-standing dedication to our readers in the analysis below.
Nvidia’s Future Hinges on Blackwell – GB200s are the Standout SKUs
Due to the cyclical nature of GPU shipments, the differences in each generation are critical for investors to track. The Hopper generation has driven immense revenue growth over the past two years, while the Blackwell generation is expected to drive revenue that exceeds 2023 and 2024 combined. Hopper brought Nvidia to a $100 billion data center segment – at $26.3 billion in fiscal Q2 2024 – yet I pointed out how Blackwell could drive the data center segment to $200 billion-plus ten months ago.
You can read more about the nuances of each generation of Nvidia’s AI accelerators here, and why Nvidia’s future generations of GPUs can help Nvidia Stock reach $10 Trillion in Market Cap.
The rack-scale GB200 NVL72 features 36 GB200s, or 72 B200 GPUs and 36 Grace CPUs, offering up to 30X inference improvement and up to 5X training improvement versus the HGX H100, with significant improvements in energy efficiency and data processing speeds. The GB200 NVL72 was designed to address and unlock real-time inference for trillion-plus parameter models.
Despite being on an accelerated timeline, Blackwell will deliver the largest leap generationally to date for Nvidia’s AI GPUs.
- The B200 GPU will deliver a 2.5X training improvement and 5X inference improvement over the H100.
- Blackwell will see a massive upgrade in chip size, at 208 billion transistors compared to the H100’s 80 billion transistors.
- The B200 will also have 20 petaflops of FP4 compared to the H100’s 4 petaflops of FP8. I’ve covered the importance of this awhile back, and more recently following DeepSeek.
Blackwell’s pricing power is one of the key factors behind its growth potential, with the GB200 expected to be priced between $60,000 to $70,000, around double the H200’s estimated $32,000 price tag. For rack-scale solutions, the price tags are much heftier – the GB200 NVL36 (featuring 18 GB200s and 18 Grace CPUs) is estimated to carry a $1.8 million price tag, while the GB200 NVL72 is estimated to command an eye-watering $3 million price tag.
Roughly 35,000 NVL72 racks are estimated to be shipped this year, that’s already up to $105 billion in revenue for Nvidia. This volume correlates to shipments of ~2.52 million GB200s, versus Hopper shipments of 2 million-plus in 2024. To note, the revenue and volume do not include the NVL36 racks, B200s, B200As, B300s, and Hopper GPUs still on deck this year.
Nvidia B200s Will Be in High Demand and Likely Ship in Volume in H1
There is certainly a scenario where Nvidia’s GPUs are in such high demand that other SKUs can help make up for a delay on the much-larger GB200 NVL systems. The B200s are expected to be in high demand and companies such as Super Micro are shipping B200 HGX systems in volume. This is one reason Super Micro has seen favorable price action despite lowering their fiscal year guidance; it’s assumed the company can make up for any delays on the GB200s with B200 and B200 HGX systems.
However, the market does not like surprises. As of now, the market is expecting the GB200s to ramp in Q1 and further ramp in volume in Q2. It will be a tall order to meet these expectations (on the dot) with lower priced GPUs and HGX systems.
SECTION TAKEAWAY: Blackwell is a significant undertaking to have 36 GPUs and up to 72 GPUs communicate as one GPU (number depends on the SKU, with complexity increasing if it’s a single rack versus multi-rack configuration). Previously, Nvidia’s AI systems combined 8 GPUs. This significant leap not only increases pricing from roughly $32,000 per GPU to up to $3 million per system, but it also greatly increases the need for new networking architecture and comes with increased power requirements.
Nvidia GB200 Delay Rumors Grow – The I/O Fund’s Take for Q1 & Q2
Dating back six months, there have been rumors that the GB200s are delayed. The first rumor came from The Information, stating machines were sitting idle at Taiwan Semiconductor, where Nvidia’s chips are made. I informed my readers at the time this was unlikely to be the case as TSMC was reporting record high-performance computing revenue. In fact, TSMC continues to report strong HPC revenue.
In January, analyst Ming-Chi Kuo reported there were “Short-term Potential Risks in Nvidia and Related Supply Chain,” stating that GB200 NVL72 shipments in 2025 could reach roughly 25,000 to 35,000 racks compared to 50,000 to 80,000 racks from last year. The same analyst reported earlier in the year that “The biggest challenges in NVL72 development mainly stem from the 132kW thermal design point (TDP) requirement, which makes it the highest-power-consuming server in history. Nvidia and its supply chain need more time to solve unprecedented technology issues.”
This post piqued my interest as the emphasis would be outside the foundry. Instead, this would suggest an entirely new delay with thermal management issues (direct liquid cooling), or perhaps with power management solutions, or even networking related — rather than a material delay related to TSMC and yield issues.
To cross-examine this possibility, below is a brief summary from some of the more well-known direct-liquid cooling suppliers, PMIC suppliers (power management integrated circuits), and a few networking companies outlining commentary that leads to a higher probability there is a delay for the GB200s. If my read-through from these management teams is correct, the delay falls somewhere on the thermal/power solutions or networking components part of the supply chain rather than on TSMC.
