Thomas Schulz

I think so, yes. Taiwan had a multi-month buffer. SK Hynix and Samsung had like 6 months worth of Helium and procured more from diversified sources to further reduce the risk and they are also working harder on recycling Helium. Micron should be totally isolated from this risk as US-based with the US being large Helium-producers. They also said on the call they see pretty much zero supply risk from the Iran war, which would be a pretty risky statement if they needed to be worried. Helium also has lower-priority uses like party balloons, which obviously get cut way before critical semiconductors. So unless the war takes forever, this should be no problem. Given how high memory margins are, who cares if they need to pay double or triple for a small input cost? If margins go from 81% to 80%, that doesn't really move the needle.

Is it fair to say that the helium risk has mostly been overstated? This was one of the reasons memory stocks crashed a few weeks back.

Link: developer.nvidia.com/blog/optimizin…

Memory stocks are down on Google's TurboQuant reducing memory demand in KVcache by "6-8x" by compressing down to 3-bit. But the paper was published a year ago. In the blog post linked below, $NVDA wrote Blackwell will soon reduce KVcache further down from 8-bit to 4-bit. So the incremental step is closer to 4 -> 3 bit in new setups. So a 25% reduction rather than 6-8x. Also, KV cache is only part of total memory (weights etc. unchanged), so total memory reduction is much smaller. And lower memory per token typically increases usage through higher ROI for hyperscaler (Jevons Paradox). $MU

Source: en.yna.co.kr/view/AEN202603…

South Korea’s semiconductor exports are up +163.9% YoY for March 1-20, after being up 134% in February and "more than double" in January. Memory profitability keeps accelerating. Good for SK Hynix. #SKHynix $MU

Jensen says H200 will be sold into China in a few weeks. Chinese companies finally gave purchase orders, signaling they got Chinese authorization. U.S. government issued the export licenses. In his eyes it is a done deal. Once the U.S. companies have Vera Rubin out, he'll ask Trump to sell Blackwell chips to China. He refuted the Reuters report though that NVIDIA will sell the new Groq chip to China. "Totally false." punchbowl.news/article/tech/n…

$NVDA will not only start selling H200 again to the Chinese, but may also sell the brand-new Groq chips. Neither Groq nor China are in the $1 T order visibility. After all the back and forth on China, this won't be the final word. But China surely would LOVE to buy a new chip and Jensen met Lutnick yesterday, so they must have talked about it. reuters.com/world/china/nv…

Samsung talked about 3-5 year contracts at their AGM. It's a big change that reduces cyclicality and suggests AI demand is structural rather than a short-term buildout that fades. bloomberg.com/news/articles/…

"SK Hynix chairman expects memory shortage to last until 2030" Memory makers now seem willing to stabilize memory prices in return for adopting a less cyclical industry structure. Hyperscalers seem to increasingly lock in 3-5-year supply with very significant prepayments. Probably memory makers commit to using some of that money to build more badly needed supply with less risk. Memory per GPU keeps rising and rapidly improving GPUs and AI capabilities increase demand, which means memory demand keeps growing strongly. Yet memory stocks trade at low forward multiples as if earnings will collapse in 2-3 years. But if demand grows rapidly and supply gets increasingly built against committed demand, the traditional boom-bust dynamic should be significantly less severe than in past cycles. I expect great earnings from $MU tomorrow and also like SK Hynix. reuters.com/world/asia-pac…

$NVDA China sales are back! Jensen also said they will use at least 50% of FCF for buybacks+dividend

@BenBajarin Yeah, $600 B is my expectation now

@ValueTom84 Yep I pointed that out as well. Very likely 600B is possible 27

I had SO many conversations with investors on this last night. There has to be more upside 2027 that manifests as the YoY growth rate was a question. Note via MS.

$NVDA increased Blackwell + Rubin order visibility from $500 B (2025-2026) to $1 T (2025-2027). Importantly, this excludes Hopper (significant in 2025) and other products (e.g. Groq, Vera CPU racks, self-driving tech). Crucially, this doesn't mean $NVDA expects to earn "only" $500 B in 2027 (~7% above estimates). This is order visibility 9 months before the year even starts! Jensen explicitly said this pipeline keeps growing. The market seems to interpret this as revenue guidance for 2027. But order visibility doubled in just 4.5 months and they can clearly take more orders. It seems VERY likely that they will get more orders for 2027 and do significantly better than $500 B.

You are worrying about genius moves. Jensen secured memory supply early before prices went parabolic. Now just a couple months later he would have to pay at least 50% more like all the others. Google fired its purchasing manager because he did not secure a LTA agreement early like Jensen. This was a great investment that already clearly is paying off. Securing future wafer capacity at TSMC helps with a 2nd key bottleneck. Others can't build as many chips as they'd like and reduce NVIDIA market share, because TSMC allocation is limited, NVIDIA secures so much ahead of time and TSMC won't expand unlimited. So $NVDA keeps growing and the growth of competitors stays bottlenecked. $AMD doesn't have the money to do this and the hyperscalers need theirs for large capex. So $NVDA is the only one that can really do this, making it a moat. So I consider both of these genius moves and hope Jensen makes more of them. Inventory growth is also not worrying when revenue is growing > 2x faster. Accounts receivable days of sales outstanding also went down from 53 days to 51.

Just a $NVDA thought. open.substack.com/pub/michaeljbu…

@mvcinvesting In the prepared remarks they said they already have 60% of $16-20 B capex funding from balance sheet, existing operations and commitments. But then later in the Q&A they said it is 60% just from the hyperscaler contracts. Would be great to know which one is correct.

“The majority of our CapEx guidance (60% or more) will be financed by the cash we’ve received and will continue to receive in 2026 from the favorable terms of our long-term contracts. We’re talking about significant amounts of cash.” — Ophir Nave, COO of $NBIS 🤯

@SupBagholder Jensen says 1 GW = $50 B. So 100 GW = $5 T. „Hundreds of GW“ would then be $10+ T. It sounds so insanely much that I cant understand how they would be using all of that themselves. And he talked about sovereigns in his post.

@ValueTom84 Why would you think that? Has meta ever shown any intrest in starting a cloud compute business?

Why is the market assuming they are selling compute ?

ASICs: Building one chip is not enough. NVIDIA builds entire racks, 3 types of switches, uses gigantic memory and context. List of diverse LLMs has exploded. The complexity keeps increasing, so tough to do all that from scratch. Very few teams in world are extraordinary at doing this.

China sales still zero. Disappointed, but will keep trying.

$NVDA Q3 - Revenue +62% YoY and +22% QoQ with a 63% EBIT-margin and they guided another +14% QoQ with even higher gross margins - They reiterated the $500 B of Blackwell + Rubin revenue through CY2026 and added something big: That number will almost certainly go up. New deals with Anthropic and Saudis are already incremental. NVIDIA always guides conservatively → even $500 B is too low. - Rubin on track to ramp in H2 2026. 7 new chips. Ecosystem ready. Will ramp it hard and expect smooth transition. - A100s shipped 6 years ago still running at full utilization. CUDA compatibility extend useful life well beyond accounting depreciation (through software upgrades). - GB200 NVLink72 is 10-15x better than H200. Certain about multi-year leadership. By the time competitors catch up, NVIDIA will be on the next generation.