Explaining the Lower Rack Shipments of NVL72 Systems
Regarding the statement that the GB200 NVL72 shipments in 2025 could reach roughly 25,000 to 35,000 racks compared to 50,000 to 80,000 racks from last year, I want to take a step back and clarify here that while that would represent a substantial drop, it’s more representative of a significant shift in NVL72 share from last year, and more reflective of what key partner TSMC can handle in terms of CoWoS capacity.
In July of 2024, estimates for GB200 racks were revised 50% higher from 40,000 to 60,000 racks, with the NVL36 at more than 80% share at 50,000 racks with the remaining 10,000 racks to be NVL72.
This would correlate to 2.52 million GPUs shipped – coincidentally, the same amount of GPUs expected to be shipped with 35,000 NVL72 racks, with more on deck from NVL36 configurations.
So, while it may look to be a large decline, GPU shipments at 35,000 NVL72 racks plus additional NVL36 racks would actually be higher than last year’s 60,000 NVL72 and NVL36 estimate due to a large shift towards the NVL72 configuration.
Additionally, the industry’s more recent, most optimistic forecasts for 50,000 to 80,000 NVL72 racks would be unrealistic and practically impossible for TSMC to meet – at the midpoint, this would require ~4.68 million B200 GPUs, or more than 292,000 in CoWoS wafer allocation simply for the NVL72 (not including NVL36).
Current estimates from Morgan Stanley place the B200’s CoWoS allocation at ~220,000 wafers for 2025, or more than 30% of TSMC’s total capacity for the year. This coincides with estimated B200 GPU shipments of 3.52 million, more than enough to handle 35,000 NVL72 racks and similar volumes of NVL36 racks.
SECTION TAKEAWAY:
With what the I/O Fund knows today, the product mix of NVL72 being higher, thus resulting in a lower number of racks being delivered, is not the issue. Rather, it’s the timing that could pose an issue due to Wall Street potentially having to face uncertainty on what quarter NVL systems would ship in volume.
Blackwell Necessitates Direct Liquid Cooling – Super Micro and Vertiv’s Comments
In June, we wrote an analysis on AI Power Consumption: Rapidly Becoming Mission Critical which stated that as the industry progresses towards a million-GPU scale, more emphasis is placed on future AI GPU generations to focus on power efficiency while delivering increasing levels of compute. Data centers are expected to adopt liquid cooling technologies to meet the cooling requirements to house these increasingly large GPU clusters, and the percentage of air-cooled systems shipped versus liquid cooled systems will change (dramatically) with Blackwell’s NVL systems.
Therefore, any changes to Blackwell’s NVL systems timing would appear in commentary from companies that provide direct-liquid cooling solutions and servers, as the previous generations are air cooled.
Super Micro Lowers FY25 Revenue Guide, Pushes Back 30% Liquid Cooling Target
Super Micro recently reported preliminary fiscal Q2 results, in which it cut its fiscal 2025 revenue guide while hinting that Blackwell GPUs could take longer to ramp in the first part of the calendar year.
At Computex in June 2024, CEO Charles Liang stated that he “expects 15 percent of racks [SMCI] sells this year to use DLC, and 30 percent to employ it [in 2025].” In August 2024, Liang slightly changed his tune, noting that they now expect “25% to 30% of the new global datacenter deployments to use DLC solutions in the next 12 months.”
I remember this comment well, as it was part of the evidence used to debunk the August rumors considering SMCI had moved up their timeline for DLC delivery (from end of year 2025 to mid-year 2025).
Two weeks ago in the February 2025 report, this timeline was shifted back with Liang stating that Super Micro now believes that “DLC or overall liquid cooling market share will grow all the way to 30% or even more in the next 12 months” — it’s not a dramatic shift in tone but rather pushes the 30% target back from mid-year 2025 noted in August, to beyond the original year-end timeline — to now technically being early 2026.
The August earnings call also housed one critical piece of information regarding Blackwell and Super Micro’s 2025 revenue. Wells Fargo’s Aaron Rakers questioned management about Blackwell and guidance for December 2024 and onwards. Liang explained that for the December quarter, Blackwell “will be very small. Engineering sample small volume. So the real volume, I believe, had to be March quarter next year. And that’s why we foresee only $26 billion to $30 billion.”
But now, SMCI implied they are not able to ship Blackwell in the March quarter:
“Do you guys have a forecast from NVIDIA? You know when you think you’re going to start to see you know supply the GPU, so that you can ship the NVL72 where visibility is still pretty low on availability of the GPUs?
Sign up for I/O Fund’s free newsletter with gains of up to 2600% because of Nvidia’s epic run – Click here
Charles Liang
We already proved pretty much everything. And now, just waiting for – and we are in some allocation, some volume, but the volume demand is way much bigger. So we are waiting for more allocation. So hopefully very soon we can ship in much higher volume.”
